| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-igemm/c8-neon-mull.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2021 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/math.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qs8_qc8w_igemm_minmax_fp32_ukernel_1x8c8__neon_mlal( |
| 19 |
size_t mr, |
| 20 |
size_t nc, |
| 21 |
size_t kc, |
| 22 |
size_t ks, |
| 23 |
const int8_t** restrict a, |
| 24 |
const void* restrict w, |
| 25 |
int8_t* restrict c, |
| 26 |
size_t cm_stride, |
| 27 |
size_t cn_stride, |
| 28 |
size_t a_offset, |
| 29 |
const int8_t* zero, |
| 30 |
const union xnn_qs8_qc8w_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 31 |
{ |
| 32 |
assert(mr != 0); |
| 33 |
assert(mr <= 1); |
| 34 |
assert(nc != 0); |
| 35 |
assert(kc != 0); |
| 36 |
assert(ks != 0); |
| 37 |
assert(ks % (1 * sizeof(void*)) == 0); |
| 38 |
assert(a_offset % sizeof(int8_t) == 0); |
| 39 |
assert(a != NULL); |
| 40 |
assert(w != NULL); |
| 41 |
assert(c != NULL); |
| 42 |
|
| 43 |
kc = round_up_po2(kc, 8 * sizeof(int8_t)); |
| 44 |
int8_t* c0 = c; |
| 45 |
|
| 46 |
do { |
| 47 |
int32x4_t vacc0x0 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 48 |
int32x4_t vacc0x1 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 49 |
int32x4_t vacc0x2 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 50 |
int32x4_t vacc0x3 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 51 |
int32x4_t vacc0x4 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 52 |
int32x4_t vacc0x5 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 53 |
int32x4_t vacc0x6 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 54 |
int32x4_t vacc0x7 = vld1q_lane_s32(w, vmovq_n_s32(0), 0); w = (const int32_t*) w + 1; |
| 55 |
|
| 56 |
size_t p = ks; |
| 57 |
do { |
| 58 |
const int8_t* restrict a0 = a[0]; |
| 59 |
if XNN_UNPREDICTABLE(a0 != zero) { |
| 60 |
a0 = (const int8_t*) ((uintptr_t) a0 + a_offset); |
| 61 |
} |
| 62 |
a += 1; |
| 63 |
|
| 64 |
size_t k = kc; |
| 65 |
// 2x partial unrolled loop to load 16 bytes at a time using MLA. |
| 66 |
while (k >= 16 * sizeof(int8_t)) { |
| 67 |
const int8x8_t va0x0 = vld1_s8(a0); a0 += 8; |
| 68 |
const int8x8_t va0x1 = vld1_s8(a0); a0 += 8; |
| 69 |
|
| 70 |
const int8x8_t vb0x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 71 |
const int8x8_t vb1x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 72 |
const int8x8_t vb2x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 73 |
const int8x8_t vb3x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 74 |
const int8x8_t vb4x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 75 |
const int8x8_t vb5x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 76 |
const int8x8_t vb6x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 77 |
const int8x8_t vb7x0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 78 |
|
| 79 |
const int8x8_t vb0x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 80 |
int16x8_t vprod0x0 = vmull_s8(vb0x0, va0x0); |
| 81 |
vprod0x0 = vmlal_s8(vprod0x0, vb0x1, va0x1); |
| 82 |
vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); |
| 83 |
const int8x8_t vb1x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 84 |
int16x8_t vprod0x1 = vmull_s8(vb1x0, va0x0); |
| 85 |
vprod0x1 = vmlal_s8(vprod0x1, vb1x1, va0x1); |
| 86 |
vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); |
| 87 |
const int8x8_t vb2x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 88 |
int16x8_t vprod0x2 = vmull_s8(vb2x0, va0x0); |
| 89 |
vprod0x2 = vmlal_s8(vprod0x2, vb2x1, va0x1); |
| 90 |
vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2); |
| 91 |
