| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/c8-neoni8mm.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2023 Google LLC |
| 6 |
// |
| 7 |
// This source code is licensed under the BSD-style license found in the |
| 8 |
// LICENSE file in the root directory of this source tree. |
| 9 |
|
| 10 |
#include <assert.h> |
| 11 |
|
| 12 |
#include <arm_neon.h> |
| 13 |
|
| 14 |
#include "xnnpack/igemm.h" |
| 15 |
#include "xnnpack/intrinsics-polyfill.h" |
| 16 |
#include "xnnpack/math.h" |
| 17 |
|
| 18 |
|
| 19 |
void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_2x16c8__neoni8mm( |
| 20 |
size_t mr, |
| 21 |
size_t nc, |
| 22 |
size_t kc, |
| 23 |
size_t ks, |
| 24 |
const int8_t** restrict a, |
| 25 |
const void* restrict w, |
| 26 |
int8_t* restrict c, |
| 27 |
size_t cm_stride, |
| 28 |
size_t cn_stride, |
| 29 |
size_t a_offset, |
| 30 |
const int8_t* zero, |
| 31 |
const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 32 |
{ |
| 33 |
assert(mr != 0); |
| 34 |
assert(mr <= 2); |
| 35 |
assert(nc != 0); |
| 36 |
assert(kc != 0); |
| 37 |
assert(ks != 0); |
| 38 |
assert(ks % (2 * sizeof(void*)) == 0); |
| 39 |
assert(a_offset % sizeof(int8_t) == 0); |
| 40 |
assert(a != NULL); |
| 41 |
assert(w != NULL); |
| 42 |
assert(c != NULL); |
| 43 |
|
| 44 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 45 |
int8_t* c0 = c; |
| 46 |
int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride); |
| 47 |
if XNN_UNPREDICTABLE(mr != 2) { |
| 48 |
c1 = c0; |
| 49 |
} |
| 50 |
|
| 51 |
do { |
| 52 |
const uint64x2x2_t vbias01x0123 = vld2q_dup_u64(w); w = (const int32_t*) w + 4; |
| 53 |
const uint64x2x2_t vbias01x4567 = vld2q_dup_u64(w); w = (const int32_t*) w + 4; |
| 54 |
const uint64x2x2_t vbias01x89AB = vld2q_dup_u64(w); w = (const int32_t*) w + 4; |
| 55 |
const uint64x2x2_t vbias01xCDEF = vld2q_dup_u64(w); w = (const int32_t*) w + 4; |
| 56 |
int32x4_t vacc01x01 = vreinterpretq_s32_u64(vbias01x0123.val[0]); |
| 57 |
int32x4_t vacc01x23 = vreinterpretq_s32_u64(vbias01x0123.val[1]); |
| 58 |
int32x4_t vacc01x45 = vreinterpretq_s32_u64(vbias01x4567.val[0]); |
| 59 |
int32x4_t vacc01x67 = vreinterpretq_s32_u64(vbias01x4567.val[1]); |
| 60 |
int32x4_t vacc01x89 = vreinterpretq_s32_u64(vbias01x89AB.val[0]); |
| 61 |
int32x4_t vacc01xAB = vreinterpretq_s32_u64(vbias01x89AB.val[1]); |
| 62 |
int32x4_t vacc01xCD = vreinterpretq_s32_u64(vbias01xCDEF.val[0]); |
| 63 |
int32x4_t vacc01xEF = vreinterpretq_s32_u64(vbias01xCDEF.val[1]); |
| 64 |
|
| 65 |
size_t p = ks; |
| 66 |
do { |
| 67 |
const int8_t* restrict a0 = a[0]; |
| 68 |
if XNN_UNPREDICTABLE(a0 != zero) { |
| 69 |
a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 70 |
} |
| 71 |
const int8_t* restrict a1 = a[1]; |
| 72 |
if XNN_UNPREDICTABLE(a1 != zero) { |
| 73 |
a1 = (const int8_t*) ((uintptr_t) a1 + a_offset); |
| 74 |
} |
| 75 |
a += 2; |
| 76 |
|
| 77 |
uint64x2x2_t va01x0123456789ABCDEF; |
| 78 |
va01x0123456789ABCDEF.val[0] = vdupq_n_u64(0); |
| 79 |
va01x0123456789ABCDEF.val[1] = vdupq_n_u64(0); |
| 80 |
|
| 81 |
// Inner accumulation loop along the 16 columns. |
| 82 |
size_t k = kc; |
| 83 |
// 2x partial unrolled loop to load 8 bytes at a time. |
| 84 |
while (k >= 16 * sizeof(int8_t)) { |
| 85 |
// Load a 2x16 block of activations. |
| 86 |
va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a0, va01x0123456789ABCDEF, 0); a0 += 16; |
| 87 |
va01x0123456789ABCDEF = vld2q_lane_u64((const void*) a1, va01x0123456789ABCDEF, 1); a1 += 16; |
| 88 |
|
| 89 |
// Load a 16x16 block of weights. |
| 90 |
