blob: e62bb1d0084424ba886f770eccdf7c1ad2bd5c76 [file] [log] [blame]
// Auto-generated file. Do not edit!
// Template: src/qs8-gemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in
// Generator: tools/xngen
//
// 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.
#include <xnnpack/assembly.h>
# void xnn_qs8_gemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53(
# 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 xnn_qs8_conv_minmax_params 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
# C0 x6 v16 v20 v24 v28
# C1 x8 v17 v21 v25 v29
# C2 x9 v18 v22 v26 v30
# C3 x7 v19 v23 v27 v31
# temp v7
# unused v8 v9 v10 v11 v12 v13 v14 v15
# x10 x17 a53 temp registers
BEGIN_FUNCTION xnn_qs8_gemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
# Clamp A and C pointers
CMP x0, 2 // if mr < 2
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
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
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
LDR x11, [sp, 8] // reload params
MOV v22.16b, v20.16b
MOV v23.16b, v20.16b
SUBS x0, x2, 8 // k = kc - 8
MOV v25.16b, v24.16b
MOV v26.16b, v24.16b
MOV v27.16b, v24.16b
MOV v29.16b, v28.16b
MOV v30.16b, v28.16b
MOV v31.16b, v28.16b
# Is there at least 8 bytes for epilogue?
B.LO 4f
# Prologue
LDR d0, [x3], 8
LDP d4, d6, [x5]
LDR d1, [x15], 8
LDR d2, [x13], 8
LDR d3, [x4], 8
SXTL v0.8h, v0.8b
LDR x17, [x5, 16]
SXTL v4.8h, v4.8b
SXTL v1.8h, v1.8b
SXTL v2.8h, v2.8b
SXTL v3.8h, v3.8b
SXTL v6.8h, v6.8b
SUBS x0, x0, 8 // k = k - 8
# Is there at least 8 bytes for main loop?
B.LO 2f
# Main loop - 8 bytes of A
.p2align 3
1:
SMLAL v16.4s, v4.4h, v0.h[0]
SMLAL2 v20.4s, v4.8h, v0.h[0]
PRFM PLDL1KEEP, [x3, 128]
SMLAL v17.4s, v4.4h, v1.h[0]
SMLAL2 v21.4s, v4.8h, v1.h[0]
PRFM PLDL1KEEP, [x15, 128]
SMLAL v18.4s, v4.4h, v2.h[0]
SMLAL2 v22.4s, v4.8h, v2.h[0]
PRFM PLDL1KEEP, [x13, 128]
SMLAL v19.4s, v4.4h, v3.h[0]
SMLAL2 v23.4s, v4.8h, v3.h[0]
PRFM PLDL1KEEP, [x4, 128]
LDR d4, [x5, 24]
INS v5.d[0], x17
SMLAL v24.4s, v6.4h, v0.h[0]
SMLAL2 v28.4s, v6.8h, v0.h[0]
PRFM PLDL1KEEP, [x5, 448]
SMLAL v25.4s, v6.4h, v1.h[0]
SMLAL2 v29.4s, v6.8h, v1.h[0]
PRFM PLDL1KEEP, [x5, 512]
SXTL v5.8h, v5.8b
SMLAL v26.4s, v6.4h, v2.h[0]
SMLAL2 v30.4s, v6.8h, v2.h[0]
SMLAL v27.4s, v6.4h, v3.h[0]
SMLAL2 v31.4s, v6.8h, v3.h[0]
LDR x17, [x5, 32]
SMLAL v16.4s, v5.4h, v0.h[1]
