| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/qs8-gemm/neon-mlal-lane.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2020 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/common.h" |
| 15 |
#include "xnnpack/gemm.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_qd8_f32_qc8w_gemm_minmax_ukernel_4x8__neon_mlal_lane( |
| 19 |
size_t mr, |
| 20 |
size_t nc, |
| 21 |
size_t kc, |
| 22 |
const int8_t* restrict a, |
| 23 |
size_t a_stride, |
| 24 |
const void* restrict w, |
| 25 |
float* restrict c, |
| 26 |
size_t cm_stride, |
| 27 |
size_t cn_stride, |
| 28 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], |
| 29 |
const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS |
| 30 |
{ |
| 31 |
assert(mr != 0); |
| 32 |
assert(mr <= 4); |
| 33 |
assert(nc != 0); |
| 34 |
assert(kc != 0); |
| 35 |
assert(kc % sizeof(int8_t) == 0); |
| 36 |
assert(a != NULL); |
| 37 |
assert(w != NULL); |
| 38 |
assert(c != NULL); |
| 39 |
|
| 40 |
const int8_t* a0 = a; |
| 41 |
float* c0 = c; |
| 42 |
const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); |
| 43 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 44 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 45 |
a1 = a0; |
| 46 |
c1 = c0; |
| 47 |
} |
| 48 |
const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride); |
| 49 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 50 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 51 |
a2 = a1; |
| 52 |
c2 = c1; |
| 53 |
} |
| 54 |
const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride); |
| 55 |
float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 56 |
if XNN_UNPREDICTABLE(mr != 4) { |
| 57 |
a3 = a2; |
| 58 |
c3 = c2; |
| 59 |
} |
| 60 |
|
| 61 |
do { |
| 62 |
int32x4_t vksum0123 = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 63 |
int32x4_t vksum4567 = vld1q_s32(w); w = (const int32_t*) w + 4; |
| 64 |
const int32x4_t vzp01 = vld1q_s32(&quantization_params[0].zero_point); |
| 65 |
int32x4_t vacc0x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vzp01), 0); |
| 66 |
int32x4_t vacc1x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vzp01), 0); |
| 67 |
int32x4_t vacc0x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vzp01), 0); |
| 68 |
int32x4_t vacc1x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vzp01), 0); |
| 69 |
const int32x4_t vzp23 = vld1q_s32(&quantization_params[2].zero_point); |
| 70 |
int32x4_t vacc2x0123 = vmulq_lane_s32(vksum0123, vget_low_s32(vzp23), 0); |
| 71 |
int32x4_t vacc3x0123 = vmulq_lane_s32(vksum0123, vget_high_s32(vzp23), 0); |
| 72 |
int32x4_t vacc2x4567 = vmulq_lane_s32(vksum4567, vget_low_s32(vzp23), 0); |
| 73 |
int32x4_t vacc3x4567 = vmulq_lane_s32(vksum4567, vget_high_s32(vzp23), 0); |
| 74 |
|
| 75 |
size_t k = kc; |
| 76 |
while (k >= 8 * sizeof(int8_t)) { |
| 77 |
const int8x8_t va0 = vld1_s8(a0); a0 += 8; |
| 78 |
const int16x8_t vxa0 = vmovl_s8(va0); |
| 79 |
const int8x8_t va1 = vld1_s8(a1); a1 += 8; |
| 80 |
const int16x8_t vxa1 = vmovl_s8(va1); |
| 81 |
const int8x8_t va2 = vld1_s8(a2); a2 += 8; |
| 82 |
const int16x8_t vxa2 = vmovl_s8(va2); |
| 83 |
const int8x8_t va3 = vld1_s8(a3); a3 += 8; |
| 84 |
const int16x8_t vxa3 = vmovl_s8(va3); |
| 85 |
|
| 86 |
const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 87 |
const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 88 |
const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 89 |
const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 90 |
const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 91 |
const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 92 |
const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 93 |
const int8x8_t vb01234567c7 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 94 |
|
| 95 |
const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0); |
| 96 |
const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1); |
| 97 |
const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2); |
| 98 |
const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3); |
| 99 |
const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4); |
| 100 |
const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5); |
| 101 |
const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6); |
| 102 |
