// 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 # void xnn_f32_igemm_minmax_ukernel_4x2__aarch64_neonfma${"_prfm" if PREFETCH else ""}_cortex_a75( # size_t mr, x0 # size_t nc, x1 # size_t kc, x2 / x0 # size_t ks, x3 / x9 # const float**restrict a, x4 # const float*restrict w, x5 # float*restrict c, x6 # size_t cm_stride, x7 # size_t cn_stride, [sp] -> x10 # size_t a_offset, [sp + 8] -> x11 # const float* zero, [sp + 16] -> x12 # const xnn_f32_minmax_params params [sp + 24] -> x8 # d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. # Register usage # A0 x20 v0 v4 # A1 x13 v1 v5 # A2 x14 v2 v6 # A3 x15 v3 v7 # B x5 v16 v17 v18 v19 v20 v21 v22 v23 # C0 x6 v24 v25 # C1 x16 v26 v27 # C2 x17 v28 v29 # C3 x7 v30 v31 # Clamp v4 v5 BEGIN_FUNCTION xnn_f32_igemm_minmax_ukernel_4x2__aarch64_neonfma${"_prfm" if PREFETCH else ""}_cortex_a75 # Load cn_stride, a_offset LDP x10, x11, [sp] # Load zero, params pointer LDP x12, x8, [sp, 16] # Load min/max values LD2R {v4.2s, v5.2s}, [x8] # Save x20 on stack STR x20, [sp, -16]! # Clamp C pointers CMP x0, 2 // if mr < 2 ADD x16, x6, x7 // c1 = c0 + cm_stride CSEL x16, x6, x16, LO // c1 = c0 ADD x17, x16, x7 // c2 = c1 + cm_stride // if mr <= 2 CSEL x17, x16, x17, LS // c2 = c1 CMP x0, 4 // if mr < 4 ADD x7, x17, x7 // c3 = c2 + cm_stride CSEL x7, x17, x7, LO // c3 = c2 0: # Load initial bias from w into accumulators LDR d24, [x5], 8 MOV v26.8b, v24.8b MOV v28.8b, v24.8b MOV v30.8b, v24.8b MOVI v25.2s, 0 $if PREFETCH: PRFM PLDL1KEEP, [x5, 64] MOVI v27.2s, 0 $if PREFETCH: PRFM PLDL1KEEP, [x5, 128] MOVI v29.2s, 0 $if PREFETCH: PRFM PLDL1KEEP, [x5, 192] MOVI v31.2s, 0 $if PREFETCH: PRFM PLDL1KEEP, [x5, 256] MOV x9, x3 // p = ks 1: # Load next 4 A pointers LDP x20, x13, [x4], 16 LDP x14, x15, [x4], 16 CMP x20, x12 // if a0 == zero ADD x20, x20, x11 // a0 += a_offset CSEL x20, x12, x20, EQ // a0 = zero, else += a0 + a_offset CMP x13, x12 // if a1 == zero ADD x13, x13, x11 // a1 += a_offset CSEL x13, x12, x13, EQ // a1 = zero, else += a1 + a_offset CMP x14, x12 // if a2 == zero ADD x14, x14, x11 // a2 += a_offset CSEL x14, x12, x14, EQ // a2 = zero, else += a2 + a_offset CMP x15, x12 // if a3 == zero ADD x15, x15, x11 // a3 += a_offset CSEL x15, x12, x15, EQ // a3 = zero, else += a3 + a_offset # Is there at least 8 floats (32 bytes) for prologue + epilogue? SUBS x0, x2, 32 // k = kc - 32 B.LO 5f # Prologue # Read first block of 4 A and B. LDR q0, [x20], 16 LDP d20, d21, [x5], 16 LDR q1, [x13], 16 LDR q2, [x14], 16 LDR q3, [x15], 16 LDP d22, d23, [x5], 16 # Is there at least 