// Copyright (c) Facebook, Inc. and its affiliates. // All rights reserved. // // Copyright 2019 Google LLC // // This source code is licensed under the BSD-style license found in the // LICENSE file in the root directory of this source tree. #include #include #include #include #include #ifdef _WIN32 #include #else #include #include #include #include #endif #ifdef _MSC_VER #include #endif #ifndef __EMSCRIPTEN__ #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifndef XNN_ENABLE_ASSEMBLY #define XNN_ENABLE_ASSEMBLY 1 #endif #if XNN_PLATFORM_WINDOWS static INIT_ONCE init_guard = INIT_ONCE_STATIC_INIT; #else static pthread_once_t init_guard = PTHREAD_ONCE_INIT; #endif #define XNN_MR_TO_INDEX(MR) (MR-1) #ifndef XNN_ENABLE_GEMM_M_SPECIALIZATION #error "XNN_ENABLE_GEMM_M_SPECIALIZATION is not defined" #endif static const struct xnn_allocator* volatile init_allocator = NULL; static void init(void) { #if XNN_ARCH_WASM || XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD // Unlike most other architectures, on x86/x86-64 when floating-point instructions // have no NaN arguments, but produce NaN output, the output NaN has sign bit set. // We use it to distinguish x86/x86-64 from other architectures, by doing subtraction // of two infinities (must produce NaN per IEEE 754 standard). static const volatile float inf = INFINITY; const bool is_wasm_x86 = signbit(inf - inf); #endif uint32_t init_flags = XNN_INIT_FLAG_XNNPACK; #if XNN_ARCH_ARM #if XNN_PLATFORM_MOBILE if (!cpuinfo_has_arm_neon()) { xnn_log_error("XNNPACK initialization failed: NEON is not supported"); return; } #else if (!cpuinfo_has_arm_v6()) { xnn_log_error("XNNPACK initialization failed: ARMv6 instructions not supported"); return; } if (!cpuinfo_has_arm_vfpv2() && !cpuinfo_has_arm_vfpv3()) { xnn_log_error("XNNPACK initialization failed: VFP is not supported"); return; } #endif if (cpuinfo_has_arm_neon()) { /**************************** QC8 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_QC8_OPERATORS init_flags |= XNN_INIT_FLAG_QC8; #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD switch (cpuinfo_get_uarch(0)->uarch) { case cpuinfo_uarch_cortex_a55: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_cortex_a55); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_cortex_a55); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c4__neondot); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 2; break; default: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c4__neondot); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 2; break; } #endif // XNN_ENABLE_ARM_DOTPROD } else { switch (cpuinfo_get_uarch(0)->uarch) { case cpuinfo_uarch_cortex_a5: case cpuinfo_uarch_cortex_a7: case cpuinfo_uarch_krait: case cpuinfo_uarch_kryo: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neon_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a32: case cpuinfo_uarch_cortex_a35: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a57: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_cortex_a53); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_cortex_a53); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_prfm_cortex_a35); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_prfm_cortex_a35); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a55r0: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a53); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a53); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a72: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 1; xnn_params.qc8.gemm.log2_sr = 2; break; case cpuinfo_uarch_exynos_m1: case cpuinfo_uarch_exynos_m2: case cpuinfo_uarch_exynos_m3: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_prfm_cortex_a35); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_prfm_cortex_a35); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; break; default: if (cpuinfo_has_arm_neon_v8()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; } else { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neon_mlal_lane_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neon_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; } break; } } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.qc8.gemm.mr; const uint32_t nr = xnn_params.qc8.gemm.nr; const uint32_t log2_kr = xnn_params.qc8.gemm.log2_kr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a55: #if XNN_ENABLE_ARM_DOTPROD if (mr == 4 && nr == 8 && log2_kr == 2 && cpuinfo_has_arm_neon_dot()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_cortex_a55; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__aarch32_neondot_cortex_a55; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c4__neondot; } #endif // XNN_ENABLE_ARM_DOTPROD break; case cpuinfo_uarch_cortex_a53: if (mr == 4 && nr == 8 && log2_kr == 0) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_cortex_a53; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_prfm_cortex_a53; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_prfm_cortex_a35; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_prfm_cortex_a35; } break; case cpuinfo_uarch_cortex_a55r0: if (mr == 4 && nr == 8 && log2_kr == 0) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a53; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8__aarch32_neonv8_mlal_lane_cortex_a53; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8__aarch32_neonv8_mlal_lane_cortex_a35; } break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x8c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x8c4__neondot); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c4__neondot); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else if (cpuinfo_has_arm_v8()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 1; xnn_params.qc8.gemm.log2_sr = 2; } else { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2s4__neon_mlal); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neon_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2s4__neon_mlal); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neon_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 1; xnn_params.qc8.gemm.log2_sr = 2; } #endif // XNN_ENABLE_ASSEMBLY if (cpuinfo_has_arm_neon_v8()) { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__aarch32_neonv8_mla8_cortex_a35; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.dwconv[0].channel_tile = 16; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neonv8_mla8_ld64; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.dwconv[1].channel_tile = 16; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neonv8_mla8_ld64; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.dwconv[2].channel_tile = 8; xnn_params.qc8.dwconv[2].primary_tile = 25; } else { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__neon_mla8_ld128; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neon_params; xnn_params.qc8.dwconv[0].channel_tile = 16; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neon_mla8_ld64; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neon_params; xnn_params.qc8.dwconv[1].channel_tile = 16; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__neon_mla8_ld64; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neon_params; xnn_params.qc8.dwconv[2].channel_tile = 8; xnn_params.qc8.dwconv[2].primary_tile = 25; } #endif // XNN_NO_QC8_OPERATORS /**************************** QS8 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QS8; #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD switch (cpuinfo_get_uarch(0)->uarch) { case cpuinfo_uarch_cortex_a55: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_cortex_a55); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_cortex_a55); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 2; break; default: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_ld64); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 2; break; } #endif // XNN_ENABLE_ARM_DOTPROD } else { switch (cpuinfo_get_uarch(0)->uarch) { case cpuinfo_uarch_cortex_a5: case cpuinfo_uarch_cortex_a7: case cpuinfo_uarch_krait: case cpuinfo_uarch_kryo: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a32: case cpuinfo_uarch_cortex_a35: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a57: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a55r0: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a72: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 1; xnn_params.qs8.gemm.log2_sr = 2; break; case cpuinfo_uarch_exynos_m1: case cpuinfo_uarch_exynos_m2: case cpuinfo_uarch_exynos_m3: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; break; default: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; break; } } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.qs8.gemm.mr; const uint32_t nr = xnn_params.qs8.gemm.nr; const uint32_t log2_kr = xnn_params.qs8.gemm.log2_kr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a55: #if XNN_ENABLE_ARM_DOTPROD if (mr == 4 && nr == 8 && log2_kr == 2 && cpuinfo_has_arm_neon_dot()) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_cortex_a55; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__aarch32_neondot_cortex_a55; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neondot; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neondot; } #endif // XNN_ENABLE_ARM_DOTPROD break; case cpuinfo_uarch_cortex_a53: if (mr == 4 && nr == 8 && log2_kr == 0) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7; } break; case cpuinfo_uarch_cortex_a55r0: if (mr == 4 && nr == 8 && log2_kr == 0) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7; } break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neondot); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 1; xnn_params.qs8.gemm.log2_sr = 2; } #endif // XNN_ENABLE_ASSEMBLY xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mla8_ld64; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.dwconv[0].channel_tile = 16; xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mla8_ld64; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.dwconv[1].channel_tile = 8; xnn_params.qs8.dwconv[1].primary_tile = 25; xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8, .init.qs8 = xnn_init_qs8_avgpool_minmax_rndnu_neon_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_rndnu_neon_params, .row_tile = 7, .channel_tile = 8, }; xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__neon_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x16, .init.qs8_add = xnn_init_qs8_add_minmax_neon_params, .element_tile = 16, }; xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .init.qs8_mul = xnn_init_qs8_mul_minmax_rndnu_neon_params, .element_tile = 16, }; xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__neon_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_neon_params, .element_tile = 32, }; #endif // XNN_NO_QS8_OPERATORS /*************************** QU8 AArch32 micro-kernels ***************************/ #ifndef XNN_NO_QU8_OPERATORS init_flags |= XNN_INIT_FLAG_QU8; #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; xnn_params.qu8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { switch (cpuinfo_get_uarch(0)->uarch) { case cpuinfo_uarch_cortex_a5: case cpuinfo_uarch_cortex_a7: case cpuinfo_uarch_krait: case cpuinfo_uarch_kryo: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a32: case cpuinfo_uarch_cortex_a35: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a57: case cpuinfo_uarch_cortex_a72: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a55r0: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; break; case cpuinfo_uarch_exynos_m1: case cpuinfo_uarch_exynos_m2: case cpuinfo_uarch_exynos_m3: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; break; default: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_ld64); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; break; } } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.qu8.gemm.mr; const uint32_t nr = xnn_params.qu8.gemm.nr; const uint32_t log2_kr = xnn_params.qu8.gemm.log2_kr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a53: if (mr == 4 && nr == 8 && log2_kr == 0) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53; xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53; xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7; xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_prfm_cortex_a7; } break; case cpuinfo_uarch_cortex_a55r0: if (mr == 4 && nr == 8 && log2_kr == 0) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53; xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53; xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7; xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__aarch32_neon_mlal_lane_cortex_a7; } break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x8c4__neondot); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8c4__neondot); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 8; xnn_params.qu8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x8__neon_mlal_lane); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 3; xnn_params.qu8.gemm.nr = 8; } #endif // XNN_ENABLE_ASSEMBLY xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul8; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.dwconv[0].channel_tile = 16; xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.dwconv[1].channel_tile = 8; xnn_params.qu8.dwconv[1].primary_tile = 25; xnn_params.qu8.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8, .init.qu8 = xnn_init_qu8_avgpool_minmax_neon_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 8, }; xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8, .init.qu8 = xnn_init_qu8_avgpool_minmax_rndnu_neon_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_rndnu_neon_params, .row_tile = 7, .channel_tile = 8, }; xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__neon_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x16, .init.qu8_add = xnn_init_qu8_add_minmax_neon_params, .element_tile = 8, }; xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .init.qu8_mul = xnn_init_qu8_mul_minmax_rndnu_neon_params, .element_tile = 16, }; xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__neon_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_neon_params, .element_tile = 32, }; #endif // XNN_NO_QU8_OPERATORS /**************************** S8 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_S8_OPERATORS init_flags |= XNN_INIT_FLAG_S8; xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__neon_x64, .init.s8_minmax = xnn_init_s8_minmax_neon_params, .element_tile = 64, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__neon_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16, .init.s8 = xnn_init_s8_minmax_neon_params, .mr = 9, .qr = 8, }; #endif // XNN_NO_S8_OPERATORS /**************************** U8 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_U8_OPERATORS init_flags |= XNN_INIT_FLAG_U8; xnn_params.u8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_u8_vclamp_ukernel__neon_x64, .init.u8_minmax = xnn_init_u8_minmax_neon_params, .element_tile = 64, }; xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__neon_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.u8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16, .init.u8 = xnn_init_u8_minmax_neon_params, .mr = 9, .qr = 8, }; xnn_params.u8.rmax = xnn_u8_rmax_ukernel__neon; xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; #endif // XNN_NO_U8_OPERATORS /**************************** X8 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_X8_OPERATORS init_flags |= XNN_INIT_FLAG_X8; xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar_x4; xnn_params.x8.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__neon, .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__neon, .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__neon, .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__neon, }; xnn_params.x8.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x8_transposec_ukernel__16x16_reuse_dec_zip_neon, .tile_size = 32, }; #endif // XNN_NO_X8_OPERATORS /**************************** X16 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_X16_OPERATORS init_flags |= XNN_INIT_FLAG_X16; xnn_params.x16.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x16_transposec_ukernel__8x8_reuse_dec_zip_neon, .tile_size = 32, }; #endif // XNN_NO_X16_OPERATORS /**************************** F32 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_F32_OPERATORS init_flags |= XNN_INIT_FLAG_F32; #if XNN_ENABLE_ASSEMBLY switch (cpuinfo_get_uarch(0)->uarch) { case cpuinfo_uarch_cortex_a5: case cpuinfo_uarch_cortex_a7: case cpuinfo_uarch_krait: case cpuinfo_uarch_kryo: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a7); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a53: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a53); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a53); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a55r0: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a32: case cpuinfo_uarch_cortex_a35: case cpuinfo_uarch_cortex_a55: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a57: case cpuinfo_uarch_cortex_a72: case cpuinfo_uarch_cortex_a73: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a75); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; break; default: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a75); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; #if XNN_ENABLE_JIT xnn_params.f32.gemm.generator.gemm = xnn_init_hmp_gemm_codegen(xnn_generate_f32_gemm_ukernel_4x8__aarch32_neon_cortex_a75); xnn_params.f32.gemm.generator.igemm = xnn_init_hmp_igemm_codegen(xnn_generate_f32_igemm_ukernel_4x8__aarch32_neon_cortex_a75); #endif break; } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.f32.gemm.mr; const uint32_t nr = xnn_params.f32.gemm.nr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a53: if (mr == 4 && nr == 8) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_prfm_cortex_a53; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64; } break; case cpuinfo_uarch_cortex_a55r0: if (mr == 4 && nr == 8) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a53; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64; } break; case cpuinfo_uarch_cortex_a55: if (mr == 4 && nr == 8) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch32_neon_cortex_a55; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64; } break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // XNN_ENABLE_ASSEMBLY xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__neon_lane_ld128); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__neon_lane_ld128); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neon_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; #endif // XNN_ENABLE_ASSEMBLY xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__neon_lane_ld64); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__neon_lane_ld64); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x3__neon; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[0].channel_tile = 8, xnn_params.f32.dwconv[0].primary_tile = 3, xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__neon; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[1].channel_tile = 8, xnn_params.f32.dwconv[1].primary_tile = 4, xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__neon; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 8; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[3].channel_tile = 8; xnn_params.f32.dwconv[3].primary_tile = 25; xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__neon_c4, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4, .init.f32 = xnn_init_f32_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .update.f32 = xnn_update_f32_scaleminmax_scalar_params, .row_tile = 7, .channel_tile = 4, }; xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4, .init.f32 = xnn_init_f32_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__neon_c4, .mr = 4, }; xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__neon_c4, .mr = 9, }; xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__neon_c4, .mr = 9, .qr = 8, }; xnn_params.f32.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__neon_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; if (cpuinfo_has_arm_neon_fma()) { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__neonfma_rr1_p6_x8, .init.f32_elu = xnn_init_f32_elu_neonfma_rr1_p6_params, .element_tile = 8, }; } else { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__neon_rr2_lut16_p3_x8, .init.f32_elu = xnn_init_f32_elu_neon_rr2_lut16_p3_params, .element_tile = 8, }; } xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__neon_x16, .init.f32_hswish = xnn_init_f32_hswish_scalar_params, .element_tile = 16, }; xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__neon_x8, .init.f32_lrelu = xnn_init_f32_lrelu_scalar_params, .element_tile = 8, }; xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__neon_x8, .element_tile = 8, }; if (cpuinfo_has_arm_neon_v8()) { xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__neonv8_x8, .element_tile = 8, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__neonv8_x8, .element_tile = 8, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__neonv8_x8, .element_tile = 8, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__neonv8_x8, .element_tile = 8, }; } else { xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__neon_x8, .element_tile = 8, }; } xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__neon_rr2_lut64_p2_nr2recps_x8, .init.f32_sigmoid = xnn_init_f32_sigmoid_neon_rr2_lut64_p2_params, .element_tile = 8, }; xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__scalar_sqrt_x1, .element_tile = 1, }; xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8, .row_tile = 2, .channel_tile = 8, }; xnn_params.f32.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f32_raddstoreexpminusmax_ukernel__neon_rr2_lut64_p2_x8, .init.f32 = xnn_init_f32_expminus_neon_rr2_lut64_p2_params, .element_tile = 8, }; xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__neon; xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__scalar_x2, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__scalar_x2, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__scalar_x2, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 2, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__neon_2x, .init.f32 = xnn_init_f32_minmax_scalar_params, .channel_tile = 4, .row_tile = 2, }; #ifndef XNN_NO_NCHW_OPERATORS init_flags |= XNN_INIT_FLAG_CHW_OPT; xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_32x1__neon, .mr = 32, .nr = 1, }; xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neon_2x2, .output_channel_tile = 4, .output_height_tile = 2, .output_width_tile = 2, }; xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__neon_2x4, .output_width_tile = 4, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neon_1x4, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__neon_1x4, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neon_1x4, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__neon_x4, .channel_tile = 4, }; xnn_params.f32.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f32_ibilinear_chw_ukernel__neon_p8, .channel_tile = 1, .pixel_tile = 8, }; #endif // XNN_NO_NCHW_OPERATORS #endif // XNN_NO_F32_OPERATORS /*************************** VCVT AArch32 micro-kernels ***************************/ #ifndef XNN_NO_VCVT_OPERATORS init_flags |= XNN_INIT_FLAG_VCVT; if (cpuinfo_has_arm_neon_fp16()) { xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__neonfp16_x16, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__neonfp16_x16, .element_tile = 16, }; } else { xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__neon_int16_x16, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_neon_params, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__neon_x8, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_neon_params, .element_tile = 8, }; } if (cpuinfo_has_arm_neon_v8()) { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__neonv8_x32, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_neonv8_params, .element_tile = 32, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__neonv8_x32, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_neonv8_params, .element_tile = 32, }; } else { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__neon_x32, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_neon_params, .element_tile = 32, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__neon_x32, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_neon_params, .element_tile = 32, }; } xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__neon_x32, .init.qs8_cvt = xnn_init_qs8_cvt_neon_params, .element_tile = 32, }; xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__neon_x32, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_neon_params, .element_tile = 32, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__neon_x32, .init.qu8_cvt = xnn_init_qu8_cvt_neon_params, .element_tile = 32, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__neon_x32, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_neon_params, .element_tile = 32, }; #endif // XNN_NO_VCVT_OPERATORS /**************************** X32 AArch32 micro-kernels ****************************/ #ifndef XNN_NO_X32_OPERATORS init_flags |= XNN_INIT_FLAG_X32; xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__neon; xnn_params.x32.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__neon, .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__neon, .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__neon, .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__neon, }; xnn_params.x32.