// Copyright 2020 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" $REWIND_DECREMENT = 3 if CHANNELWISE else {"RNDNU": 15, "FP32": 7}[REQUANTIZATION] # void xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_4x16c4__aarch64_neondot_ld32( # 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] -> x12 # const union ${PARAMS_UNION} params) [sp + 8] -> x11 # d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS. # Register usage # A0 x3 v0 # A1 x15 v1 # A2 x13 v2 # A3 x4 v3 # B x5 v4 v5 v6 v7 # C0 x6 v16 v20 v24 v28 # C1 x8 v17 v21 v25 v29 # C2 x9 v18 v22 v26 v30 # C3 x7 v19 v23 v27 v31 # unused v8 v9 v10 v11 v12 v13 v14 v15 BEGIN_FUNCTION xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_4x16c4__aarch64_neondot_ld32 # Clamp A and C pointers CMP x0, 2 // if mr < 2 ADD x2, x2, 3 // kc = (kc + 3) & ~3 ADD x15, x3, x4 // a1 = a0 + a_stride ADD x8, x6, x7 // c1 = c0 + cm_stride CSEL x15, x3, x15, LO // a1 = a0 CSEL x8, x6, x8, LO // c1 = c0 BIC x2, x2, 3 ADD x13, x15, x4 // a2 = a1 + a_stride ADD x9, x8, x7 // c2 = c1 + cm_stride // if mr <= 2 CSEL x13, x15, x13, LS // a2 = a1 CSEL x9, x8, x9, LS // c2 = c1 LDP x12, x11, [sp] // cn_stride, params CMP x0, 4 // if mr < 4 ADD x4, x13, x4 // a3 = a2 + a_stride ADD x7, x9, x7 // c3 = c2 + cm_stride CSEL x4, x13, x4, LO // a3 = a2 CSEL x7, x9, x7, LO // c3 = c2 .p2align 3 0: # Load initial bias from w into accumulators LDP q16, q20, [x5], 32 MOV v17.16b, v16.16b MOV v18.16b, v16.16b LDP q24, q28, [x5], 32 MOV v19.16b, v16.16b MOV v21.16b, v20.16b MOV v22.16b, v20.16b MOV v23.16b, v20.16b MOV v25.16b, v24.16b MOV v26.16b, v24.16b MOV x0, x2 // k = kc. assumes kc > 0 MOV v27.16b, v24.16b MOV v29.16b, v28.16b MOV v30.16b, v28.16b MOV v31.16b, v28.16b # Main loop - 4 bytes of A .p2align 3 1: LD1R {v0.4s}, [x3], 4 LDR q4, [x5], 16 LD1R {v1.4s}, [x15], 4 LD1R {v2.4s}, [x13], 4 LD1R {v3.4s}, [x4], 4 SDOT v16.4s, v4.16b, v0.16b SDOT v17.4s, v4.16b, v1.16b LDR q5, [x5], 16 SDOT v18.4s, v4.16b, v2.16b SDOT v19.4s, v4.16b, v3.16b LDR q6, [x5], 16 SDOT v20.4s, v5.16b, v0.16b SDOT v21.4s, v5.16b, v1.16b LDR q7, [x5], 16 SDOT v22.4s, v5.16b, v2.16b SDOT v23.4s, v5.16b, v3.16b SUBS x0, x0, 4 SDOT v24.4s, v6.16b, v0.16b SDOT v25.4s, v6.16b, v1.16b SDOT v26.4s, v6.16b, v2.16b SDOT v27.4s, v6.16b, v3.16b SDOT v28.4s, v7.16b, v0.16b SDOT v29.4s, v7.16b, v1.16b SDOT v30.4s, v7.16b, v2.16b SDOT v31.4s, v7.16b, v3.16b B.HI 1b $if REQUANTIZATION == "RNDNU": # Apply params - preshift, scale, postshift, bias and clamp LD1R {v4.4s}, [x11], 4 SQSHL v16.4s, v16.4s, v4.4s // shift to upper bits SQSHL v17.4s, v17.4s, v4.4s SQSHL v18.4s, v18.4s, v4.4s SQSHL v19.4s, v19.4s, v4.4s SQSHL