const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 91 |
const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 92 |
const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 93 |
const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 94 |
const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 95 |
const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 96 |
const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 97 |
const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 98 |
const int8x16_t vb01x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 99 |
const int8x16_t vb23x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 100 |
const int8x16_t vb45x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 101 |
const int8x16_t vb67x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 102 |
const int8x16_t vb89x89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 103 |
const int8x16_t vbABx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 104 |
const int8x16_t vbCDx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 105 |
const int8x16_t vbEFx89ABCDEF = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 106 |
|
| 107 |
// Multiply-accumulate: 2x16 * 16x16 --> 2x16. |
| 108 |
vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb01x01234567); |
| 109 |
vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb23x01234567); |
| 110 |
vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb45x01234567); |
| 111 |
vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb67x01234567); |
| 112 |
vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vb89x01234567); |
| 113 |
vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbABx01234567); |
| 114 |
vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbCDx01234567); |
| 115 |
vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[0]), vbEFx01234567); |
| 116 |
vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb01x89ABCDEF); |
| 117 |
vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb23x89ABCDEF); |
| 118 |
vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb45x89ABCDEF); |
| 119 |
vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb67x89ABCDEF); |
| 120 |
vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vb89x89ABCDEF); |
| 121 |
vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbABx89ABCDEF); |
| 122 |
vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbCDx89ABCDEF); |
| 123 |
vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x0123456789ABCDEF.val[1]), vbEFx89ABCDEF); |
| 124 |
|
| 125 |
k -= 16 * sizeof(int8_t); |
| 126 |
} |
| 127 |
// Handle up to 8 final positions of `k` |
| 128 |
if XNN_UNLIKELY(k != 0) { |
| 129 |
// Load a 2x8 block of activations. |
| 130 |
uint64x2_t va01x01234567 = vld1q_dup_u64((const void*) a0); a0 += 8; |
| 131 |
va01x01234567 = vld1q_lane_u64((const void*) a1, va01x01234567, 1); a1 += 8; |
| 132 |
|
| 133 |
// Load a 16x16 block of weights. |
| 134 |
const int8x16_t vb01x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 135 |
const int8x16_t vb23x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 136 |
const int8x16_t vb45x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 137 |
const int8x16_t vb67x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 138 |
const int8x16_t vb89x01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 139 |
const int8x16_t vbABx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 140 |
const int8x16_t vbCDx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 141 |
const int8x16_t vbEFx01234567 = vld1q_s8(w); w = (const int8_t*) w + 16; |