const int16x8_t vxb01234567c7 = vmovl_s8(vb01234567c7); |
| 103 |
|
| 104 |
|
| 105 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 106 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 107 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); |
| 108 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); |
| 109 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); |
| 110 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); |
| 111 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); |
| 112 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); |
| 113 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 114 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 115 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); |
| 116 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); |
| 117 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1); |
| 118 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); |
| 119 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); |
| 120 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); |
| 121 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 122 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 123 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2); |
| 124 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); |
| 125 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2); |
| 126 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); |
| 127 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); |
| 128 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); |
| 129 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 130 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 131 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3); |
| 132 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); |
| 133 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3); |
| 134 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); |
| 135 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); |
| 136 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); |
| 137 |
|
| 138 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 139 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 140 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); |
| 141 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); |
| 142 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); |
| 143 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); |
| 144 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); |
| 145 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); |
| 146 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 147 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 148 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); |
| 149 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); |
| 150 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); |
| 151 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); |
| 152 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); |
| 153 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); |
| 154 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 155 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 156 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); |
| 157 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); |
| 158 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); |
| 159 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); |
| 160 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); |
| 161 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); |
| 162 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa0), 3); |
| 163 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa0), 3); |
| 164 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa1), 3); |
| 165 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa1), 3); |
| 166 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa2), 3); |
| 167 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa2), 3); |
| 168 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c7), vget_high_s16(vxa3), 3); |
| 169 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c7), vget_high_s16(vxa3), 3); |
| 170 |
|
| 171 |
k -= 8 * sizeof(int8_t); |
| 172 |
} |
| 173 |
