File size: 14,851 Bytes
8b7c501 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 |
// Auto-generated file. Do not edit!
// Template: src/f16-gemm/neonfp16arith-ld64.c.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 <assert.h>
#include <arm_neon.h>
#include <xnnpack/common.h>
#include <xnnpack/gemm.h>
#include <xnnpack/intrinsics-polyfill.h>
void xnn_f16_gemminc_minmax_ukernel_4x16__neonfp16arith_ld64(
size_t mr,
size_t nc,
size_t kc,
const void* restrict a,
size_t a_stride,
const void* restrict w,
void* restrict c,
size_t cm_stride,
size_t cn_stride,
const void* restrict acc,
const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
assert(mr != 0);
assert(mr <= 4);
assert(nc != 0);
assert(kc != 0);
assert(kc % sizeof(uint16_t) == 0);
assert(a != NULL);
assert(w != NULL);
assert(c != NULL);
assert(acc != NULL);
const uint16_t* a0 = (const uint16_t*) a;
uint16_t* c0 = (uint16_t*) c;
const uint16_t* a1 = (const uint16_t*) ((uintptr_t) a0 + a_stride);
uint16_t* c1 = (uint16_t*) ((uintptr_t) c0 + cm_stride);
if XNN_UNPREDICTABLE(mr < 2) {
a1 = a0;
c1 = c0;
}
const uint16_t* a2 = (const uint16_t*) ((uintptr_t) a1 + a_stride);
uint16_t* c2 = (uint16_t*) ((uintptr_t) c1 + cm_stride);
if XNN_UNPREDICTABLE(mr <= 2) {
a2 = a1;
c2 = c1;
}
const uint16_t* a3 = (const uint16_t*) ((uintptr_t) a2 + a_stride);
uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride);
if XNN_UNPREDICTABLE(mr != 4) {
a3 = a2;
c3 = c2;
}
do {
float16x8_t vacc0x01234567 = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
float16x8_t vacc0x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
float16x8_t vacc1x01234567 = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
float16x8_t vacc1x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
float16x8_t vacc2x01234567 = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
float16x8_t vacc2x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
float16x8_t vacc3x01234567 = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
float16x8_t vacc3x89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(acc)); acc = (const void*) ((uintptr_t) acc + sizeof(float16x8_t));
size_t k = kc;
while (k >= 4 * sizeof(uint16_t)) {
const float16x4_t va0 = vreinterpret_f16_u16(vld1_u16(a0)); a0 += 4;
const float16x4_t va1 = vreinterpret_f16_u16(vld1_u16(a1)); a1 += 4;
const float16x4_t va2 = vreinterpret_f16_u16(vld1_u16(a2)); a2 += 4;
const float16x4_t va3 = vreinterpret_f16_u16(vld1_u16(a3)); a3 += 4;
const float16x8_t vb01234567c0 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
const float16x8_t vb89ABCDEFc0 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
#if XNN_ARCH_ARM64
vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc0, va1, 0);
vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc0, va2, 0);
vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);
#else
vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c0, va0, 0);
vacc1x01234567 = vmlaq_lane_f16(vacc1x01234567, vb01234567c0, va1, 0);
vacc2x01234567 = vmlaq_lane_f16(vacc2x01234567, vb01234567c0, va2, 0);
vacc3x01234567 = vmlaq_lane_f16(vacc3x01234567, vb01234567c0, va3, 0);
vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc0, va0, 0);
vacc1x89ABCDEF = vmlaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc0, va1, 0);
vacc2x89ABCDEF = vmlaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc0, va2, 0);
vacc3x89ABCDEF = vmlaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc0, va3, 0);
#endif
const float16x8_t vb01234567c1 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
const float16x8_t vb89ABCDEFc1 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
#if XNN_ARCH_ARM64
vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc1, va1, 1);
vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc1, va2, 1);
vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);
#else
vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c1, va0, 1);
