| 1 |
// Auto-generated file. Do not edit! |
| 2 |
// Template: src/f32-ppmm/neon.c.in |
| 3 |
// Generator: tools/xngen |
| 4 |
// |
| 5 |
// Copyright 2019 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/ppmm.h" |
| 16 |
|
| 17 |
|
| 18 |
void xnn_f32_ppmm_minmax_ukernel_4x16__aarch64_neonfma( |
| 19 |
size_t mr, |
| 20 |
size_t nc, |
| 21 |
size_t kc, |
| 22 |
const float* restrict a, |
| 23 |
const float* restrict w, |
| 24 |
float* restrict c, |
| 25 |
size_t cm_stride, |
| 26 |
size_t cn_stride, |
| 27 |
const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) |
| 28 |
{ |
| 29 |
assert(mr != 0); |
| 30 |
assert(mr <= 4); |
| 31 |
assert(nc != 0); |
| 32 |
assert(kc != 0); |
| 33 |
assert(kc % sizeof(float) == 0); |
| 34 |
|
| 35 |
#if XNN_ARCH_ARM64 |
| 36 |
const float32x4x2_t vminmax = vld2q_dup_f32(¶ms->scalar.min); |
| 37 |
const float32x4_t vmin = vminmax.val[0]; |
| 38 |
const float32x4_t vmax = vminmax.val[1]; |
| 39 |
#else |
| 40 |
const float32x2x2_t vminmax = vld2_dup_f32(¶ms->scalar.min); |
| 41 |
const float32x4_t vmin = vcombine_f32(vminmax.val[0], vminmax.val[0]); |
| 42 |
const float32x4_t vmax = vcombine_f32(vminmax.val[1], vminmax.val[1]); |
| 43 |
#endif |
| 44 |
|
| 45 |
float* c0 = c; |
| 46 |
float* c1 = (float*) ((uintptr_t) c0 + cm_stride); |
| 47 |
if XNN_UNPREDICTABLE(mr < 2) { |
| 48 |
c1 = c0; |
| 49 |
} |
| 50 |
float* c2 = (float*) ((uintptr_t) c1 + cm_stride); |
| 51 |
if XNN_UNPREDICTABLE(mr <= 2) { |
| 52 |
c2 = c1; |
| 53 |
} |
| 54 |
float* c3 = (float*) ((uintptr_t) c2 + cm_stride); |
| 55 |
if XNN_UNPREDICTABLE(mr != 4) { |
| 56 |
c3 = c2; |
| 57 |
} |
| 58 |
|
| 59 |
|
| 60 |
do { |
| 61 |
float32x4_t vacc0x0123 = vld1q_f32(w); w += 4; |
| 62 |
float32x4_t vacc0x4567 = vld1q_f32(w); w += 4; |
| 63 |
float32x4_t vacc0x89AB = vld1q_f32(w); w += 4; |
| 64 |
float32x4_t vacc0xCDEF = vld1q_f32(w); w += 4; |
| 65 |
float32x4_t vacc1x0123 = vacc0x0123; |
| 66 |
float32x4_t vacc1x4567 = vacc0x4567; |
| 67 |
float32x4_t vacc1x89AB = vacc0x89AB; |
| 68 |
float32x4_t vacc1xCDEF = vacc0xCDEF; |
| 69 |
float32x4_t vacc2x0123 = vacc0x0123; |
| 70 |
float32x4_t vacc2x4567 = vacc0x4567; |
| 71 |
float32x4_t vacc2x89AB = vacc0x89AB; |
| 72 |
float32x4_t vacc2xCDEF = vacc0xCDEF; |
| 73 |
float32x4_t vacc3x0123 = vacc0x0123; |
| 74 |
float32x4_t vacc3x4567 = vacc0x4567; |
| 75 |
float32x4_t vacc3x89AB = vacc0x89AB; |
| 76 |
float32x4_t vacc3xCDEF = vacc0xCDEF; |
| 77 |
|
| 78 |
size_t k = kc; |
| 79 |
do { |
| 80 |
const float32x4_t va0123 = vld1q_f32(a); a += 4; |
| 81 |
|
| 82 |