SMLAL2 v20.4s, v5.8h, v0.h[1]
SMLAL v17.4s, v5.4h, v1.h[1]
SMLAL2 v21.4s, v5.8h, v1.h[1]
SXTL v4.8h, v4.8b
SMLAL v18.4s, v5.4h, v2.h[1]
SMLAL2 v22.4s, v5.8h, v2.h[1]
SMLAL v19.4s, v5.4h, v3.h[1]
SMLAL2 v23.4s, v5.8h, v3.h[1]
LDR d5, [x5, 40]
INS v6.d[0], x17
SMLAL v24.4s, v4.4h, v0.h[1]
SMLAL2 v28.4s, v4.8h, v0.h[1]
SMLAL v25.4s, v4.4h, v1.h[1]
SMLAL2 v29.4s, v4.8h, v1.h[1]
SXTL v6.8h, v6.8b
SMLAL v26.4s, v4.4h, v2.h[1]
SMLAL2 v30.4s, v4.8h, v2.h[1]
SMLAL v27.4s, v4.4h, v3.h[1]
SMLAL2 v31.4s, v4.8h, v3.h[1]
LDR x17, [x5, 48]
SMLAL v16.4s, v6.4h, v0.h[2]
SMLAL2 v20.4s, v6.8h, v0.h[2]
SMLAL v17.4s, v6.4h, v1.h[2]
SXTL v5.8h, v5.8b
SMLAL2 v21.4s, v6.8h, v1.h[2]
SMLAL v18.4s, v6.4h, v2.h[2]
SMLAL2 v22.4s, v6.8h, v2.h[2]
SMLAL v19.4s, v6.4h, v3.h[2]
SMLAL2 v23.4s, v6.8h, v3.h[2]
LDR d6, [x5, 56]
INS v4.d[0], x17
SMLAL v24.4s, v5.4h, v0.h[2]
SMLAL2 v28.4s, v5.8h, v0.h[2]
SMLAL v25.4s, v5.4h, v1.h[2]
SMLAL2 v29.4s, v5.8h, v1.h[2]
SXTL v4.8h, v4.8b
SMLAL v26.4s, v5.4h, v2.h[2]
SMLAL2 v30.4s, v5.8h, v2.h[2]
SMLAL v27.4s, v5.4h, v3.h[2]
SMLAL2 v31.4s, v5.8h, v3.h[2]
LDR x17, [x5, 64]
SMLAL v16.4s, v4.4h, v0.h[3]
SMLAL2 v20.4s, v4.8h, v0.h[3]
SMLAL v17.4s, v4.4h, v1.h[3]
SMLAL2 v21.4s, v4.8h, v1.h[3]
SXTL v6.8h, v6.8b
SMLAL v18.4s, v4.4h, v2.h[3]
SMLAL2 v22.4s, v4.8h, v2.h[3]
SMLAL v19.4s, v4.4h, v3.h[3]
SMLAL2 v23.4s, v4.8h, v3.h[3]
LDR d4, [x5, 72]
INS v5.d[0], x17
SMLAL v24.4s, v6.4h, v0.h[3]
SMLAL2 v28.4s, v6.8h, v0.h[3]
SXTL v5.8h, v5.8b
SMLAL v25.4s, v6.4h, v1.h[3]
SMLAL2 v29.4s, v6.8h, v1.h[3]
SMLAL v26.4s, v6.4h, v2.h[3]
SMLAL2 v30.4s, v6.8h, v2.h[3]
SMLAL v27.4s, v6.4h, v3.h[3]
SMLAL2 v31.4s, v6.8h, v3.h[3]
LDR x17, [x5, 80]
SMLAL v16.4s, v5.4h, v0.h[4]
SMLAL2 v20.4s, v5.8h, v0.h[4]
SMLAL v17.4s, v5.4h, v1.h[4]
SMLAL2 v21.4s, v5.8h, v1.h[4]
SXTL v4.8h, v4.8b
SMLAL v18.4s, v5.4h, v2.h[4]
SMLAL2 v22.4s, v5.8h, v2.h[4]
SMLAL v19.4s, v5.4h, v3.h[4]
SMLAL2 v23.4s, v5.8h, v3.h[4]
LDR d5, [x5, 88]
INS v6.d[0], x17
SMLAL v24.4s, v4.4h, v0.h[4]
SMLAL2 v28.4s, v4.8h, v0.h[4]
SMLAL v25.4s, v4.4h, v1.h[4]
SMLAL2 v29.4s, v4.8h, v1.h[4]
SXTL v6.8h, v6.8b
SMLAL v26.4s, v4.4h, v2.h[4]
SMLAL2 v30.4s, v4.8h, v2.h[4]
SMLAL v27.4s, v4.4h, v3.h[4]
SMLAL2 v31.4s, v4.8h, v3.h[4]
LDR x17, [x5, 96]
SMLAL v16.4s, v6.4h, v0.h[5]
SMLAL2 v20.4s, v6.8h, v0.h[5]
SMLAL v17.4s, v6.4h, v1.h[5]
SMLAL2 v21.4s, v6.8h, v1.h[5]
SXTL v5.8h, v5.8b
