// Copyright 2021 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. $assert REQUANTIZATION in ["FP32", "RNDNU"] $assert not CHANNELWISE or REQUANTIZATION == "FP32" #include $DATATYPE = "qc8" if CHANNELWISE else "qs8" $PARAMS_UNION = "xnn_qs8_minmax_params" if CHANNELWISE else "xnn_qs8_conv_minmax_params" # void xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c8__aarch64_neon_mull( # size_t mr, x0 # size_t nc, x1 # size_t kc, x2 / x0 # const int8_t* restrict a, x3 # size_t a_stride, x4 # const void* restrict w, x5 # int8_t* restrict c, x6 # size_t cm_stride, x7 # size_t cn_stride, [sp] -> x10 # const union ${PARAMS_UNION} params) [sp + 8] -> x11 # d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. # Register usage # A0 x3 v0 # A1 x4 v1 # B x5 v4 v5 v6 v7 # C0 x7 v16 v18 v20 v22 v24 v26 v28 v30 # C1 x8 v17 v19 v21 v23 v25 v27 v29 v31 # temp0 v2 v10 v12 v14 # temp1 v3 v11 v13 v15 # unused v8 v9 BEGIN_FUNCTION xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c8__aarch64_neon_mull # Clamp A and C pointers CMP x0, 2 // if mr < 2 STP d10, d11, [sp, -48]! ADD x4, x3, x4 // a1 = a0 + a_stride STP d12, d13, [sp, 16] ADD x7, x6, x7 // c1 = c0 + cm_stride STP d14, d15, [sp, 32] CSEL x4, x3, x4, LO // a1 = a0 ADD x2, x2, 7 // kc = (kc + 7) & ~7 CSEL x7, x6, x7, LO // c1 = c0 BIC x2, x2, 7 .p2align 3 0: # Load initial bias from w into accumulators MOV x0, x2 // k = kc LDP s16, s18, [x5], 8 MOV v17.16b, v16.16b MOV v19.16b, v18.16b LDP s20, s22, [x5], 8 MOV v21.16b, v20.16b MOV v23.16b, v22.16b LDP s24, s26, [x5], 8 MOV v25.16b, v24.16b MOV v27.16b, v26.16b LDP s28, s30, [x5], 8 MOV v29.16b, v28.16b LDP x10, x11, [sp, 48] // cn_stride, params MOV v31.16b, v30.16b # Main loop - 8 bytes of A .p2align 3 1: LDR d0, [x3], 8 LDP d4, d5, [x5] LDR d1, [x4], 8 LDP d6, d7, [x5, 16] SMULL v2.8h, v4.8b, v0.8b SMULL v3.8h, v4.8b, v1.8b SMULL v10.8h, v5.8b, v0.8b SMULL v11.8h, v5.8b, v1.8b SMULL v12.8h, v6.8b, v0.8b SADALP v16.4s, v2.8h SMULL v13.8h, v6.8b, v1.8b SADALP v17.4s, v3.8h SMULL v14.8h, v7.8b, v0.8b SADALP v18.4s, v10.8h SMULL v15.8h, v7.8b, v1.8b SADALP v19.4s, v11.8h LDP d4, d5, [x5, 32] SMULL v2.8h, v4.8b, v0.8b SADALP v20.4s, v12.8h SMULL v3.8h, v4.8b, v1.8b SADALP v21.4s, v13.8h SMULL v10.8h, v5.8b, v0.8b SADALP v22.4s, v14.8h SMULL v11.8h, v5.8b, v1.8b SADALP v23.4s, v15.8h LDP d6, d7, [x5, 48] SMULL v12.8h, v6.8b, v0.8b SADALP v24.4s, v2.8h SMULL v13.8h, v6.8b, v1.8b SADALP v25.4s, v3.8h SMULL v14.8h, v7.8b, v0.8b SADALP v26.4s, v10.8h SMULL v15.8h, v7.8b, v1.8b SADALP v27.4s, v11.8h ADD x5, x5, 64 SADALP v28.4s, v12.8h SADALP v29.4s, v13.8h SUBS x0, x0, 8 SADALP v30.4s, v14.8h SADALP v31.4s, v15.8h