const float32x4_t vb0123 = vld1q_f32(w); w += 4; |
| 83 |
const float32x4_t vb4567 = vld1q_f32(w); w += 4; |
| 84 |
const float32x4_t vb89AB = vld1q_f32(w); w += 4; |
| 85 |
const float32x4_t vbCDEF = vld1q_f32(w); w += 4; |
| 86 |
|
| 87 |
#if XNN_ARCH_ARM64 |
| 88 |
vacc0x0123 = vfmaq_laneq_f32(vacc0x0123, vb0123, va0123, 0); |
| 89 |
vacc1x0123 = vfmaq_laneq_f32(vacc1x0123, vb0123, va0123, 1); |
| 90 |
vacc2x0123 = vfmaq_laneq_f32(vacc2x0123, vb0123, va0123, 2); |
| 91 |
vacc3x0123 = vfmaq_laneq_f32(vacc3x0123, vb0123, va0123, 3); |
| 92 |
vacc0x4567 = vfmaq_laneq_f32(vacc0x4567, vb4567, va0123, 0); |
| 93 |
vacc1x4567 = vfmaq_laneq_f32(vacc1x4567, vb4567, va0123, 1); |
| 94 |
vacc2x4567 = vfmaq_laneq_f32(vacc2x4567, vb4567, va0123, 2); |
| 95 |
vacc3x4567 = vfmaq_laneq_f32(vacc3x4567, vb4567, va0123, 3); |
| 96 |
vacc0x89AB = vfmaq_laneq_f32(vacc0x89AB, vb89AB, va0123, 0); |
| 97 |
vacc1x89AB = vfmaq_laneq_f32(vacc1x89AB, vb89AB, va0123, 1); |
| 98 |
vacc2x89AB = vfmaq_laneq_f32(vacc2x89AB, vb89AB, va0123, 2); |
| 99 |
vacc3x89AB = vfmaq_laneq_f32(vacc3x89AB, vb89AB, va0123, 3); |
| 100 |
vacc0xCDEF = vfmaq_laneq_f32(vacc0xCDEF, vbCDEF, va0123, 0); |
| 101 |
vacc1xCDEF = vfmaq_laneq_f32(vacc1xCDEF, vbCDEF, va0123, 1); |
| 102 |
vacc2xCDEF = vfmaq_laneq_f32(vacc2xCDEF, vbCDEF, va0123, 2); |
| 103 |
vacc3xCDEF = vfmaq_laneq_f32(vacc3xCDEF, vbCDEF, va0123, 3); |
| 104 |
#else |
| 105 |
const float32x4_t va0000 = vdupq_lane_f32(vget_low_f32(va0123), 0); |
| 106 |
const float32x4_t va1111 = vdupq_lane_f32(vget_low_f32(va0123), 1); |
| 107 |
const float32x4_t va2222 = vdupq_lane_f32(vget_high_f32(va0123), 0); |
| 108 |
const float32x4_t va3333 = vdupq_lane_f32(vget_high_f32(va0123), 1); |
| 109 |
|
| 110 |
vacc0x0123 = vfmaq_f32(vacc0x0123, va0000, vb0123); |
| 111 |
vacc1x0123 = vfmaq_f32(vacc1x0123, va1111, vb0123); |
| 112 |
vacc2x0123 = vfmaq_f32(vacc2x0123, va2222, vb0123); |
| 113 |
vacc3x0123 = vfmaq_f32(vacc3x0123, va3333, vb0123); |
| 114 |
vacc0x4567 = vfmaq_f32(vacc0x4567, va0000, vb4567); |
| 115 |
vacc1x4567 = vfmaq_f32(vacc1x4567, va1111, vb4567); |
| 116 |
vacc2x4567 = vfmaq_f32(vacc2x4567, va2222, vb4567); |
| 117 |
vacc3x4567 = vfmaq_f32(vacc3x4567, va3333, vb4567); |
| 118 |
vacc0x89AB = vfmaq_f32(vacc0x89AB, va0000, vb89AB); |
| 119 |
vacc1x89AB = vfmaq_f32(vacc1x89AB, va1111, vb89AB); |
| 120 |
vacc2x89AB = vfmaq_f32(vacc2x89AB, va2222, vb89AB); |
| 121 |
vacc3x89AB = vfmaq_f32(vacc3x89AB, va3333, vb89AB); |
| 122 |
vacc0xCDEF = vfmaq_f32(vacc0xCDEF, va0000, vbCDEF); |
| 123 |
vacc1xCDEF = vfmaq_f32(vacc1xCDEF, va1111, vbCDEF); |
| 124 |