vacc1x01234567 = vmlaq_lane_f16(vacc1x01234567, vb01234567c1, va1, 1);
vacc2x01234567 = vmlaq_lane_f16(vacc2x01234567, vb01234567c1, va2, 1);
vacc3x01234567 = vmlaq_lane_f16(vacc3x01234567, vb01234567c1, va3, 1);
vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc1, va0, 1);
vacc1x89ABCDEF = vmlaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc1, va1, 1);
vacc2x89ABCDEF = vmlaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc1, va2, 1);
vacc3x89ABCDEF = vmlaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc1, va3, 1);
#endif
const float16x8_t vb01234567c2 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
const float16x8_t vb89ABCDEFc2 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
#if XNN_ARCH_ARM64
vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc2, va1, 2);
vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc2, va2, 2);
vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);
#else
vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c2, va0, 2);
vacc1x01234567 = vmlaq_lane_f16(vacc1x01234567, vb01234567c2, va1, 2);
vacc2x01234567 = vmlaq_lane_f16(vacc2x01234567, vb01234567c2, va2, 2);
vacc3x01234567 = vmlaq_lane_f16(vacc3x01234567, vb01234567c2, va3, 2);
vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc2, va0, 2);
vacc1x89ABCDEF = vmlaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc2, va1, 2);
vacc2x89ABCDEF = vmlaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc2, va2, 2);
vacc3x89ABCDEF = vmlaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc2, va3, 2);
#endif
const float16x8_t vb01234567c3 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
const float16x8_t vb89ABCDEFc3 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
#if XNN_ARCH_ARM64
vacc0x01234567 = vfmaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
vacc1x01234567 = vfmaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
vacc2x01234567 = vfmaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
vacc3x01234567 = vfmaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
vacc0x89ABCDEF = vfmaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
vacc1x89ABCDEF = vfmaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc3, va1, 3);
vacc2x89ABCDEF = vfmaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc3, va2, 3);
vacc3x89ABCDEF = vfmaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);
#else
vacc0x01234567 = vmlaq_lane_f16(vacc0x01234567, vb01234567c3, va0, 3);
vacc1x01234567 = vmlaq_lane_f16(vacc1x01234567, vb01234567c3, va1, 3);
vacc2x01234567 = vmlaq_lane_f16(vacc2x01234567, vb01234567c3, va2, 3);
vacc3x01234567 = vmlaq_lane_f16(vacc3x01234567, vb01234567c3, va3, 3);
vacc0x89ABCDEF = vmlaq_lane_f16(vacc0x89ABCDEF, vb89ABCDEFc3, va0, 3);
vacc1x89ABCDEF = vmlaq_lane_f16(vacc1x89ABCDEF, vb89ABCDEFc3, va1, 3);
vacc2x89ABCDEF = vmlaq_lane_f16(vacc2x89ABCDEF, vb89ABCDEFc3, va2, 3);
vacc3x89ABCDEF = vmlaq_lane_f16(vacc3x89ABCDEF, vb89ABCDEFc3, va3, 3);
#endif
k -= 4 * sizeof(uint16_t);
}
if XNN_UNLIKELY(k != 0) {
do {
const float16x8_t va0 = vreinterpretq_f16_u16(vld1q_dup_u16(a0)); a0 += 1;
const float16x8_t va1 = vreinterpretq_f16_u16(vld1q_dup_u16(a1)); a1 += 1;
const float16x8_t va2 = vreinterpretq_f16_u16(vld1q_dup_u16(a2)); a2 += 1;
const float16x8_t va3 = vreinterpretq_f16_u16(vld1q_dup_u16(a3)); a3 += 1;
const float16x8_t vb01234567 = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
const float16x8_t vb89ABCDEF = vreinterpretq_f16_u16(vld1q_u16(w)); w = (const void*) ((uintptr_t) w + sizeof(float16x8_t));
vacc0x01234567 = vfmaq_f16(vacc0x01234567, va0, vb01234567);
vacc1x01234567 = vfmaq_f16(vacc1x01234567, va1, vb01234567);
vacc2x01234567 = vfmaq_f16(vacc2x01234567, va2, vb01234567);
vacc3x01234567 = vfmaq_f16(vacc3x01234567, va3, vb01234567);
vacc0x89ABCDEF = vfmaq_f16(vacc0x89ABCDEF, va0, vb89ABCDEF);
vacc1x89ABCDEF = vfmaq_f16(vacc1x89ABCDEF, va1, vb89ABCDEF);