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x32_transposec_ukernel__4x4_reuse_dec_zip_neon, .tile_size = 32, }; #endif // XNN_NO_X32_OPERATORS /**************************** XX AArch32 micro-kernels ****************************/ #ifndef XNN_NO_XX_OPERATORS init_flags |= XNN_INIT_FLAG_XX; xnn_params.xx.copy = (xnn_vunary_ukernel_function) xnn_xx_copy_ukernel__memcpy; xnn_params.xx.fill = (struct fill_parameters) { .ukernel = (xnn_fill_ukernel_function) xnn_xx_fill_ukernel__neon_x64, .row_tile = 1, }; xnn_params.xx.pad = (struct pad_parameters) { .ukernel = (xnn_pad_ukernel_function) xnn_xx_pad_ukernel__neon, .row_tile = 1, }; xnn_params.xx.transpose = (struct transpose_parameters) { .variable_size_ukernel = xnn_xx_transposev_ukernel__1x1_memcpy, .tile_size = 32, }; #endif // XNN_NO_XX_OPERATORS } else if (!XNN_PLATFORM_MOBILE) { /*************************** QC8 AArch32 Pre-NEON micro-kernels ***************************/ #ifndef XNN_NO_QC8_OPERATORS init_flags |= XNN_INIT_FLAG_QC8; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x2c4__armsimd32); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x2c4__armsimd32); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x2c4__armsimd32); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x2c4__armsimd32); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_armsimd32_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 2; xnn_params.qc8.gemm.log2_kr = 2; xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up1x3__scalar_fmagic; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qc8.dwconv[0].channel_tile = 1; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qc8.dwconv[1].channel_tile = 1; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_fmagic; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qc8.dwconv[2].channel_tile = 1; xnn_params.qc8.dwconv[2].primary_tile = 25; #endif // XNN_NO_QS8_OPERATORS /*************************** QS8 AArch32 Pre-NEON micro-kernels ***************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QS8; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_2x2c4__armsimd32); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_2x2c4__armsimd32); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x2c4__armsimd32); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x2c4__armsimd32); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_armsimd32_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 2; xnn_params.qs8.gemm.log2_kr = 2; xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qs8.dwconv[0].channel_tile = 1; xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_fmagic; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qs8.dwconv[1].channel_tile = 1; xnn_params.qs8.dwconv[1].primary_tile = 25; xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c1, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c1, .init.qs8 = xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_fp32_scalar_imagic_params, .row_tile = 7, .channel_tile = 1, }; xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__scalar_x1, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__scalar_x1, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__scalar_x1, .init.qs8_add = xnn_init_qs8_add_minmax_scalar_params, .element_tile = 1, }; xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_fp32_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4, .init.qs8_mul = xnn_init_qs8_mul_minmax_fp32_scalar_params, .element_tile = 4, }; xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__armsimd32_x4, .init.qs8_lrelu = xnn_init_qs8_lrelu_armsimd32_params, .element_tile = 4, }; #endif // XNN_NO_QS8_OPERATORS /*************************** QU8 AArch32 Pre-NEON micro-kernels ***************************/ #ifndef XNN_NO_QU8_OPERATORS init_flags |= XNN_INIT_FLAG_QU8; xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_2x2c4__armsimd32); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_2x2c4__armsimd32); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x2c4__armsimd32); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x2c4__armsimd32); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_armsimd32_params; xnn_params.qu8.gemm.mr = 2; xnn_params.qu8.gemm.nr = 2; xnn_params.qu8.gemm.log2_kr = 2; xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up1x9__scalar_fmagic; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qu8.dwconv[0].channel_tile = 1; xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_fmagic; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qu8.dwconv[1].channel_tile = 1; xnn_params.qu8.dwconv[1].primary_tile = 25; xnn_params.qu8.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1, .init.qu8 = xnn_init_qu8_avgpool_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c1, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c1, .init.qu8 = xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_fp32_scalar_imagic_params, .row_tile = 7, .channel_tile = 1, }; xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__scalar_x1, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__scalar_x1, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__scalar_x1, .init.qu8_add = xnn_init_qu8_add_minmax_scalar_params, .element_tile = 1, }; xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4, .init.qu8_mul = xnn_init_qu8_mul_minmax_fp32_scalar_params, .element_tile = 4, }; xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__armsimd32_x4, .init.qu8_lrelu = xnn_init_qu8_lrelu_armsimd32_params, .element_tile = 4, }; #endif // XNN_NO_QU8_OPERATORS /**************************** S8 AArch32 Pre-NEON micro-kernels ****************************/ #ifndef XNN_NO_S8_OPERATORS init_flags |= XNN_INIT_FLAG_S8; xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__scalar_x4, .init.s8_minmax = xnn_init_s8_minmax_scalar_params, .element_tile = 4, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__scalar_c1, .pixel_tile = 1, .channel_tile = 1, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.s8 = xnn_init_s8_minmax_scalar_params, .mr = 9, .qr = 8, }; #endif // XNN_NO_S8_OPERATORS /**************************** U8 AArch32 Pre-NEON micro-kernels ****************************/ #ifndef XNN_NO_U8_OPERATORS init_flags |= XNN_INIT_FLAG_U8; xnn_params.u8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_u8_vclamp_ukernel__scalar_x4, .init.u8_minmax = xnn_init_u8_minmax_scalar_params, .element_tile = 4, }; xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__scalar_c1, .pixel_tile = 1, .channel_tile = 1, }; xnn_params.u8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.u8 = xnn_init_u8_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; #endif // XNN_NO_U8_OPERATORS /**************************** X8 AArch32 Pre-NEON micro-kernels ****************************/ #ifndef XNN_NO_X8_OPERATORS init_flags |= XNN_INIT_FLAG_X8; xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar_x4; xnn_params.x8.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, }; xnn_params.x8.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x8_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X8_OPERATORS /**************************** X16 AArch32 Pre-NEON micro-kernels ****************************/ #ifndef XNN_NO_X16_OPERATORS init_flags |= XNN_INIT_FLAG_X16; xnn_params.x16.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x16_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X16_OPERATORS /**************************** F32 AArch32 Pre-NEON micro-kernels ****************************/ #ifndef XNN_NO_F32_OPERATORS init_flags |= XNN_INIT_FLAG_F32; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x4__scalar); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x4__scalar); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x4__scalar); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x4__scalar); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x4__scalar); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x4__scalar); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x4__scalar); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x4__scalar); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x4__scalar); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__scalar); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 4; xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__scalar); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__scalar); xnn_params.f32.gemm2.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2__scalar); xnn_params.f32.gemm2.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__scalar); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x3__scalar_acc2; xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x3__scalar_acc2; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[0].channel_tile = 1; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x4__scalar_acc2; xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x4__scalar_acc2; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[1].channel_tile = 1; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x9__scalar_acc2; xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x9__scalar_acc2; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 1; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x25__scalar_acc2; xnn_params.f32.dwconv[3].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x25__scalar_acc2; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[3].channel_tile = 1; xnn_params.f32.dwconv[3].primary_tile = 25; xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1, .init.f32 = xnn_init_f32_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__scalar_c1, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .update.f32 = xnn_update_f32_scaleminmax_scalar_params, .row_tile = 7, .channel_tile = 1, }; xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.f32 = xnn_init_f32_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__scalar_c1, .mr = 4, }; xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__scalar_c1, .mr = 9, }; xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1, .mr = 9, .qr = 8, }; xnn_params.f32.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__scalar_c2, .pixel_tile = 1, .channel_tile = 2, }; xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__scalar_x4, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 4, }; xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x4, .init.f32_elu = xnn_init_f32_elu_scalar_rr2_lut16_p3_params, .element_tile = 4, }; xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__scalar_x4, .init.f32_hswish = xnn_init_f32_hswish_scalar_params, .element_tile = 4, }; xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__scalar_x4, .init.f32_lrelu = xnn_init_f32_lrelu_scalar_params, .element_tile = 4, }; xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x2, .init.f32_sigmoid = xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, .element_tile = 2, }; xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__scalar_sqrt_x1, .element_tile = 1, }; xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__scalar_2x4, .row_tile = 4, .channel_tile = 4, }; xnn_params.f32.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f32_raddstoreexpminusmax_ukernel__scalar_rr2_p5_x4_acc2, .init.f32 = xnn_init_f32_expminus_scalar_rr2_p5_params, .element_tile = 4, }; xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__scalar; xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__scalar_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__scalar_x2, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__scalar_x2, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__scalar_x2, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 2, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__scalar_x8, .element_tile = 8, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__scalar_x8, .element_tile = 8, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__scalar_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__scalar_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, .element_tile = 8, }; xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c1__scalar_2x, .init.f32 = xnn_init_f32_minmax_scalar_params, .channel_tile = 1, .row_tile = 2, }; #ifndef XNN_NO_NCHW_OPERATORS init_flags |= XNN_INIT_FLAG_CHW_OPT; xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x1__scalar, .mr = 8, .nr = 1, }; xnn_params.f32.spmm2 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x2__scalar, .mr = 8, .nr = 2, }; xnn_params.f32.spmm4 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x4__scalar, .mr = 8, .nr = 4, }; xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1, .output_channel_tile = 4, .output_height_tile = 1, .output_width_tile = 1, }; xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_4x1, .output_width_tile = 1, .output_height_tile = 4, }; xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_2x1_acc2, .output_width_tile = 1, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_2x1_acc2, .output_width_tile = 1, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_2x1_acc2, .output_width_tile = 1, .output_height_tile = 2, }; xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__scalar_x1, .channel_tile = 1, }; xnn_params.f32.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f32_ibilinear_chw_ukernel__scalar_p4, .channel_tile = 1, .pixel_tile = 4, }; #endif // XNN_NO_NCHW_OPERATORS #endif // XNN_NO_F32_OPERATORS /*************************** VCVT AArch32 Pre-NEON micro-kernels ***************************/ #ifndef XNN_NO_VCVT_OPERATORS init_flags |= XNN_INIT_FLAG_VCVT; xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__scalar_x4, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_scalar_params, .element_tile = 4, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x2, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_scalar_fabsf_params, .element_tile = 2, }; xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__scalar_imagic_x4, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_scalar_imagic_params, .element_tile = 4, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x4, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_scalar_imagic_params, .element_tile = 4, }; xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__armsimd32_x8, .init.qs8_cvt = xnn_init_qs8_cvt_scalar_params, .element_tile = 8, }; xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__scalar_x4, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_scalar_params, .element_tile = 4, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__armsimd32_x8, .init.qu8_cvt = xnn_init_qu8_cvt_scalar_params, .element_tile = 8, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__scalar_x4, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_scalar_params, .element_tile = 4, }; #endif // XNN_NO_VCVT_OPERATORS /**************************** X32 AArch32 Pre-NEON micro-kernels ****************************/ #ifndef XNN_NO_X32_OPERATORS init_flags |= XNN_INIT_FLAG_X32; xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__scalar; xnn_params.x32.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__scalar, .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__scalar, .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__scalar, .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__scalar, }; xnn_params.x32.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x32_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X32_OPERATORS /**************************** XX AArch32 Pre-NEON micro-kernels ****************************/ #ifndef XNN_NO_XX_OPERATORS init_flags |= XNN_INIT_FLAG_XX; xnn_params.xx.copy = (xnn_vunary_ukernel_function) xnn_xx_copy_ukernel__memcpy; xnn_params.xx.fill = (struct fill_parameters) { .ukernel = (xnn_fill_ukernel_function) xnn_xx_fill_ukernel__scalar_x16, .row_tile = 1, }; xnn_params.xx.pad = (struct pad_parameters) { .ukernel = (xnn_pad_ukernel_function) xnn_xx_pad_ukernel__scalar, .row_tile = 1, }; xnn_params.xx.transpose = (struct transpose_parameters) { .variable_size_ukernel = xnn_xx_transposev_ukernel__1x1_memcpy, .tile_size = 32, }; #endif // XNN_NO_XX_OPERATORS } #elif XNN_ARCH_ARM64 /**************************** QC8 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_QC8_OPERATORS init_flags |= XNN_INIT_FLAG_QC8; #if XNN_PLATFORM_IOS || XNN_PLATFORM_MAC #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 16; xnn_params.qc8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 3; } #else // !XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 16; xnn_params.qc8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 1; xnn_params.qc8.gemm.log2_sr = 2; } #endif // XNN_ENABLE_ASSEMBLY #else // !XNN_PLATFORM_IOS && !XNN_PLATFORM_MAC #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a55: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55); break; case cpuinfo_uarch_cortex_x1: case cpuinfo_uarch_cortex_a78: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld128); break; default: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_ld64); break; } xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 16; xnn_params.qc8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a35: case cpuinfo_uarch_kryo: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 16; break; case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a55r0: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16__neonv8_mlal_lane); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 16; break; case cpuinfo_uarch_cortex_a72: case cpuinfo_uarch_cortex_a73: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 3; break; default: xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 3; break; } } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.qc8.gemm.mr; const uint32_t nr = xnn_params.qc8.gemm.nr; const uint32_t log2_kr = xnn_params.qc8.gemm.log2_kr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a55r0: if (mr == 2 && nr == 8 && log2_kr == 3) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53; } break; case cpuinfo_uarch_cortex_a55: #if XNN_ENABLE_ARM_DOTPROD if (mr == 4 && nr == 16 && log2_kr == 2 && cpuinfo_has_arm_neon_dot()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__aarch64_neondot_cortex_a55; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot; xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16c4__neondot; } #endif // XNN_ENABLE_ARM_DOTPROD break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // !XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c4__neondot); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16c4__neondot); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 16; xnn_params.qc8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c2s4__neonv8_mlal); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 1; xnn_params.qc8.gemm.log2_sr = 2; } #endif // XNN_ENABLE_ASSEMBLY #endif // XNN_PLATFORM_IOS || XNN_PLATFORM_MAC xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__neonv8_mla8_ld128; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.dwconv[0].channel_tile = 16; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__neonv8_mla8_ld64; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.dwconv[1].channel_tile = 16; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__neonv8_mla8_ld64; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_neonv8_params; xnn_params.qc8.dwconv[2].channel_tile = 16; xnn_params.qc8.dwconv[2].primary_tile = 25; #endif // XNN_NO_QC8_OPERATORS /**************************** QS8 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QS8; #if XNN_PLATFORM_IOS || XNN_PLATFORM_MAC #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 16; xnn_params.qs8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 3; } #else // !XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 16; xnn_params.qs8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 1; xnn_params.qs8.gemm.log2_sr = 2; } #endif // XNN_ENABLE_ASSEMBLY #else // !XNN_PLATFORM_IOS && !XNN_PLATFORM_MAC #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a55: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55); break; case cpuinfo_uarch_cortex_x1: case cpuinfo_uarch_cortex_a78: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128); break; default: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64); break; } xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 16; xnn_params.qs8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a35: case cpuinfo_uarch_kryo: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_ld64); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 16; break; case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a55r0: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 16; break; case cpuinfo_uarch_cortex_a72: case cpuinfo_uarch_cortex_a73: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 3; break; default: xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 3; break; } } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.qs8.gemm.mr; const uint32_t nr = xnn_params.qs8.gemm.nr; const uint32_t log2_kr = xnn_params.qs8.gemm.log2_kr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a55r0: if (mr == 2 && nr == 8 && log2_kr == 3) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm_cortex_a53; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal_prfm_cortex_a53; } break; case cpuinfo_uarch_cortex_a55: #if XNN_ENABLE_ARM_DOTPROD if (mr == 4 && nr == 16 && log2_kr == 2 && cpuinfo_has_arm_neon_dot()) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot; xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neondot; } #endif // XNN_ENABLE_ARM_DOTPROD break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // !XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__neondot); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 16; xnn_params.qs8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 1; xnn_params.qs8.gemm.log2_sr = 2; } #endif // XNN_ENABLE_ASSEMBLY #endif // XNN_PLATFORM_IOS || XNN_PLATFORM_MAC xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mla8_ld64; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.dwconv[0].channel_tile = 16; xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_rndnu_ukernel_up16x25__neon_mla8_ld64; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_rndnu_neon_params; xnn_params.qs8.dwconv[1].channel_tile = 16; xnn_params.qs8.dwconv[1].primary_tile = 25; xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8, .init.qs8 = xnn_init_qs8_avgpool_minmax_rndnu_neon_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_rndnu_neon_params, .row_tile = 7, .channel_tile = 8, }; xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__neon_ld64_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__neon_ld64_x32, .init.qs8_add = xnn_init_qs8_add_minmax_neon_params, .element_tile = 32, }; xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .init.qs8_mul = xnn_init_qs8_mul_minmax_rndnu_neon_params, .element_tile = 16, }; xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__neon_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_neon_params, .element_tile = 32, }; #endif // XNN_NO_QS8_OPERATORS /**************************** QU8 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_QU8_OPERATORS init_flags |= XNN_INIT_FLAG_QU8; #if XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a55: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; xnn_params.qu8.gemm.log2_kr = 2; break; default: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; xnn_params.qu8.gemm.log2_kr = 2; break; } #endif // XNN_ENABLE_ARM_DOTPROD } else { switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a55r0: case cpuinfo_uarch_kryo: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; break; case cpuinfo_uarch_cortex_a57: case cpuinfo_uarch_cortex_a72: case cpuinfo_uarch_cortex_a73: case cpuinfo_uarch_cortex_a75: case cpuinfo_uarch_cortex_a76: case cpuinfo_uarch_exynos_m1: case cpuinfo_uarch_exynos_m2: case cpuinfo_uarch_exynos_m3: case cpuinfo_uarch_exynos_m4: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a75); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; break; default: xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; break; } } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.qu8.gemm.mr; const uint32_t nr = xnn_params.qu8.gemm.nr; const uint32_t log2_kr = xnn_params.qu8.gemm.log2_kr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a55r0: if (mr == 4 && nr == 16 && log2_kr == 0) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53; xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53; } break; case cpuinfo_uarch_cortex_a55: #if XNN_ENABLE_ARM_DOTPROD if (mr == 4 && nr == 16 && log2_kr == 2 && cpuinfo_has_arm_neon_dot()) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55; xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55; } #endif // XNN_ENABLE_ARM_DOTPROD break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // !XNN_ENABLE_ASSEMBLY if (XNN_ENABLE_ARM_DOTPROD && cpuinfo_has_arm_neon_dot()) { #if XNN_ENABLE_ARM_DOTPROD xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16c4__neondot); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x16c4__neondot); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; xnn_params.qu8.gemm.log2_kr = 2; #endif // XNN_ENABLE_ARM_DOTPROD } else { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; } #endif // XNN_ENABLE_ASSEMBLY xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_rndnu_ukernel_up16x9__neon_mul8; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.dwconv[0].channel_tile = 16; xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_rndnu_ukernel_up8x25__neon_mul8; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_rndnu_neon_params; xnn_params.qu8.dwconv[1].channel_tile = 8; xnn_params.qu8.dwconv[1].primary_tile = 25; xnn_params.qu8.