SMLAL v18.4s, v6.4h, v2.h[5]
SMLAL2 v22.4s, v6.8h, v2.h[5]
SMLAL v19.4s, v6.4h, v3.h[5]
SMLAL2 v23.4s, v6.8h, v3.h[5]
LDR d6, [x5, 104]
INS v4.d[0], x17
SMLAL v24.4s, v5.4h, v0.h[5]
SMLAL2 v28.4s, v5.8h, v0.h[5]
SMLAL v25.4s, v5.4h, v1.h[5]
SMLAL2 v29.4s, v5.8h, v1.h[5]
SXTL v4.8h, v4.8b
SMLAL v26.4s, v5.4h, v2.h[5]
SMLAL2 v30.4s, v5.8h, v2.h[5]
SMLAL v27.4s, v5.4h, v3.h[5]
SMLAL2 v31.4s, v5.8h, v3.h[5]
SXTL v6.8h, v6.8b
LDR x17, [x5, 112]
SMLAL v16.4s, v4.4h, v0.h[6]
SMLAL2 v20.4s, v4.8h, v0.h[6]
SMLAL v17.4s, v4.4h, v1.h[6]
SMLAL2 v21.4s, v4.8h, v1.h[6]
SMLAL v18.4s, v4.4h, v2.h[6]
SMLAL2 v22.4s, v4.8h, v2.h[6]
SMLAL v19.4s, v4.4h, v3.h[6]
SMLAL2 v23.4s, v4.8h, v3.h[6]
LDR d5, [x5, 120]
INS v4.d[0], x17
SMLAL v24.4s, v6.4h, v0.h[6]
SMLAL2 v28.4s, v6.8h, v0.h[6]
SMLAL v25.4s, v6.4h, v1.h[6]
SMLAL2 v29.4s, v6.8h, v1.h[6]
SXTL v4.8h, v4.8b
ADD x5, x5, 128
SMLAL v26.4s, v6.4h, v2.h[6]
SMLAL2 v30.4s, v6.8h, v2.h[6]
LDR x17, [x5]
SMLAL v27.4s, v6.4h, v3.h[6]
SMLAL2 v31.4s, v6.8h, v3.h[6]
SXTL v5.8h, v5.8b
LDR x10, [x3], 8
SMLAL v16.4s, v4.4h, v0.h[7]
SMLAL2 v20.4s, v4.8h, v0.h[7]
SMLAL v17.4s, v4.4h, v1.h[7]
SMLAL2 v21.4s, v4.8h, v1.h[7]
SMLAL v18.4s, v4.4h, v2.h[7]
SMLAL2 v22.4s, v4.8h, v2.h[7]
SMLAL v19.4s, v4.4h, v3.h[7]
SMLAL2 v23.4s, v4.8h, v3.h[7]
LDR d6, [x5, 8]
INS v4.d[0], x17
SMLAL v24.4s, v5.4h, v0.h[7]
SMLAL2 v28.4s, v5.8h, v0.h[7]
LDR x17, [x13], 8
SMLAL v25.4s, v5.4h, v1.h[7]
SMLAL2 v29.4s, v5.8h, v1.h[7]
LDR d1, [x15], 8
INS v0.d[0], x10
SMLAL v26.4s, v5.4h, v2.h[7]
SMLAL2 v30.4s, v5.8h, v2.h[7]
SMLAL v27.4s, v5.4h, v3.h[7]
SMLAL2 v31.4s, v5.8h, v3.h[7]
LDR d3, [x4], 8
INS v2.d[0], x17
SXTL v0.8h, v0.8b
SXTL v1.8h, v1.8b
LDR x17, [x5, 16]
SXTL v4.8h, v4.8b
SXTL v2.8h, v2.8b
SUBS x0, x0, 8
SXTL v3.8h, v3.8b
SXTL v6.8h, v6.8b
B.HS 1b
# Epilogue. Same as main loop but no preloads in final group
.p2align 3
2:
SMLAL v16.4s, v4.4h, v0.h[0]
SMLAL2 v20.4s, v4.8h, v0.h[0]
SMLAL v17.4s, v4.4h, v1.h[0]
SMLAL2 v21.4s, v4.8h, v1.h[0]
SMLAL v18.4s, v4.4h, v2.h[0]
SMLAL2 v22.4s, v4.8h, v2.h[0]
SMLAL v19.4s, v4.4h, v3.h[0]
SMLAL2 v23.4s, v4.8h, v3.h[0]
LDR d4, [x5, 24]
INS v5.d[0], x17
SMLAL v24.4s, v6.4h, v0.h[0]
SMLAL2 v28.4s, v6.8h, v0.h[0]
SMLAL v25.4s, v6.4h, v1.h[0]
SMLAL2 v29.4s, v6.8h, v1.h[0]
SXTL v5.8h, v5.8b
SMLAL v26.4s, v6.4h, v2.h[0]
SMLAL2 v30.4s, v6.8h, v2.h[0]
SMLAL v27.4s, v6.4h, v3.h[0]
