/* * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * * Permission is hereby granted, free of charge, to any person obtaining a copy * of this software and associated documentation files (the "Software"), to * deal in the Software without restriction, including without limitation the * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or * sell copies of the Software, and to permit persons to whom the Software is * furnished to do so, subject to the following conditions: * * The above copyright notice and this permission notice shall be included in all * copies or substantial portions of the Software. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include "src/cpu/kernels/CpuGemmLowpOffsetContributionOutputStageKernel.h" #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Utils.h" #include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" #include "src/core/NEON/NEAsymm.h" #include "src/core/NEON/wrapper/wrapper.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include namespace arm_compute { namespace cpu { namespace kernels { namespace { inline int32x4x4_t load_results_input(const Iterator &mm_result_it, int32_t x) { return { { vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 0), vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 4), vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 8), vld1q_s32(reinterpret_cast(mm_result_it.ptr()) + x + 12) } }; } inline int32x4x4_t load(const int32_t *ptr, int32_t x) { return { { vld1q_s32(ptr + x + 0), vld1q_s32(ptr + x + 4), vld1q_s32(ptr + x + 8), vld1q_s32(ptr + x + 12) } }; } inline int32x4x4_t add_s32(int32x4x4_t a, int32x4_t b) { return { { vaddq_s32(a.val[0], b), vaddq_s32(a.val[1], b), vaddq_s32(a.val[2], b), vaddq_s32(a.val[3], b) } }; } inline int32x4x4_t add_s32(int32x4x4_t a, int32x4x4_t b) { return { { vaddq_s32(a.val[0], b.val[0]), vaddq_s32(a.val[1], b.val[1]), vaddq_s32(a.val[2], b.val[2]), vaddq_s32(a.val[3], b.val[3]) } }; } inline int32x4x4_t mul_s32(int32x4x4_t &a, int32_t mul_scalar) { return { { vmulq_n_s32(a.val[0], mul_scalar), vmulq_n_s32(a.val[1], mul_scalar), vmulq_n_s32(a.val[2], mul_scalar), vmulq_n_s32(a.val[3], mul_scalar) } }; } inline int32x4x4_t mul_s32(int32x4x4_t &a, const int32_t *multilpier) { return { { vmulq_s32(a.val[0], vld1q_s32(multilpier)), vmulq_s32(a.val[1], vld1q_s32(multilpier + 4)), vmulq_s32(a.val[2], vld1q_s32(multilpier + 8)), vmulq_s32(a.val[3], vld1q_s32(multilpier + 12)) } }; } inline int32x4x4_t get_a_offset(const int32_t *vector_sum_col_ptr, int32_t a_offset, int32_t x) { int32x4x4_t a_offset_term_s32 = load(vector_sum_col_ptr, x); a_offset_term_s32.val[0] = vmulq_n_s32(a_offset_term_s32.val[0], a_offset); a_offset_term_s32.val[1] = vmulq_n_s32(a_offset_term_s32.val[1], a_offset); a_offset_term_s32.val[2] = vmulq_n_s32(a_offset_term_s32.val[2], a_offset); a_offset_term_s32.val[3] = vmulq_n_s32(a_offset_term_s32.val[3], a_offset); return a_offset_term_s32; } inline int32x4_t get_b_offset(const int32_t *vector_sum_row_ptr, int32_t b_offset) { int32x4_t b_offset_term_s32 = vld1q_dup_s32(vector_sum_row_ptr); b_offset_term_s32 = vmulq_n_s32(b_offset_term_s32, b_offset); return b_offset_term_s32; } inline int32x4x4_t get_k_offset(int32_t k_offset) { return { { vdupq_n_s32(k_offset), vdupq_n_s32(k_offset), vdupq_n_s32(k_offset), vdupq_n_s32(k_offset) } }; } inline uint8x16_t finalize_quantization_floating_point(int32x4x4_t &in_s32, int32x4_t result_shift_s32, uint8x16_t min_u8, uint8x16_t max_u8, bool is_bounded_relu) { const static int32x4_t zero_s32 = vdupq_n_s32(0); // Shift final result (negative value shift