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__neon_c8, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__neon_c8, .init.qu8 = xnn_init_qu8_avgpool_minmax_neon_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 8, }; xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_rndnu_ukernel_7x__neon_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_rndnu_ukernel_7p7x__neon_c8, .init.qu8 = xnn_init_qu8_avgpool_minmax_rndnu_neon_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_rndnu_neon_params, .row_tile = 7, .channel_tile = 8, }; xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__neon_ld64_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x32, .init.qu8_add = xnn_init_qu8_add_minmax_neon_params, .element_tile = 8, }; xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_rndnu_ukernel__neon_ld64_x16, .init.qu8_mul = xnn_init_qu8_mul_minmax_rndnu_neon_params, .element_tile = 16, }; xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__neon_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_neon_params, .element_tile = 32, }; #endif // XNN_NO_QU8_OPERATORS /**************************** S8 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_S8_OPERATORS init_flags |= XNN_INIT_FLAG_S8; xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__neon_x64, .init.s8_minmax = xnn_init_s8_minmax_neon_params, .element_tile = 64, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__neon_c16, .pixel_tile = 1, .channel_tile = 16, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__neon_c16, .init.s8 = xnn_init_s8_minmax_neon_params, .mr = 9, .qr = 8, }; #endif // XNN_NO_S8_OPERATORS /**************************** U8 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_U8_OPERATORS init_flags |= XNN_INIT_FLAG_U8; xnn_params.u8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_u8_vclamp_ukernel__neon_x64, .init.u8_minmax = xnn_init_u8_minmax_neon_params, .element_tile = 64, }; xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__neon_c16, .pixel_tile = 1, .channel_tile = 16, }; xnn_params.u8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__neon_c16, .init.u8 = xnn_init_u8_minmax_neon_params, .mr = 9, .qr = 8, }; xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; xnn_params.u8.rmax = xnn_u8_rmax_ukernel__neon; #endif // XNN_NO_U8_OPERATORS /**************************** X8 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_X8_OPERATORS init_flags |= XNN_INIT_FLAG_X8; xnn_params.x8.lut = xnn_x8_lut_ukernel__neon_tbx128x4_x64; xnn_params.x8.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__neon, .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__neon, .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__neon, .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__neon, }; xnn_params.x8.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x8_transposec_ukernel__16x16_reuse_dec_zip_neon, .tile_size = 32, }; #endif // XNN_NO_X8_OPERATORS /**************************** X16 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_X16_OPERATORS init_flags |= XNN_INIT_FLAG_X16; xnn_params.x16.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x16_transposec_ukernel__8x8_reuse_dec_zip_neon, .tile_size = 32, }; #endif // XNN_NO_X16_OPERATORS /**************************** F16 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_F16_OPERATORS #if XNN_ENABLE_ARM_FP16 if (cpuinfo_has_arm_neon_fp16_arith()) { init_flags |= XNN_INIT_FLAG_F16 | XNN_INIT_FLAG_F16_NATIVE; #if XNN_ENABLE_ASSEMBLY switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a55: xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55); xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.gemm.mr = 6; xnn_params.f16.gemm.nr = 16; break; case cpuinfo_uarch_cortex_a55r0: case cpuinfo_uarch_cortex_a75: xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55r0); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55r0); xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.gemm.mr = 6; xnn_params.f16.gemm.nr = 16; break; case cpuinfo_uarch_exynos_m5: xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.gemm.mr = 4; xnn_params.f16.gemm.nr = 16; break; case cpuinfo_uarch_exynos_m4: xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.gemm.mr = 6; xnn_params.f16.gemm.nr = 16; break; default: case cpuinfo_uarch_cortex_a76: case cpuinfo_uarch_cortex_a77: case cpuinfo_uarch_cortex_a78: case cpuinfo_uarch_cortex_x1: xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a75); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a75); xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__aarch64_neonfp16arith_ld64); xnn_params.f16.gemm.init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.gemm.mr = 6; xnn_params.f16.gemm.nr = 16; break; } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.f16.gemm.mr; const uint32_t nr = xnn_params.f16.gemm.nr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a55: if (mr == 6 && nr == 16) { xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55; xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55; } break; case cpuinfo_uarch_cortex_a55r0: case cpuinfo_uarch_cortex_a75: if (mr == 6 && nr == 16) { xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55r0; xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__aarch64_neonfp16arith_cortex_a55r0; } break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 #else // XNN_ENABLE_ASSEMBLY xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_6x16__neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_6x16__neonfp16arith_ld64); xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__neonfp16arith_ld64); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__neonfp16arith_ld64); xnn_params.f16.gemm.init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.gemm.mr = 6; xnn_params.f16.gemm.nr = 16; #endif // XNN_ENABLE_ASSEMBLY xnn_params.f16.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x3__neonfp16arith; xnn_params.f16.dwconv[0].init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.dwconv[0].channel_tile = 16; xnn_params.f16.dwconv[0].primary_tile = 3; xnn_params.f16.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x4__neonfp16arith; xnn_params.f16.dwconv[1].init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.dwconv[1].channel_tile = 16; xnn_params.f16.dwconv[1].primary_tile = 4; xnn_params.f16.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x9__neonfp16arith; xnn_params.f16.dwconv[2].init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.dwconv[2].channel_tile = 16; xnn_params.f16.dwconv[2].primary_tile = 9; xnn_params.f16.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2; xnn_params.f16.dwconv[3].init.f16 = xnn_init_f16_minmax_neon_params; xnn_params.f16.dwconv[3].channel_tile = 8; xnn_params.f16.dwconv[3].primary_tile = 25; xnn_params.f16.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f16_avgpool_minmax_ukernel_9x__neonfp16arith_c8, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f16_avgpool_minmax_ukernel_9p8x__neonfp16arith_c8, .init.f16 = xnn_init_f16_scaleminmax_neon_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 8, }; xnn_params.f16.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f16_pavgpool_minmax_ukernel_9x__neonfp16arith_c8, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f16_pavgpool_minmax_ukernel_9p8x__neonfp16arith_c8, .init.f16 = xnn_init_f16_minmax_neon_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 8, }; xnn_params.f16.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f16_gavgpool_minmax_ukernel_7x__neonfp16arith_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f16_gavgpool_minmax_ukernel_7p7x__neonfp16arith_c8, .init.f16 = xnn_init_f16_scaleminmax_neon_params, .update.f16 = xnn_update_f16_scaleminmax_neon_params, .row_tile = 7, .channel_tile = 8, }; xnn_params.f16.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f16_maxpool_minmax_ukernel_9p8x__neonfp16arith_c8, .init.f16 = xnn_init_f16_minmax_neon_params, .mr = 9, .qr = 8, }; xnn_params.f16.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f16_ibilinear_ukernel__neonfp16arith_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.f16.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f16_prelu_ukernel__neonfp16arith_2x16, .row_tile = 2, .channel_tile = 16, }; xnn_params.f16.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f16_raddstoreexpminusmax_ukernel__neonfp16arith_rr2_p2_x40, .init.f16 = xnn_init_f16_expminus_neonfp16arith_rr2_p2_params, .element_tile = 40, }; xnn_params.f16.rmax = (xnn_rmax_ukernel_function) xnn_f16_rmax_ukernel__neonfp16arith; xnn_params.f16.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vadd_minmax_ukernel__neonfp16arith_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vaddc_minmax_ukernel__neonfp16arith_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vaddc_minmax_ukernel__neonfp16arith_x16, .init.f16_minmax = xnn_init_f16_minmax_neon_params, .element_tile = 16, }; xnn_params.f16.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vdiv_minmax_ukernel__neonfp16arith_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vdivc_minmax_ukernel__neonfp16arith_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vrdivc_minmax_ukernel__neonfp16arith_x8, .init.f16_minmax = xnn_init_f16_minmax_neon_params, .element_tile = 8, }; xnn_params.f16.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmax_ukernel__neonfp16arith_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmaxc_ukernel__neonfp16arith_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmaxc_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmin_ukernel__neonfp16arith_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vminc_ukernel__neonfp16arith_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vminc_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmul_minmax_ukernel__neonfp16arith_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmulc_minmax_ukernel__neonfp16arith_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmulc_minmax_ukernel__neonfp16arith_x16, .init.f16_minmax = xnn_init_f16_minmax_neon_params, .element_tile = 16, }; xnn_params.f16.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsub_minmax_ukernel__neonfp16arith_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsubc_minmax_ukernel__neonfp16arith_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vrsubc_minmax_ukernel__neonfp16arith_x16, .init.f16_minmax = xnn_init_f16_minmax_neon_params, .element_tile = 16, }; xnn_params.f16.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsqrdiff_ukernel__neonfp16arith_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsqrdiffc_ukernel__neonfp16arith_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsqrdiffc_ukernel__neonfp16arith_x16, .init.f16_minmax = xnn_init_f16_minmax_neon_params, .element_tile = 16, }; xnn_params.f16.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f16_vmulcaddc_minmax_ukernel_c8__neonfp16arith_2x, .init.f16 = xnn_init_f16_minmax_neon_params, .channel_tile = 8, .row_tile = 2, }; xnn_params.f16.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vabs_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vclamp_ukernel__neonfp16arith_x16, .init.f16_minmax = xnn_init_f16_minmax_neon_params, .element_tile = 16, }; xnn_params.f16.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_velu_ukernel__neonfp16arith_rr1_p3_x16, .init.f16_elu = xnn_init_f16_elu_neonfp16arith_rr1_p3_params, .element_tile = 16, }; xnn_params.f16.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vhswish_ukernel__neonfp16arith_x16, .init.f16_hswish = xnn_init_f16_hswish_neon_params, .element_tile = 16, }; xnn_params.f16.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vlrelu_ukernel__neonfp16arith_x16, .init.f16_lrelu = xnn_init_f16_lrelu_neon_params, .element_tile = 16, }; xnn_params.f16.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vneg_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndne_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndz_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndu_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndd_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vsigmoid_ukernel__neonfp16arith_rr2_p2_nr1fma_x40, .init.f16_sigmoid = xnn_init_f16_sigmoid_neonfp16arith_rr2_p2_params, .element_tile = 40, }; xnn_params.f16.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vsqr_ukernel__neonfp16arith_x16, .element_tile = 16, }; xnn_params.f16.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vsqrt_ukernel__neonfp16arith_sqrt_x8, .element_tile = 8, }; #ifndef XNN_NO_NCHW_OPERATORS init_flags |= XNN_INIT_FLAG_CHW_OPT; xnn_params.f16.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f16_spmm_minmax_ukernel_32x1__neonfp16arith, .mr = 32, .nr = 1, }; xnn_params.f16.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f16_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfp16arith_2x2, .output_channel_tile = 4, .output_height_tile = 2, .output_width_tile = 2, }; xnn_params.f16.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f16_dwconv2d_chw_ukernel_3x3p1__neonfp16arith_2x8, .output_width_tile = 8, .output_height_tile = 2, }; xnn_params.f16.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f16_dwconv2d_chw_ukernel_3x3s2p1__neonfp16arith_1x4, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f16.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f16_dwconv2d_chw_ukernel_5x5p2__neonfp16arith_1x4, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f16.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f16_dwconv2d_chw_ukernel_5x5s2p2__neonfp16arith_1x4, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f16.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f16_gavgpool_cw_ukernel__neonfp16arith_x4, .channel_tile = 4, }; xnn_params.f16.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f16_ibilinear_chw_ukernel__neonfp16arith_p8, .channel_tile = 1, .pixel_tile = 8, }; #endif // XNN_NO_NCHW_OPERATORS } #endif // XNN_ENABLE_ARM_FP16 #endif // XNN_NO_F16_OPERATORS /**************************** F32 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_F32_OPERATORS init_flags |= XNN_INIT_FLAG_F32; #if XNN_ENABLE_ASSEMBLY && !XNN_PLATFORM_IOS && !XNN_PLATFORM_MAC switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_cortex_a72: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a57: case cpuinfo_uarch_cortex_a75: case cpuinfo_uarch_cortex_a76: case cpuinfo_uarch_exynos_m3: case cpuinfo_uarch_exynos_m4: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75); #if XNN_ENABLE_GEMM_M_SPECIALIZATION xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a75); #endif xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; #if XNN_ENABLE_JIT xnn_params.f32.gemm.generator.gemm = xnn_init_hmp_gemm_codegen(xnn_generate_f32_gemm_ukernel_upto6x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.generator.igemm = xnn_init_hmp_igemm_codegen(xnn_generate_f32_igemm_ukernel_upto6x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.generator.gemm1 = xnn_init_hmp_gemm_codegen(xnn_generate_f32_gemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.generator.igemm1 = xnn_init_hmp_igemm_codegen(xnn_generate_f32_igemm_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); #endif break; case cpuinfo_uarch_exynos_m1: case cpuinfo_uarch_exynos_m2: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8s4__neonfma); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8s4__neonfma); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8s4__neonfma); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8s4__neonfma); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; xnn_params.f32.gemm.log2_sr = 2; break; case cpuinfo_uarch_cortex_a53: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a53); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a53); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a55r0: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a35: case cpuinfo_uarch_cortex_a55: case cpuinfo_uarch_kryo: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a73: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a73); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a77: case cpuinfo_uarch_exynos_m5: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a75); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a75); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; break; case cpuinfo_uarch_cortex_a78: case cpuinfo_uarch_cortex_x1: default: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_ld128); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_ld128); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neonfma_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; #if XNN_ENABLE_JIT xnn_params.f32.gemm.generator.gemm = xnn_init_hmp_gemm_codegen(xnn_generate_f32_gemm_ukernel_6x8__aarch64_neonfma_ld128); xnn_params.f32.gemm.generator.igemm = xnn_init_hmp_igemm_codegen(xnn_generate_f32_igemm_ukernel_6x8__aarch64_neonfma_ld128); #endif break; } #if XNN_MAX_UARCH_TYPES > 1 { /* Choose micro-kernels for little cores according to micro-kernel specification for the big core */ const uint32_t mr = xnn_params.f32.gemm.mr; const uint32_t nr = xnn_params.f32.gemm.nr; const uint32_t log2_sr = xnn_params.f32.gemm.log2_sr; for (size_t i = 1; i < XNN_MAX_UARCH_TYPES; i++) { const struct cpuinfo_uarch_info* uarch_info = cpuinfo_get_uarch(i); if (uarch_info == NULL) { /* No more microarchitectures in the system */ break; } switch (uarch_info->uarch) { case cpuinfo_uarch_cortex_a53: if (mr == 6 && nr == 8 && log2_sr == 0) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a53; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53; } else if (mr == 4 && nr == 8 && log2_sr == 0) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_prfm_cortex_a53; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a53; } break; case cpuinfo_uarch_cortex_a55r0: if (mr == 6 && nr == 8 && log2_sr == 0) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; } else if (mr == 4 && nr == 8 && log2_sr == 0) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; } break; case cpuinfo_uarch_cortex_a55: if (mr == 6 && nr == 8 && log2_sr == 0) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_cortex_a55; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; } else if (mr == 4 && nr == 8 && log2_sr == 0) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__aarch64_neonfma_cortex_a55; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)].function[i] = (xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_cortex_a53; } break; default: break; } } } #endif // XNN_MAX_UARCH_TYPES > 1 xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; #else // XNN_ENABLE_ASSEMBLY && !XNN_PLATFORM_IOS && !XNN_PLATFORM_MAC #if XNN_ENABLE_ASSEMBLY xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__aarch64_neonfma_prfm_cortex_a75); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; #else // !XNN_ENABLE_ASSEMBLY xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_6x8__neonfma_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(6)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_6x8__neonfma_lane_ld64); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__neonfma_lane_ld64); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__neonfma_lane_ld64); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 6; xnn_params.f32.gemm.nr = 8; xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__neonfma_lane_ld64); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__neonfma_lane_ld64); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; #endif // XNN_ENABLE_ASSEMBLY #endif // XNN_ENABLE_ASSEMBLY && !XNN_PLATFORM_IOS && !XNN_PLATFORM_MAC xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x3__neonfma; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[0].channel_tile = 8; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__neonfma; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[1].channel_tile = 8; xnn_params.f32.dwconv[1].primary_tile = 4; #if XNN_PLATFORM_IOS || XNN_PLATFORM_MAC xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 8; xnn_params.f32.dwconv[2].primary_tile = 9; #else // !XNN_PLATFORM_IOS && !XNN_PLATFORM_MAC switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_kryo: xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 8; xnn_params.f32.dwconv[2].primary_tile = 9; break; #if XNN_ENABLE_ASSEMBLY case cpuinfo_uarch_cortex_a53: case cpuinfo_uarch_cortex_a55r0: case cpuinfo_uarch_cortex_a55: xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x9__aarch64_neonfma_cortex_a55; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 4; xnn_params.f32.dwconv[2].primary_tile = 9; break; #endif // XNN_ENABLE_ASSEMBLY default: xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__neonfma; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 8; xnn_params.f32.dwconv[2].primary_tile = 9; break; } #endif // XNN_PLATFORM_IOS && XNN_PLATFORM_MAC xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[3].channel_tile = 8; xnn_params.f32.dwconv[3].primary_tile = 25; xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__neon_c4, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__neon_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__neon_c4, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__neon_c4, .init.f32 = xnn_init_f32_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__neon_c4, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__neon_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .update.f32 = xnn_update_f32_scaleminmax_scalar_params, .row_tile = 7, .channel_tile = 4, }; xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__neon_c4, .init.f32 = xnn_init_f32_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__neon_c4, .mr = 4, }; xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__neon_c4, .mr = 9, }; xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__neon_c4, .mr = 9, .qr = 8, }; xnn_params.f32.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__neonfma_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__neonfma_rr1_lut16_p3_x16, .init.f32_elu = xnn_init_f32_elu_neonfma_rr1_lut16_p3_params, .element_tile = 16, }; xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__neon_x16, .init.f32_hswish = xnn_init_f32_hswish_scalar_params, .element_tile = 16, }; xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__neon_x8, .init.f32_lrelu = xnn_init_f32_lrelu_scalar_params, .element_tile = 8, }; xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__neonv8_x8, .element_tile = 8, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__neonv8_x8, .element_tile = 8, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__neonv8_x8, .element_tile = 8, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__neonv8_x8, .element_tile = 8, }; xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__neonfma_rr1_lut64_p2_nr2recps_x16, .init.f32_sigmoid = xnn_init_f32_sigmoid_neonfma_rr1_lut64_p2_params, .element_tile = 16, }; xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__neon_sqrt_x4, .element_tile = 4, }; xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__neon_2x8, .row_tile = 2, .channel_tile = 8, }; xnn_params.f32.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f32_raddstoreexpminusmax_ukernel__neonfma_rr1_lut64_p2_x16, .init.f32 = xnn_init_f32_expminus_neonfma_rr1_lut64_p2_params, .element_tile = 16, }; xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__neon; xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__neon_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__neon_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__neon_x8, .element_tile = 8, }; xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__neonfma_2x, .init.f32 = xnn_init_f32_minmax_scalar_params, .channel_tile = 4, .row_tile = 2, }; #ifndef XNN_NO_NCHW_OPERATORS init_flags |= XNN_INIT_FLAG_CHW_OPT; xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_32x1__neonfma_pipelined, .mr = 32, .nr = 1, }; xnn_params.f32.spmm2 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_32x2__neonfma, .mr = 32, .nr = 2, }; xnn_params.f32.spmm4 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_32x4__neonfma, .mr = 32, .nr = 4, }; xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__neonfma_2x2, .output_channel_tile = 4, .output_height_tile = 2, .output_width_tile = 2, }; xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__neonfma_3x4, .output_width_tile = 4, .output_height_tile = 3, }; xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__neonfma_2x4_acc2, .output_width_tile = 4, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__neonfma_4x4, .output_width_tile = 4, .output_height_tile = 4, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__neonfma_1x4_acc2, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__neon_x4, .channel_tile = 4, }; xnn_params.f32.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f32_ibilinear_chw_ukernel__neonfma_p8, .channel_tile = 1, .pixel_tile = 8, }; #endif // XNN_NO_NCHW_OPERATORS #endif // XNN_NO_F32_OPERATORS /*************************** VCVT AArch64 micro-kernels ***************************/ #ifndef XNN_NO_VCVT_OPERATORS init_flags |= XNN_INIT_FLAG_VCVT; xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__neonfp16_x16, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__neonfp16_x16, .element_tile = 16, }; xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__neonv8_x32, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_neonv8_params, .element_tile = 32, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__neonv8_x32, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_neonv8_params, .element_tile = 32, }; xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__neon_x32, .init.qs8_cvt = xnn_init_qs8_cvt_neon_params, .element_tile = 32, }; xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__neon_x32, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_neon_params, .element_tile = 32, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__neon_x32, .init.qu8_cvt = xnn_init_qu8_cvt_neon_params, .element_tile = 32, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__neon_x32, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_neon_params, .element_tile = 32, }; #endif // XNN_NO_VCVT_OPERATORS /**************************** X32 AArch64 micro-kernels ****************************/ #ifndef XNN_NO_X32_OPERATORS init_flags |= XNN_INIT_FLAG_X32; xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__neon; xnn_params.x32.