B.HI 1b # Add columns ADDP v16.4s, v16.4s, v18.4s ADDP v20.4s, v20.4s, v22.4s $if REQUANTIZATION == "RNDNU": LD1R {v4.4s}, [x11], 4 ADDP v24.4s, v24.4s, v26.4s ADDP v28.4s, v28.4s, v30.4s $if REQUANTIZATION == "RNDNU": LD1R {v7.4s}, [x11], 4 ADDP v17.4s, v17.4s, v19.4s ADDP v21.4s, v21.4s, v23.4s ADDP v25.4s, v25.4s, v27.4s ADDP v29.4s, v29.4s, v31.4s ADDP v0.4s, v16.4s, v20.4s ADDP v1.4s, v24.4s, v28.4s ADDP v2.4s, v17.4s, v21.4s ADDP v3.4s, v25.4s, v29.4s $if REQUANTIZATION == "RNDNU": # Apply params - preshift, scale, postshift, bias and clamp LD1R {v5.4s}, [x11], 4 SQSHL v0.4s, v0.4s, v4.4s // shift to upper bits SQSHL v1.4s, v1.4s, v4.4s SQSHL v2.4s, v2.4s, v4.4s SQSHL v3.4s, v3.4s, v4.4s SQDMULH v0.4s, v0.4s, v7.4s // scale without rounding SQDMULH v1.4s, v1.4s, v7.4s SQDMULH v2.4s, v2.4s, v7.4s SQDMULH v3.4s, v3.4s, v7.4s SRSHL v0.4s, v0.4s, v5.4s // signed rounding shift left SRSHL v1.4s, v1.4s, v5.4s SRSHL v2.4s, v2.4s, v5.4s SRSHL v3.4s, v3.4s, v5.4s $elif REQUANTIZATION == "FP32": $if not CHANNELWISE: # Apply params - scale, bias and clamp SCVTF v0.4s, v0.4s LD1R {v4.4s}, [x11], 4 SCVTF v1.4s, v1.4s SCVTF v2.4s, v2.4s SCVTF v3.4s, v3.4s FMUL v0.4s, v0.4s, v4.4s FMUL v1.4s, v1.4s, v4.4s FMUL v2.4s, v2.4s, v4.4s FMUL v3.4s, v3.4s, v4.4s $else: # Load per channel scale values from weights SCVTF v0.4s, v0.4s LDR q4, [x5], 16 SCVTF v1.4s, v1.4s LDR q5, [x5], 16 SCVTF v2.4s, v2.4s SCVTF v3.4s, v3.4s FMUL v0.4s, v0.4s, v4.4s FMUL v1.4s, v1.4s, v5.4s FMUL v2.4s, v2.4s, v4.4s FMUL v3.4s, v3.4s, v5.4s FCVTNS v0.4s, v0.4s FCVTNS v1.4s, v1.4s FCVTNS v2.4s, v2.4s FCVTNS v3.4s, v3.4s LD1R {v5.8h}, [x11], 2 SQXTN v0.4h, v0.4s SQXTN v2.4h, v2.4s SQXTN2 v0.8h, v1.4s SQXTN2 v2.8h, v3.4s SUBS x1, x1, 8 SQADD v0.8h, v0.8h, v5.8h SQADD v1.8h, v2.8h, v5.8h SQXTN v0.8b, v0.8h SQXTN2 v0.16b, v1.8h LD1R {v1.16b}, [x11], 1 LD1R {v2.16b}, [x11] SMAX v0.16b, v0.16b, v1.16b SMIN v0.16b, v0.16b, v2.16b B.LO 2f # Store full 2 x 8 ST1 {v0.8b}, [x6], x10 SUB x3, x3, x2 // a0 -= kc ST1 {v0.d}[1], [x7], x10 SUB x4, x4, x2 // a1 -= kc B.HI 0b # Restore d10-d15 from stack LDP d14, d15, [sp, 32] LDP d12, d13, [sp, 16] LDP d10, d11, [sp], 48 RET # Store odd width .p2align 3 2: TBZ x1, 2, 3f STR s0, [x6], 4 ST1 {v0.s}[2], [x7], 4 EXT v0.16b, v0.16b, v0.16b, 4 3: TBZ x1, 1, 4f STR h0, [x6], 2 ST1 {v0.h}[4], [x7], 2 EXT v0.16b, v0.16b, v0.16b, 2 4: TBZ x1, 0, 5f STR b0, [x6] ST1 {v0.b}[8], [x7] 5: # Restore d10-d15 from stack LDP d14, d15, [sp, 32] LDP d12, d13, [sp, 16] LDP d10, d11, [sp], 48 RET END_FUNCTION xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c8__aarch64_neon_mull #ifdef __ELF__ .section ".note.GNU-stack","",%progbits #endif