right) in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32); in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32); in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32); in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32); // Saturate negative values in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); // Convert S32 to S16 const int16x8x2_t in_s16 = { { vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) } }; // Convert S16 to U8 uint8x16_t out_u8 = vcombine_u8(vqmovun_s16(in_s16.val[0]), vqmovun_s16(in_s16.val[1])); if(is_bounded_relu) { out_u8 = vmaxq_u8(out_u8, min_u8); out_u8 = vminq_u8(out_u8, max_u8); } return out_u8; } inline int8x16_t finalize_quantization_floating_point(int32x4x4_t &in_s32, int32x4_t result_shift_s32, int8x16_t min_s8, int8x16_t max_s8, bool is_bounded_relu) { const static int32x4_t zero_s32 = vdupq_n_s32(0); // Shift final result (negative value shift right) in_s32.val[0] = vshlq_s32(in_s32.val[0], result_shift_s32); in_s32.val[1] = vshlq_s32(in_s32.val[1], result_shift_s32); in_s32.val[2] = vshlq_s32(in_s32.val[2], result_shift_s32); in_s32.val[3] = vshlq_s32(in_s32.val[3], result_shift_s32); // Saturate negative values in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); // Convert S32 to S16 const int16x8x2_t in_s16 = { { vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) } }; // Convert S16 to S8 int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); if(is_bounded_relu) { out_s8 = vmaxq_s8(out_s8, min_s8); out_s8 = vminq_s8(out_s8, max_s8); } return out_s8; } inline int8x16_t finalize_quantization_floating_point(int32x4x4_t &in_s32, int32x4x4_t result_shift_s32, int8x16_t min_s8, int8x16_t max_s8, bool is_bounded_relu) { const static int32x4_t zero_s32 = vdupq_n_s32(0); // Shift final result (negative value shift right) in_s32.val[0] = vshlq_s32(in_s32.val[0], vnegq_s32(result_shift_s32.val[0])); in_s32.val[1] = vshlq_s32(in_s32.val[1], vnegq_s32(result_shift_s32.val[1])); in_s32.val[2] = vshlq_s32(in_s32.val[2], vnegq_s32(result_shift_s32.val[2])); in_s32.val[3] = vshlq_s32(in_s32.val[3], vnegq_s32(result_shift_s32.val[3])); // Saturate negative values in_s32.val[0] = vmaxq_s32(in_s32.val[0], zero_s32); in_s32.val[1] = vmaxq_s32(in_s32.val[1], zero_s32); in_s32.val[2] = vmaxq_s32(in_s32.val[2], zero_s32); in_s32.val[3] = vmaxq_s32(in_s32.val[3], zero_s32); // Convert S32 to S16 const int16x8x2_t in_s16 = { { vcombine_s16(vqmovn_s32(in_s32.val[0]), vqmovn_s32(in_s32.val[1])), vcombine_s16(vqmovn_s32(in_s32.val[2]), vqmovn_s32(in_s32.val[3])) } }; // Convert S16 to S8 int8x16_t out_s8 = vcombine_s8(vqmovn_s16(in_s16.val[0]), vqmovn_s16(in_s16.val[1])); if(is_bounded_relu) { out_s8 = vmaxq_s8(out_s8, min_s8); out_s8 = vminq_s8(out_s8, max_s8); } return out_s8; } template struct VectorTyper { using stype = T; using vtype = typename wrapper::traits::neon_bitvector_t; }; inline Window get_win_vector_sum(const Window &window) { Window win_vector_sum(window); win_vector_sum.set(Window::DimY, Window::Dimension(0, 0, 0)); win_vector_sum.set(Window::DimZ, Window::Dimension(0, 0, 0)); return win_vector_sum; } inline Iterator get_vector_sum_col_it(const Window &window, const ITensor *vector_sum_col) { Iterator vector_sum_col_it(vector_sum_col, get_win_vector_sum(window)); return vector_sum_col_it; } inline Iterator get_vector_sum_row_it(const Window &window, const ITensor *vector_sum_row) { Window win_vector_sum_row = get_win_vector_sum(window); win_vector_sum_row.set(Window::DimX, Window::Dimension(0, 0, 0)); Iterator vector_sum_row_it(vector_sum_row, win_vector_sum_row); return vector_sum_row_it; } inline Iterator get_bias_it(const Window &window, const ITensor *bias) { Window