SMLAL2 v31.4s, v6.8h, v3.h[0]
LDR x17, [x5, 32]
SMLAL v16.4s, v5.4h, v0.h[1]
SMLAL2 v20.4s, v5.8h, v0.h[1]
SMLAL v17.4s, v5.4h, v1.h[1]
SMLAL2 v21.4s, v5.8h, v1.h[1]
SXTL v4.8h, v4.8b
SMLAL v18.4s, v5.4h, v2.h[1]
SMLAL2 v22.4s, v5.8h, v2.h[1]
SMLAL v19.4s, v5.4h, v3.h[1]
SMLAL2 v23.4s, v5.8h, v3.h[1]
LDR d5, [x5, 40]
INS v6.d[0], x17
SMLAL v24.4s, v4.4h, v0.h[1]
SMLAL2 v28.4s, v4.8h, v0.h[1]
SMLAL v25.4s, v4.4h, v1.h[1]
SMLAL2 v29.4s, v4.8h, v1.h[1]
SXTL v6.8h, v6.8b
SMLAL v26.4s, v4.4h, v2.h[1]
SMLAL2 v30.4s, v4.8h, v2.h[1]
SMLAL v27.4s, v4.4h, v3.h[1]
SMLAL2 v31.4s, v4.8h, v3.h[1]
LDR x17, [x5, 48]
SMLAL v16.4s, v6.4h, v0.h[2]
SMLAL2 v20.4s, v6.8h, v0.h[2]
SMLAL v17.4s, v6.4h, v1.h[2]
SXTL v5.8h, v5.8b
SMLAL2 v21.4s, v6.8h, v1.h[2]
SMLAL v18.4s, v6.4h, v2.h[2]
SMLAL2 v22.4s, v6.8h, v2.h[2]
SMLAL v19.4s, v6.4h, v3.h[2]
SMLAL2 v23.4s, v6.8h, v3.h[2]
LDR d6, [x5, 56]
INS v4.d[0], x17
SMLAL v24.4s, v5.4h, v0.h[2]
SMLAL2 v28.4s, v5.8h, v0.h[2]
SMLAL v25.4s, v5.4h, v1.h[2]
SMLAL2 v29.4s, v5.8h, v1.h[2]
SXTL v4.8h, v4.8b
SMLAL v26.4s, v5.4h, v2.h[2]
SMLAL2 v30.4s, v5.8h, v2.h[2]
SMLAL v27.4s, v5.4h, v3.h[2]
SMLAL2 v31.4s, v5.8h, v3.h[2]
LDR x17, [x5, 64]
SMLAL v16.4s, v4.4h, v0.h[3]
SMLAL2 v20.4s, v4.8h, v0.h[3]
SMLAL v17.4s, v4.4h, v1.h[3]
SMLAL2 v21.4s, v4.8h, v1.h[3]
SXTL v6.8h, v6.8b
SMLAL v18.4s, v4.4h, v2.h[3]
SMLAL2 v22.4s, v4.8h, v2.h[3]
SMLAL v19.4s, v4.4h, v3.h[3]
SMLAL2 v23.4s, v4.8h, v3.h[3]
LDR d4, [x5, 72]
INS v5.d[0], x17
SMLAL v24.4s, v6.4h, v0.h[3]
SMLAL2 v28.4s, v6.8h, v0.h[3]
SXTL v5.8h, v5.8b
SMLAL v25.4s, v6.4h, v1.h[3]
SMLAL2 v29.4s, v6.8h, v1.h[3]
SMLAL v26.4s, v6.4h, v2.h[3]
SMLAL2 v30.4s, v6.8h, v2.h[3]
SMLAL v27.4s, v6.4h, v3.h[3]
SMLAL2 v31.4s, v6.8h, v3.h[3]
LDR x17, [x5, 80]
SMLAL v16.4s, v5.4h, v0.h[4]
SMLAL2 v20.4s, v5.8h, v0.h[4]
SMLAL v17.4s, v5.4h, v1.h[4]
SMLAL2 v21.4s, v5.8h, v1.h[4]
SXTL v4.8h, v4.8b
SMLAL v18.4s, v5.4h, v2.h[4]
SMLAL2 v22.4s, v5.8h, v2.h[4]
SMLAL v19.4s, v5.4h, v3.h[4]
SMLAL2 v23.4s, v5.8h, v3.h[4]
LDR d5, [x5, 88]
INS v6.d[0], x17
SMLAL v24.4s, v4.4h, v0.h[4]
SMLAL2 v28.4s, v4.8h, v0.h[4]
SMLAL v25.4s, v4.4h, v1.h[4]
SMLAL2 v29.4s, v4.8h, v1.h[4]
SXTL v6.8h, v6.8b
SMLAL v26.4s, v4.4h, v2.h[4]
SMLAL2 v30.4s, v4.8h, v2.h[4]
SMLAL v27.4s, v4.4h, v3.h[4]
SMLAL2 v31.4s, v4.8h, v3.h[4]
LDR x17, [x5, 96]
SMLAL v16.4s, v6.4h, v0.h[5]
SMLAL2 v20.4s, v6.8h, v0.h[5]