32. yes do main loop SUBS x0, x0, 32 B.LO 3f # Main loop - 8 floats of A (32 bytes) 2: # First block of 4. FMA for first 4, loads for 2nd block of 4. FMLA v24.2s, v20.2s, v0.s[0] LDR q4, [x20], 16 FMLA v26.2s, v20.2s, v1.s[0] FMLA v28.2s, v20.2s, v2.s[0] LDR d16, [x5, 0] FMLA v30.2s, v20.2s, v3.s[0] FMLA v25.2s, v21.2s, v0.s[1] LDR q5, [x13], 16 FMLA v27.2s, v21.2s, v1.s[1] FMLA v29.2s, v21.2s, v2.s[1] LDR q6, [x14], 16 FMLA v31.2s, v21.2s, v3.s[1] FMLA v24.2s, v22.2s, v0.s[2] LDR q7, [x15], 16 FMLA v26.2s, v22.2s, v1.s[2] FMLA v28.2s, v22.2s, v2.s[2] LDR d17, [x5, 8] FMLA v30.2s, v22.2s, v3.s[2] FMLA v25.2s, v23.2s, v0.s[3] LDR d18, [x5, 16] FMLA v27.2s, v23.2s, v1.s[3] FMLA v29.2s, v23.2s, v2.s[3] LDR d19, [x5, 24] FMLA v31.2s, v23.2s, v3.s[3] $if PREFETCH: PRFM PLDL1KEEP, [x5, 320] # Second block of 4. FMA for second 4, loads for 1st block of 4. FMLA v24.2s, v16.2s, v4.s[0] LDR q0, [x20], 16 FMLA v26.2s, v16.2s, v5.s[0] FMLA v28.2s, v16.2s, v6.s[0] LDR d20, [x5, 32] FMLA v30.2s, v16.2s, v7.s[0] FMLA v25.2s, v17.2s, v4.s[1] LDR q1, [x13], 16 FMLA v27.2s, v17.2s, v5.s[1] FMLA v29.2s, v17.2s, v6.s[1] LDR q2, [x14], 16 FMLA v31.2s, v17.2s, v7.s[1] FMLA v24.2s, v18.2s, v4.s[2] LDR q3, [x15], 16 FMLA v26.2s, v18.2s, v5.s[2] FMLA v28.2s, v18.2s, v6.s[2] LDR d21, [x5, 40] FMLA v30.2s, v18.2s, v7.s[2] SUBS x0, x0, 32 FMLA v25.2s, v19.2s, v4.s[3] LDR d22, [x5, 48] FMLA v27.2s, v19.2s, v5.s[3] LDR d23, [x5, 56] FMLA v29.2s, v19.2s, v6.s[3] ADD x5, x5, 64 FMLA v31.2s, v19.2s, v7.s[3] B.HS 2b 3: # Epilogue # First block of 4. FMA for first 4, loads for 2nd block of 4. FMLA v24.2s, v20.2s, v0.s[0] LDR q4, [x20], 16 FMLA v26.2s, v20.2s, v1.s[0] FMLA v28.2s, v20.2s, v2.s[0] LDR d16, [x5, 0] FMLA v30.2s, v20.2s, v3.s[0] FMLA v25.2s, v21.2s, v0.s[1] LDR q5, [x13], 16 FMLA v27.2s, v21.2s, v1.s[1] FMLA v29.2s, v21.2s, v2.s[1] LDR q6, [x14], 16 FMLA v31.2s, v21.2s, v3.s[1] FMLA v24.2s, v22.2s, v0.s[2] LDR q7, [x15], 16 FMLA v26.2s, v22.2s, v1.s[2] FMLA v28.2s, v22.2s, v2.s[2] LDR d17, [x5, 8] FMLA v30.2s, v22.2s, v3.s[2] FMLA v25.2s, v23.2s, v0.s[3] LDR d18, [x5, 16] FMLA v27.2s, v23.2s, v1.s[3] FMLA v29.2s, v23.2s, v2.s[3] LDR d19, [x5, 24] FMLA v31.2s, v23.2s, v3.s[3] $if PREFETCH: PRFM PLDL1KEEP, [x5, 320] # Second block of 4. FMA for second 4, no loads FMLA v24.2s, v16.2s, v4.s[0] FMLA v26.2s, v16.2s, v5.s[0] FMLA v28.2s, v16.2s, v6.s[0] FMLA v30.2s, v16.2s, v7.s[0] FMLA v25.2s, v17.2s, v4.s[1] FMLA v27.2s, v17.2s, v5.s[1] FMLA v29.2s, v17.2s, v6.s[1] FMLA v31.2s, v17.2s, v7.s[1] FMLA v24.2s, v18.2s, v4.s[2] FMLA