win_bias(window); win_bias.set(Window::DimY, Window::Dimension(0, 1, 1)); win_bias.set(Window::DimZ, Window::Dimension(0, 1, 1)); Iterator bias_it(bias, win_bias); return bias_it; } template inline void run_offset_contribution_output_stage_window(const int32_t *vector_sum_col_ptr, const int32_t *vector_sum_row_ptr, const int32_t *bias_ptr, Iterator mm_result_it, Iterator out_it, const int32x4_t result_offset_s32, const int32x4_t result_shift_s32, typename VT::vtype min_vec, typename VT::vtype max_vec, int32_t a_offset, int32_t b_offset, int32_t k_offset, int32_t multiplier, int32_t shift, int32_t offset, int32_t min_bound, int32_t max_bound, int window_step_x, int window_start_x, int window_end_x, bool has_a_offset, bool has_b_offset, bool has_bias, bool is_bounded_relu, bool is_fixed_point) { int32x4x4_t offset_term_s32 = { 0, 0, 0, 0 }; if(!is_fixed_point) { // Combine quantization offset with other offsets. offset_term_s32 = add_s32(offset_term_s32, result_offset_s32); } if(has_a_offset && has_b_offset) { offset_term_s32 = add_s32(offset_term_s32, get_k_offset(k_offset)); } if(has_b_offset) { offset_term_s32 = add_s32(offset_term_s32, get_b_offset(vector_sum_row_ptr, b_offset)); } int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { int32x4x4_t in_s32 = load_results_input(mm_result_it, x); if(has_a_offset) { in_s32 = add_s32(in_s32, get_a_offset(vector_sum_col_ptr, a_offset, x)); } if(has_bias) { in_s32 = add_s32(in_s32, load(bias_ptr, x)); } if(!is_fixed_point || has_b_offset) { in_s32 = add_s32(in_s32, offset_term_s32); } if(!is_fixed_point) { in_s32 = mul_s32(in_s32, multiplier); } if(is_fixed_point) { wrapper::vstore(reinterpret_cast(out_it.ptr() + x), finalize_quantization(in_s32, multiplier, shift, result_offset_s32, min_vec, max_vec, is_bounded_relu)); } else { wrapper::vstore(reinterpret_cast(out_it.ptr() + x), finalize_quantization_floating_point(in_s32, result_shift_s32, min_vec, max_vec, is_bounded_relu)); } } // Compute left-over elements for(; x < window_end_x; ++x) { int32_t in_value = *(reinterpret_cast(mm_result_it.ptr()) + x) + wrapper::vgetlane(offset_term_s32.val[0], 0); if(has_a_offset) { in_value += (*(vector_sum_col_ptr + x) * a_offset); } if(has_bias) { in_value += *(bias_ptr + x); } if(is_fixed_point) { // Finalize and store the result *reinterpret_cast(out_it.ptr() + x) = finalize_quantization(in_value, multiplier, shift, offset, static_cast(min_bound), static_cast(max_bound), is_bounded_relu); } else { // Finalize quantization in_value = (in_value * multiplier) >> shift; // Bound and store the result if(is_bounded_relu) { in_value = static_cast(std::max(min_bound, std::min(max_bound, in_value))); } *reinterpret_cast(out_it.ptr() + x) = static_cast(std::max(static_cast(std::numeric_limits::lowest()), std::min(static_cast(std::numeric_limits::max()), in_value))); } } } inline void run_offset_contribution_output_stage_window_symm(const int32_t *vector_sum_col_ptr, const int32_t *bias_ptr, Iterator mm_result_it, Iterator out_it, const int32_t *result_multipliers, const int32_t *result_shifts, const int32x4_t result_offset, int8x16_t min_s8, int8x16_t max_s8, int32_t a_offset, int32_t offset, int32_t min_bound, int32_t max_bound, int window_step_x, int window_start_x, int window_end_x, bool has_a_offset, bool has_bias, bool is_bounded_relu, bool is_fixed_point) { int32x4x4_t offset_term_s32 = { 0, 0, 0, 0 }; if(!is_fixed_point) { // Combine quantization offset with other offsets. offset_term_s32 = add_s32(offset_term_s32, result_offset); } int x = window_start_x; for(; x <= (window_end_x - window_step_x); x += window_step_x) { int32x4x4_t in_s32 = load_results_input(mm_result_it, x); if(has_a_offset) { in_s32 = add_s32(in_s32, get_a_offset(vector_sum_col_ptr, a_offset, x)); } if(has_bias) { in_s32 = add_s32(in_s32, load(bias_ptr, x)); } if(!is_fixed_point) { in_s32 = add_s32(in_s32, offset_term_s32); in_s32 = mul_s32(in_s32, result_multipliers + x); } if(is_fixed_point) { vst1q_s8(reinterpret_cast(out_it.ptr() + x), finalize_quantization_symm(in_s32, load(result_multipliers, x), load(result_shifts, x), result_offset, min_s8, max_s8, is_bounded_relu)); } else { vst1q_s8(reinterpret_cast(out_it.ptr() + x), finalize_quantization_floating_point(in_s32, load(result_shifts, x), min_s8, max_s8, is_bounded_relu)); } } // Compute left-over elements for(; x < window_end_x; ++x) { int32_t in_value = *(reinterpret_cast(mm_result_it.ptr()) + x) + wrapper::vgetlane(offset_term_s32.val[0], 0); if(has_a_offset) { in_value += (*(vector_sum_col_ptr + x) * a_offset); } if(has_bias) { in_value += *(bias_ptr + x); } if(is_fixed_point) { // Finalize and store the result *(out_it.ptr() + x) = finalize_quantization(in_value, result_multipliers[x], result_shifts[x], offset, static_cast(min_bound), static_cast(max_bound), is_bounded_relu); } else { // Finalize quantization in_value = (in_value * result_multipliers[x]) >> (-result_shifts[x]); // Bound and store the result if(is_bounded_relu) { in_value = static_cast(std::max(min_bound, std::min(max_bound, in_value))); } *(out_it.ptr() + x) = static_cast(std::max(-128, std::min(127, in_value))); } } } template void run_offset_contribution_output_stage(const Window &window, const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, int32_t a_offset, int32_t b_offset, int32_t k_offset, bool is_vector_sum_col_batched, GEMMLowpOutputStageInfo output_stage, bool is_gemm3d, bool is_bounded_relu, bool is_fixed_point) { // Semantics of XYZW Explained for each tensor // // | Tensor | XYZW when is_gemm3d == false | XYZW when is_gemm3d == true | // ------------------------------------------------------------------------------------------------------------------- // | mm_result | x -> width, y -> height, z -> batch | x -> width, y -> height, z -> depth, w -> batch | // | collapsed window | x -> width, y -> height, z -> batch | x -> width, y -> height, z -> depth * batch | // | vector_sum_row | x -> height, y -> batch | x -> height * depth, y -> batch | // | Vector_sum_col | x -> width, y -> batch | x -> width, y -> batch | using ExactTagType = typename wrapper::traits::neon_bitvector_tag_t; using Typer = VectorTyper; const int height_input = is_gemm3d ? mm_result->info()->dimension(1) : 0; const int depth_input = is_gemm3d ? mm_result->info()->dimension(2) : 1; const int32_t multiplier = output_stage.gemmlowp_multiplier; const int32_t shift = output_stage.gemmlowp_shift; const int32_t offset = output_stage.gemmlowp_offset; const int32_t min_bound = output_stage.gemmlowp_min_bound; const int32_t max_bound = output_stage.gemmlowp_max_bound; const int32x4_t result_offset_s32 = vdupq_n_s32(offset); const int32x4_t result_shift_s32 = vdupq_n_s32(is_fixed_point ? shift : -shift); const auto min_vec = wrapper::vdup_n(static_cast(min_bound), ExactTagType{}); const auto max_vec = wrapper::vdup_n(static_cast(max_bound), ExactTagType{}); const int window_step_x = 16; const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); Window win(window); win.set(Window::DimX, Window::Dimension(0, 1, 1)); Window collapsed_window = win.collapse_if_possible(win, Window::DimZ); Iterator mm_result_it(mm_result, win); Iterator out_it(output, win); if((a_offset != 0) && (b_offset != 0)) { ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row); Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); Iterator