SMLAL v17.4s, v6.4h, v1.h[5]
SMLAL2 v21.4s, v6.8h, v1.h[5]
SXTL v5.8h, v5.8b
SMLAL v18.4s, v6.4h, v2.h[5]
SMLAL2 v22.4s, v6.8h, v2.h[5]
SMLAL v19.4s, v6.4h, v3.h[5]
SMLAL2 v23.4s, v6.8h, v3.h[5]
LDR d6, [x5, 104]
INS v4.d[0], x17
SMLAL v24.4s, v5.4h, v0.h[5]
SMLAL2 v28.4s, v5.8h, v0.h[5]
SMLAL v25.4s, v5.4h, v1.h[5]
SMLAL2 v29.4s, v5.8h, v1.h[5]
SXTL v4.8h, v4.8b
SMLAL v26.4s, v5.4h, v2.h[5]
SMLAL2 v30.4s, v5.8h, v2.h[5]
SMLAL v27.4s, v5.4h, v3.h[5]
SMLAL2 v31.4s, v5.8h, v3.h[5]
SXTL v6.8h, v6.8b
SMLAL v16.4s, v4.4h, v0.h[6]
SMLAL2 v20.4s, v4.8h, v0.h[6]
SMLAL v17.4s, v4.4h, v1.h[6]
SMLAL2 v21.4s, v4.8h, v1.h[6]
SMLAL v18.4s, v4.4h, v2.h[6]
SMLAL2 v22.4s, v4.8h, v2.h[6]
SMLAL v19.4s, v4.4h, v3.h[6]
SMLAL2 v23.4s, v4.8h, v3.h[6]
LDR x17, [x5, 112]
SMLAL v24.4s, v6.4h, v0.h[6]
SMLAL2 v28.4s, v6.8h, v0.h[6]
SMLAL v25.4s, v6.4h, v1.h[6]
SMLAL2 v29.4s, v6.8h, v1.h[6]
LDR d5, [x5, 120]
INS v4.d[0], x17
SXTL v4.8h, v4.8b
SMLAL v26.4s, v6.4h, v2.h[6]
SMLAL2 v30.4s, v6.8h, v2.h[6]
SMLAL v27.4s, v6.4h, v3.h[6]
SMLAL2 v31.4s, v6.8h, v3.h[6]
SMLAL v16.4s, v4.4h, v0.h[7]
SMLAL2 v20.4s, v4.8h, v0.h[7]
SMLAL v17.4s, v4.4h, v1.h[7]
SMLAL2 v21.4s, v4.8h, v1.h[7]
SXTL v5.8h, v5.8b
SMLAL v18.4s, v4.4h, v2.h[7]
SMLAL2 v22.4s, v4.8h, v2.h[7]
SMLAL v19.4s, v4.4h, v3.h[7]
SMLAL2 v23.4s, v4.8h, v3.h[7]
ADD x5, x5, 128
SMLAL v24.4s, v5.4h, v0.h[7]
SMLAL2 v28.4s, v5.8h, v0.h[7]
SMLAL v25.4s, v5.4h, v1.h[7]
SMLAL2 v29.4s, v5.8h, v1.h[7]
AND x0, x2, 7 // kc remainder 0 to 7
SMLAL v26.4s, v5.4h, v2.h[7]
SMLAL2 v30.4s, v5.8h, v2.h[7]
SMLAL v27.4s, v5.4h, v3.h[7]
SMLAL2 v31.4s, v5.8h, v3.h[7]
# Is there a remainder?- 1 to 7 bytes of A
CBNZ x0, 4f
3:
# Apply params - scale, shift, bias and clamp
LD2R {v0.4s, v1.4s}, [x11], 8
CMEQ v2.4s, v1.4s, 0
BIC v4.16b, v16.16b, v2.16b
BIC v5.16b, v17.16b, v2.16b
BIC v6.16b, v18.16b, v2.16b
BIC v7.16b, v19.16b, v2.16b
SQRDMULH v16.4s, v16.4s, v0.4s
SQRDMULH v17.4s, v17.4s, v0.4s
SQRDMULH v18.4s, v18.4s, v0.4s
SQRDMULH v19.4s, v19.4s, v0.4s
SSRA v16.4s, v4.4s, 31 // signed shift right accumulate
SSRA v17.4s, v5.4s, 31
SSRA v18.4s, v6.4s, 31
SSRA v19.4s, v7.4s, 31
BIC v4.16b, v20.16b, v2.16b
BIC v5.16b, v21.16b, v2.16b
BIC v6.16b, v22.16b, v2.16b
BIC v7.16b, v23.16b, v2.16b
SQRDMULH v20.4s, v20.4s, v0.4s
SQRDMULH v21.4s, v21.4s, v0.4s
SQRDMULH v22.4s, v22.4s, v0.4s
SQRDMULH v23.4s, v23.4s, v0.4s