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__neon, .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__neon, .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__neon, .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__neon, }; xnn_params.x32.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x32_transposec_ukernel__4x4_aarch64_neon_tbl, .tile_size = 32, }; #endif // XNN_NO_X32_OPERATORS /**************************** XX AArch64 micro-kernels ****************************/ #ifndef XNN_NO_XX_OPERATORS init_flags |= XNN_INIT_FLAG_XX; xnn_params.xx.copy = (xnn_vunary_ukernel_function) xnn_xx_copy_ukernel__memcpy; xnn_params.xx.fill = (struct fill_parameters) { .ukernel = (xnn_fill_ukernel_function) xnn_xx_fill_ukernel__neon_x64, .row_tile = 1, }; xnn_params.xx.pad = (struct pad_parameters) { .ukernel = (xnn_pad_ukernel_function) xnn_xx_pad_ukernel__neon, .row_tile = 1, }; xnn_params.xx.transpose = (struct transpose_parameters) { .variable_size_ukernel = xnn_xx_transposev_ukernel__1x1_memcpy, .tile_size = 32, }; #endif #elif XNN_ARCH_X86 || XNN_ARCH_X86_64 if (!cpuinfo_has_x86_sse2()) { xnn_log_error("XNNPACK initialization failed: SSE2 is not supported"); return; } /**************************** QC8 x86 micro-kernels ****************************/ #ifndef XNN_NO_QC8_OPERATORS init_flags |= XNN_INIT_FLAG_QC8; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx512_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 16; xnn_params.qc8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_xop()) { // XOP should be checked before AVX2: AMD Excavator supports both, but performs better with XOP microkernels xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld64); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 4; xnn_params.qc8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_3x8c8__avx2); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_3x8c8__avx2); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x8c8__avx2); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x8c8__avx2); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx2_params; xnn_params.qc8.gemm.mr = 3; xnn_params.qc8.gemm.nr = 8; xnn_params.qc8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_avx()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x4c8__avx_ld128); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld128); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 4; xnn_params.qc8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld64); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.gemm.mr = 3; xnn_params.qc8.gemm.nr = 4; xnn_params.qc8.gemm.log2_kr = 3; } else { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse2_params; xnn_params.qc8.gemm.mr = 3; xnn_params.qc8.gemm.nr = 4; xnn_params.qc8.gemm.log2_kr = 3; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up32x3__avx512skx_mul32; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx512_params; xnn_params.qc8.dwconv[0].channel_tile = 32; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx512_params; xnn_params.qc8.dwconv[1].channel_tile = 32; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx512_params; xnn_params.qc8.dwconv[2].channel_tile = 32; } else if (cpuinfo_has_x86_xop()) { // XOP should be checked before AVX2: AMD Excavator supports both, but performs better with XOP microkernels xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__xop_mul16_add16; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[0].channel_tile = 16; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[1].channel_tile = 16; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul16_add16; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[2].channel_tile = 16; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx2_mul32; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx2_params; xnn_params.qc8.dwconv[0].channel_tile = 16; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx2_params; xnn_params.qc8.dwconv[1].channel_tile = 16; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_avx2_params; xnn_params.qc8.dwconv[2].channel_tile = 16; } else if (cpuinfo_has_x86_avx()) { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__avx_mul16_add16; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[0].channel_tile = 16; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16_add16; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[1].channel_tile = 16; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul16_add16; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[2].channel_tile = 16; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__sse41_mul16; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[0].channel_tile = 8; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[1].channel_tile = 8; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul16; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse4_params; xnn_params.qc8.dwconv[2].channel_tile = 8; } else if (cpuinfo_has_x86_sse2()) { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x3__sse2_mul16; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse2_params; xnn_params.qc8.dwconv[0].channel_tile = 8; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse2_params; xnn_params.qc8.dwconv[1].channel_tile = 8; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_sse2_params; xnn_params.qc8.dwconv[2].channel_tile = 8; } xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].primary_tile = 25; #endif // XNN_NO_QC8_OPERATORS /**************************** QS8 x86 micro-kernels ****************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QS8; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_avx512_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 16; xnn_params.qs8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_xop()) { // XOP should be checked before AVX2: AMD Excavator supports both, but performs better with XOP microkernels xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld64); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld64); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 4; xnn_params.qs8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_3x8c8__avx2); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_3x8c8__avx2); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x8c8__avx2); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x8c8__avx2); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_avx2_params; xnn_params.qs8.gemm.mr = 3; xnn_params.qs8.gemm.nr = 8; xnn_params.qs8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_avx()) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_2x4c8__avx_ld128); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld128); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 4; xnn_params.qs8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld64); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld64); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.gemm.mr = 3; xnn_params.qs8.gemm.nr = 4; xnn_params.qs8.gemm.log2_kr = 3; } else { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse2_params; xnn_params.qs8.gemm.mr = 3; xnn_params.qs8.gemm.nr = 4; xnn_params.qs8.gemm.log2_kr = 3; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_avx512_params; xnn_params.qs8.dwconv[0].channel_tile = 32; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_avx512_params; xnn_params.qs8.dwconv[1].channel_tile = 32; } else if (cpuinfo_has_x86_xop()) { // XOP should be checked before AVX2: AMD Excavator supports both, but performs better with XOP microkernels xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul16_add16; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.dwconv[0].channel_tile = 16; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul16_add16; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.dwconv[1].channel_tile = 16; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_avx2_params; xnn_params.qs8.dwconv[0].channel_tile = 16; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_avx2_params; xnn_params.qs8.dwconv[1].channel_tile = 16; } else if (cpuinfo_has_x86_avx()) { xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16_add16; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.dwconv[0].channel_tile = 16; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul16_add16; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.dwconv[1].channel_tile = 16; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16_add16; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.dwconv[0].channel_tile = 8; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul16_add16; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse4_params; xnn_params.qs8.dwconv[1].channel_tile = 8; } else if (cpuinfo_has_x86_sse2()) { xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16_add16; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse2_params; xnn_params.qs8.dwconv[0].channel_tile = 8; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16_add16; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_sse2_params; xnn_params.qs8.dwconv[1].channel_tile = 8; } xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].primary_tile = 25; if (cpuinfo_has_x86_sse4_1()) { xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c8, .init.qs8 = xnn_init_qs8_avgpool_minmax_fp32_sse4_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_fp32_sse4_params, .row_tile = 7, .channel_tile = 8, }; } else { xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c8, .init.qs8 = xnn_init_qs8_avgpool_minmax_fp32_sse2_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_fp32_sse2_params, .row_tile = 7, .channel_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__avx512skx_mul32_ld128_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x16, .init.qs8_add = xnn_init_qs8_add_minmax_avx512_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_xop()) { xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__xop_mul32_ld32_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__xop_mul32_ld32_x8, .init.qs8_add = xnn_init_qs8_add_minmax_sse4_mul32_params, .element_tile = 8, }; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16, .init.qs8_add = xnn_init_qs8_add_minmax_avx2_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__avx_mul32_ld32_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__avx_mul32_ld32_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__avx_mul32_ld32_x8, .init.qs8_add = xnn_init_qs8_add_minmax_sse4_mul32_params, .element_tile = 8, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__sse41_mul16_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8, .init.qs8_add = xnn_init_qs8_add_minmax_sse4_mul16_params, .element_tile = 8, }; } else { xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__sse2_mul16_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8, .init.qs8_add = xnn_init_qs8_add_minmax_sse2_params, .element_tile = 8, }; } if (cpuinfo_has_x86_avx()) { xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16, .init.qs8_mul = xnn_init_qs8_mul_minmax_fp32_sse4_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16, .init.qs8_mul = xnn_init_qs8_mul_minmax_fp32_sse4_params, .element_tile = 16, }; } else { xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8, .init.qs8_mul = xnn_init_qs8_mul_minmax_fp32_sse2_params, .element_tile = 8, }; } if (cpuinfo_has_x86_avx2()) { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__avx2_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_avx2_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__avx_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_avx_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__sse41_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_sse2_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__ssse3_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_sse2_params, .element_tile = 32, }; } else { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__sse2_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_sse2_params, .element_tile = 32, }; } #endif // XNN_NO_QS8_OPERATORS /**************************** QU8 x86 micro-kernels ****************************/ #ifndef XNN_NO_QU8_OPERATORS init_flags |= XNN_INIT_FLAG_QU8; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_4x16c8__avx512skx); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_4x16c8__avx512skx); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x16c8__avx512skx); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x16c8__avx512skx); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_avx512_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 16; xnn_params.qu8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_xop()) { // XOP should be checked before AVX2: AMD Excavator supports both, but performs better with XOP microkernels xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__xop_ld64); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__xop_ld64); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.gemm.mr = 2; xnn_params.qu8.gemm.nr = 4; xnn_params.qu8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_3x8c8__avx2); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_3x8c8__avx2); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x8c8__avx2); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x8c8__avx2); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_avx2_params; xnn_params.qu8.gemm.mr = 3; xnn_params.qu8.gemm.nr = 8; xnn_params.qu8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_avx()) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_2x4c8__avx_ld128); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__avx_ld128); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.gemm.mr = 2; xnn_params.qu8.gemm.nr = 4; xnn_params.qu8.gemm.log2_kr = 3; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse41_ld64); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse41_ld64); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.gemm.mr = 3; xnn_params.qu8.gemm.nr = 4; xnn_params.qu8.gemm.log2_kr = 3; } else { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_3x4c8__sse2_ld64); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x4c8__sse2_ld64); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.gemm.mr = 3; xnn_params.qu8.gemm.nr = 4; xnn_params.qu8.gemm.log2_kr = 3; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up32x9__avx512skx_mul32; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_avx512_params; xnn_params.qu8.dwconv[0].channel_tile = 32; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up32x25__avx512skx_mul32; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_avx512_params; xnn_params.qu8.dwconv[1].channel_tile = 32; } else if (cpuinfo_has_x86_xop()) { // XOP should be checked before AVX2: AMD Excavator supports both, but performs better with XOP microkernels xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__xop_mul32; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[0].channel_tile = 16; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__xop_mul32; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[1].channel_tile = 16; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx2_mul32; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_avx2_params; xnn_params.qu8.dwconv[0].channel_tile = 16; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx2_mul32; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_avx2_params; xnn_params.qu8.dwconv[1].channel_tile = 16; } else if (cpuinfo_has_x86_avx()) { xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up16x9__avx_mul16; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[0].channel_tile = 16; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up16x25__avx_mul16; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[1].channel_tile = 16; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse41_mul16; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[0].channel_tile = 8; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse41_mul16; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[1].channel_tile = 8; } else if (cpuinfo_has_x86_sse2()) { xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__sse2_mul16; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[0].channel_tile = 8; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__sse2_mul16; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_sse2_params; xnn_params.qu8.dwconv[1].channel_tile = 8; } xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].primary_tile = 25; xnn_params.qu8.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__sse2_c8, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__sse2_c8, .init.qu8 = xnn_init_qu8_avgpool_minmax_sse2_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 8, }; if (cpuinfo_has_x86_sse4_1()) { xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse41_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse41_c8, .init.qu8 = xnn_init_qu8_avgpool_minmax_fp32_sse4_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_fp32_sse4_params, .row_tile = 7, .channel_tile = 8, }; } else { xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__sse2_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__sse2_c8, .init.qu8 = xnn_init_qu8_avgpool_minmax_fp32_sse2_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_fp32_sse2_params, .row_tile = 7, .channel_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__avx512skx_mul32_ld128_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__avx512skx_mul32_ld128_x16, .init.qu8_add = xnn_init_qu8_add_minmax_avx512_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_xop()) { xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__xop_mul32_ld32_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__xop_mul32_ld32_x8, .init.qu8_add = xnn_init_qu8_add_minmax_sse4_params, .element_tile = 8, }; } else if (cpuinfo_has_x86_avx2()) { xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__avx2_mul32_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16, .init.qu8_add = xnn_init_qu8_add_minmax_avx2_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__avx_mul32_ld32_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__avx_mul32_ld32_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__avx_mul32_ld32_x8, .init.qu8_add = xnn_init_qu8_add_minmax_sse4_params, .element_tile = 8, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__sse41_mul16_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__sse41_mul16_ld64_x8, .init.qu8_add = xnn_init_qu8_add_minmax_sse2_params, .element_tile = 8, }; } else { xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__sse2_mul16_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__sse2_mul16_ld64_x8, .init.qu8_add = xnn_init_qu8_add_minmax_sse2_params, .element_tile = 8, }; } if (cpuinfo_has_x86_avx()) { xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_fp32_ukernel__avx_mul16_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__avx_mul16_ld64_x16, .init.qu8_mul = xnn_init_qu8_mul_minmax_fp32_sse2_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_fp32_ukernel__sse41_mul16_ld64_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__sse41_mul16_ld64_x16, .init.qu8_mul = xnn_init_qu8_mul_minmax_fp32_sse2_params, .element_tile = 16, }; } else { xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_fp32_ukernel__sse2_mul16_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__sse2_mul16_ld64_x8, .init.qu8_mul = xnn_init_qu8_mul_minmax_fp32_sse2_params, .element_tile = 8, }; } if (cpuinfo_has_x86_avx2()) { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__avx2_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_avx2_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__avx_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_avx_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__sse41_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_sse2_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__ssse3_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_sse2_params, .element_tile = 32, }; } else { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__sse2_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_sse2_params, .element_tile = 32, }; } #endif // XNN_NO_QU8_OPERATORS /**************************** U8 x86 micro-kernels ****************************/ #ifndef XNN_NO_S8_OPERATORS init_flags |= XNN_INIT_FLAG_S8; if (cpuinfo_has_x86_sse4_1()) { xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__sse41_x64, .init.s8_minmax = xnn_init_s8_minmax_sse4_params, .element_tile = 64, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__sse41_c16, .pixel_tile = 1, .channel_tile = 16, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__sse41_c16, .init.s8 = xnn_init_s8_minmax_sse4_params, .mr = 9, .qr = 8, }; } else { xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__sse2_x64, .init.s8_minmax = xnn_init_s8_minmax_sse2_params, .element_tile = 64, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__sse2_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__sse2_c16, .init.s8 = xnn_init_s8_minmax_sse2_params, .mr = 9, .qr = 8, }; } #endif // XNN_NO_S8_OPERATORS /**************************** U8 x86 micro-kernels ****************************/ #ifndef XNN_NO_U8_OPERATORS init_flags |= XNN_INIT_FLAG_U8; xnn_params.u8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_u8_vclamp_ukernel__sse2_x64, .init.u8_minmax = xnn_init_u8_minmax_sse2_params, .element_tile = 64, }; if (cpuinfo_has_x86_sse4_1()) { xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__sse41_c16, .pixel_tile = 1, .channel_tile = 16, }; } else { xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__sse2_c8, .pixel_tile = 1, .channel_tile = 8, }; } xnn_params.u8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__sse2_c16, .init.u8 = xnn_init_u8_minmax_sse2_params, .mr = 9, .qr = 8, }; xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; xnn_params.u8.rmax = xnn_u8_rmax_ukernel__sse2; #endif // XNN_NO_U8_OPERATORS /**************************** X8 x86 micro-kernels ****************************/ #ifndef XNN_NO_X8_OPERATORS init_flags |= XNN_INIT_FLAG_X8; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.x8.lut = xnn_x8_lut_ukernel__avx512skx_vpshufb_x64; } else if (cpuinfo_has_x86_avx2()) { xnn_params.x8.lut = xnn_x8_lut_ukernel__avx2_x128; } else if (cpuinfo_has_x86_avx()) { xnn_params.x8.lut = xnn_x8_lut_ukernel__avx_x64; } else { // Note: SSSE3 version is usually slower than scalar xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar_x4; } xnn_params.x8.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__sse2, .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__sse2, .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__sse2, .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__sse2, }; xnn_params.x8.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x8_transposec_ukernel__16x16_reuse_mov_sse2, .tile_size = 32, }; #endif // XNN_NO_X8_OPERATORS /**************************** X16 x86 micro-kernels ****************************/ #ifndef XNN_NO_X16_OPERATORS init_flags |= XNN_INIT_FLAG_X16; xnn_params.x16.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x16_transposec_ukernel__8x8_reuse_multi_sse2, .tile_size = 32, }; #endif // XNN_NO_X16_OPERATORS /**************************** F16 x86 micro-kernels ****************************/ #ifndef XNN_NO_F16_OPERATORS if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx2()) { init_flags |= XNN_INIT_FLAG_F16; xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_4x16__avx2_broadcast); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_4x16__avx2_broadcast); xnn_params.f16.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f16_gemm_minmax_ukernel_1x16__avx2_broadcast); xnn_params.f16.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f16_igemm_minmax_ukernel_1x16__avx2_broadcast); xnn_params.f16.gemm.init.f16 = xnn_init_f16_minmax_avx_params; xnn_params.f16.gemm.mr = 4; xnn_params.f16.gemm.nr = 16; xnn_params.f16.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x3__fma3; xnn_params.f16.dwconv[0].init.f16 = xnn_init_f16_minmax_avx_params; xnn_params.f16.dwconv[0].channel_tile = 16; xnn_params.f16.dwconv[0].primary_tile = 3; xnn_params.f16.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x4__fma3; xnn_params.f16.dwconv[1].init.f16 = xnn_init_f16_minmax_avx_params; xnn_params.f16.dwconv[1].channel_tile = 16; xnn_params.f16.dwconv[1].primary_tile = 4; xnn_params.f16.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up16x9__fma3; xnn_params.f16.dwconv[2].init.f16 = xnn_init_f16_minmax_avx_params; xnn_params.f16.dwconv[2].channel_tile = 16; xnn_params.f16.dwconv[2].primary_tile = 9; xnn_params.f16.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f16_dwconv_minmax_ukernel_up8x25__fma3_acc2; xnn_params.f16.dwconv[3].init.f16 = xnn_init_f16_minmax_avx_params; xnn_params.f16.dwconv[3].channel_tile = 8; xnn_params.f16.dwconv[3].primary_tile = 25; xnn_params.f16.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f16_avgpool_minmax_ukernel_9x__f16c_c8, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f16_avgpool_minmax_ukernel_9p8x__f16c_c8, .init.f16 = xnn_init_f16_scaleminmax_avx_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 8, }; xnn_params.f16.