vacc2xCDEF = vfmaq_f32(vacc2xCDEF, va2222, vbCDEF); |
| 125 |
vacc3xCDEF = vfmaq_f32(vacc3xCDEF, va3333, vbCDEF); |
| 126 |
#endif |
| 127 |
|
| 128 |
k -= sizeof(float); |
| 129 |
} while (k != 0); |
| 130 |
|
| 131 |
vacc0x0123 = vminq_f32(vacc0x0123, vmax); |
| 132 |
vacc1x0123 = vminq_f32(vacc1x0123, vmax); |
| 133 |
vacc2x0123 = vminq_f32(vacc2x0123, vmax); |
| 134 |
vacc3x0123 = vminq_f32(vacc3x0123, vmax); |
| 135 |
vacc0x4567 = vminq_f32(vacc0x4567, vmax); |
| 136 |
vacc1x4567 = vminq_f32(vacc1x4567, vmax); |
| 137 |
vacc2x4567 = vminq_f32(vacc2x4567, vmax); |
| 138 |
vacc3x4567 = vminq_f32(vacc3x4567, vmax); |
| 139 |
vacc0x89AB = vminq_f32(vacc0x89AB, vmax); |
| 140 |
vacc1x89AB = vminq_f32(vacc1x89AB, vmax); |
| 141 |
vacc2x89AB = vminq_f32(vacc2x89AB, vmax); |
| 142 |
vacc3x89AB = vminq_f32(vacc3x89AB, vmax); |
| 143 |
vacc0xCDEF = vminq_f32(vacc0xCDEF, vmax); |
| 144 |
vacc1xCDEF = vminq_f32(vacc1xCDEF, vmax); |
| 145 |
vacc2xCDEF = vminq_f32(vacc2xCDEF, vmax); |
| 146 |
vacc3xCDEF = vminq_f32(vacc3xCDEF, vmax); |
| 147 |
|
| 148 |
vacc0x0123 = vmaxq_f32(vacc0x0123, vmin); |
| 149 |
vacc1x0123 = vmaxq_f32(vacc1x0123, vmin); |
| 150 |
vacc2x0123 = vmaxq_f32(vacc2x0123, vmin); |
| 151 |
vacc3x0123 = vmaxq_f32(vacc3x0123, vmin); |
| 152 |
vacc0x4567 = vmaxq_f32(vacc0x4567, vmin); |
| 153 |
vacc1x4567 = vmaxq_f32(vacc1x4567, vmin); |
| 154 |
vacc2x4567 = vmaxq_f32(vacc2x4567, vmin); |
| 155 |
vacc3x4567 = vmaxq_f32(vacc3x4567, vmin); |
| 156 |
vacc0x89AB = vmaxq_f32(vacc0x89AB, vmin); |
| 157 |
vacc1x89AB = vmaxq_f32(vacc1x89AB, vmin); |
| 158 |
vacc2x89AB = vmaxq_f32(vacc2x89AB, vmin); |
| 159 |
vacc3x89AB = vmaxq_f32(vacc3x89AB, vmin); |
| 160 |
vacc0xCDEF = vmaxq_f32(vacc0xCDEF, vmin); |
| 161 |
vacc1xCDEF = vmaxq_f32(vacc1xCDEF, vmin); |
| 162 |
vacc2xCDEF = vmaxq_f32(vacc2xCDEF, vmin); |
| 163 |
vacc3xCDEF = vmaxq_f32(vacc3xCDEF, vmin); |
| 164 |
|
| 165 |
if XNN_LIKELY(nc >= 16) { |
| 166 |
vst1q_f32(c3, vacc3x0123); |
| 167 |
vst1q_f32(c3 + 4, vacc3x4567); |
| 168 |
vst1q_f32(c3 + 8, vacc3x89AB); |
| 169 |
vst1q_f32(c3 + 12, vacc3xCDEF); |
| 170 |
c3 = (float*) ((uintptr_t) c3 + cn_stride); |
| 171 |
vst1q_f32(c2, vacc2x0123); |
| 172 |
vst1q_f32(c2 + 4, vacc2x4567); |
| 173 |
vst1q_f32(c2 + 8, vacc2x89AB); |
| 174 |
vst1q_f32(c2 + 12, vacc2xCDEF); |
| 175 |
c2 = (float*) ((uintptr_t) c2 + cn_stride); |
| 176 |
vst1q_f32(c1, vacc1x0123); |
| 177 |
vst1q_f32(c1 + 4, vacc1x4567); |
| 178 |
vst1q_f32(c1 + 8, vacc1x89AB); |
| 179 |
vst1q_f32(c1 + 12, vacc1xCDEF); |
| 180 |
c1 = (float*) ((uintptr_t) c1 + cn_stride); |
| 181 |