SSRA v20.4s, v4.4s, 31
SSRA v21.4s, v5.4s, 31
SSRA v22.4s, v6.4s, 31
SSRA v23.4s, v7.4s, 31
BIC v4.16b, v24.16b, v2.16b
BIC v5.16b, v25.16b, v2.16b
BIC v6.16b, v26.16b, v2.16b
BIC v7.16b, v27.16b, v2.16b
SQRDMULH v24.4s, v24.4s, v0.4s
SQRDMULH v25.4s, v25.4s, v0.4s
SQRDMULH v26.4s, v26.4s, v0.4s
SQRDMULH v27.4s, v27.4s, v0.4s
SSRA v24.4s, v4.4s, 31
SSRA v25.4s, v5.4s, 31
SSRA v26.4s, v6.4s, 31
SSRA v27.4s, v7.4s, 31
BIC v4.16b, v28.16b, v2.16b
BIC v5.16b, v29.16b, v2.16b
BIC v6.16b, v30.16b, v2.16b
BIC v7.16b, v31.16b, v2.16b
SQRDMULH v28.4s, v28.4s, v0.4s
SQRDMULH v29.4s, v29.4s, v0.4s
SQRDMULH v30.4s, v30.4s, v0.4s
SQRDMULH v31.4s, v31.4s, v0.4s
SSRA v28.4s, v4.4s, 31
SSRA v29.4s, v5.4s, 31
SSRA v30.4s, v6.4s, 31
SSRA v31.4s, v7.4s, 31
SRSHL v16.4s, v16.4s, v1.4s // signed rounding shift left
SRSHL v17.4s, v17.4s, v1.4s
SRSHL v18.4s, v18.4s, v1.4s
SRSHL v19.4s, v19.4s, v1.4s
SRSHL v20.4s, v20.4s, v1.4s
SRSHL v21.4s, v21.4s, v1.4s
SRSHL v22.4s, v22.4s, v1.4s
SRSHL v23.4s, v23.4s, v1.4s
SRSHL v24.4s, v24.4s, v1.4s
SRSHL v25.4s, v25.4s, v1.4s
SRSHL v26.4s, v26.4s, v1.4s
SRSHL v27.4s, v27.4s, v1.4s
SRSHL v28.4s, v28.4s, v1.4s
SRSHL v29.4s, v29.4s, v1.4s
SRSHL v30.4s, v30.4s, v1.4s
SRSHL v31.4s, v31.4s, v1.4s
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 {v2.8h}, [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, v2.8h
SQADD v17.8h, v17.8h, v2.8h
SQADD v18.8h, v18.8h, v2.8h
SQADD v19.8h, v19.8h, v2.8h
SQADD v24.8h, v24.8h, v2.8h
SQADD v25.8h, v25.8h, v2.8h
SQADD v26.8h, v26.8h, v2.8h
SQADD v27.8h, v27.8h, v2.8h
LD1R {v0.16b}, [x11], 1 // clamp min value
SQXTN v4.8b, v16.8h
SQXTN v5.8b, v17.8h
SQXTN v6.8b, v18.8h
SQXTN v7.8b, v19.8h
LD1R {v1.16b}, [x11] // clamp max value
SQXTN2 v4.16b, v24.8h
SQXTN2 v5.16b, v25.8h
SQXTN2 v6.16b, v26.8h
SQXTN2 v7.16b, v27.8h
LDR x12, [sp, 0] // cn_stride
SMAX v4.16b, v4.16b, v0.16b
SMAX v5.16b, v5.16b, v0.16b
SMAX v6.16b, v6.16b, v0.16b
SMAX v7.16b, v7.16b, v0.16b
SUBS x1, x1, 16
SMIN v4.16b, v4.16b, v1.16b
SMIN v5.16b, v5.16b, v1.16b
SMIN v6.16b, v6.16b, v1.16b
SMIN v7.16b, v7.16b, v1.16b
B.LO 5f
# Store full 4 x 16
ST1 {v4.16b}, [x6], x12
SUB x3, x3, x2 // a0 -= kc
ST1 {v5.16b}, [x8], x12
SUB x15, x15, x2 // a1 -= kc
ST1 {v6.16b}, [x9], x12
SUB x13, x13, x2 // a2 -= kc
ST1 {v7.16b}, [x7], x12
SUB x4, x4, x2 // a3 -= kc
B.NE 0b
RET
# Remainder- 1 to 7 bytes of A
.p2align 3