| 142 |
|
| 143 |
// Multiply-accumulate: 2x4 * 4x16 --> 2x16. |
| 144 |
vacc01x01 = vmmlaq_s32(vacc01x01, vreinterpretq_s8_u64(va01x01234567), vb01x01234567); |
| 145 |
vacc01x23 = vmmlaq_s32(vacc01x23, vreinterpretq_s8_u64(va01x01234567), vb23x01234567); |
| 146 |
vacc01x45 = vmmlaq_s32(vacc01x45, vreinterpretq_s8_u64(va01x01234567), vb45x01234567); |
| 147 |
vacc01x67 = vmmlaq_s32(vacc01x67, vreinterpretq_s8_u64(va01x01234567), vb67x01234567); |
| 148 |
vacc01x89 = vmmlaq_s32(vacc01x89, vreinterpretq_s8_u64(va01x01234567), vb89x01234567); |
| 149 |
vacc01xAB = vmmlaq_s32(vacc01xAB, vreinterpretq_s8_u64(va01x01234567), vbABx01234567); |
| 150 |
vacc01xCD = vmmlaq_s32(vacc01xCD, vreinterpretq_s8_u64(va01x01234567), vbCDx01234567); |
| 151 |
vacc01xEF = vmmlaq_s32(vacc01xEF, vreinterpretq_s8_u64(va01x01234567), vbEFx01234567); |
| 152 |
} |
| 153 |
|
| 154 |
p -= 2 * sizeof(void*); |
| 155 |
} while (p != 0); |
| 156 |
|
| 157 |
int32x4_t vacc0x0123 = vcombine_s32(vget_low_s32(vacc01x01), vget_low_s32(vacc01x23)); |
| 158 |
int32x4_t vacc1x0123 = vcombine_s32(vget_high_s32(vacc01x01), vget_high_s32(vacc01x23)); |
| 159 |
int32x4_t vacc0x4567 = vcombine_s32(vget_low_s32(vacc01x45), vget_low_s32(vacc01x67)); |
| 160 |
int32x4_t vacc1x4567 = vcombine_s32(vget_high_s32(vacc01x45), vget_high_s32(vacc01x67)); |
| 161 |
int32x4_t vacc0x89AB = vcombine_s32(vget_low_s32(vacc01x89), vget_low_s32(vacc01xAB)); |
| 162 |
int32x4_t vacc1x89AB = vcombine_s32(vget_high_s32(vacc01x89), vget_high_s32(vacc01xAB)); |
| 163 |
int32x4_t vacc0xCDEF = vcombine_s32(vget_low_s32(vacc01xCD), vget_low_s32(vacc01xEF)); |
| 164 |
int32x4_t vacc1xCDEF = vcombine_s32(vget_high_s32(vacc01xCD), vget_high_s32(vacc01xEF)); |
| 165 |
|
| 166 |
float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123); |
| 167 |
float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567); |
| 168 |
float32x4_t vfpacc0x89AB = vcvtq_f32_s32(vacc0x89AB); |
| 169 |
float32x4_t vfpacc0xCDEF = vcvtq_f32_s32(vacc0xCDEF); |
| 170 |
float32x4_t vfpacc1x0123 = vcvtq_f32_s32(vacc1x0123); |
| 171 |
float32x4_t vfpacc1x4567 = vcvtq_f32_s32(vacc1x4567); |
| 172 |
float32x4_t vfpacc1x89AB = vcvtq_f32_s32(vacc1x89AB); |
| 173 |
float32x4_t vfpacc1xCDEF = vcvtq_f32_s32(vacc1xCDEF); |
| 174 |
|
| 175 |
const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 176 |
vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123); |
| 177 |
vfpacc1x0123 = vmulq_f32(vfpacc1x0123, vscale0123); |
| 178 |
const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 179 |
vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567); |
| 180 |
vfpacc1x4567 = vmulq_f32(vfpacc1x4567, vscale4567); |
| 181 |
const float32x4_t vscale89AB = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 182 |
vfpacc0x89AB = vmulq_f32(vfpacc0x89AB, vscale89AB); |
| 183 |
vfpacc1x89AB = vmulq_f32(vfpacc1x89AB, vscale89AB); |
| 184 |
const float32x4_t vscaleCDEF = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 185 |
vfpacc0xCDEF = vmulq_f32(vfpacc0xCDEF, vscaleCDEF); |
| 186 |
vfpacc1xCDEF = vmulq_f32(vfpacc1xCDEF, vscaleCDEF); |
| 187 |
|
| 188 |
vacc0x0123 = vcvtnq_s32_f32(vfpacc0x0123); |
| 189 |
vacc0x4567 = vcvtnq_s32_f32(vfpacc0x4567); |
| 190 |
vacc0x89AB = vcvtnq_s32_f32(vfpacc0x89AB); |
| 191 |
vacc0xCDEF = vcvtnq_s32_f32(vfpacc0xCDEF); |
| 192 |
vacc1x0123 = vcvtnq_s32_f32(vfpacc1x0123); |
| 193 |
vacc1x4567 = vcvtnq_s32_f32(vfpacc1x4567); |
| 194 |
vacc1x89AB = vcvtnq_s32_f32(vfpacc1x89AB); |
| 195 |
vacc1xCDEF = vcvtnq_s32_f32(vfpacc1xCDEF); |
| 196 |
|
| 197 |