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f16_pavgpool_minmax_ukernel_9x__avx2_c8, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f16_pavgpool_minmax_ukernel_9p8x__avx2_c8, .init.f16 = xnn_init_f16_minmax_avx_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 8, }; xnn_params.f16.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f16_gavgpool_minmax_ukernel_7x__f16c_c8, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f16_gavgpool_minmax_ukernel_7p7x__f16c_c8, .init.f16 = xnn_init_f16_scaleminmax_avx_params, .update.f16 = xnn_update_f16_scaleminmax_avx_params, .row_tile = 7, .channel_tile = 8, }; xnn_params.f16.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f16_maxpool_minmax_ukernel_9p8x__f16c_c8, .init.f16 = xnn_init_f16_minmax_avx_params, .mr = 9, .qr = 8, }; xnn_params.f16.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f16_ibilinear_ukernel__fma3_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.f16.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f16_prelu_ukernel__f16c_2x16, .row_tile = 2, .channel_tile = 16, }; xnn_params.f16.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f16_raddstoreexpminusmax_ukernel__avx2_rr1_p2_x40, .init.f16 = xnn_init_f16_expminus_avx2_rr1_p2_params, .element_tile = 40, }; xnn_params.f16.rmax = (xnn_rmax_ukernel_function) xnn_f16_rmax_ukernel__f16c; xnn_params.f16.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vadd_minmax_ukernel__f16c_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vaddc_minmax_ukernel__f16c_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vaddc_minmax_ukernel__f16c_x16, .init.f16_minmax = xnn_init_f16_minmax_avx_params, .element_tile = 16, }; xnn_params.f16.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vdiv_minmax_ukernel__f16c_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vdivc_minmax_ukernel__f16c_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vrdivc_minmax_ukernel__f16c_x8, .init.f16_minmax = xnn_init_f16_minmax_avx_params, .element_tile = 8, }; xnn_params.f16.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmax_ukernel__f16c_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmaxc_ukernel__f16c_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmaxc_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.f16.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmin_ukernel__f16c_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vminc_ukernel__f16c_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vminc_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.f16.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmul_minmax_ukernel__f16c_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmulc_minmax_ukernel__f16c_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vmulc_minmax_ukernel__f16c_x16, .init.f16_minmax = xnn_init_f16_minmax_avx_params, .element_tile = 16, }; xnn_params.f16.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsub_minmax_ukernel__f16c_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsubc_minmax_ukernel__f16c_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vrsubc_minmax_ukernel__f16c_x16, .init.f16_minmax = xnn_init_f16_minmax_avx_params, .element_tile = 16, }; xnn_params.f16.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsqrdiff_ukernel__f16c_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsqrdiffc_ukernel__f16c_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f16_vsqrdiffc_ukernel__f16c_x16, .init.f16_minmax = xnn_init_f16_minmax_avx_params, .element_tile = 16, }; xnn_params.f16.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f16_vmulcaddc_minmax_ukernel_c8__fma3_2x, .init.f16 = xnn_init_f16_minmax_avx_params, .channel_tile = 8, .row_tile = 2, }; xnn_params.f16.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vabs_ukernel__sse2_x16, .init.f16_abs = xnn_init_f16_abs_sse_params, .element_tile = 16, }; xnn_params.f16.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vclamp_ukernel__f16c_x16, .init.f16_minmax = xnn_init_f16_minmax_avx_params, .element_tile = 16, }; xnn_params.f16.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_velu_ukernel__avx2_rr1_p3_x16, .init.f16_elu = xnn_init_f16_elu_avx2_rr1_p3_params, .element_tile = 16, }; xnn_params.f16.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vhswish_ukernel__f16c_x16, .init.f16_hswish = xnn_init_f16_hswish_avx_params, .element_tile = 16, }; xnn_params.f16.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vlrelu_ukernel__f16c_x16, .init.f16_lrelu = xnn_init_f16_lrelu_avx_params, .element_tile = 16, }; xnn_params.f16.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vneg_ukernel__sse2_x16, .init.f16_neg = xnn_init_f16_neg_sse_params, .element_tile = 16, }; xnn_params.f16.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndne_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.f16.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndz_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.f16.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndu_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.f16.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vrndd_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.f16.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vsigmoid_ukernel__avx2_rr1_p2_rcp_x32, .init.f16_sigmoid = xnn_init_f16_sigmoid_avx2_rr1_p2_params, .element_tile = 32, }; xnn_params.f16.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vsqr_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.f16.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_vsqrt_ukernel__f16c_sqrt_x8, .element_tile = 8, }; } #endif // XNN_NO_F16_OPERATORS /**************************** F32 x86 micro-kernels ****************************/ #ifndef XNN_NO_F32_OPERATORS init_flags |= XNN_INIT_FLAG_F32; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(7)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_7x16__avx512f_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(7)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_7x16__avx512f_broadcast); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 7; xnn_params.f32.gemm.nr = 16; } else if (cpuinfo_has_x86_fma3()) { switch (cpuinfo_get_core(0)->uarch) { case cpuinfo_uarch_zen: case cpuinfo_uarch_dhyana: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x16s4__fma3_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x16s4__fma3_broadcast); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16s4__fma3_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16s4__fma3_broadcast); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 16; xnn_params.f32.gemm.log2_sr = 2; break; default: xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_5x16__fma3_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_5x16__fma3_broadcast); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16__fma3_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16__fma3_broadcast); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.gemm.mr = 5; xnn_params.f32.gemm.nr = 16; break; } } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_5x16__avx_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_5x16__avx_broadcast); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x16__avx_broadcast); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x16__avx_broadcast); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.gemm.mr = 5; xnn_params.f32.gemm.nr = 16; } else { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__sse_load1); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__sse_load1); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__sse_load1); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__sse_load1); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_sse_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; } xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__sse); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__sse); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_sse_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; xnn_params.f32.gemm2.log2_kr = 2; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x3__avx512f; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[0].channel_tile = 16; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[1].channel_tile = 16; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 16; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[3].channel_tile = 16; xnn_params.f32.dwconv[3].primary_tile = 25; } else if (cpuinfo_has_x86_fma3()) { xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x3__fma3; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[0].channel_tile = 16; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x4__fma3; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[1].channel_tile = 16; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x9__fma3; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[2].channel_tile = 16; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__fma3; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[3].channel_tile = 8; xnn_params.f32.dwconv[3].primary_tile = 25; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x3__avx; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[0].channel_tile = 16; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x4__avx; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[1].channel_tile = 16; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up16x9__avx; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[2].channel_tile = 16; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__avx; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_avx_params; xnn_params.f32.dwconv[3].channel_tile = 8; xnn_params.f32.dwconv[3].primary_tile = 25; } else { xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x3__sse; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_sse_params; xnn_params.f32.dwconv[0].channel_tile = 8; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__sse; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_sse_params; xnn_params.f32.dwconv[1].channel_tile = 8; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__sse; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_sse_params; xnn_params.f32.dwconv[2].channel_tile = 8; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__sse; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_sse_params; xnn_params.f32.dwconv[3].channel_tile = 8; xnn_params.f32.dwconv[3].primary_tile = 25; } xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__sse_c4, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__sse_c4, .init.f32 = xnn_init_f32_scaleminmax_sse_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__sse_c4, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__sse_c4, .init.f32 = xnn_init_f32_minmax_sse_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__sse_c4, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__sse_c4, .init.f32 = xnn_init_f32_scaleminmax_sse_params, .update.f32 = xnn_update_f32_scaleminmax_sse_params, .row_tile = 7, .channel_tile = 4, }; xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__sse_c4, .init.f32 = xnn_init_f32_minmax_sse_params, .mr = 9, .qr = 8, }; xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__sse2_c4, .mr = 4, }; xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__sse2_c4, .mr = 9, }; xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__sse2_c4, .mr = 9, .qr = 8, }; xnn_params.f32.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__sse_c8, .pixel_tile = 1, .channel_tile = 8, }; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__avx512f_x16, .init.f32_abs = xnn_init_f32_abs_avx512_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__avx_x16, .init.f32_abs = xnn_init_f32_abs_avx_params, .element_tile = 16, }; } else { xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__sse_x8, .init.f32_abs = xnn_init_f32_abs_sse_params, .element_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__avx512f_x16, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__avx_x16, .init.f32_minmax = xnn_init_f32_minmax_avx_params, .element_tile = 16, }; } else { xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__sse_x8, .init.f32_minmax = xnn_init_f32_minmax_sse_params, .element_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__avx512f_rr1_lut16_p3_perm_x64, .init.f32_elu = xnn_init_f32_elu_avx512_rr1_lut16_p3_params, .element_tile = 64, }; } else if (cpuinfo_has_x86_avx2()) { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__avx2_rr1_lut4_p4_perm_x56, .init.f32_elu = xnn_init_f32_elu_avx2_rr1_lut4_p4_params, .element_tile = 56, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__avx_rr2_lut4_p4_perm_x32, .init.f32_elu = xnn_init_f32_elu_avx_rr2_lut4_p4_params, .element_tile = 32, }; } else { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__sse2_rr2_lut16_p3_x12, .init.f32_elu = xnn_init_f32_elu_sse2_rr2_lut16_p3_params, .element_tile = 12, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__avx512f_x16, .init.f32_hswish = xnn_init_f32_hswish_avx512_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_fma3()) { xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__fma3_x16, .init.f32_hswish = xnn_init_f32_hswish_avx_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__avx_x16, .init.f32_hswish = xnn_init_f32_hswish_avx_params, .element_tile = 16, }; } else { xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__sse_x8, .init.f32_hswish = xnn_init_f32_hswish_sse_params, .element_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__avx512f_x16, .init.f32_lrelu = xnn_init_f32_lrelu_scalar_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__avx_x16, .init.f32_lrelu = xnn_init_f32_lrelu_avx_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__sse41_x8, .init.f32_lrelu = xnn_init_f32_lrelu_sse_params, .element_tile = 8, }; } else { xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__sse_x8, .init.f32_lrelu = xnn_init_f32_lrelu_sse_params, .element_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__avx512f_x16, .init.f32_neg = xnn_init_f32_neg_avx512_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__avx_x16, .init.f32_neg = xnn_init_f32_neg_avx_params, .element_tile = 16, }; } else { xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__sse_x8, .init.f32_neg = xnn_init_f32_neg_sse_params, .element_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__avx512f_x16, .element_tile = 16, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__avx512f_x16, .element_tile = 16, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__avx512f_x16, .element_tile = 16, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__avx512f_x16, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__avx_x16, .init.f32_rnd = xnn_init_f32_rnd_avx_params, .element_tile = 16, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__avx_x16, .init.f32_rnd = xnn_init_f32_rnd_avx_params, .element_tile = 16, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__avx_x16, .init.f32_rnd = xnn_init_f32_rnd_avx_params, .element_tile = 16, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__avx_x16, .init.f32_rnd = xnn_init_f32_rnd_avx_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__sse41_x8, .element_tile = 8, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__sse41_x8, .element_tile = 8, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__sse41_x8, .element_tile = 8, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__sse41_x8, .element_tile = 8, }; } else { xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__sse2_x8, .init.f32_rnd = xnn_init_f32_rnd_sse2_params, .element_tile = 8, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__sse2_x8, .init.f32_rnd = xnn_init_f32_rnd_sse2_params, .element_tile = 8, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__sse2_x8, .init.f32_rnd = xnn_init_f32_rnd_sse2_params, .element_tile = 8, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__sse2_x8, .init.f32_rnd = xnn_init_f32_rnd_sse2_params, .element_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__avx512f_rr2_lut32_p2_perm2_scalef_div_x64, .init.f32_sigmoid = xnn_init_f32_sigmoid_avx512_rr2_lut32_p2_params, .element_tile = 64, }; } else if (cpuinfo_has_x86_avx2()) { xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_div_x40, .init.f32_sigmoid = xnn_init_f32_sigmoid_avx2_rr1_p5_params, .element_tile = 40, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__avx_rr2_p5_nr2_x40, .init.f32_sigmoid = xnn_init_f32_sigmoid_avx_rr2_p5_params, .element_tile = 40, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__sse41_rr2_lut64_p2_div_x8, .init.f32_sigmoid = xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, .element_tile = 8, }; } else { xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__sse2_rr2_lut64_p2_div_x8, .init.f32_sigmoid = xnn_init_f32_sigmoid_sse2_rr2_lut64_p2_params, .element_tile = 8, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__avx512f_x16, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__avx_x16, .init.f32_default = xnn_init_f32_default_avx_params, .element_tile = 16, }; } else { xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__sse_x8, .element_tile = 8, }; } if (cpuinfo_has_x86_avx()) { xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__avx_sqrt_x8, .init.f32_sqrt = xnn_init_f32_sqrt_avx_params, .element_tile = 8, }; } else { xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__sse_sqrt_x4, .element_tile = 4, }; } if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__avx512f_2x16, .row_tile = 2, .channel_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__avx_2x16, .row_tile = 2, .channel_tile = 16, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__sse41_2x8, .row_tile = 2, .channel_tile = 8, }; } else { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__sse2_2x8, .row_tile = 2, .channel_tile = 8, }; } xnn_params.f32.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f32_raddstoreexpminusmax_ukernel__sse2_rr2_p5_x20_acc2, .init.f32 = xnn_init_f32_expminus_sse2_rr2_p5_params, .element_tile = 20, }; xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__sse; if (!XNN_PLATFORM_MOBILE && cpuinfo_has_x86_avx512f()) { xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__avx512f_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx512f_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx512f_x32, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 32, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__avx512f_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__avx512f_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__avx512f_x32, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 32, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__avx512f_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx512f_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx512f_x32, .element_tile = 32, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__avx512f_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx512f_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx512f_x32, .element_tile = 32, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__avx512f_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx512f_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx512f_x32, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 32, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__avx512f_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__avx512f_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__avx512f_x32, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 32, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__avx512f_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx512f_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx512f_x32, .element_tile = 32, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__avx_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__avx_x16, .init.f32_minmax = xnn_init_f32_minmax_avx_params, .element_tile = 16, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__avx_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__avx_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__avx_x16, .init.f32_minmax = xnn_init_f32_minmax_avx_params, .element_tile = 16, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__avx_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__avx_x16, .init.f32_default = xnn_init_f32_default_avx_params, .element_tile = 16, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__avx_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__avx_x16, .init.f32_default = xnn_init_f32_default_avx_params, .element_tile = 16, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__avx_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__avx_x16, .init.f32_minmax = xnn_init_f32_minmax_avx_params, .element_tile = 16, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__avx_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__avx_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__avx_x16, .init.f32_minmax = xnn_init_f32_minmax_avx_params, .element_tile = 16, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__avx_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__avx_x16, .init.f32_default = xnn_init_f32_default_avx_params, .element_tile = 16, }; } else { xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__sse_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__sse_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__sse_x8, .init.f32_minmax = xnn_init_f32_minmax_sse_params, .element_tile = 8, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__sse_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__sse_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__sse_x8, .init.f32_minmax = xnn_init_f32_minmax_sse_params, .element_tile = 8, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__sse_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__sse_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__sse_x8, .element_tile = 8, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__sse_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__sse_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__sse_x8, .element_tile = 8, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__sse_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__sse_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__sse_x8, .init.f32_minmax = xnn_init_f32_minmax_sse_params, .element_tile = 8, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__sse_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__sse_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__sse_x8, .init.f32_minmax = xnn_init_f32_minmax_sse_params, .element_tile = 8, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__sse_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__sse_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__sse_x8, .element_tile = 8, }; } xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__sse_2x, .init.f32 = xnn_init_f32_minmax_sse_params, .channel_tile = 4, .row_tile = 2, }; #ifndef XNN_NO_NCHW_OPERATORS // Sparse microkernels on x86 currently target only SSE, and on processors // with AVX ISA dense inference is expected to be faster than sparse. if (!cpuinfo_has_x86_avx()) { init_flags |= XNN_INIT_FLAG_CHW_OPT; } xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_32x1__sse, .mr = 32, .nr = 1, }; xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__sse_2x2, .output_channel_tile = 4, .output_height_tile = 2, .output_width_tile = 2, }; if (cpuinfo_has_x86_ssse3()) { xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2, .output_width_tile = 4, .output_height_tile = 2, }; } else { xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__sse_2x4_acc2, .output_width_tile = 4, .output_height_tile = 2, }; } xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__sse_1x4_acc3, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__sse_4x4, .output_width_tile = 4, .output_height_tile = 4, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__sse_2x4, .output_width_tile = 4, .output_height_tile = 2, }; xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__sse_x4, .channel_tile = 4, }; xnn_params.f32.