v20.4s, v20.4s, v4.4s SQSHL v21.4s, v21.4s, v4.4s SQSHL v22.4s, v22.4s, v4.4s SQSHL v23.4s, v23.4s, v4.4s LD1R {v5.4s}, [x11], 4 SQSHL v24.4s, v24.4s, v4.4s SQSHL v25.4s, v25.4s, v4.4s SQSHL v26.4s, v26.4s, v4.4s SQSHL v27.4s, v27.4s, v4.4s SQSHL v28.4s, v28.4s, v4.4s SQSHL v29.4s, v29.4s, v4.4s SQSHL v30.4s, v30.4s, v4.4s SQSHL v31.4s, v31.4s, v4.4s LD1R {v6.4s}, [x11], 4 SQDMULH v16.4s, v16.4s, v5.4s // scale without rounding SQDMULH v17.4s, v17.4s, v5.4s SQDMULH v18.4s, v18.4s, v5.4s SQDMULH v19.4s, v19.4s, v5.4s SQDMULH v20.4s, v20.4s, v5.4s SQDMULH v21.4s, v21.4s, v5.4s SQDMULH v22.4s, v22.4s, v5.4s SQDMULH v23.4s, v23.4s, v5.4s SQDMULH v24.4s, v24.4s, v5.4s SQDMULH v25.4s, v25.4s, v5.4s SQDMULH v26.4s, v26.4s, v5.4s SQDMULH v27.4s, v27.4s, v5.4s SQDMULH v28.4s, v28.4s, v5.4s SQDMULH v29.4s, v29.4s, v5.4s SQDMULH v30.4s, v30.4s, v5.4s SQDMULH v31.4s, v31.4s, v5.4s SRSHL v16.4s, v16.4s, v6.4s // signed rounding shift left SRSHL v17.4s, v17.4s, v6.4s SRSHL v18.4s, v18.4s, v6.4s SRSHL v19.4s, v19.4s, v6.4s SRSHL v20.4s, v20.4s, v6.4s SRSHL v21.4s, v21.4s, v6.4s SRSHL v22.4s, v22.4s, v6.4s SRSHL v23.4s, v23.4s, v6.4s SRSHL v24.4s, v24.4s, v6.4s SRSHL v25.4s, v25.4s, v6.4s SRSHL v26.4s, v26.4s, v6.4s SRSHL v27.4s, v27.4s, v6.4s SRSHL v28.4s, v28.4s, v6.4s SRSHL v29.4s, v29.4s, v6.4s SRSHL v30.4s, v30.4s, v6.4s SRSHL v31.4s, v31.4s, v6.4s $elif REQUANTIZATION == "FP32": SCVTF v16.4s, v16.4s SCVTF v17.4s, v17.4s $if not CHANNELWISE: # Apply params - scale, bias and clamp LD1R {v4.4s}, [x11], 4 SCVTF v18.4s, v18.4s SCVTF v19.4s, v19.4s $else: # Load per channel scale values from weights LDR q4, [x5], 16 SCVTF v18.4s, v18.4s SCVTF v19.4s, v19.4s LDR q5, [x5], 16 SCVTF v20.4s, v20.4s SCVTF v21.4s, v21.4s SCVTF v22.4s, v22.4s SCVTF v23.4s, v23.4s SCVTF v24.4s, v24.4s SCVTF v25.4s, v25.4s SCVTF v26.4s, v26.4s SCVTF v27.4s, v27.4s SCVTF v28.4s, v28.4s SCVTF v29.4s, v29.4s SCVTF v30.4s, v30.4s SCVTF v31.4s, v31.4s $if CHANNELWISE: LDR q6, [x5], 16 FMUL v16.4s, v16.4s, v4.4s FMUL v17.4s, v17.4s, v4.4s FMUL v18.4s, v18.4s, v4.4s FMUL v19.4s, v19.4s, v4.4s FMUL v20.4s, v20.4s, v5.4s LDR q4, [x5], 16 FMUL v21.4s, v21.4s, v5.4s FMUL v22.4s, v22.4s, v5.4s FMUL v23.4s, v23.4s, v5.4s FMUL v24.4s, v24.4s, v6.4s FMUL v25.4s, v25.4s, v6.4s FMUL v26.4s, v26.4s, v6.4s FMUL v27.4s, v27.4s, v6.4s FMUL v28.4s, v28.4s, v4.4s FMUL v29.4s, v29.4s, v4.4s FMUL v30.4s, v30.4s, v4.4s FMUL v31.4s, v31.4s, v4.4s $else: FMUL v16.4s, v16.4s, v4.4s FMUL v17.4s, v17.4s, v4.4s FMUL v18.4s, v18.4s, v4.4s FMUL v19.4s, v19.4s, v4.4s FMUL v20.4s, v20.4s, v4.4s FMUL v21.4s, v21.4s, v4.4s FMUL v22.4s, v22.4s, v4.4s FMUL v23.4s, v23.4s, v4.4s FMUL