const int8x8_t vb3x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 92 |
int16x8_t vprod0x3 = vmull_s8(vb3x0, va0x0); |
| 93 |
vprod0x3 = vmlal_s8(vprod0x3, vb3x1, va0x1); |
| 94 |
vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3); |
| 95 |
const int8x8_t vb4x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 96 |
int16x8_t vprod0x4 = vmull_s8(vb4x0, va0x0); |
| 97 |
vprod0x4 = vmlal_s8(vprod0x4, vb4x1, va0x1); |
| 98 |
vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); |
| 99 |
const int8x8_t vb5x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 100 |
int16x8_t vprod0x5 = vmull_s8(vb5x0, va0x0); |
| 101 |
vprod0x5 = vmlal_s8(vprod0x5, vb5x1, va0x1); |
| 102 |
vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); |
| 103 |
const int8x8_t vb6x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 104 |
int16x8_t vprod0x6 = vmull_s8(vb6x0, va0x0); |
| 105 |
vprod0x6 = vmlal_s8(vprod0x6, vb6x1, va0x1); |
| 106 |
vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); |
| 107 |
const int8x8_t vb7x1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 108 |
int16x8_t vprod0x7 = vmull_s8(vb7x0, va0x0); |
| 109 |
vprod0x7 = vmlal_s8(vprod0x7, vb7x1, va0x1); |
| 110 |
vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); |
| 111 |
|
| 112 |
k -= 16 * sizeof(int8_t); |
| 113 |
} |
| 114 |
|
| 115 |
// Handle 8 bytes at a time using MUL. |
| 116 |
if (k != 0) { |
| 117 |
const int8x8_t va0 = vld1_s8(a0); a0 += 8; |
| 118 |
|
| 119 |
const int8x8_t vb0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 120 |
const int16x8_t vprod0x0 = vmull_s8(vb0, va0); |
| 121 |
vacc0x0 = vpadalq_s16(vacc0x0, vprod0x0); |
| 122 |
const int8x8_t vb1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 123 |
const int16x8_t vprod0x1 = vmull_s8(vb1, va0); |
| 124 |
vacc0x1 = vpadalq_s16(vacc0x1, vprod0x1); |
| 125 |
const int8x8_t vb2 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 126 |
const int16x8_t vprod0x2 = vmull_s8(vb2, va0); |
| 127 |
vacc0x2 = vpadalq_s16(vacc0x2, vprod0x2); |
| 128 |
const int8x8_t vb3 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 129 |
const int16x8_t vprod0x3 = vmull_s8(vb3, va0); |
| 130 |
vacc0x3 = vpadalq_s16(vacc0x3, vprod0x3); |
| 131 |
const int8x8_t vb4 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 132 |
const int16x8_t vprod0x4 = vmull_s8(vb4, va0); |
| 133 |
vacc0x4 = vpadalq_s16(vacc0x4, vprod0x4); |
| 134 |
const int8x8_t vb5 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 135 |
const int16x8_t vprod0x5 = vmull_s8(vb5, va0); |
| 136 |
vacc0x5 = vpadalq_s16(vacc0x5, vprod0x5); |
| 137 |
const int8x8_t vb6 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 138 |
const int16x8_t vprod0x6 = vmull_s8(vb6, va0); |
| 139 |
vacc0x6 = vpadalq_s16(vacc0x6, vprod0x6); |
| 140 |
const int8x8_t vb7 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 141 |
const int16x8_t vprod0x7 = vmull_s8(vb7, va0); |
| 142 |
vacc0x7 = vpadalq_s16(vacc0x7, vprod0x7); |
| 143 |
|
| 144 |
k -= 8 * sizeof(int8_t); |
| 145 |
} |
| 146 |
|
| 147 |
p -= 1 * sizeof(void*); |
| 148 |
} while (p != 0); |
| 149 |
|
| 150 |
#if XNN_ARCH_ARM64 |
| 151 |
const int32x4_t vsum0x01 = vpaddq_s32(vacc0x0, vacc0x1); |
| 152 |
const int32x4_t vsum0x23 = vpaddq_s32(vacc0x2, vacc0x3); |
| 153 |
const int32x4_t vsum0x45 = vpaddq_s32(vacc0x4, vacc0x5); |
| 154 |
const int32x4_t vsum0x67 = vpaddq_s32(vacc0x6, vacc0x7); |
| 155 |
|
| 156 |
int32x4_t vacc0x0123 = vpaddq_s32(vsum0x01, vsum0x23); |
| 157 |
int32x4_t vacc0x4567 = vpaddq_s32(vsum0x45, vsum0x67); |
| 158 |
#else |
| 159 |
const int32x2_t vpsum0x0 = vadd_s32(vget_low_s32(vacc0x0), vget_high_s32(vacc0x0)); |
| 160 |