vacc2x89ABCDEF = vfmaq_f16(vacc2x89ABCDEF, va2, vb89ABCDEF);
vacc3x89ABCDEF = vfmaq_f16(vacc3x89ABCDEF, va3, vb89ABCDEF);
k -= sizeof(uint16_t);
} while (k != 0);
}
const float16x8_t vmin = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.min));
vacc0x01234567 = vmaxq_f16(vacc0x01234567, vmin);
vacc1x01234567 = vmaxq_f16(vacc1x01234567, vmin);
vacc2x01234567 = vmaxq_f16(vacc2x01234567, vmin);
vacc3x01234567 = vmaxq_f16(vacc3x01234567, vmin);
vacc0x89ABCDEF = vmaxq_f16(vacc0x89ABCDEF, vmin);
vacc1x89ABCDEF = vmaxq_f16(vacc1x89ABCDEF, vmin);
vacc2x89ABCDEF = vmaxq_f16(vacc2x89ABCDEF, vmin);
vacc3x89ABCDEF = vmaxq_f16(vacc3x89ABCDEF, vmin);
const float16x8_t vmax = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.max));
vacc0x01234567 = vminq_f16(vacc0x01234567, vmax);
vacc1x01234567 = vminq_f16(vacc1x01234567, vmax);
vacc2x01234567 = vminq_f16(vacc2x01234567, vmax);
vacc3x01234567 = vminq_f16(vacc3x01234567, vmax);
vacc0x89ABCDEF = vminq_f16(vacc0x89ABCDEF, vmax);
vacc1x89ABCDEF = vminq_f16(vacc1x89ABCDEF, vmax);
vacc2x89ABCDEF = vminq_f16(vacc2x89ABCDEF, vmax);
vacc3x89ABCDEF = vminq_f16(vacc3x89ABCDEF, vmax);
if XNN_LIKELY(nc >= 16) {
vst1q_u16(c0, vreinterpretq_u16_f16(vacc0x01234567));
vst1q_u16(c0 + 8, vreinterpretq_u16_f16(vacc0x89ABCDEF));
c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride);
vst1q_u16(c1, vreinterpretq_u16_f16(vacc1x01234567));
vst1q_u16(c1 + 8, vreinterpretq_u16_f16(vacc1x89ABCDEF));
c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride);
vst1q_u16(c2, vreinterpretq_u16_f16(vacc2x01234567));
vst1q_u16(c2 + 8, vreinterpretq_u16_f16(vacc2x89ABCDEF));
c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride);
vst1q_u16(c3, vreinterpretq_u16_f16(vacc3x01234567));
vst1q_u16(c3 + 8, vreinterpretq_u16_f16(vacc3x89ABCDEF));
c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride);
a0 = (const uint16_t*) ((uintptr_t) a0 - kc);
a1 = (const uint16_t*) ((uintptr_t) a1 - kc);
a2 = (const uint16_t*) ((uintptr_t) a2 - kc);
a3 = (const uint16_t*) ((uintptr_t) a3 - kc);
nc -= 16;
} else {
if (nc & 8) {
vst1q_u16(c0, vreinterpretq_u16_f16(vacc0x01234567)); c0 += 8;
vst1q_u16(c1, vreinterpretq_u16_f16(vacc1x01234567)); c1 += 8;
vst1q_u16(c2, vreinterpretq_u16_f16(vacc2x01234567)); c2 += 8;
vst1q_u16(c3, vreinterpretq_u16_f16(vacc3x01234567)); c3 += 8;
vacc0x01234567 = vacc0x89ABCDEF;
vacc1x01234567 = vacc1x89ABCDEF;
vacc2x01234567 = vacc2x89ABCDEF;
vacc3x01234567 = vacc3x89ABCDEF;
}
float16x4_t vacc0x0123 = vget_low_f16(vacc0x01234567);
float16x4_t vacc1x0123 = vget_low_f16(vacc1x01234567);
float16x4_t vacc2x0123 = vget_low_f16(vacc2x01234567);
float16x4_t vacc3x0123 = vget_low_f16(vacc3x01234567);
if (nc & 4) {
vst1_u16(c0, vreinterpret_u16_f16(vacc0x0123)); c0 += 4;
vst1_u16(c1, vreinterpret_u16_f16(vacc1x0123)); c1 += 4;
vst1_u16(c2, vreinterpret_u16_f16(vacc2x0123)); c2 += 4;
vst1_u16(c3, vreinterpret_u16_f16(vacc3x0123)); c3 += 4;
vacc0x0123 = vget_high_f16(vacc0x01234567);
vacc1x0123 = vget_high_f16(vacc1x01234567);
vacc2x0123 = vget_high_f16(vacc2x01234567);
vacc3x0123 = vget_high_f16(vacc3x01234567);
}
if (nc & 2) {
vst1_lane_u32((void*) c0, vreinterpret_u32_f16(vacc0x0123), 0); c0 += 2;
vst1_lane_u32((void*) c1, vreinterpret_u32_f16(vacc1x0123), 0); c1 += 2;
vst1_lane_u32((void*) c2, vreinterpret_u32_f16(vacc2x0123), 0); c2 += 2;
vst1_lane_u32((void*) c3, vreinterpret_u32_f16(vacc3x0123), 0); c3 += 2;
vacc0x0123 = vext_f16(vacc0x0123, vacc0x0123, 2);
vacc1x0123 = vext_f16(vacc1x0123, vacc1x0123, 2);
vacc2x0123 = vext_f16(vacc2x0123, vacc2x0123, 2);
vacc3x0123 = vext_f16(vacc3x0123, vacc3x0123, 2);
}
if (nc & 1) {
vst1_lane_u16(c0, vreinterpret_u16_f16(vacc0x0123), 0);
vst1_lane_u16(c1, vreinterpret_u16_f16(vacc1x0123), 0);
vst1_lane_u16(c2, vreinterpret_u16_f16(vacc2x0123), 0);
vst1_lane_u16(c3, vreinterpret_u16_f16(vacc3x0123), 0);
}
nc = 0;
}
} while (nc != 0);
}
|