if XNN_UNLIKELY(k != 0) { |
| 174 |
const int8x8_t va0 = vld1_s8(a0); a0 = (const int8_t*) ((uintptr_t) a0 + k); |
| 175 |
const int16x8_t vxa0 = vmovl_s8(va0); |
| 176 |
const int8x8_t va1 = vld1_s8(a1); a1 = (const int8_t*) ((uintptr_t) a1 + k); |
| 177 |
const int16x8_t vxa1 = vmovl_s8(va1); |
| 178 |
const int8x8_t va2 = vld1_s8(a2); a2 = (const int8_t*) ((uintptr_t) a2 + k); |
| 179 |
const int16x8_t vxa2 = vmovl_s8(va2); |
| 180 |
const int8x8_t va3 = vld1_s8(a3); a3 = (const int8_t*) ((uintptr_t) a3 + k); |
| 181 |
const int16x8_t vxa3 = vmovl_s8(va3); |
| 182 |
|
| 183 |
const int8x8_t vb01234567c0 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 184 |
|
| 185 |
const int16x8_t vxb01234567c0 = vmovl_s8(vb01234567c0); |
| 186 |
|
| 187 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 188 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa0), 0); |
| 189 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa1), 0); |
| 190 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa1), 0); |
| 191 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa2), 0); |
| 192 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa2), 0); |
| 193 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c0), vget_low_s16(vxa3), 0); |
| 194 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c0), vget_low_s16(vxa3), 0); |
| 195 |
|
| 196 |
if (k >= 2 * sizeof(int8_t)) { |
| 197 |
const int8x8_t vb01234567c1 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 198 |
const int16x8_t vxb01234567c1 = vmovl_s8(vb01234567c1); |
| 199 |
|
| 200 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 201 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa0), 1); |
| 202 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa1), 1); |
| 203 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa1), 1); |
| 204 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa2), 1); |
| 205 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa2), 1); |
| 206 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c1), vget_low_s16(vxa3), 1); |
| 207 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c1), vget_low_s16(vxa3), 1); |
| 208 |
|
| 209 |
if (k > 2 * sizeof(int8_t)) { |
| 210 |
const int8x8_t vb01234567c2 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 211 |
const int16x8_t vxb01234567c2 = vmovl_s8(vb01234567c2); |
| 212 |
|
| 213 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 214 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa0), 2); |
| 215 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa1), 2); |
| 216 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa1), 2); |
| 217 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa2), 2); |
| 218 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa2), 2); |
| 219 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c2), vget_low_s16(vxa3), 2); |
| 220 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c2), vget_low_s16(vxa3), 2); |
| 221 |
|
| 222 |
if (k >= 4 * sizeof(int8_t)) { |
| 223 |
const int8x8_t vb01234567c3 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 224 |
const int16x8_t vxb01234567c3 = vmovl_s8(vb01234567c3); |
| 225 |
|
| 226 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 227 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa0), 3); |
| 228 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa1), 3); |
| 229 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa1), 3); |
| 230 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa2), 3); |
| 231 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa2), 3); |
| 232 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c3), vget_low_s16(vxa3), 3); |
| 233 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c3), vget_low_s16(vxa3), 3); |
| 234 |
|
| 235 |
if (k > 4 * sizeof(int8_t)) { |
| 236 |
const int8x8_t vb01234567c4 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 237 |
const int16x8_t vxb01234567c4 = vmovl_s8(vb01234567c4); |
| 238 |
|
| 239 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 240 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa0), 0); |
| 241 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa1), 0); |
| 242 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa1), 0); |
| 243 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa2), 0); |