const int32x2_t vpsum0x1 = vadd_s32(vget_low_s32(vacc0x1), vget_high_s32(vacc0x1)); |
| 161 |
const int32x2_t vpsum0x2 = vadd_s32(vget_low_s32(vacc0x2), vget_high_s32(vacc0x2)); |
| 162 |
const int32x2_t vpsum0x3 = vadd_s32(vget_low_s32(vacc0x3), vget_high_s32(vacc0x3)); |
| 163 |
const int32x2_t vsum0x01 = vpadd_s32(vpsum0x0, vpsum0x1); |
| 164 |
const int32x2_t vsum0x23 = vpadd_s32(vpsum0x2, vpsum0x3); |
| 165 |
int32x4_t vacc0x0123 = vcombine_s32(vsum0x01, vsum0x23 ); |
| 166 |
const int32x2_t vpsum0x4 = vadd_s32(vget_low_s32(vacc0x4), vget_high_s32(vacc0x4)); |
| 167 |
const int32x2_t vpsum0x5 = vadd_s32(vget_low_s32(vacc0x5), vget_high_s32(vacc0x5)); |
| 168 |
const int32x2_t vpsum0x6 = vadd_s32(vget_low_s32(vacc0x6), vget_high_s32(vacc0x6)); |
| 169 |
const int32x2_t vpsum0x7 = vadd_s32(vget_low_s32(vacc0x7), vget_high_s32(vacc0x7)); |
| 170 |
const int32x2_t vsum0x45 = vpadd_s32(vpsum0x4, vpsum0x5); |
| 171 |
const int32x2_t vsum0x67 = vpadd_s32(vpsum0x6, vpsum0x7); |
| 172 |
int32x4_t vacc0x4567 = vcombine_s32(vsum0x45, vsum0x67 ); |
| 173 |
#endif |
| 174 |
|
| 175 |
float32x4_t vfpacc0x0123 = vcvtq_f32_s32(vacc0x0123); |
| 176 |
float32x4_t vfpacc0x4567 = vcvtq_f32_s32(vacc0x4567); |
| 177 |
|
| 178 |
const float32x4_t vscale0123 = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 179 |
vfpacc0x0123 = vmulq_f32(vfpacc0x0123, vscale0123); |
| 180 |
const float32x4_t vscale4567 = vld1q_f32((const float*) w); w = (const float*) w + 4; |
| 181 |
vfpacc0x4567 = vmulq_f32(vfpacc0x4567, vscale4567); |
| 182 |
|
| 183 |
const float32x4_t vmagic_bias = vld1q_dup_f32(¶ms->fp32_neon.magic_bias); |
| 184 |
vacc0x0123 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x0123, vmagic_bias)); |
| 185 |
vacc0x4567 = vreinterpretq_s32_f32(vaddq_f32(vfpacc0x4567, vmagic_bias)); |
| 186 |
|
| 187 |
const int32x4_t vmagic_bias_less_output_zero_point = vld1q_dup_s32(¶ms->fp32_neon.magic_bias_less_output_zero_point); |
| 188 |
vacc0x0123 = vqsubq_s32(vacc0x0123, vmagic_bias_less_output_zero_point); |
| 189 |
vacc0x4567 = vqsubq_s32(vacc0x4567, vmagic_bias_less_output_zero_point); |
| 190 |
|
| 191 |
#if XNN_ARCH_ARM64 |
| 192 |
int16x8_t vacc0x01234567 = vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567); |
| 193 |
|
| 194 |
|
| 195 |
int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567); |
| 196 |
#else |
| 197 |
int16x8_t vacc0x01234567 = vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)); |
| 198 |
|
| 199 |
|
| 200 |
int8x8_t vout0x01234567 = vqmovn_s16(vacc0x01234567); |
| 201 |
#endif |
| 202 |
|
| 203 |
const int8x8_t voutput_min = vld1_dup_s8(¶ms->fp32_neon.output_min); |
| 204 |
vout0x01234567 = vmax_s8(vout0x01234567, voutput_min); |
| 205 |
|
| 206 |
const int8x8_t voutput_max = vld1_dup_s8(¶ms->fp32_neon.output_max); |
| 207 |
vout0x01234567 = vmin_s8(vout0x01234567, voutput_max); |
| 208 |
|
| 209 |
if (nc >= 8) { |
| 210 |
vst1_s8(c0 + 0, vout0x01234567); |
| 211 |
|
| 212 |
c0 = (int8_t*) ((uintptr_t) c0 + cn_stride); |
| 213 |
|
| 214 |
a = (const int8_t**restrict) ((uintptr_t) a - ks); |
| 215 |
|
| 216 |
nc -= 8; |
| 217 |
} else { |
| 218 |
if (nc & 4) { |
| 219 |
vst1_lane_u32((void*) c0, vreinterpret_u32_s8(vout0x01234567), 0); c0 += 4; |
| 220 |
vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 4); |
| 221 |
} |
| 222 |
if (nc & 2) { |
| 223 |
vst1_lane_u16((void*) c0, vreinterpret_u16_s8(vout0x01234567), 0); c0 += 2; |
| 224 |
vout0x01234567 = vext_s8(vout0x01234567, vout0x01234567, 2); |
| 225 |
} |
| 226 |
if (nc & 1) { |
| 227 |
vst1_lane_s8(c0, vout0x01234567, 0); |
| 228 |
} |
| 229 |
|
| 230 |
nc = 0; |
| 231 |
} |
| 232 |
} while (nc != 0); |
| 233 |
} |