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f32_ibilinear_chw_ukernel__sse_p8, .channel_tile = 1, .pixel_tile = 8, }; #endif // XNN_NO_NCHW_OPERATORS #endif // XNN_NO_F32_OPERATORS /*************************** VCVT x86 micro-kernels ***************************/ #ifndef XNN_NO_VCVT_OPERATORS init_flags |= XNN_INIT_FLAG_VCVT; if (cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__avx512skx_x16, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__avx512skx_x16, .element_tile = 16, }; } else if (cpuinfo_has_x86_f16c()) { xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__f16c_x16, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__f16c_x16, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_f16c_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__avx_int16_x16, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_sse_int16_params, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__avx_x24, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_sse2_params, .element_tile = 24, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__sse41_int16_x16, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_sse_int16_params, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__sse41_x8, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_sse2_params, .element_tile = 8, }; } else { xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__sse2_int16_x32, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_sse_int16_params, .element_tile = 32, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__sse2_x16, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_sse2_params, .element_tile = 16, }; } if (cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__avx512skx_x128, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_avx512_params, .element_tile = 128, }; } else if (cpuinfo_has_x86_avx2()) { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__avx2_x64, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_avx2_params, .element_tile = 64, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__avx_x32, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_avx_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__sse41_x32, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_sse4_params, .element_tile = 32, }; } else { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__sse2_x32, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_sse2_params, .element_tile = 32, }; } if (cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__avx512skx_x128, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_avx512_params, .element_tile = 128, }; } else if (cpuinfo_has_x86_avx2()) { xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__avx2_x64, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_avx2_params, .element_tile = 64, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__avx_x32, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_avx_params, .element_tile = 32, }; } else { xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__sse2_x32, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_sse2_params, .element_tile = 32, }; } if (cpuinfo_has_x86_avx2()) { xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__avx2_x32, .init.qs8_cvt = xnn_init_qs8_cvt_avx2_params, .element_tile = 32, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__avx2_x32, .init.qu8_cvt = xnn_init_qu8_cvt_avx2_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__avx_x32, .init.qs8_cvt = xnn_init_qs8_cvt_ssse3_params, .element_tile = 32, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__avx_x32, .init.qu8_cvt = xnn_init_qu8_cvt_ssse3_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__sse41_x32, .init.qs8_cvt = xnn_init_qs8_cvt_ssse3_params, .element_tile = 32, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__sse41_x32, .init.qu8_cvt = xnn_init_qu8_cvt_ssse3_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_ssse3()) { xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__ssse3_x32, .init.qs8_cvt = xnn_init_qs8_cvt_ssse3_params, .element_tile = 32, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__ssse3_x32, .init.qu8_cvt = xnn_init_qu8_cvt_ssse3_params, .element_tile = 32, }; } else { xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__sse2_x32, .init.qs8_cvt = xnn_init_qs8_cvt_sse2_params, .element_tile = 32, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__sse2_x32, .init.qu8_cvt = xnn_init_qu8_cvt_sse2_params, .element_tile = 32, }; } if (cpuinfo_has_x86_avx512f() && cpuinfo_has_x86_avx512bw() && cpuinfo_has_x86_avx512dq() && cpuinfo_has_x86_avx512vl()) { xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__avx512skx_x32, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_avx512_params, .element_tile = 32, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__avx512skx_x32, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_avx512_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_avx2()) { xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__avx2_x16, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_avx_params, .element_tile = 16, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__avx2_x16, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_avx_params, .element_tile = 16, }; } else if (cpuinfo_has_x86_avx()) { xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__avx_x32, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_avx_params, .element_tile = 32, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__avx_x32, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_avx_params, .element_tile = 32, }; } else if (cpuinfo_has_x86_sse4_1()) { xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__sse41_x16, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_sse4_params, .element_tile = 16, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__sse41_x16, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_sse4_params, .element_tile = 16, }; } else { xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__sse2_x32, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_sse2_params, .element_tile = 32, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__sse2_x32, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_sse2_params, .element_tile = 32, }; } #endif // XNN_NO_VCVT_OPERATORS /**************************** X32 x86 micro-kernels ****************************/ #ifndef XNN_NO_X32_OPERATORS init_flags |= XNN_INIT_FLAG_X32; xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__sse2; xnn_params.x32.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__sse2, .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__sse2, .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__sse2, .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__sse2, }; xnn_params.x32.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x32_transposec_ukernel__4x4_sse, .tile_size = 32, }; #endif // XNN_NO_X32_OPERATORS /**************************** XX x86 micro-kernels ****************************/ #ifndef XNN_NO_XX_OPERATORS init_flags |= XNN_INIT_FLAG_XX; xnn_params.xx.copy = (xnn_vunary_ukernel_function) xnn_xx_copy_ukernel__memcpy; xnn_params.xx.fill = (struct fill_parameters) { .ukernel = (xnn_fill_ukernel_function) xnn_xx_fill_ukernel__sse2_x64, .row_tile = 1, }; xnn_params.xx.pad = (struct pad_parameters) { .ukernel = (xnn_pad_ukernel_function) xnn_xx_pad_ukernel__sse2, .row_tile = 1, }; xnn_params.xx.transpose = (struct transpose_parameters) { .variable_size_ukernel = xnn_xx_transposev_ukernel__1x1_memcpy, .tile_size = 32, }; #endif #elif XNN_ARCH_WASMSIMD || XNN_ARCH_WASMRELAXEDSIMD /**************************** QC8 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QC8; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_wasmsimd_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 4; xnn_params.qc8.gemm.log2_kr = 1; xnn_params.qc8.gemm.log2_sr = 2; xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x3__wasmsimd_mul16_add16; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_wasmsimd_params; xnn_params.qc8.dwconv[0].channel_tile = 16; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x9__wasmsimd_mul16_add16; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_wasmsimd_params; xnn_params.qc8.dwconv[1].channel_tile = 16; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16_add16; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_wasmsimd_params; xnn_params.qc8.dwconv[2].channel_tile = 16; xnn_params.qc8.dwconv[2].primary_tile = 25; #endif // XNN_NO_QC8_OPERATORS /**************************** QS8 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QS8; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_wasmsimd_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 4; xnn_params.qs8.gemm.log2_kr = 1; xnn_params.qs8.gemm.log2_sr = 2; xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x9__wasmsimd_mul16_add16; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_wasmsimd_params; xnn_params.qs8.dwconv[0].channel_tile = 16; xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up16x25__wasmsimd_mul16_add16; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_wasmsimd_params; xnn_params.qs8.dwconv[1].channel_tile = 16; xnn_params.qs8.dwconv[1].primary_tile = 25; xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__wasmsimd_c16, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16, .init.qs8 = xnn_init_qs8_avgpool_minmax_fp32_wasmsimd_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_fp32_wasmsimd_params, .row_tile = 7, .channel_tile = 16, }; xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__wasmsimd_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__wasmsimd_x32, .init.qs8_add = xnn_init_qs8_add_minmax_wasmsimd_params, .element_tile = 32, }; xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8, .init.qs8_mul = xnn_init_qs8_mul_minmax_fp32_wasmsimd_params, .element_tile = 8, }; #if XNN_ARCH_WASMRELAXEDSIMD if (is_wasm_x86) { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__wasmrelaxedsimd_x86_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_wasmsimd_x86_params, .element_tile = 32, }; } else { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__wasmrelaxedsimd_arm_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_wasmsimd_arm_params, .element_tile = 32, }; } #else if (is_wasm_x86) { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__wasmsimd_x86_x16, .init.qs8_lrelu = xnn_init_qs8_lrelu_wasmsimd_x86_params, .element_tile = 16, }; } else { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__wasmsimd_arm_x32, .init.qs8_lrelu = xnn_init_qs8_lrelu_wasmsimd_arm_params, .element_tile = 32, }; } #endif #endif // XNN_NO_QS8_OPERATORS /**************************** QU8 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_QU8_OPERATORS init_flags |= XNN_INIT_FLAG_QU8; xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x4c2s4__wasmsimd_dot16x2_ld128); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_wasmsimd_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 4; xnn_params.qu8.gemm.log2_kr = 1; xnn_params.qu8.gemm.log2_sr = 2; xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up8x9__wasmsimd_mul16; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_wasmsimd_params; xnn_params.qu8.dwconv[0].channel_tile = 8; xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up8x25__wasmsimd_mul16; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_wasmsimd_params; xnn_params.qu8.dwconv[1].channel_tile = 8; xnn_params.qu8.dwconv[1].primary_tile = 25; xnn_params.qu8.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1, .init.qu8 = xnn_init_qu8_avgpool_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__wasmsimd_c16, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__wasmsimd_c16, .init.qu8 = xnn_init_qu8_avgpool_minmax_fp32_wasmsimd_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_fp32_wasmsimd_params, .row_tile = 7, .channel_tile = 16, }; xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__wasmsimd_x32, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x32, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__wasmsimd_x32, .init.qu8_add = xnn_init_qu8_add_minmax_wasmsimd_params, .element_tile = 32, }; xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__wasmsimd_mul32_ld64_x8, .init.qu8_mul = xnn_init_qu8_mul_minmax_fp32_wasmsimd_params, .element_tile = 8, }; #if XNN_ARCH_WASMRELAXEDSIMD if (is_wasm_x86) { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__wasmrelaxedsimd_x86_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_wasmsimd_x86_params, .element_tile = 32, }; } else { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__wasmrelaxedsimd_arm_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_wasmsimd_arm_params, .element_tile = 32, }; } #else if (is_wasm_x86) { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__wasmsimd_x86_x16, .init.qu8_lrelu = xnn_init_qu8_lrelu_wasmsimd_x86_params, .element_tile = 16, }; } else { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__wasmsimd_arm_x32, .init.qu8_lrelu = xnn_init_qu8_lrelu_wasmsimd_arm_params, .element_tile = 32, }; } #endif #endif // XNN_NO_QU8_OPERATORS /**************************** S8 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_S8_OPERATORS init_flags |= XNN_INIT_FLAG_S8; xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__wasmsimd_x64, .init.s8_minmax = xnn_init_s8_minmax_wasmsimd_params, .element_tile = 64, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__wasmsimd_dot16x2_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16, .init.s8 = xnn_init_s8_minmax_wasmsimd_params, .mr = 9, .qr = 8, }; #endif // XNN_NO_S8_OPERATORS /**************************** U8 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_U8_OPERATORS init_flags |= XNN_INIT_FLAG_U8; xnn_params.u8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_u8_vclamp_ukernel__wasmsimd_x64, .init.u8_minmax = xnn_init_u8_minmax_wasmsimd_params, .element_tile = 64, }; xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__wasmsimd_dot16x2_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.u8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__wasmsimd_c16, .init.u8 = xnn_init_u8_minmax_wasmsimd_params, .mr = 9, .qr = 8, }; xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; #endif // XNN_NO_U8_OPERATORS /**************************** X8 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_X8_OPERATORS init_flags |= XNN_INIT_FLAG_X8; xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar_x4; xnn_params.x8.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, }; xnn_params.x8.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x8_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X8_OPERATORS /**************************** X16 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_X16_OPERATORS init_flags |= XNN_INIT_FLAG_X16; xnn_params.x16.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x16_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X16_OPERATORS /**************************** F32 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_F32_OPERATORS init_flags |= XNN_INIT_FLAG_F32; if (is_wasm_x86) { #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__wasmrelaxedsimd_fma_loadsplat); #else xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x8__wasmsimd_x86_splat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x8__wasmsimd_x86_splat); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_x86_splat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__wasmsimd_x86_splat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x8__wasmsimd_splat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x8__wasmsimd_splat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x8__wasmsimd_splat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x8__wasmsimd_splat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x8__wasmsimd_splat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x8__wasmsimd_splat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__wasmsimd_splat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__wasmsimd_splat); #endif xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 8; #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__wasmrelaxedsimd_fma); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__wasmrelaxedsimd_fma); xnn_params.f32.gemm2.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2c4__wasmrelaxedsimd_fma); xnn_params.f32.gemm2.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__wasmrelaxedsimd_fma); #else xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_x86); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__wasmsimd_x86); xnn_params.f32.gemm2.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2c4__wasmsimd); xnn_params.f32.gemm2.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__wasmsimd); #endif xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; xnn_params.f32.gemm2.log2_kr = 2; } else { #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_5x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_5x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_5x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_5x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_5x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_5x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__wasmrelaxedsimd_fma_splat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__wasmrelaxedsimd_fma_splat); #else xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_5x8__wasmsimd_arm_splat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_5x8__wasmsimd_arm_splat); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x8__wasmsimd_arm_splat); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x8__wasmsimd_arm_splat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_5x8__wasmsimd_splat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_5x8__wasmsimd_splat); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x8__wasmsimd_splat); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x8__wasmsimd_splat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_5x8__wasmsimd_splat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(5)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_5x8__wasmsimd_splat); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x8__wasmsimd_splat); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x8__wasmsimd_splat); #endif xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.gemm.mr = 5; xnn_params.f32.gemm.nr = 8; #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__wasmrelaxedsimd_fma); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__wasmrelaxedsimd_fma); xnn_params.f32.gemm2.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2c4__wasmrelaxedsimd_fma); xnn_params.f32.gemm2.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__wasmrelaxedsimd_fma); #else xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2c4__wasmsimd_arm); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2c4__wasmsimd_arm); xnn_params.f32.gemm2.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2c4__wasmsimd); xnn_params.f32.gemm2.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2c4__wasmsimd); #endif xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; xnn_params.f32.gemm2.log2_kr = 2; } #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x3__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x3__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[0].channel_tile = 8; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x4__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[1].channel_tile = 8; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x9__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[2].channel_tile = 8; xnn_params.f32.dwconv[2].primary_tile = 9; #else if (is_wasm_x86) { xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x3__wasmsimd_x86; xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x3__wasmsimd; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[0].channel_tile = 8; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x4__wasmsimd_x86; xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x4__wasmsimd; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[1].channel_tile = 8; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x9__wasmsimd_x86; xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x9__wasmsimd; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[2].channel_tile = 8; xnn_params.f32.dwconv[2].primary_tile = 9; } else { xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x3__wasmsimd_arm; xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up4x3__wasmsimd; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[0].channel_tile = 4; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x4__wasmsimd_arm; xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up4x4__wasmsimd; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[1].channel_tile = 4; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x9__wasmsimd_arm; xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up4x9__wasmsimd; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[2].channel_tile = 4; xnn_params.f32.dwconv[2].primary_tile = 9; } #endif #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up8x25__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[3].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up8x25__wasmrelaxedsimd_fma; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[3].channel_tile = 8; xnn_params.f32.dwconv[3].primary_tile = 25; #else xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm; xnn_params.f32.dwconv[3].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up4x25__wasmsimd; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_wasmsimd_params; xnn_params.f32.dwconv[3].channel_tile = 4; xnn_params.f32.dwconv[3].primary_tile = 25; #endif if (is_wasm_x86) { xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_x86_c4, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_x86_c4, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_x86_c4, .init.f32 = xnn_init_f32_minmax_wasmsimd_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_x86_c4, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_x86_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .update.f32 = xnn_update_f32_scaleminmax_scalar_params, .row_tile = 7, .channel_tile = 4, }; } else { xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__wasmsimd_arm_c4, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__wasmsimd_arm_c4, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__wasmsimd_arm_c4, .init.f32 = xnn_init_f32_minmax_wasmsimd_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 4, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__wasmsimd_arm_c4, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__wasmsimd_arm_c4, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .update.f32 = xnn_update_f32_scaleminmax_scalar_params, .row_tile = 7, .channel_tile = 4, }; } if (is_wasm_x86) { xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_x86_c4, .init.f32 = xnn_init_f32_minmax_wasmsimd_params, .mr = 9, .qr = 8, }; } else { xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__wasmsimd_arm_c4, .init.f32 = xnn_init_f32_minmax_wasmsimd_params, .mr = 9, .qr = 8, }; } xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__wasmsimd_c4, .mr = 4, }; xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__wasmsimd_c4, .mr = 9, }; xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__wasmsimd_c4, .mr = 9, .qr = 8, }; xnn_params.f32.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__wasmsimd_c8, .pixel_tile = 1, .channel_tile = 8, }; xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__wasmsimd_x8, .init.f32_abs = xnn_init_f32_abs_wasmsimd_params, .element_tile = 16, }; if (is_wasm_x86) { xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__wasmsimd_x86_x8, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 8, }; } else { xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__wasmsimd_arm_x8, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 8, }; } if (is_wasm_x86) { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__wasmsimd_x86_rr2_p6_x20, .init.f32_elu = xnn_init_f32_elu_wasmsimd_rr2_p6_params, .element_tile = 20, }; } else { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__wasmsimd_arm_rr2_p6_x20, .init.f32_elu = xnn_init_f32_elu_wasmsimd_rr2_p6_params, .element_tile = 20, }; } xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__wasmsimd_x16, .init.f32_hswish = xnn_init_f32_hswish_wasmsimd_params, .element_tile = 16, }; if (is_wasm_x86) { xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__wasmsimd_minmax_x8, .init.f32_lrelu = xnn_init_f32_lrelu_wasmsimd_params, .element_tile = 8, }; } else { xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__wasmsimd_bitselect_x8, .init.f32_lrelu = xnn_init_f32_lrelu_wasmsimd_params, .element_tile = 8, }; } xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__wasmsimd_x8, .init.f32_neg = xnn_init_f32_neg_wasmsimd_params, .element_tile = 16, }; xnn_params.f32.relu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrelu_ukernel__wasmsimd_x16, .element_tile = 16, }; xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__wasmsimd_x8, .element_tile = 8, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__wasmsimd_x8, .element_tile = 8, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__wasmsimd_x8, .element_tile = 8, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__wasmsimd_x8, .element_tile = 8, }; xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__wasmsimd_rr2_p5_div_x16, .init.f32_sigmoid = xnn_init_f32_sigmoid_wasmsimd_rr2_p5_params, .element_tile = 16, }; xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__wasmsimd_x8, .element_tile = 16, }; xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__wasmsimd_sqrt_x8, .element_tile = 8, }; if (is_wasm_x86) { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__wasmsimd_minmax_2x8, .row_tile = 2, .channel_tile = 8, }; } else { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__wasmsimd_bitselect_2x8, .row_tile = 2, .channel_tile = 8, }; } xnn_params.f32.