vector_sum_row_it = get_vector_sum_row_it(collapsed_window, vector_sum_row); const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); // Offset in case vector_sum_col is batched in y dimension const int vector_sum_col_stride_batch = is_vector_sum_col_batched ? vector_sum_col->info()->strides_in_bytes().y() : 0; if(bias != nullptr) { Iterator bias_it = get_bias_it(collapsed_window, bias); execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y() + (id.z() % depth_input) * height_input; run_offset_contribution_output_stage_window(vector_sum_col_ptr, vector_sum_row_ptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, true, true, true, is_bounded_relu, is_fixed_point); }, vector_sum_col_it, vector_sum_row_it, bias_it, mm_result_it, out_it); } else { execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y() + (id.z() % depth_input) * height_input; run_offset_contribution_output_stage_window(vector_sum_col_ptr, vector_sum_row_ptr, nullptr, mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, true, true, false, is_bounded_relu, is_fixed_point); }, vector_sum_col_it, vector_sum_row_it, mm_result_it, out_it); } } else if((a_offset == 0) && (b_offset != 0)) { ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_row); Iterator vector_sum_row_it = get_vector_sum_row_it(collapsed_window, vector_sum_row); const size_t sum_row_stride_y = vector_sum_row->info()->strides_in_bytes().y(); if(bias != nullptr) { Iterator bias_it = get_bias_it(collapsed_window, bias); execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y() + (id.z() % depth_input) * height_input; run_offset_contribution_output_stage_window(nullptr, vector_sum_row_ptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, false, true, true, is_bounded_relu, is_fixed_point); }, vector_sum_row_it, bias_it, mm_result_it, out_it); } else { execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_row_ptr = reinterpret_cast(vector_sum_row_it.ptr() + batch_id * sum_row_stride_y) + id.y() + (id.z() % depth_input) * height_input; run_offset_contribution_output_stage_window(nullptr, vector_sum_row_ptr, nullptr, mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, false, true, false, is_bounded_relu, is_fixed_point); }, vector_sum_row_it, mm_result_it, out_it); } } else if((a_offset != 0) && (b_offset == 0)) { ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); // Offset in case vector_sum_col is batched in y dimension const int vector_sum_col_stride_batch = is_vector_sum_col_batched ? vector_sum_col->info()->strides_in_bytes().y() : 0; if(bias != nullptr) { Iterator bias_it = get_bias_it(collapsed_window, bias); execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); run_offset_contribution_output_stage_window(vector_sum_col_ptr, nullptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, true, false, true, is_bounded_relu, is_fixed_point); }, vector_sum_col_it, bias_it, mm_result_it, out_it); } else { execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); run_offset_contribution_output_stage_window(vector_sum_col_ptr, nullptr, nullptr, mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, true, false, false, is_bounded_relu, is_fixed_point); }, vector_sum_col_it, mm_result_it, out_it); } } else { if(bias != nullptr) { Iterator bias_it = get_bias_it(collapsed_window, bias); execute_window_loop(collapsed_window, [&](const Coordinates &) { run_offset_contribution_output_stage_window(nullptr, nullptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, false, false, true, is_bounded_relu, is_fixed_point); }, bias_it, mm_result_it, out_it); } else { execute_window_loop(collapsed_window, [&](const Coordinates &) { run_offset_contribution_output_stage_window(nullptr, nullptr, nullptr, mm_result_it, out_it, result_offset_s32, result_shift_s32, min_vec, max_vec, a_offset, b_offset, k_offset, multiplier, shift, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, false, false, false, is_bounded_relu, is_fixed_point); }, mm_result_it, out_it); } return; } } void run_offset_contribution_output_stage_symm(const Window &window, const ITensor *mm_result, const ITensor *vector_sum_col, const ITensor *vector_sum_row, const ITensor *bias, ITensor *output, int32_t a_offset, int32_t b_offset, int32_t k_offset, bool is_vector_sum_col_batched, GEMMLowpOutputStageInfo output_stage, bool is_gemm3d, bool is_bounded_relu, bool is_fixed_point) { ARM_COMPUTE_UNUSED(vector_sum_row, b_offset, k_offset); const int depth_input = is_gemm3d ? mm_result->info()->dimension(2) : 1; const int32_t offset = output_stage.gemmlowp_offset; const int32_t min_bound = output_stage.gemmlowp_min_bound; const int32_t max_bound = output_stage.gemmlowp_max_bound; const int32_t *result_multipliers = output_stage.gemmlowp_multipliers.data(); const int32_t *result_shifts = output_stage.gemmlowp_shifts.data(); const int32x4_t result_offset_s32 = vdupq_n_s32(offset); const int8x16_t min_s8 = vdupq_n_s8(static_cast(min_bound)); const int8x16_t max_s8 = vdupq_n_s8(static_cast(max_bound)); const int window_step_x = 16; const auto window_start_x = static_cast(window.x().start()); const auto window_end_x = static_cast(window.x().end()); Window win(window); win.set(Window::DimX, Window::Dimension(0, 1, 1)); Window collapsed_window = win.collapse_if_possible(win, Window::DimZ); Iterator mm_result_it(mm_result, win); Iterator out_it(output, win); if(a_offset != 0) { ARM_COMPUTE_ERROR_ON_NULLPTR(vector_sum_col); Iterator vector_sum_col_it = get_vector_sum_col_it(collapsed_window, vector_sum_col); // Offset in case vector_sum_col is batched in y dimension const int vector_sum_col_stride_batch = is_vector_sum_col_batched ? vector_sum_col->info()->strides_in_bytes().y() : 0; if(bias != nullptr) { Iterator bias_it = get_bias_it(collapsed_window, bias); execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); run_offset_contribution_output_stage_window_symm(vector_sum_col_ptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, result_multipliers, result_shifts, result_offset_s32, min_s8, max_s8, a_offset, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, true, true, is_bounded_relu, is_fixed_point); }, vector_sum_col_it, bias_it, mm_result_it, out_it); } else { execute_window_loop(collapsed_window, [&](const Coordinates & id) { const int batch_id = id.z() / depth_input; const auto vector_sum_col_ptr = reinterpret_cast(vector_sum_col_it.ptr() + batch_id * vector_sum_col_stride_batch); run_offset_contribution_output_stage_window_symm(vector_sum_col_ptr, nullptr, mm_result_it, out_it, result_multipliers, result_shifts, result_offset_s32, min_s8, max_s8, a_offset, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, true, false, is_bounded_relu, is_fixed_point); }, vector_sum_col_it, mm_result_it, out_it); } } else { if(bias != nullptr) { Iterator bias_it = get_bias_it(collapsed_window, bias); execute_window_loop(collapsed_window, [&](const Coordinates &) { run_offset_contribution_output_stage_window_symm(nullptr, reinterpret_cast(bias_it.ptr()), mm_result_it, out_it, result_multipliers, result_shifts, result_offset_s32, min_s8, max_s8, a_offset, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, false, true, is_bounded_relu, is_fixed_point); }, bias_it, mm_result_it, out_it); } else { execute_window_loop(collapsed_window, [&](const