v24.4s, v24.4s, v4.4s FMUL v25.4s, v25.4s, v4.4s FMUL v26.4s, v26.4s, v4.4s FMUL v27.4s, v27.4s, v4.4s FMUL v28.4s, v28.4s, v4.4s FMUL v29.4s, v29.4s, v4.4s FMUL v30.4s, v30.4s, v4.4s FMUL v31.4s, v31.4s, v4.4s FCVTNS v16.4s, v16.4s FCVTNS v17.4s, v17.4s FCVTNS v18.4s, v18.4s FCVTNS v19.4s, v19.4s FCVTNS v20.4s, v20.4s FCVTNS v21.4s, v21.4s FCVTNS v22.4s, v22.4s FCVTNS v23.4s, v23.4s FCVTNS v24.4s, v24.4s FCVTNS v25.4s, v25.4s FCVTNS v26.4s, v26.4s FCVTNS v27.4s, v27.4s FCVTNS v28.4s, v28.4s FCVTNS v29.4s, v29.4s FCVTNS v30.4s, v30.4s FCVTNS v31.4s, v31.4s SQXTN v16.4h, v16.4s SQXTN v17.4h, v17.4s SQXTN v18.4h, v18.4s SQXTN v19.4h, v19.4s SQXTN v24.4h, v24.4s SQXTN v25.4h, v25.4s SQXTN v26.4h, v26.4s SQXTN v27.4h, v27.4s LD1R {v6.8h}, [x11], 2 // add bias SQXTN2 v16.8h, v20.4s SQXTN2 v17.8h, v21.4s SQXTN2 v18.8h, v22.4s SQXTN2 v19.8h, v23.4s SQXTN2 v24.8h, v28.4s SQXTN2 v25.8h, v29.4s SQXTN2 v26.8h, v30.4s SQXTN2 v27.8h, v31.4s SQADD v16.8h, v16.8h, v6.8h SQADD v17.8h, v17.8h, v6.8h SQADD v18.8h, v18.8h, v6.8h SQADD v19.8h, v19.8h, v6.8h SQADD v24.8h, v24.8h, v6.8h SQADD v25.8h, v25.8h, v6.8h SQADD v26.8h, v26.8h, v6.8h SQADD v27.8h, v27.8h, v6.8h LD1R {v4.16b}, [x11], 1 // clamp min value SQXTN v0.8b, v16.8h SQXTN v1.8b, v17.8h SQXTN v2.8b, v18.8h SQXTN v3.8b, v19.8h LD1R {v5.16b}, [x11] // clamp max value SQXTN2 v0.16b, v24.8h SQXTN2 v1.16b, v25.8h SQXTN2 v2.16b, v26.8h SQXTN2 v3.16b, v27.8h SUB x11, x11, ${REWIND_DECREMENT} // rewind params pointer SMAX v0.16b, v0.16b, v4.16b SMAX v1.16b, v1.16b, v4.16b SMAX v2.16b, v2.16b, v4.16b SMAX v3.16b, v3.16b, v4.16b SUBS x1, x1, 16 SMIN v0.16b, v0.16b, v5.16b SMIN v1.16b, v1.16b, v5.16b SMIN v2.16b, v2.16b, v5.16b SMIN v3.16b, v3.16b, v5.16b B.LO 2f # Store full 4 x 16 ST1 {v0.16b}, [x6], x12 SUB x3, x3, x2 // a0 -= kc ST1 {v1.16b}, [x8], x12 SUB x15, x15, x2 // a1 -= kc ST1 {v2.16b}, [x9], x12 SUB x13, x13, x2 // a2 -= kc ST1 {v3.16b}, [x7], x12 SUB x4, x4, x2 // a3 -= kc B.NE 0b RET # Store odd width .p2align 3 2: TBZ x1, 3, 3f STR d0, [x6], 8 STR d1, [x8], 8 DUP d0, v0.d[1] DUP d1, v1.d[1] STR d2, [x9], 8 STR d3, [x7], 8 DUP d2, v2.d[1] DUP d3, v3.d[1] 3: TBZ x1, 2, 4f STR s0, [x6], 4 STR s1, [x8], 4 DUP s0, v0.s[1] DUP s1, v1.s[1] STR s2, [x9], 4 STR s3, [x7], 4 DUP s2, v2.s[1] DUP s3, v3.s[1] 4: TBZ x1, 1, 5f STR h0, [x6], 2 STR h1, [x8], 2 DUP h0, v0.h[1] DUP h1, v1.h[1] STR h2, [x9], 2 STR h3, [x7], 2 DUP h2, v2.h[1] DUP h3, v3.h[1] 5: TBZ x1, 0, 6f STR b0, [x6] STR b1, [x8] STR b2, [x9] STR b3, [x7] 6: RET END_FUNCTION xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_4x16c4__aarch64_neondot_ld32 #ifdef __ELF__ .section ".note.GNU-stack","",%progbits #endif