v26.2s, v18.2s, v5.s[2] FMLA v28.2s, v18.2s, v6.s[2] ADDS x0, x0, 32 FMLA v30.2s, v18.2s, v7.s[2] FMLA v25.2s, v19.2s, v4.s[3] ADD x5, x5, 32 FMLA v27.2s, v19.2s, v5.s[3] FMLA v29.2s, v19.2s, v6.s[3] LD2R {v4.2s, v5.2s}, [x8] // Load min/max values FMLA v31.2s, v19.2s, v7.s[3] # Is there a remainder? up to 8 floats (32 bytes) B.NE 5f 4: # ks loop SUBS x9, x9, 32 // ks -= MR * sizeof(void*) B.HI 1b FADD v24.2s, v24.2s, v25.2s FADD v26.2s, v26.2s, v27.2s FADD v28.2s, v28.2s, v29.2s FADD v30.2s, v30.2s, v31.2s # Clamp FMAX v24.2s, v24.2s, v4.2s FMAX v26.2s, v26.2s, v4.2s FMAX v28.2s, v28.2s, v4.2s FMAX v30.2s, v30.2s, v4.2s SUBS x1, x1, 2 FMIN v24.2s, v24.2s, v5.2s FMIN v26.2s, v26.2s, v5.2s FMIN v28.2s, v28.2s, v5.2s FMIN v30.2s, v30.2s, v5.2s # Store full 4 x 2 B.LO 8f STR d30, [x7] ADD x7, x7, x10 STR d28, [x17] ADD x17, x17, x10 STR d26, [x16] ADD x16, x16, x10 STR d24, [x6] ADD x6, x6, x10 SUB x4, x4, x3 // a -= ks # nc loop B.HI 0b # Restore x20 from stack LDR x20, [sp], 16 RET 5: # Remainder- 4 floats of A (16 bytes) TBZ x0, 4, 6f LDR q0, [x20], 16 LDP d20, d21, [x5], 16 LDR q1, [x13], 16 LDR q2, [x14], 16 LDR q3, [x15], 16 LDP d22, d23, [x5], 16 FMLA v24.2s, v20.2s, v0.s[0] FMLA v26.2s, v20.2s, v1.s[0] FMLA v28.2s, v20.2s, v2.s[0] FMLA v30.2s, v20.2s, v3.s[0] FMLA v25.2s, v21.2s, v0.s[1] FMLA v27.2s, v21.2s, v1.s[1] FMLA v29.2s, v21.2s, v2.s[1] FMLA v31.2s, v21.2s, v3.s[1] FMLA v24.2s, v22.2s, v0.s[2] FMLA v26.2s, v22.2s, v1.s[2] FMLA v28.2s, v22.2s, v2.s[2] FMLA v30.2s, v22.2s, v3.s[2] FMLA v25.2s, v23.2s, v0.s[3] FMLA v27.2s, v23.2s, v1.s[3] FMLA v29.2s, v23.2s, v2.s[3] FMLA v31.2s, v23.2s, v3.s[3] 6: # Remainder- 2 floats of A (8 bytes) TBZ x0, 3, 7f LDR d0, [x20], 8 LDP d20, d21, [x5], 16 LDR d1, [x13], 8 LDR d2, [x14], 8 LDR d3, [x15], 8 FMLA v24.2s, v20.2s, v0.s[0] FMLA v26.2s, v20.2s, v1.s[0] FMLA v28.2s, v20.2s, v2.s[0] FMLA v30.2s, v20.2s, v3.s[0] FMLA v25.2s, v21.2s, v0.s[1] FMLA v27.2s, v21.2s, v1.s[1] FMLA v29.2s, v21.2s, v2.s[1] FMLA v31.2s, v21.2s, v3.s[1] 7: # Remainder- 1 float of A (4 bytes) TBZ x0, 2, 4b LDR s0, [x20], 4 LDR d20, [x5], 8 LDR s1, [x13], 4 LDR s2, [x14], 4 LDR s3, [x15], 4 FMLA v24.2s, v20.2s, v0.s[0] FMLA v26.2s, v20.2s, v1.s[0] FMLA v28.2s, v20.2s, v2.s[0] FMLA v30.2s, v20.2s, v3.s[0] B 4b # Store odd width 8: STR s30, [x7] STR s28, [x17] STR s26, [x16] STR s24, [x6] # Restore x20 from stack LDR x20, [sp], 16 RET END_FUNCTION xnn_f32_igemm_minmax_ukernel_4x2__aarch64_neonfma${"_prfm" if PREFETCH else ""}_cortex_a75 #ifdef __ELF__ .section ".note.GNU-stack","",%progbits #endif