Coordinates &) { run_offset_contribution_output_stage_window_symm(nullptr, nullptr, mm_result_it, out_it, result_multipliers, result_shifts, result_offset_s32, min_s8, max_s8, a_offset, offset, min_bound, max_bound, window_step_x, window_start_x, window_end_x, false, false, is_bounded_relu, is_fixed_point); }, mm_result_it, out_it); } return; } } Status validate_arguments(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mm_result, 1, DataType::S32); if(output->data_type() != DataType::QASYMM8) { ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) > 1 && output_stage.gemmlowp_multipliers.size() > 1 && b_offset != 0); } ARM_COMPUTE_RETURN_ERROR_ON(output_stage.gemmlowp_min_bound > output_stage.gemmlowp_max_bound); ARM_COMPUTE_RETURN_ERROR_ON(output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN && output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT); if(bias != nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1); ARM_COMPUTE_RETURN_ERROR_ON(mm_result->dimension(0) != bias->dimension(0)); } // If a_offset == 0, vector_sum_col can be a nullptr if(a_offset != 0) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->dimension(0) != mm_result->dimension(0)); ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_col->num_dimensions() > 2); } // If b_offset == 0, vector_sum_row can be a nullptr if(b_offset != 0) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); // Check if input is a 3D reinterpretation const bool reinterpret_as_3d = mm_result->num_dimensions() > 1 && mm_result->tensor_shape().y() != vector_sum_row->tensor_shape().x(); // Validate input ARM_COMPUTE_RETURN_ERROR_ON(reinterpret_as_3d && vector_sum_row->dimension(0) != (mm_result->dimension(1) * mm_result->dimension(2))); ARM_COMPUTE_RETURN_ERROR_ON(!reinterpret_as_3d && vector_sum_row->dimension(0) != mm_result->dimension(1)); TensorShape output_shape = output->tensor_shape(); if(output_shape.num_dimensions() > 1) { const unsigned int output_batch_idx = reinterpret_as_3d ? 3 : 2; TensorShape vector_sum_row_shape = vector_sum_row->tensor_shape(); vector_sum_row_shape.collapse_from(1); output_shape.collapse_from(output_batch_idx); ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_row_shape[1] != output_shape[output_batch_idx], "mm_result tensor must have the same number of batches of output tensor"); if(a_offset != 0) { TensorShape vector_sum_col_shape = vector_sum_col->tensor_shape(); vector_sum_col_shape.collapse_from(1); ARM_COMPUTE_RETURN_ERROR_ON_MSG(vector_sum_col_shape[1] != 1 && vector_sum_col_shape[1] != vector_sum_row_shape[1], "vector_sum_col tensor must have the same number of batches of vector_sum_row_shape or the number of batches must be set to 1"); } } // Check Tensor Rank of vector_sum_row ARM_COMPUTE_RETURN_ERROR_ON(vector_sum_row->num_dimensions() > 3); } if(output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(mm_result, output); } return Status{}; } } // namespace void CpuGemmLowpOffsetContributionOutputStageKernel::configure(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, ITensorInfo *dst, int32_t k, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) { ARM_COMPUTE_UNUSED(vector_sum_row, bias); // Perform validate step ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, dst); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, dst, a_offset, b_offset, output_stage)); _a_offset = a_offset; _b_offset = b_offset; _k_offset = a_offset * b_offset * k; _output_stage = output_stage; // If a_offset == 0, vector_sum_col can be a nullptr if(a_offset != 0) { // Check if vector_sum_col_shape should be slidden or not // Don't slide vector_sum_col_shape