const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->fp32_neonv8.output_zero_point); |
| 198 |
#if XNN_ARCH_ARM64 |
| 199 |
const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point); |
| 200 |
const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x89AB), vacc0xCDEF), voutput_zero_point); |
| 201 |
const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point); |
| 202 |
const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x89AB), vacc1xCDEF), voutput_zero_point); |
| 203 |
|
| 204 |
int8x16_t vout0x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc0x89ABCDEF); |
| 205 |
int8x16_t vout1x0123456789ABCDEF = vqmovn_high_s16(vqmovn_s16(vacc1x01234567), vacc1x89ABCDEF); |
| 206 |
#else |
| 207 |
const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point); |
| 208 |
const int16x8_t vacc0x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x89AB), vqmovn_s32(vacc0xCDEF)), voutput_zero_point); |
| 209 |
const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point); |
| 210 |
const int16x8_t vacc1x89ABCDEF = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x89AB), vqmovn_s32(vacc1xCDEF)), voutput_zero_point); |
| 211 |
|
| 212 |
int8x16_t vout0x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc0x89ABCDEF)); |
| 213 |
int8x16_t vout1x0123456789ABCDEF = vcombine_s8(vqmovn_s16(vacc1x01234567), vqmovn_s16(vacc1x89ABCDEF)); |
| 214 |
#endif |
| 215 |
const int8x16_t voutput_min = vld1q_dup_s8(¶ms->fp32_neonv8.output_min); |
| 216 |
const int8x16_t voutput_max = vld1q_dup_s8(¶ms->fp32_neonv8.output_max); |
| 217 |
|
| 218 |
vout1x0123456789ABCDEF = vmaxq_s8(vout1x0123456789ABCDEF, voutput_min); |
| 219 |
vout0x0123456789ABCDEF = vmaxq_s8(vout0x0123456789ABCDEF, voutput_min); |
| 220 |
|
| 221 |
vout1x0123456789ABCDEF = vminq_s8(vout1x0123456789ABCDEF, voutput_max); |
| 222 |
vout0x0123456789ABCDEF = vminq_s8(vout0x0123456789ABCDEF, voutput_max); |
| 223 |
|
| 224 |
if (nc >= 16) { |
| 225 |
vst1q_s8(c1, vout1x0123456789ABCDEF); |
| 226 |
vst1q_s8(c0, vout0x0123456789ABCDEF); |
| 227 |
|
| 228 |
c1 = (int8_t*) ((uintptr_t) c1 + cn_stride); |
| 229 |
c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 230 |
|
| 231 |
a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 232 |
|
| 233 |
nc -= 16; |
| 234 |
} else { |
| 235 |
int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vget_low_s8(vout0x0123456789ABCDEF), vget_low_s8(vout1x0123456789ABCDEF)); |
| 236 |
if (nc & 8) { |
| 237 |
vst1_s8(c1, vget_high_s8(vout0x01234567_1x01234567)); c1 += 8; |
| 238 |
vst1_s8(c0, vget_low_s8(vout0x01234567_1x01234567)); c0 += 8; |
| 239 |
vout0x01234567_1x01234567 = vcombine_s8(vget_high_s8(vout0x0123456789ABCDEF), vget_high_s8(vout1x0123456789ABCDEF)); |
| 240 |
} |
| 241 |
if (nc & 4) { |
| 242 |
vst1q_lane_u32((void*) c1, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4; |
| 243 |
vst1q_lane_u32((void*) c0, vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4; |
| 244 |
vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4); |
| 245 |
} |
| 246 |
if (nc & 2) { |
| 247 |
vst1q_lane_u16((void*) c1, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2; |
| 248 |
vst1q_lane_u16((void*) c0, vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2; |
| 249 |
vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2); |
| 250 |
} |
| 251 |
if (nc & 1) { |
| 252 |
vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8); |
| 253 |
vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0); |
| 254 |
} |
| 255 |
|
| 256 |
nc = 0; |
| 257 |
} |
| 258 |
} while (nc != 0); |
| 259 |
} |