| 244 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa2), 0); |
| 245 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c4), vget_high_s16(vxa3), 0); |
| 246 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c4), vget_high_s16(vxa3), 0); |
| 247 |
|
| 248 |
if (k >= 6 * sizeof(int8_t)) { |
| 249 |
const int8x8_t vb01234567c5 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 250 |
const int16x8_t vxb01234567c5 = vmovl_s8(vb01234567c5); |
| 251 |
|
| 252 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 253 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa0), 1); |
| 254 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa1), 1); |
| 255 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa1), 1); |
| 256 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa2), 1); |
| 257 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa2), 1); |
| 258 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c5), vget_high_s16(vxa3), 1); |
| 259 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c5), vget_high_s16(vxa3), 1); |
| 260 |
|
| 261 |
if (k > 6 * sizeof(int8_t)) { |
| 262 |
const int8x8_t vb01234567c6 = vld1_s8(w); w = (const int8_t*) w + 8; |
| 263 |
const int16x8_t vxb01234567c6 = vmovl_s8(vb01234567c6); |
| 264 |
|
| 265 |
vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 266 |
vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa0), 2); |
| 267 |
vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa1), 2); |
| 268 |
vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa1), 2); |
| 269 |
vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa2), 2); |
| 270 |
vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa2), 2); |
| 271 |
vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567c6), vget_high_s16(vxa3), 2); |
| 272 |
vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567c6), vget_high_s16(vxa3), 2); |
| 273 |
} |
| 274 |
} |
| 275 |
} |
| 276 |
} |
| 277 |
} |
| 278 |
} |
| 279 |
} |
| 280 |
|
| 281 |
float32x4_t vout0x0123 = vcvtq_f32_s32(vacc0x0123); |
| 282 |
float32x4_t vout0x4567 = vcvtq_f32_s32(vacc0x4567); |
| 283 |
float32x4_t vout1x0123 = vcvtq_f32_s32(vacc1x0123); |
| 284 |
float32x4_t vout1x4567 = vcvtq_f32_s32(vacc1x4567); |
| 285 |
float32x4_t vout2x0123 = vcvtq_f32_s32(vacc2x0123); |
| 286 |
float32x4_t vout2x4567 = vcvtq_f32_s32(vacc2x4567); |
| 287 |
float32x4_t vout3x0123 = vcvtq_f32_s32(vacc3x0123); |
| 288 |
float32x4_t vout3x4567 = vcvtq_f32_s32(vacc3x4567); |
| 289 |
|
| 290 |
const float32x4_t vinput_scale01 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[0].zero_point)); |
| 291 |
vout0x0123 = vmulq_lane_f32(vout0x0123, vget_low_f32(vinput_scale01), 1); |
| 292 |
vout1x0123 = vmulq_lane_f32(vout1x0123, vget_high_f32(vinput_scale01), 1); |
| 293 |
vout0x4567 = vmulq_lane_f32(vout0x4567, vget_low_f32(vinput_scale01), 1); |
| 294 |
vout1x4567 = vmulq_lane_f32(vout1x4567, vget_high_f32(vinput_scale01), 1); |
| 295 |
const float32x4_t vinput_scale23 = vreinterpretq_f32_s32(vld1q_s32(&quantization_params[2].zero_point)); |
| 296 |
vout2x0123 = vmulq_lane_f32(vout2x0123, vget_low_f32(vinput_scale23), 1); |
| 297 |
vout3x0123 = vmulq_lane_f32(vout3x0123, vget_high_f32(vinput_scale23), 1); |
| 298 |
vout2x4567 = vmulq_lane_f32(vout2x4567, vget_low_f32(vinput_scale23), 1); |
| 299 |
vout3x4567 = vmulq_lane_f32(vout3x4567, vget_high_f32(vinput_scale23), 1); |
| 300 |
|
| 301 |
const float32x4_t vfilter_output_scale0123 = vld1q_f32(w); w = (const float*) w + 4; |
| 302 |
const float32x4_t vfilter_output_scale4567 = vld1q_f32(w); w = (const float*) w + 4; |
| 303 |
|
| 304 |
const float32x4_t vbias0123 = vld1q_f32(w); w = (const float*) w + 4; |
| 305 |
#if XNN_ARCH_ARM64 |
| 306 |
vout0x0123 = vfmaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123); |
| 307 |
vout1x0123 = vfmaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123); |
| 308 |
vout2x0123 = vfmaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123); |
| 309 |
vout3x0123 = vfmaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123); |
| 310 |
#else |
| 311 |
vout0x0123 = vmlaq_f32(vbias0123, vout0x0123, vfilter_output_scale0123); |
| 312 |
vout1x0123 = vmlaq_f32(vbias0123, vout1x0123, vfilter_output_scale0123); |
| 313 |