4:
AND x0, x2, 7 // kc remainder 1 to 7
LD1 {v0.8b}, [x3], x0
LDP d4, d5, [x5], 16
LD1 {v1.8b}, [x15], x0
LD1 {v2.8b}, [x13], x0
LD1 {v3.8b}, [x4], x0
SXTL v0.8h, v0.8b
SXTL v4.8h, v4.8b
SXTL v5.8h, v5.8b
SXTL v1.8h, v1.8b
SXTL v2.8h, v2.8b
SXTL v3.8h, v3.8b
SMLAL v16.4s, v4.4h, v0.h[0]
SMLAL2 v20.4s, v4.8h, v0.h[0]
SMLAL v24.4s, v5.4h, v0.h[0]
SMLAL2 v28.4s, v5.8h, v0.h[0]
SMLAL v17.4s, v4.4h, v1.h[0]
SMLAL2 v21.4s, v4.8h, v1.h[0]
SMLAL v25.4s, v5.4h, v1.h[0]
SMLAL2 v29.4s, v5.8h, v1.h[0]
SMLAL v18.4s, v4.4h, v2.h[0]
SMLAL2 v22.4s, v4.8h, v2.h[0]
SMLAL v26.4s, v5.4h, v2.h[0]
SMLAL2 v30.4s, v5.8h, v2.h[0]
SMLAL v19.4s, v4.4h, v3.h[0]
SMLAL2 v23.4s, v4.8h, v3.h[0]
SMLAL v27.4s, v5.4h, v3.h[0]
SMLAL2 v31.4s, v5.8h, v3.h[0]
CMP x0, 2
B.LO 3b
LDP d4, d5, [x5], 16
SXTL v4.8h, v4.8b
SXTL v5.8h, v5.8b
SMLAL v16.4s, v4.4h, v0.h[1]
SMLAL2 v20.4s, v4.8h, v0.h[1]
SMLAL v24.4s, v5.4h, v0.h[1]
SMLAL2 v28.4s, v5.8h, v0.h[1]
SMLAL v17.4s, v4.4h, v1.h[1]
SMLAL2 v21.4s, v4.8h, v1.h[1]
SMLAL v25.4s, v5.4h, v1.h[1]
SMLAL2 v29.4s, v5.8h, v1.h[1]
SMLAL v18.4s, v4.4h, v2.h[1]
SMLAL2 v22.4s, v4.8h, v2.h[1]
SMLAL v26.4s, v5.4h, v2.h[1]
SMLAL2 v30.4s, v5.8h, v2.h[1]
SMLAL v19.4s, v4.4h, v3.h[1]
SMLAL2 v23.4s, v4.8h, v3.h[1]
SMLAL v27.4s, v5.4h, v3.h[1]
SMLAL2 v31.4s, v5.8h, v3.h[1]
B.EQ 3b
LDP d4, d5, [x5], 16
SXTL v4.8h, v4.8b
SXTL v5.8h, v5.8b
SMLAL v16.4s, v4.4h, v0.h[2]
SMLAL2 v20.4s, v4.8h, v0.h[2]
SMLAL v24.4s, v5.4h, v0.h[2]
SMLAL2 v28.4s, v5.8h, v0.h[2]
SMLAL v17.4s, v4.4h, v1.h[2]
SMLAL2 v21.4s, v4.8h, v1.h[2]
SMLAL v25.4s, v5.4h, v1.h[2]
SMLAL2 v29.4s, v5.8h, v1.h[2]
SMLAL v18.4s, v4.4h, v2.h[2]
SMLAL2 v22.4s, v4.8h, v2.h[2]
SMLAL v26.4s, v5.4h, v2.h[2]
SMLAL2 v30.4s, v5.8h, v2.h[2]
SMLAL v19.4s, v4.4h, v3.h[2]
SMLAL2 v23.4s, v4.8h, v3.h[2]
SMLAL v27.4s, v5.4h, v3.h[2]
SMLAL2 v31.4s, v5.8h, v3.h[2]
CMP x0, 4
B.LO 3b
LDP d4, d5, [x5], 16
SXTL v4.8h, v4.8b
SXTL v5.8h, v5.8b
SMLAL v16.4s, v4.4h, v0.h[3]
SMLAL2 v20.4s, v4.8h, v0.h[3]
SMLAL v24.4s, v5.4h, v0.h[3]
SMLAL2 v28.4s, v5.8h, v0.h[3]
SMLAL v17.4s, v4.4h, v1.h[3]
SMLAL2 v21.4s, v4.8h, v1.h[3]
SMLAL v25.4s, v5.4h, v1.h[3]
SMLAL2 v29.4s, v5.8h, v1.h[3]
SMLAL v18.4s, v4.4h, v2.h[3]
SMLAL2 v22.4s, v4.8h, v2.h[3]
SMLAL v26.4s, v5.4h, v2.h[3]
SMLAL2 v30.4s, v5.8h, v2.h[3]
SMLAL v19.4s, v4.4h, v3.h[3]
SMLAL2 v23.4s, v4.8h, v3.h[3]
SMLAL v27.4s, v5.4h, v3.h[3]