vst1q_f32(c0, vacc0x0123); |
| 182 |
vst1q_f32(c0 + 4, vacc0x4567); |
| 183 |
vst1q_f32(c0 + 8, vacc0x89AB); |
| 184 |
vst1q_f32(c0 + 12, vacc0xCDEF); |
| 185 |
c0 = (float*) ((uintptr_t) c0 + cn_stride); |
| 186 |
|
| 187 |
a = (const float*) ((uintptr_t) a - kc * 4); |
| 188 |
|
| 189 |
nc -= 16; |
| 190 |
} else { |
| 191 |
if (nc & 8) { |
| 192 |
vst1q_f32(c3, vacc3x0123); c3 += 4; |
| 193 |
vst1q_f32(c3, vacc3x4567); c3 += 4; |
| 194 |
vst1q_f32(c2, vacc2x0123); c2 += 4; |
| 195 |
vst1q_f32(c2, vacc2x4567); c2 += 4; |
| 196 |
vst1q_f32(c1, vacc1x0123); c1 += 4; |
| 197 |
vst1q_f32(c1, vacc1x4567); c1 += 4; |
| 198 |
vst1q_f32(c0, vacc0x0123); c0 += 4; |
| 199 |
vst1q_f32(c0, vacc0x4567); c0 += 4; |
| 200 |
|
| 201 |
vacc3x0123 = vacc3x89AB; |
| 202 |
vacc3x4567 = vacc3xCDEF; |
| 203 |
vacc2x0123 = vacc2x89AB; |
| 204 |
vacc2x4567 = vacc2xCDEF; |
| 205 |
vacc1x0123 = vacc1x89AB; |
| 206 |
vacc1x4567 = vacc1xCDEF; |
| 207 |
vacc0x0123 = vacc0x89AB; |
| 208 |
vacc0x4567 = vacc0xCDEF; |
| 209 |
} |
| 210 |
if (nc & 4) { |
| 211 |
vst1q_f32(c3, vacc3x0123); c3 += 4; |
| 212 |
vst1q_f32(c2, vacc2x0123); c2 += 4; |
| 213 |
vst1q_f32(c1, vacc1x0123); c1 += 4; |
| 214 |
vst1q_f32(c0, vacc0x0123); c0 += 4; |
| 215 |
|
| 216 |
vacc3x0123 = vacc3x4567; |
| 217 |
vacc3x4567 = vacc3x89AB; |
| 218 |
vacc3x89AB = vacc3xCDEF; |
| 219 |
vacc2x0123 = vacc2x4567; |
| 220 |
vacc2x4567 = vacc2x89AB; |
| 221 |
vacc2x89AB = vacc2xCDEF; |
| 222 |
vacc1x0123 = vacc1x4567; |
| 223 |
vacc1x4567 = vacc1x89AB; |
| 224 |
vacc1x89AB = vacc1xCDEF; |
| 225 |
vacc0x0123 = vacc0x4567; |
| 226 |
vacc0x4567 = vacc0x89AB; |
| 227 |
vacc0x89AB = vacc0xCDEF; |
| 228 |
} |
| 229 |
float32x2_t vacc3x01 = vget_low_f32(vacc3x0123); |
| 230 |
float32x2_t vacc2x01 = vget_low_f32(vacc2x0123); |
| 231 |
float32x2_t vacc1x01 = vget_low_f32(vacc1x0123); |
| 232 |
float32x2_t vacc0x01 = vget_low_f32(vacc0x0123); |
| 233 |
if (nc & 2) { |
| 234 |
vst1_f32(c3, vacc3x01); c3 += 2; |
| 235 |
vst1_f32(c2, vacc2x01); c2 += 2; |
| 236 |
vst1_f32(c1, vacc1x01); c1 += 2; |
| 237 |
vst1_f32(c0, vacc0x01); c0 += 2; |
| 238 |
|
| 239 |
vacc3x01 = vget_high_f32(vacc3x0123); |
| 240 |
vacc2x01 = vget_high_f32(vacc2x0123); |
| 241 |
vacc1x01 = vget_high_f32(vacc1x0123); |
| 242 |
vacc0x01 = vget_high_f32(vacc0x0123); |
| 243 |
} |
| 244 |
if (nc & 1) { |
| 245 |
vst1_lane_f32(c3, vacc3x01, 0); |
| 246 |
vst1_lane_f32(c2, vacc2x01, 0); |
| 247 |
vst1_lane_f32(c1, vacc1x01, 0); |
| 248 |
vst1_lane_f32(c0, vacc0x01, 0); |
| 249 |
} |
| 250 |
|
| 251 |
nc = 0; |
| 252 |
} |
| 253 |
} while (nc != 0); |
| 254 |
} |