along the y dimension if vector_sum_col_shape has just 1 dimension and vector_sum_row_shape more than 1 // This scenario can happen when the the matrix multiplication is used to perform a convolution operation _is_vector_sum_col_batched = vector_sum_col->tensor_shape().num_dimensions() > 1; } // Output auto inizialitation if not yet initialized auto_init_if_empty(*dst, mm_result->clone()->set_data_type(DataType::QASYMM8)); // Configure kernel window Window win = calculate_max_window(*mm_result, Steps()); // Note: This kernel performs 16 elements per iteration. // However, since we use a left-over for loop, we cannot have any read or write out of memory // For this reason num_elems_processed_per_iteration is 1 and so update_window_and_padding() can be skipped ICpuKernel::configure(win); } Status CpuGemmLowpOffsetContributionOutputStageKernel::validate(const ITensorInfo *mm_result, const ITensorInfo *vector_sum_col, const ITensorInfo *vector_sum_row, const ITensorInfo *bias, const ITensorInfo *output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) { ARM_COMPUTE_ERROR_ON_NULLPTR(mm_result, output); ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(mm_result, vector_sum_col, vector_sum_row, bias, output, a_offset, b_offset, output_stage)); return Status{}; } void CpuGemmLowpOffsetContributionOutputStageKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) { ARM_COMPUTE_UNUSED(info); ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICpuKernel::window(), window); auto mm_result = tensors.get_const_tensor(TensorType::ACL_SRC_0); auto vector_sum_col = tensors.get_const_tensor(TensorType::ACL_SRC_1); auto vector_sum_row = tensors.get_const_tensor(TensorType::ACL_SRC_2); auto bias = tensors.get_const_tensor(TensorType::ACL_SRC_3); auto dst = tensors.get_tensor(TensorType::ACL_DST); PixelValue type_min{}; PixelValue type_max{}; std::tie(type_min, type_max) = get_min_max(dst->info()->data_type()); int32_t type_min_int = type_min.get(); int32_t type_max_int = type_max.get(); const bool reinterpret_as_3d = vector_sum_row != nullptr && mm_result->info()->num_dimensions() > 1 && mm_result->info()->tensor_shape().y() != vector_sum_row->info()->tensor_shape().x(); const bool is_bounded_relu = !(_output_stage.gemmlowp_min_bound <= type_min_int && _output_stage.gemmlowp_max_bound >= type_max_int); // Check if we need to perform fixed point requantization const bool is_fixed_point = _output_stage.type != GEMMLowpOutputStageType::QUANTIZE_DOWN; // Check if symmetric per-channel execution const bool is_signed = dst->info()->data_type() == DataType::QASYMM8_SIGNED; // Check if symmetric per-channel execution const bool is_symm = _output_stage.is_quantized_per_channel; if(is_symm) { run_offset_contribution_output_stage_symm(window, mm_result, vector_sum_col, vector_sum_row, bias, dst, _a_offset, _b_offset, _k_offset, _is_vector_sum_col_batched, _output_stage, reinterpret_as_3d, is_bounded_relu, is_fixed_point); } else { if(is_signed) { run_offset_contribution_output_stage(window, mm_result, vector_sum_col, vector_sum_row, bias, dst, _a_offset, _b_offset, _k_offset, _is_vector_sum_col_batched, _output_stage, reinterpret_as_3d, is_bounded_relu, is_fixed_point); } else { run_offset_contribution_output_stage(window, mm_result, vector_sum_col, vector_sum_row, bias, dst, _a_offset, _b_offset, _k_offset, _is_vector_sum_col_batched, _output_stage, reinterpret_as_3d, is_bounded_relu, is_fixed_point); } } } const char *CpuGemmLowpOffsetContributionOutputStageKernel::name() const { return "CpuGemmLowpOffsetContributionOutputStageKernel"; } } // namespace kernels } // namespace cpu } // namespace arm_compute