SMLAL2 v31.4s, v5.8h, v3.h[3]
B.EQ 3b
LDP d4, d5, [x5], 16
SXTL v4.8h, v4.8b
SXTL v5.8h, v5.8b
SMLAL v16.4s, v4.4h, v0.h[4]
SMLAL2 v20.4s, v4.8h, v0.h[4]
SMLAL v24.4s, v5.4h, v0.h[4]
SMLAL2 v28.4s, v5.8h, v0.h[4]
SMLAL v17.4s, v4.4h, v1.h[4]
SMLAL2 v21.4s, v4.8h, v1.h[4]
SMLAL v25.4s, v5.4h, v1.h[4]
SMLAL2 v29.4s, v5.8h, v1.h[4]
SMLAL v18.4s, v4.4h, v2.h[4]
SMLAL2 v22.4s, v4.8h, v2.h[4]
SMLAL v26.4s, v5.4h, v2.h[4]
SMLAL2 v30.4s, v5.8h, v2.h[4]
SMLAL v19.4s, v4.4h, v3.h[4]
SMLAL2 v23.4s, v4.8h, v3.h[4]
SMLAL v27.4s, v5.4h, v3.h[4]
SMLAL2 v31.4s, v5.8h, v3.h[4]
CMP x0, 6
B.LO 3b
LDP d4, d5, [x5], 16
SXTL v4.8h, v4.8b
SXTL v5.8h, v5.8b
SMLAL v16.4s, v4.4h, v0.h[5]
SMLAL2 v20.4s, v4.8h, v0.h[5]
SMLAL v24.4s, v5.4h, v0.h[5]
SMLAL2 v28.4s, v5.8h, v0.h[5]
SMLAL v17.4s, v4.4h, v1.h[5]
SMLAL2 v21.4s, v4.8h, v1.h[5]
SMLAL v25.4s, v5.4h, v1.h[5]
SMLAL2 v29.4s, v5.8h, v1.h[5]
SMLAL v18.4s, v4.4h, v2.h[5]
SMLAL2 v22.4s, v4.8h, v2.h[5]
SMLAL v26.4s, v5.4h, v2.h[5]
SMLAL2 v30.4s, v5.8h, v2.h[5]
SMLAL v19.4s, v4.4h, v3.h[5]
SMLAL2 v23.4s, v4.8h, v3.h[5]
SMLAL v27.4s, v5.4h, v3.h[5]
SMLAL2 v31.4s, v5.8h, v3.h[5]
B.EQ 3b
LDP d4, d5, [x5], 16
SXTL v4.8h, v4.8b
SXTL v5.8h, v5.8b
SMLAL v16.4s, v4.4h, v0.h[6]
SMLAL2 v20.4s, v4.8h, v0.h[6]
SMLAL v24.4s, v5.4h, v0.h[6]
SMLAL2 v28.4s, v5.8h, v0.h[6]
SMLAL v17.4s, v4.4h, v1.h[6]
SMLAL2 v21.4s, v4.8h, v1.h[6]
SMLAL v25.4s, v5.4h, v1.h[6]
SMLAL2 v29.4s, v5.8h, v1.h[6]
SMLAL v18.4s, v4.4h, v2.h[6]
SMLAL2 v22.4s, v4.8h, v2.h[6]
SMLAL v26.4s, v5.4h, v2.h[6]
SMLAL2 v30.4s, v5.8h, v2.h[6]
SMLAL v19.4s, v4.4h, v3.h[6]
SMLAL2 v23.4s, v4.8h, v3.h[6]
SMLAL v27.4s, v5.4h, v3.h[6]
SMLAL2 v31.4s, v5.8h, v3.h[6]
B 3b
# Store odd width
.p2align 3
5:
TBZ x1, 3, 6f
STR d4, [x6], 8
DUP d4, v4.d[1]
STR d5, [x8], 8
DUP d5, v5.d[1]
STR d6, [x9], 8
DUP d6, v6.d[1]
STR d7, [x7], 8
DUP d7, v7.d[1]
6:
TBZ x1, 2, 7f
STR s4, [x6], 4
DUP s4, v4.s[1]
STR s5, [x8], 4
DUP s5, v5.s[1]
STR s6, [x9], 4
DUP s6, v6.s[1]
STR s7, [x7], 4
DUP s7, v7.s[1]
7:
TBZ x1, 1, 8f
ST1 {v4.h}[0], [x6], 2
DUP h4, v4.h[1]
ST1 {v5.h}[0], [x8], 2
DUP h5, v5.h[1]
ST1 {v6.h}[0], [x9], 2
DUP h6, v6.h[1]
ST1 {v7.h}[0], [x7], 2
DUP h7, v7.h[1]
8:
TBZ x1, 0, 9f
ST1 {v4.b}[0], [x6]
ST1 {v5.b}[0], [x8]
ST1 {v6.b}[0], [x9]
ST1 {v7.b}[0], [x7]
9:
RET
END_FUNCTION xnn_qs8_gemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
#endif