vout2x0123 = vmlaq_f32(vbias0123, vout2x0123, vfilter_output_scale0123); |
| 314 |
vout3x0123 = vmlaq_f32(vbias0123, vout3x0123, vfilter_output_scale0123); |
| 315 |
#endif |
| 316 |
const float32x4_t vbias4567 = vld1q_f32(w); w = (const float*) w + 4; |
| 317 |
#if XNN_ARCH_ARM64 |
| 318 |
vout0x4567 = vfmaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567); |
| 319 |
vout1x4567 = vfmaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567); |
| 320 |
vout2x4567 = vfmaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567); |
| 321 |
vout3x4567 = vfmaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567); |
| 322 |
#else |
| 323 |
vout0x4567 = vmlaq_f32(vbias4567, vout0x4567, vfilter_output_scale4567); |
| 324 |
vout1x4567 = vmlaq_f32(vbias4567, vout1x4567, vfilter_output_scale4567); |
| 325 |
vout2x4567 = vmlaq_f32(vbias4567, vout2x4567, vfilter_output_scale4567); |
| 326 |
vout3x4567 = vmlaq_f32(vbias4567, vout3x4567, vfilter_output_scale4567); |
| 327 |
#endif |
| 328 |
|
| 329 |
const float32x4_t voutput_min = vld1q_dup_f32(¶ms->scalar.min); |
| 330 |
vout0x0123 = vmaxq_f32(vout0x0123, voutput_min); |
| 331 |
vout0x4567 = vmaxq_f32(vout0x4567, voutput_min); |
| 332 |
vout1x0123 = vmaxq_f32(vout1x0123, voutput_min); |
| 333 |
vout1x4567 = vmaxq_f32(vout1x4567, voutput_min); |
| 334 |
vout2x0123 = vmaxq_f32(vout2x0123, voutput_min); |
| 335 |
vout2x4567 = vmaxq_f32(vout2x4567, voutput_min); |
| 336 |
vout3x0123 = vmaxq_f32(vout3x0123, voutput_min); |
| 337 |
vout3x4567 = vmaxq_f32(vout3x4567, voutput_min); |
| 338 |
|
| 339 |
const float32x4_t voutput_max = vld1q_dup_f32(¶ms->scalar.max); |
| 340 |
vout0x0123 = vminq_f32(vout0x0123, voutput_max); |
| 341 |
vout0x4567 = vminq_f32(vout0x4567, voutput_max); |
| 342 |
vout1x0123 = vminq_f32(vout1x0123, voutput_max); |
| 343 |
vout1x4567 = vminq_f32(vout1x4567, voutput_max); |
| 344 |
vout2x0123 = vminq_f32(vout2x0123, voutput_max); |
| 345 |
vout2x4567 = vminq_f32(vout2x4567, voutput_max); |
| 346 |
vout3x0123 = vminq_f32(vout3x0123, voutput_max); |
| 347 |
vout3x4567 = vminq_f32(vout3x4567, voutput_max); |
| 348 |
|
| 349 |
if XNN_LIKELY(nc >= 8) { |
| 350 |
vst1q_f32(&c0[0], vout0x0123); |
| 351 |
vst1q_f32(&c0[4], vout0x4567); |
| 352 |
vst1q_f32(&c1[0], vout1x0123); |
| 353 |
vst1q_f32(&c1[4], vout1x4567); |
| 354 |
vst1q_f32(&c2[0], vout2x0123); |
| 355 |
vst1q_f32(&c2[4], vout2x4567); |
| 356 |
vst1q_f32(&c3[0], vout3x0123); |
| 357 |
vst1q_f32(&c3[4], vout3x4567); |
| 358 |
|
| 359 |
a0 = (const int8_t*) ((uintptr_t) a0 - kc); |
| 360 |
a1 = (const int8_t*) ((uintptr_t) a1 - kc); |
| 361 |
a2 = (const int8_t*) ((uintptr_t) a2 - kc); |
| 362 |
a3 = (const int8_t*) ((uintptr_t) a3 - kc); |
| 363 |
|
| 364 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 365 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 366 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 367 |
c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 368 |
|
| 369 |
nc -= 8; |
| 370 |
} else { |
| 371 |
if (nc & 4) { |
| 372 |
vst1q_f32(c0, vout0x0123); c0 += 4; |
| 373 |
vout0x0123 = vout0x4567; |
| 374 |
vst1q_f32(c1, vout1x0123); c1 += 4; |
| 375 |
vout1x0123 = vout1x4567; |
| 376 |
vst1q_f32(c2, vout2x0123); c2 += 4; |
| 377 |
vout2x0123 = vout2x4567; |
| 378 |
vst1q_f32(c3, vout3x0123); c3 += 4; |
| 379 |
vout3x0123 = vout3x4567; |
| 380 |
} |
| 381 |
float32x2_t vout0x01 = vget_low_f32(vout0x0123); |
| 382 |
float32x2_t vout1x01 = vget_low_f32(vout1x0123); |
| 383 |
float32x2_t vout2x01 = vget_low_f32(vout2x0123); |
| 384 |
float32x2_t vout3x01 = vget_low_f32(vout3x0123); |
| 385 |
if (nc & 2) { |
| 386 |
vst1_f32(c0, vout0x01); c0 += 2; |
| 387 |
vst1_f32(c1, vout1x01); c1 += 2; |
| 388 |
vst1_f32(c2, vout2x01); c2 += 2; |
| 389 |
vst1_f32(c3, vout3x01); c3 += 2; |
| 390 |
vout0x01 = vget_high_f32(vout0x0123); |
| 391 |
vout1x01 = vget_high_f32(vout1x0123); |
| 392 |
vout2x01 = vget_high_f32(vout2x0123); |
| 393 |
vout3x01 = vget_high_f32(vout3x0123); |
| 394 |
} |
| 395 |
if (nc & 1) { |
| 396 |
vst1_lane_f32(c0, vout0x01, 0); |
| 397 |
vst1_lane_f32(c1, vout1x01, 0); |
| 398 |
vst1_lane_f32(c2, vout2x01, 0); |
| 399 |
vst1_lane_f32(c3, vout3x01, 0); |
| 400 |
} |
| 401 |
nc = 0; |
| 402 |
} |
| 403 |
} while (nc != 0); |
| 404 |
} |