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f32_raddstoreexpminusmax_ukernel__wasmsimd_rr2_p5_x16_acc2, .init.f32 = xnn_init_f32_expminus_wasmsimd_rr2_p5_params, .element_tile = 16, }; if (is_wasm_x86) { xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__wasmsimd_x86; xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__wasmsimd_x86_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_x86_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__wasmsimd_x86_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__wasmsimd_x86_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__wasmsimd_x86_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__wasmsimd_x86_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_x86_x16, .element_tile = 16, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__wasmsimd_x86_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_x86_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_x86_x16, .element_tile = 16, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__wasmsimd_x86_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_x86_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__wasmsimd_x86_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__wasmsimd_x86_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__wasmsimd_x86_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; } else { xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__wasmsimd_arm; xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__wasmsimd_arm_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_arm_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasmsimd_arm_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__wasmsimd_arm_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__wasmsimd_arm_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__wasmsimd_arm_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__wasmsimd_arm_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_arm_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasmsimd_arm_x16, .element_tile = 16, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__wasmsimd_arm_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_arm_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasmsimd_arm_x16, .element_tile = 16, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__wasmsimd_arm_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_arm_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasmsimd_arm_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__wasmsimd_arm_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__wasmsimd_arm_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__wasmsimd_arm_x16, .linear.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_ukernel__wasmsimd_x16, .linear.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_ukernel__wasmsimd_x16, .linear.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_ukernel__wasmsimd_x16, .init.f32_minmax = xnn_init_f32_minmax_wasmsimd_params, .element_tile = 16, }; } xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__wasmsimd_x16, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__wasmsimd_x16, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__wasmsimd_x16, .element_tile = 16, }; if (is_wasm_x86) { xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__wasmsimd_x86_2x, .init.f32 = xnn_init_f32_minmax_wasmsimd_params, .channel_tile = 4, .row_tile = 2, }; } else { xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c4__wasmsimd_arm_2x, .init.f32 = xnn_init_f32_minmax_wasmsimd_params, .channel_tile = 4, .row_tile = 2, }; } #ifndef XNN_NO_NCHW_OPERATORS init_flags |= XNN_INIT_FLAG_CHW_OPT; if (is_wasm_x86) { xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_x86, .mr = 32, .nr = 1, }; } else { xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_32x1__wasmsimd_arm, .mr = 32, .nr = 1, }; } xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__wasmsimd_2x2, .output_channel_tile = 4, .output_height_tile = 2, .output_width_tile = 2, }; if (is_wasm_x86) { xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_x86_loadsplat_2x4, .output_width_tile = 4, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_x86_splat_1x4_acc2, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_x86_splat_3x4, .output_width_tile = 4, .output_height_tile = 3, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_x86_splat_1x4_acc2, .output_width_tile = 4, .output_height_tile = 1, }; } else { xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__wasmsimd_arm_loadsplat_2x4, .output_width_tile = 4, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__wasmsimd_arm_splat_1x4_acc4, .output_width_tile = 4, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__wasmsimd_arm_splat_3x4, .output_width_tile = 4, .output_height_tile = 3, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__wasmsimd_arm_splat_1x4_acc2, .output_width_tile = 4, .output_height_tile = 1, }; } if (is_wasm_x86) { xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__wasmsimd_x86_x4, .channel_tile = 4, }; } else { xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__wasmsimd_arm_x4, .channel_tile = 4, }; } xnn_params.f32.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f32_ibilinear_chw_ukernel__wasmsimd_p8, .channel_tile = 1, .pixel_tile = 8, }; #endif // XNN_NO_NCHW_OPERATORS #endif // XNN_NO_F32_OPERATORS /*************************** VCVT WAsm SIMD micro-kernels***************************/ #ifndef XNN_NO_VCVT_OPERATORS init_flags |= XNN_INIT_FLAG_VCVT; xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__wasmsimd_int16_x16, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_wasmsimd_int16_params, .element_tile = 16, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__wasmsimd_x24, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_wasmsimd_params, .element_tile = 24, }; xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__wasmsimd_magic_x32, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_wasmsimd_magic_params, .element_tile = 32, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__wasmsimd_magic_x32, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_wasmsimd_magic_params, .element_tile = 32, }; #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__wasmrelaxedsimd_x32, .init.qs8_cvt = xnn_init_qs8_cvt_wasmsimd_params, .element_tile = 32, }; #else xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__wasmsimd_x16, .init.qs8_cvt = xnn_init_qs8_cvt_wasmsimd_params, .element_tile = 16, }; #endif xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__wasmsimd_x32, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_wasmsimd_params, .element_tile = 32, }; #if XNN_ARCH_WASMRELAXEDSIMD xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__wasmrelaxedsimd_x32, .init.qu8_cvt = xnn_init_qu8_cvt_wasmsimd_params, .element_tile = 32, }; #else xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__wasmsimd_x16, .init.qu8_cvt = xnn_init_qu8_cvt_wasmsimd_params, .element_tile = 16, }; #endif xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__wasmsimd_x32, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_wasmsimd_params, .element_tile = 32, }; #endif // XNN_NO_VCVT_OPERATORS /**************************** X32 WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_X32_OPERATORS init_flags |= XNN_INIT_FLAG_X32; xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__wasmsimd; xnn_params.x32.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__wasmsimd, .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__wasmsimd, .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__wasmsimd, .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__wasmsimd, }; xnn_params.x32.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x32_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X32_OPERATORS /**************************** XX WAsm SIMD micro-kernels****************************/ #ifndef XNN_NO_XX_OPERATORS init_flags |= XNN_INIT_FLAG_XX; xnn_params.xx.copy = (xnn_vunary_ukernel_function) xnn_xx_copy_ukernel__memcpy; xnn_params.xx.fill = (struct fill_parameters) { .ukernel = (xnn_fill_ukernel_function) xnn_xx_fill_ukernel__wasmsimd_x64, .row_tile = 1, }; xnn_params.xx.pad = (struct pad_parameters) { .ukernel = (xnn_pad_ukernel_function) xnn_xx_pad_ukernel__wasmsimd, .row_tile = 1, }; xnn_params.xx.transpose = (struct transpose_parameters) { .variable_size_ukernel = xnn_xx_transposev_ukernel__1x1_memcpy, .tile_size = 32, }; #endif #elif XNN_ARCH_WASM /**************************** QC8 WAsm micro-kernels****************************/ #ifndef XNN_NO_QC8_OPERATORS init_flags |= XNN_INIT_FLAG_QC8; if (is_wasm_x86) { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qc8.gemm.mr = 2; xnn_params.qc8.gemm.nr = 2; } else { xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_4x4__wasm_fmagic); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x4__wasm_fmagic); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qc8.gemm.mr = 4; xnn_params.qc8.gemm.nr = 4; } if (is_wasm_x86) { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x3__scalar_imagic; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qc8.dwconv[0].channel_tile = 2; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qc8.dwconv[1].channel_tile = 2; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qc8.dwconv[2].channel_tile = 1; xnn_params.qc8.dwconv[2].primary_tile = 25; } else { xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x3__wasm_fmagic; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qc8.dwconv[0].channel_tile = 2; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qc8.dwconv[1].channel_tile = 2; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qc8.dwconv[2].channel_tile = 2; xnn_params.qc8.dwconv[2].primary_tile = 25; } #endif // XNN_NO_QC8_OPERATORS /**************************** QS8 WAsm micro-kernels****************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QS8; if (is_wasm_x86) { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qs8.gemm.mr = 2; xnn_params.qs8.gemm.nr = 2; } else { xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_4x4__wasm_fmagic); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x4__wasm_fmagic); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qs8.gemm.mr = 4; xnn_params.qs8.gemm.nr = 4; } if (is_wasm_x86) { xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qs8.dwconv[0].channel_tile = 2; xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qs8.dwconv[1].channel_tile = 1; xnn_params.qs8.dwconv[1].primary_tile = 25; } else { xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qs8.dwconv[0].channel_tile = 2; xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qs8.dwconv[1].channel_tile = 2; xnn_params.qs8.dwconv[1].primary_tile = 25; } xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c4, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4, .init.qs8 = xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_fp32_scalar_imagic_params, .row_tile = 7, .channel_tile = 4, }; xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__scalar_x4, .init.qs8_add = xnn_init_qs8_add_minmax_scalar_params, .element_tile = 4, }; xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_fp32_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4, .init.qs8_mul = xnn_init_qs8_mul_minmax_fp32_scalar_params, .element_tile = 4, }; if (is_wasm_x86) { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__scalar_select_x4, .init.qs8_lrelu = xnn_init_qs8_lrelu_scalar_select_params, .element_tile = 4, }; } else { xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__scalar_andxor_x4, .init.qs8_lrelu = xnn_init_qs8_lrelu_scalar_andxor_params, .element_tile = 4, }; } #endif // XNN_NO_QS8_OPERATORS /**************************** QU8 WAsm micro-kernels****************************/ #ifndef XNN_NO_QU8_OPERATORS init_flags |= XNN_INIT_FLAG_QU8; if (is_wasm_x86) { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_2x2__scalar_imagic); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_2x2__scalar_imagic); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x2__scalar_imagic); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x2__scalar_imagic); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qu8.gemm.mr = 2; xnn_params.qu8.gemm.nr = 2; } else { xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_4x4__wasm_fmagic); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_4x4__wasm_fmagic); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x4__wasm_fmagic); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x4__wasm_fmagic); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qu8.gemm.mr = 4; xnn_params.qu8.gemm.nr = 4; } if (is_wasm_x86) { xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_imagic; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qu8.dwconv[0].channel_tile = 2; xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up1x25__scalar_imagic; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_imagic_params; xnn_params.qu8.dwconv[1].channel_tile = 1; xnn_params.qu8.dwconv[1].primary_tile = 25; } else { xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__wasm_fmagic; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qu8.dwconv[0].channel_tile = 2; xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__wasm_fmagic; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_fmagic_params; xnn_params.qu8.dwconv[1].channel_tile = 2; xnn_params.qu8.dwconv[1].primary_tile = 25; } xnn_params.qu8.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1, .init.qu8 = xnn_init_qu8_avgpool_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c4, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c4, .init.qu8 = xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_fp32_scalar_imagic_params, .row_tile = 7, .channel_tile = 4, }; xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__scalar_x4, .init.qu8_add = xnn_init_qu8_add_minmax_scalar_params, .element_tile = 4, }; xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4, .init.qu8_mul = xnn_init_qu8_mul_minmax_fp32_scalar_params, .element_tile = 4, }; if (is_wasm_x86) { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__scalar_select_x4, .init.qu8_lrelu = xnn_init_qu8_lrelu_scalar_select_params, .element_tile = 4, }; } else { xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__scalar_andxor_x4, .init.qu8_lrelu = xnn_init_qu8_lrelu_scalar_andxor_params, .element_tile = 4, }; } #endif // XNN_NO_QU8_OPERATORS /**************************** S8 WAsm micro-kernels****************************/ #ifndef XNN_NO_S8_OPERATORS init_flags |= XNN_INIT_FLAG_S8; xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__scalar_x4, .init.s8_minmax = xnn_init_s8_minmax_scalar_params, .element_tile = 4, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__scalar_c1, .pixel_tile = 1, .channel_tile = 1, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.s8 = xnn_init_s8_minmax_scalar_params, .mr = 9, .qr = 8, }; #endif // XNN_NO_S8_OPERATORS /**************************** U8 WAsm micro-kernels****************************/ #ifndef XNN_NO_U8_OPERATORS init_flags |= XNN_INIT_FLAG_U8; xnn_params.u8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_u8_vclamp_ukernel__scalar_x4, .init.u8_minmax = xnn_init_u8_minmax_scalar_params, .element_tile = 4, }; xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__scalar_c1, .pixel_tile = 1, .channel_tile = 1, }; xnn_params.u8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.u8 = xnn_init_u8_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; #endif // XNN_NO_U8_OPERATORS /**************************** X8 WAsm micro-kernels****************************/ #ifndef XNN_NO_X8_OPERATORS init_flags |= XNN_INIT_FLAG_X8; xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar_x4; xnn_params.x8.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, }; xnn_params.x8.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x8_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X8_OPERATORS /**************************** X16 WAsm micro-kernels****************************/ #ifndef XNN_NO_X16_OPERATORS init_flags |= XNN_INIT_FLAG_X16; xnn_params.x16.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x16_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X16_OPERATORS /**************************** F32 WAsm micro-kernels****************************/ #ifndef XNN_NO_F32_OPERATORS init_flags |= XNN_INIT_FLAG_F32; if (is_wasm_x86) { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_2x4__scalar); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_2x4__scalar); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x4__wasm); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x4__wasm); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_2x4__scalar); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_2x4__scalar); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x4__wasm); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x4__wasm); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_2x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(2)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_2x4__scalar); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__scalar); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 2; xnn_params.f32.gemm.nr = 4; } else { xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x4__wasm); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x4__wasm); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x4__wasm); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x4__wasm); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x4__wasm); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x4__wasm); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x4__wasm); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x4__wasm); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x4__scalar); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__scalar); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 4; } xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__wasm); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__wasm); xnn_params.f32.gemm2.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2__scalar); xnn_params.f32.gemm2.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__scalar); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x3__wasm_acc2; xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x3__scalar_acc2; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[0].channel_tile = 1; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x4__wasm_acc2; xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x4__scalar_acc2; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[1].channel_tile = 1; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x9__wasm_acc2; xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x9__scalar_acc2; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 1; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x25__wasm_acc2; xnn_params.f32.dwconv[3].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x25__scalar_acc2; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[3].channel_tile = 1; xnn_params.f32.dwconv[3].primary_tile = 25; xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__wasm_c1, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__wasm_c1, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__wasm_c1, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__wasm_c1, .init.f32 = xnn_init_f32_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__wasm_c1, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__wasm_c1, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .update.f32 = xnn_update_f32_scaleminmax_scalar_params, .row_tile = 7, .channel_tile = 1, }; xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__wasm_c1, .init.f32 = xnn_init_f32_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__scalar_c1, .mr = 4, }; xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__scalar_c1, .mr = 9, }; xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1, .mr = 9, .qr = 8, }; xnn_params.f32.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__scalar_c2, .pixel_tile = 1, .channel_tile = 2, }; xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__wasm_x4, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 4, }; if (is_wasm_x86) { xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__scalar_x4, .init.f32_hswish = xnn_init_f32_hswish_scalar_params, .element_tile = 4, }; } else { xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__wasm_x4, .init.f32_hswish = xnn_init_f32_hswish_scalar_params, .element_tile = 4, }; } if (is_wasm_x86) { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x2, .init.f32_elu = xnn_init_f32_elu_scalar_rr2_lut16_p3_params, .element_tile = 2, }; } else { xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__wasm_rr2_p6_x6, .init.f32_elu = xnn_init_f32_elu_scalar_rr2_p6_params, .element_tile = 6, }; } xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__scalar_x4, .init.f32_lrelu = xnn_init_f32_lrelu_scalar_params, .element_tile = 4, }; xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__scalar_x4, .element_tile = 4, }; if (is_wasm_x86) { xnn_params.f32.relu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrelu_ukernel__scalar_x8, .element_tile = 8, }; } else { xnn_params.f32.relu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrelu_ukernel__wasm_x8, .element_tile = 8, }; } xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__scalar_libm_x4, .element_tile = 4, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__scalar_libm_x4, .element_tile = 4, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__scalar_libm_x4, .element_tile = 4, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__scalar_libm_x4, .element_tile = 4, }; xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x2, .init.f32_sigmoid = xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, .element_tile = 2, }; xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__scalar_sqrt_x1, .element_tile = 1, }; if (is_wasm_x86) { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__scalar_2x4, .row_tile = 2, .channel_tile = 4, }; } else { xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__wasm_2x4, .row_tile = 2, .channel_tile = 4, }; } xnn_params.f32.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f32_raddstoreexpminusmax_ukernel__scalar_rr2_p5_x4_acc2, .init.f32 = xnn_init_f32_expminus_scalar_rr2_p5_params, .element_tile = 4, }; xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__scalar; xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__wasm_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasm_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__wasm_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__wasm_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__wasm_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__wasm_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__wasm_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasm_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__wasm_x8, .element_tile = 8, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__wasm_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasm_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__wasm_x8, .element_tile = 8, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__wasm_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasm_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__wasm_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__wasm_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__wasm_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__wasm_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, .element_tile = 8, }; xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c1__wasm_2x, .init.f32 = xnn_init_f32_minmax_scalar_params, .channel_tile = 1, .row_tile = 2, }; #ifndef XNN_NO_NCHW_OPERATORS init_flags |= XNN_INIT_FLAG_CHW_OPT; xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x1__scalar, .mr = 8, .nr = 1, }; xnn_params.f32.spmm2 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x2__scalar, .mr = 8, .nr = 2, }; xnn_params.f32.spmm4 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x4__scalar, .mr = 8, .nr = 4, }; xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1, .output_channel_tile = 4, .output_height_tile = 1, .output_width_tile = 1, }; xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2, .output_width_tile = 1, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2, .output_width_tile = 1, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5, .output_width_tile = 1, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5, .output_width_tile = 1, .output_height_tile = 1, }; xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__scalar_x1, .channel_tile = 1, }; xnn_params.f32.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f32_ibilinear_chw_ukernel__scalar_p4, .channel_tile = 1, .pixel_tile = 4, }; #endif // XNN_NO_NCHW_OPERATORS #endif // XNN_NO_F32_OPERATORS /*************************** VCVT WAsm micro-kernels***************************/ #ifndef XNN_NO_VCVT_OPERATORS init_flags |= XNN_INIT_FLAG_VCVT; xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__scalar_x1, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_scalar_params, .element_tile = 1, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__scalar_bitcast_x4, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_scalar_bitcast_params, .element_tile = 4, }; if (is_wasm_x86) { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x1, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_scalar_imagic_params, .element_tile = 1, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__scalar_imagic_x1, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_scalar_imagic_params, .element_tile = 1, }; } else { xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__wasm_fmagic_x4, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_scalar_fmagic_params, .element_tile = 4, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__wasm_fmagic_x4, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_scalar_fmagic_params, .element_tile = 4, }; } if (is_wasm_x86) { xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__scalar_x1, .init.qs8_cvt = xnn_init_qs8_cvt_scalar_params, .element_tile = 1, }; } else { xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__scalar_x4, .init.qs8_cvt = xnn_init_qs8_cvt_scalar_params, .element_tile = 4, }; } xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__scalar_x1, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_scalar_params, .element_tile = 1, }; if (is_wasm_x86) { xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__scalar_x1, .init.qu8_cvt = xnn_init_qu8_cvt_scalar_params, .element_tile = 1, }; } else { xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__scalar_x4, .init.qu8_cvt = xnn_init_qu8_cvt_scalar_params, .element_tile = 4, }; } xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__scalar_x1, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_scalar_params, .element_tile = 1, }; #endif // XNN_NO_VCVT_OPERATORS /**************************** X32 WAsm micro-kernels****************************/ #ifndef XNN_NO_X32_OPERATORS init_flags |= XNN_INIT_FLAG_X32; xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__scalar; xnn_params.x32.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__scalar, .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__scalar, .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__scalar, .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__scalar, }; xnn_params.x32.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x32_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X32_OPERATORS /**************************** XX WAsm micro-kernels****************************/ #ifndef XNN_NO_XX_OPERATORS init_flags |= XNN_INIT_FLAG_XX; xnn_params.xx.copy = (xnn_vunary_ukernel_function) xnn_xx_copy_ukernel__memcpy; xnn_params.xx.fill = (struct fill_parameters) { .ukernel = (xnn_fill_ukernel_function) xnn_xx_fill_ukernel__scalar_x16, .row_tile = 1, }; xnn_params.xx.pad = (struct pad_parameters) { .ukernel = (xnn_pad_ukernel_function) xnn_xx_pad_ukernel__scalar, .row_tile = 1, }; xnn_params.xx.transpose = (struct transpose_parameters) { .variable_size_ukernel = xnn_xx_transposev_ukernel__1x1_memcpy, .tile_size = 32, }; #endif #elif XNN_ARCH_RISCV /************************** QC8 RISC-V micro-kernels **************************/ #ifndef XNN_NO_QC8_OPERATORS init_flags |= XNN_INIT_FLAG_QC8; xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf); xnn_params.qc8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qc8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf); xnn_params.qc8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qc8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf); xnn_params.qc8.gemm.init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qc8.gemm.mr = 3; xnn_params.qc8.gemm.nr = 4; xnn_params.qc8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x3__scalar_lrintf; xnn_params.qc8.dwconv[0].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qc8.dwconv[0].channel_tile = 2; xnn_params.qc8.dwconv[0].primary_tile = 3; xnn_params.qc8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf; xnn_params.qc8.dwconv[1].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qc8.dwconv[1].channel_tile = 2; xnn_params.qc8.dwconv[1].primary_tile = 9; xnn_params.qc8.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qc8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf; xnn_params.qc8.dwconv[2].init.qc8 = xnn_init_qc8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qc8.dwconv[2].channel_tile = 2; xnn_params.qc8.dwconv[2].primary_tile = 25; #endif // XNN_NO_QS8_OPERATORS /************************** QS8 RISC-V micro-kernels **************************/ #ifndef XNN_NO_QS8_OPERATORS init_flags |= XNN_INIT_FLAG_QS8; xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf); xnn_params.qs8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qs8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf); xnn_params.qs8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qs8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf); xnn_params.qs8.gemm.init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qs8.gemm.mr = 3; xnn_params.qs8.gemm.nr = 4; xnn_params.qs8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf; xnn_params.qs8.dwconv[0].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qs8.dwconv[0].channel_tile = 2; xnn_params.qs8.dwconv[0].primary_tile = 9; xnn_params.qs8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qs8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf; xnn_params.qs8.dwconv[1].init.qs8 = xnn_init_qs8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qs8.dwconv[1].channel_tile = 2; xnn_params.qs8.dwconv[1].primary_tile = 25; xnn_params.qs8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c1, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qs8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c1, .init.qs8 = xnn_init_qs8_avgpool_minmax_fp32_scalar_imagic_params, .update.qs8 = xnn_update_qs8_avgpool_minmax_fp32_scalar_imagic_params, .row_tile = 7, .channel_tile = 1, }; xnn_params.qs8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vadd_minmax_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vaddc_minmax_ukernel__scalar_x4, .init.qs8_add = xnn_init_qs8_add_minmax_scalar_params, .element_tile = 4, }; xnn_params.qs8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmul_minmax_fp32_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qs8_vmulc_minmax_fp32_ukernel__scalar_x4, .init.qs8_mul = xnn_init_qs8_mul_minmax_fp32_scalar_params, .element_tile = 4, }; xnn_params.qs8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vlrelu_ukernel__scalar_andxor_x4, .init.qs8_lrelu = xnn_init_qs8_lrelu_scalar_andxor_params, .element_tile = 4, }; #endif // XNN_NO_QS8_OPERATORS /************************** QU8 RISC-V micro-kernels **************************/ #ifndef XNN_NO_QU8_OPERATORS init_flags |= XNN_INIT_FLAG_QU8; xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_3x4__scalar_lrintf); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(3)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_3x4__scalar_lrintf); xnn_params.qu8.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_qu8_gemm_minmax_fp32_ukernel_1x4__scalar_lrintf); xnn_params.qu8.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_qu8_igemm_minmax_fp32_ukernel_1x4__scalar_lrintf); xnn_params.qu8.gemm.init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qu8.gemm.mr = 3; xnn_params.qu8.gemm.nr = 4; xnn_params.qu8.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up2x9__scalar_lrintf; xnn_params.qu8.dwconv[0].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qu8.dwconv[0].channel_tile = 2; xnn_params.qu8.dwconv[0].primary_tile = 9; xnn_params.qu8.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_qu8_dwconv_minmax_fp32_ukernel_up2x25__scalar_lrintf; xnn_params.qu8.dwconv[1].init.qu8 = xnn_init_qu8_conv_minmax_fp32_scalar_lrintf_params; xnn_params.qu8.dwconv[1].channel_tile = 2; xnn_params.qu8.dwconv[1].primary_tile = 25; xnn_params.qu8.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_qu8_avgpool_minmax_ukernel_9p8x__scalar_c1, .init.qu8 = xnn_init_qu8_avgpool_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.qu8.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7x__scalar_imagic_c1, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_qu8_gavgpool_minmax_fp32_ukernel_7p7x__scalar_imagic_c1, .init.qu8 = xnn_init_qu8_avgpool_minmax_fp32_scalar_imagic_params, .update.qu8 = xnn_update_qu8_avgpool_minmax_fp32_scalar_imagic_params, .row_tile = 7, .channel_tile = 1, }; xnn_params.qu8.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vadd_minmax_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vaddc_minmax_ukernel__scalar_x4, .init.qu8_add = xnn_init_qu8_add_minmax_scalar_params, .element_tile = 4, }; xnn_params.qu8.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmul_minmax_fp32_ukernel__scalar_x4, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_qu8_vmulc_minmax_fp32_ukernel__scalar_x4, .init.qu8_mul = xnn_init_qu8_mul_minmax_fp32_scalar_params, .element_tile = 4, }; xnn_params.qu8.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vlrelu_ukernel__scalar_andxor_x4, .init.qu8_lrelu = xnn_init_qu8_lrelu_scalar_andxor_params, .element_tile = 4, }; #endif // XNN_NO_QU8_OPERATORS /************************** S8 RISC-V micro-kernels ***************************/ #ifndef XNN_NO_S8_OPERATORS init_flags |= XNN_INIT_FLAG_S8; xnn_params.s8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_s8_vclamp_ukernel__scalar_x4, .init.s8_minmax = xnn_init_s8_minmax_scalar_params, .element_tile = 4, }; xnn_params.s8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_s8_ibilinear_ukernel__scalar_c1, .pixel_tile = 1, .channel_tile = 1, }; xnn_params.s8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_s8_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.s8 = xnn_init_s8_minmax_scalar_params, .mr = 9, .qr = 8, }; #endif // XNN_NO_S8_OPERATORS /************************** U8 RISC-V micro-kernels ***************************/ #ifndef XNN_NO_U8_OPERATORS init_flags |= XNN_INIT_FLAG_U8; xnn_params.u8.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_u8_vclamp_ukernel__scalar_x4, .init.u8_minmax = xnn_init_u8_minmax_scalar_params, .element_tile = 4, }; xnn_params.u8.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_u8_ibilinear_ukernel__scalar_c1, .pixel_tile = 1, .channel_tile = 1, }; xnn_params.u8.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_u8_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.u8 = xnn_init_u8_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.u8.lut32norm = xnn_u8_lut32norm_ukernel__scalar; xnn_params.u8.rmax = xnn_u8_rmax_ukernel__scalar; #endif // XNN_NO_U8_OPERATORS /************************** X8 RISC-V micro-kernels ***************************/ #ifndef XNN_NO_X8_OPERATORS init_flags |= XNN_INIT_FLAG_X8; xnn_params.x8.lut = xnn_x8_lut_ukernel__scalar_x4; xnn_params.x8.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x8_zip_x2_ukernel__scalar, .x3 = (xnn_zipc_ukernel_function) xnn_x8_zip_x3_ukernel__scalar, .x4 = (xnn_zipc_ukernel_function) xnn_x8_zip_x4_ukernel__scalar, .xm = (xnn_zipv_ukernel_function) xnn_x8_zip_xm_ukernel__scalar, }; xnn_params.x8.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x8_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X8_OPERATORS /************************** X16 RISC-V micro-kernels ***************************/ #ifndef XNN_NO_X16_OPERATORS init_flags |= XNN_INIT_FLAG_X16; xnn_params.x16.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x16_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X16_OPERATORS /************************** F32 RISC-V micro-kernels **************************/ #ifndef XNN_NO_F32_OPERATORS init_flags |= XNN_INIT_FLAG_F32; xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x4__scalar); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x4__scalar); xnn_params.f32.gemm.minmax.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_1x4__scalar); xnn_params.f32.gemm.minmax.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_1x4__scalar); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_4x4__scalar); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_4x4__scalar); xnn_params.f32.gemm.relu.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_relu_ukernel_1x4__scalar); xnn_params.f32.gemm.relu.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_relu_ukernel_1x4__scalar); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x4__scalar); xnn_params.f32.gemm.linear.gemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_1x4__scalar); xnn_params.f32.gemm.linear.igemm[XNN_MR_TO_INDEX(1)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_1x4__scalar); xnn_params.f32.gemm.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm.mr = 4; xnn_params.f32.gemm.nr = 4; xnn_params.f32.gemm2.minmax.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_minmax_ukernel_4x2__scalar); xnn_params.f32.gemm2.minmax.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_minmax_ukernel_4x2__scalar); xnn_params.f32.gemm2.linear.gemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_gemm_ukernel((xnn_gemm_ukernel_function) xnn_f32_gemm_ukernel_4x2__scalar); xnn_params.f32.gemm2.linear.igemm[XNN_MR_TO_INDEX(4)] = xnn_init_hmp_igemm_ukernel((xnn_igemm_ukernel_function) xnn_f32_igemm_ukernel_4x2__scalar); xnn_params.f32.gemm2.init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.gemm2.mr = 4; xnn_params.f32.gemm2.nr = 2; xnn_params.f32.dwconv[0].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x3__scalar_acc2; xnn_params.f32.dwconv[0].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x3__scalar_acc2; xnn_params.f32.dwconv[0].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[0].channel_tile = 1; xnn_params.f32.dwconv[0].primary_tile = 3; xnn_params.f32.dwconv[1].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x4__scalar_acc2; xnn_params.f32.dwconv[1].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x4__scalar_acc2; xnn_params.f32.dwconv[1].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[1].channel_tile = 1; xnn_params.f32.dwconv[1].primary_tile = 4; xnn_params.f32.dwconv[2].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x9__scalar_acc2; xnn_params.f32.dwconv[2].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x9__scalar_acc2; xnn_params.f32.dwconv[2].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[2].channel_tile = 1; xnn_params.f32.dwconv[2].primary_tile = 9; xnn_params.f32.dwconv[3].minmax.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_minmax_ukernel_up1x25__scalar_acc2; xnn_params.f32.dwconv[3].linear.unipass = (xnn_dwconv_unipass_ukernel_function) xnn_f32_dwconv_ukernel_up1x25__scalar_acc2; xnn_params.f32.dwconv[3].init.f32 = xnn_init_f32_minmax_scalar_params; xnn_params.f32.dwconv[3].channel_tile = 1; xnn_params.f32.dwconv[3].primary_tile = 25; xnn_params.f32.avgpool = (struct avgpool_parameters) { .unipass = (xnn_avgpool_unipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_avgpool_multipass_ukernel_function) xnn_f32_avgpool_minmax_ukernel_9p8x__scalar_c1, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.f32.pavgpool = (struct pavgpool_parameters) { .unipass = (xnn_pavgpool_unipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9x__scalar_c1, .multipass = (xnn_pavgpool_multipass_ukernel_function) xnn_f32_pavgpool_minmax_ukernel_9p8x__scalar_c1, .init.f32 = xnn_init_f32_minmax_scalar_params, .primary_tile = 9, .incremental_tile = 8, .channel_tile = 1, }; xnn_params.f32.gavgpool = (struct gavgpool_parameters) { .unipass = (xnn_gavgpool_unipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7x__scalar_c1, .multipass = (xnn_gavgpool_multipass_ukernel_function) xnn_f32_gavgpool_minmax_ukernel_7p7x__scalar_c1, .init.f32 = xnn_init_f32_scaleminmax_scalar_params, .update.f32 = xnn_update_f32_scaleminmax_scalar_params, .row_tile = 7, .channel_tile = 1, }; xnn_params.f32.maxpool = (struct maxpool_parameters) { .ukernel = (xnn_maxpool_ukernel_function) xnn_f32_maxpool_minmax_ukernel_9p8x__scalar_c1, .init.f32 = xnn_init_f32_minmax_scalar_params, .mr = 9, .qr = 8, }; xnn_params.f32.argmaxpool[0] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_4x__scalar_c1, .mr = 4, }; xnn_params.f32.argmaxpool[1] = (struct argmaxpool_parameters) { .up = (xnn_argmaxpool_unipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9x__scalar_c1, .mr = 9, }; xnn_params.f32.argmaxpool[2] = (struct argmaxpool_parameters) { .mp = (xnn_argmaxpool_multipass_ukernel_function) xnn_f32_argmaxpool_ukernel_9p8x__scalar_c1, .mr = 9, .qr = 8, }; xnn_params.f32.ibilinear = (struct ibilinear_parameters) { .ukernel = (xnn_ibilinear_ukernel_function) xnn_f32_ibilinear_ukernel__scalar_c2, .pixel_tile = 1, .channel_tile = 2, }; xnn_params.f32.abs = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vabs_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.clamp = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vclamp_ukernel__scalar_x4, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 4, }; xnn_params.f32.elu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_velu_ukernel__scalar_rr2_lut16_p3_x4, .init.f32_elu = xnn_init_f32_elu_scalar_rr2_lut16_p3_params, .element_tile = 4, }; xnn_params.f32.hswish = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vhswish_ukernel__scalar_x4, .init.f32_hswish = xnn_init_f32_hswish_scalar_params, .element_tile = 4, }; xnn_params.f32.lrelu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vlrelu_ukernel__scalar_x4, .init.f32_lrelu = xnn_init_f32_lrelu_scalar_params, .element_tile = 4, }; xnn_params.f32.neg = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vneg_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.rndne = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndne_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.rndz = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndz_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.rndu = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndu_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.rndd = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vrndd_ukernel__scalar_libm_x1, .element_tile = 1, }; xnn_params.f32.sigmoid = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsigmoid_ukernel__scalar_rr2_lut64_p2_div_x2, .init.f32_sigmoid = xnn_init_f32_sigmoid_scalar_rr2_lut64_p2_params, .element_tile = 2, }; xnn_params.f32.sqr = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqr_ukernel__scalar_x4, .element_tile = 4, }; xnn_params.f32.sqrt = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_vsqrt_ukernel__scalar_sqrt_x1, .element_tile = 1, }; xnn_params.f32.prelu = (struct prelu_parameters) { .ukernel = (xnn_prelu_ukernel_function) xnn_f32_prelu_ukernel__scalar_2x4, .row_tile = 4, .channel_tile = 4, }; xnn_params.f32.raddstoreexpminusmax = (struct raddstoreexpminusmax_parameters) { .ukernel = (xnn_raddstoreexpminusmax_ukernel_function) xnn_f32_raddstoreexpminusmax_ukernel__scalar_rr2_p5_x4_acc2, .init.f32 = xnn_init_f32_expminus_scalar_rr2_p5_params, .element_tile = 4, }; xnn_params.f32.rmax = (xnn_rmax_ukernel_function) xnn_f32_rmax_ukernel__scalar; xnn_params.f32.vadd = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vadd_minmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vaddc_minmax_ukernel__scalar_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vdiv = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdiv_minmax_ukernel__scalar_x2, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vdivc_minmax_ukernel__scalar_x2, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrdivc_minmax_ukernel__scalar_x2, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 2, }; xnn_params.f32.vmax = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmaxc_ukernel__scalar_x8, .element_tile = 8, }; xnn_params.f32.vmin = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmin_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vminc_ukernel__scalar_x8, .element_tile = 8, }; xnn_params.f32.vmul = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmul_minmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vmulc_minmax_ukernel__scalar_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsub = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsub_minmax_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsubc_minmax_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vrsubc_minmax_ukernel__scalar_x8, .init.f32_minmax = xnn_init_f32_minmax_scalar_params, .element_tile = 8, }; xnn_params.f32.vsqrdiff = (struct vbinary_parameters) { .minmax.op_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiff_ukernel__scalar_x8, .minmax.opc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, .minmax.ropc_ukernel = (xnn_vbinary_ukernel_function) xnn_f32_vsqrdiffc_ukernel__scalar_x8, .element_tile = 8, }; xnn_params.f32.vmulcaddc = (struct vmulcaddc_parameters) { .ukernel = (xnn_vmulcaddc_ukernel_function) xnn_f32_vmulcaddc_minmax_ukernel_c1__scalar_2x, .init.f32 = xnn_init_f32_minmax_scalar_params, .channel_tile = 1, .row_tile = 2, }; #ifndef XNN_NO_NCHW_OPERATORS init_flags |= XNN_INIT_FLAG_CHW_OPT; xnn_params.f32.spmm = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x1__scalar, .mr = 8, .nr = 1, }; xnn_params.f32.spmm2 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x2__scalar, .mr = 8, .nr = 2, }; xnn_params.f32.spmm4 = (struct spmm_parameters) { .ukernel = (xnn_spmm_ukernel_function) xnn_f32_spmm_minmax_ukernel_8x4__scalar, .mr = 8, .nr = 4, }; xnn_params.f32.conv_hwc2chw_3x3c3s2 = (struct conv_hwc2chw_parameters) { .ukernel_with_symm_padding = (xnn_conv_hwc2chw_ukernel_function) xnn_f32_conv_hwc2chw_ukernel_3x3s2p1c3x4__scalar_1x1, .output_channel_tile = 4, .output_height_tile = 1, .output_width_tile = 1, }; xnn_params.f32.dwconv2d_chw_3x3 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3p1__scalar_2x1_acc2, .output_width_tile = 1, .output_height_tile = 2, }; xnn_params.f32.dwconv2d_chw_3x3s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_3x3s2p1__scalar_1x1_acc2, .output_width_tile = 1, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5p2__scalar_1x1_acc5, .output_width_tile = 1, .output_height_tile = 1, }; xnn_params.f32.dwconv2d_chw_5x5s2 = (struct dwconv2d_chw_parameters) { .ukernel = (xnn_dwconv2d_chw_ukernel_function) xnn_f32_dwconv2d_chw_ukernel_5x5s2p2__scalar_1x1_acc5, .output_width_tile = 1, .output_height_tile = 1, }; xnn_params.f32.gavgpool_cw = (struct gavgpool_cw_parameters) { .ukernel = (xnn_gavgpool_cw_ukernel_function) xnn_f32_gavgpool_cw_ukernel__scalar_x1, .channel_tile = 1, }; xnn_params.f32.ibilinear_chw = (struct ibilinear_chw_parameters) { .ukernel = (xnn_ibilinear_chw_ukernel_function) xnn_f32_ibilinear_chw_ukernel__scalar_p4, .channel_tile = 1, .pixel_tile = 4, }; #endif // XNN_NO_NCHW_OPERATORS #endif // XNN_NO_F32_OPERATORS /************************** VCVT RISC-V micro-kernels *************************/ #ifndef XNN_NO_VCVT_OPERATORS init_flags |= XNN_INIT_FLAG_VCVT; xnn_params.vcvt.f16_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f16_f32_vcvt_ukernel__scalar_x4, .init.f16_f32_cvt = xnn_init_f16_f32_cvt_scalar_params, .element_tile = 4, }; xnn_params.vcvt.f32_to_f16 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_f16_vcvt_ukernel__scalar_fabsf_x2, .init.f32_f16_cvt = xnn_init_f32_f16_cvt_scalar_fabsf_params, .element_tile = 2, }; xnn_params.vcvt.f32_to_qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qs8_vcvt_ukernel__scalar_lrintf_x4, .init.f32_qs8_cvt = xnn_init_f32_qs8_cvt_scalar_lrintf_params, .element_tile = 4, }; xnn_params.vcvt.f32_to_qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_f32_qu8_vcvt_ukernel__scalar_lrintf_x4, .init.f32_qu8_cvt = xnn_init_f32_qu8_cvt_scalar_lrintf_params, .element_tile = 4, }; xnn_params.vcvt.qs8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_vcvt_ukernel__scalar_x4, .init.qs8_cvt = xnn_init_qs8_cvt_scalar_params, .element_tile = 4, }; xnn_params.vcvt.qs8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qs8_f32_vcvt_ukernel__scalar_x4, .init.qs8_f32_cvt = xnn_init_qs8_f32_cvt_scalar_params, .element_tile = 4, }; xnn_params.vcvt.qu8 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_vcvt_ukernel__scalar_x4, .init.qu8_cvt = xnn_init_qu8_cvt_scalar_params, .element_tile = 4, }; xnn_params.vcvt.qu8_to_f32 = (struct vunary_parameters) { .ukernel = (xnn_vunary_ukernel_function) xnn_qu8_f32_vcvt_ukernel__scalar_x4, .init.qu8_f32_cvt = xnn_init_qu8_f32_cvt_scalar_params, .element_tile = 4, }; #endif // XNN_NO_VCVT_OPERATORS /************************** X32 RISC-V micro-kernels **************************/ #ifndef XNN_NO_X32_OPERATORS init_flags |= XNN_INIT_FLAG_X32; xnn_params.x32.unpool = (xnn_unpool_ukernel_function) xnn_x32_unpool_ukernel__scalar; xnn_params.x32.zip = (struct zip_parameters) { .x2 = (xnn_zipc_ukernel_function) xnn_x32_zip_x2_ukernel__scalar, .x3 = (xnn_zipc_ukernel_function) xnn_x32_zip_x3_ukernel__scalar, .x4 = (xnn_zipc_ukernel_function) xnn_x32_zip_x4_ukernel__scalar, .xm = (xnn_zipv_ukernel_function) xnn_x32_zip_xm_ukernel__scalar, }; xnn_params.x32.transpose = (struct transpose_parameters) { .const_size_ukernel = (xnn_transposec_ukernel_function) xnn_x32_transposec_ukernel__2x4_scalar_int, .tile_size = 32, }; #endif // XNN_NO_X32_OPERATORS /************************** XX RISC-V micro-kernels ***************************/ #ifndef XNN_NO_XX_OPERATORS init_flags |= XNN_INIT_FLAG_XX; xnn_params.xx.copy = (xnn_vunary_ukernel_function) xnn_xx_copy_ukernel__memcpy; xnn_params.xx.fill = (struct fill_parameters) { .ukernel = (xnn_fill_ukernel_function) xnn_xx_fill_ukernel__scalar_x16, .row_tile = 1, }; xnn_params.xx.pad = (struct pad_parameters) { .ukernel = (xnn_pad_ukernel_function) xnn_xx_pad_ukernel__scalar, .row_tile = 1, }; xnn_params.xx.transpose = (struct transpose_parameters) { .variable_size_ukernel = xnn_xx_transposev_ukernel__1x1_memcpy, .tile_size = 32, }; #endif // XNN_NO_XX_OPERATORS #else #error "Unsupported architecture" #endif // Get page size. #if XNN_PLATFORM_WINDOWS SYSTEM_INFO sysinfo; GetSystemInfo(&sysinfo); xnn_params.page_size = sysinfo.dwPageSize; #else const long res = sysconf(_SC_PAGESIZE); if (res == -1) { xnn_log_error("failed to get page size, error code: %d", errno); return; } xnn_params.page_size = res; #endif memcpy(&xnn_params.allocator, init_allocator, sizeof(struct xnn_allocator)); xnn_params.init_flags = init_flags; } #if XNN_PLATFORM_WINDOWS static BOOL CALLBACK init_windows(PINIT_ONCE init_once, PVOID parameter, PVOID* context) { init(); return TRUE; } #endif enum xnn_status xnn_initialize(const struct xnn_allocator* allocator) { #if !XNN_PLATFORM_WEB && !XNN_ARCH_RISCV if (!cpuinfo_initialize()) { return xnn_status_out_of_memory; } #endif // !XNN_PLATFORM_WEB && !XNN_ARCH_RISCV if (allocator == NULL) { allocator = &xnn_default_allocator; } #ifdef _MSC_VER _InterlockedCompareExchangePointer((PVOID volatile*) &init_allocator, (PVOID) allocator, NULL); #else __sync_bool_compare_and_swap(&init_allocator, NULL, allocator); #endif #if XNN_PLATFORM_WINDOWS InitOnceExecuteOnce(&init_guard, &init_windows, NULL, NULL); #else pthread_once(&init_guard, &init); #endif if ((xnn_params.init_flags & XNN_INIT_FLAG_XNNPACK) != 0) { return xnn_status_success; } else { return xnn_status_unsupported_hardware; } } enum xnn_status xnn_deinitialize(void) { #if !XNN_PLATFORM_WEB && !XNN_ARCH_RISCV cpuinfo_deinitialize(); #endif // !XNN_PLATFORM_WEB && !XNN_ARCH_RISCV return xnn_status_success; }