File size: 14,133 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 |
// Copyright 2021 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/microparams.h>
#include <xnnpack/vbinary.h>
#include <xnnpack/vunary.h>
void xnn_f16_vdiv_minmax_ukernel__aarch64_neonfp16arith_x8(
size_t batch,
const void* restrict input_a,
const void* restrict input_b,
void* restrict output,
const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(uint16_t) == 0);
assert(input_a != NULL);
assert(input_b != NULL);
assert(output != NULL);
const uint16_t* a = (const uint16_t*) input_a;
const uint16_t* b = (const uint16_t*) input_b;
uint16_t* o = (uint16_t*) output;
const float16x8_t vy_min = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.min));
const float16x8_t vy_max = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.max));
for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(a)); a += 8;
const float16x8_t vb01234567 = vreinterpretq_f16_u16(vld1q_u16(b)); b += 8;
float16x8_t vy01234567 = vdivq_f16(va01234567, vb01234567);
vy01234567 = vmaxq_f16(vy01234567, vy_min);
vy01234567 = vminq_f16(vy01234567, vy_max);
vst1q_u16(o, vreinterpretq_u16_f16(vy01234567)); o += 8;
}
if XNN_UNLIKELY(batch != 0) {
const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(a));
const float16x8_t vb01234567 = vreinterpretq_f16_u16(vld1q_u16(b));
float16x8_t vy01234567 = vdivq_f16(va01234567, vb01234567);
vy01234567 = vmaxq_f16(vy01234567, vy_min);
vy01234567 = vminq_f16(vy01234567, vy_max);
float16x4_t vy0123 = vget_low_f16(vy01234567);
if (batch & (4 * sizeof(uint16_t))) {
vst1_u16(o, vreinterpret_u16_f16(vy0123)); o += 4;
vy0123 = vget_high_f16(vy01234567);
}
if (batch & (2 * sizeof(uint16_t))) {
vst1_lane_u32((void*) o, vreinterpret_u32_f16(vy0123), 0); o += 2;
vy0123 = vext_f16(vy0123, vy0123, 2);
}
if (batch & (1 * sizeof(uint16_t))) {
vst1_lane_u16(o, vreinterpret_u16_f16(vy0123), 0);
}
}
}
void xnn_f16_vdivc_minmax_ukernel__aarch64_neonfp16arith_x8(
size_t batch,
const void* restrict input_a,
const void* restrict input_b,
void* restrict output,
const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(uint16_t) == 0);
assert(input_a != NULL);
assert(input_b != NULL);
assert(output != NULL);
const uint16_t* a = (const uint16_t*) input_a;
const uint16_t* b = (const uint16_t*) input_b;
uint16_t* o = (uint16_t*) output;
const float16x8_t vy_min = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.min));
const float16x8_t vy_max = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.max));
const float16x8_t vb = vreinterpretq_f16_u16(vld1q_dup_u16(b));
for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(a)); a += 8;
float16x8_t vy01234567 = vdivq_f16(va01234567, vb);
vy01234567 = vmaxq_f16(vy01234567, vy_min);
vy01234567 = vminq_f16(vy01234567, vy_max);
vst1q_u16(o, vreinterpretq_u16_f16(vy01234567)); o += 8;
}
if XNN_UNLIKELY(batch != 0) {
const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(a));
float16x8_t vy01234567 = vdivq_f16(va01234567, vb);
vy01234567 = vmaxq_f16(vy01234567, vy_min);
vy01234567 = vminq_f16(vy01234567, vy_max);
float16x4_t vy0123 = vget_low_f16(vy01234567);
if (batch & (4 * sizeof(uint16_t))) {
vst1_u16(o, vreinterpret_u16_f16(vy0123)); o += 4;
vy0123 = vget_high_f16(vy01234567);
}
if (batch & (2 * sizeof(uint16_t))) {
vst1_lane_u32((void*) o, vreinterpret_u32_f16(vy0123), 0); o += 2;
vy0123 = vext_f16(vy0123, vy0123, 2);
}
if (batch & (1 * sizeof(uint16_t))) {
vst1_lane_u16(o, vreinterpret_u16_f16(vy0123), 0);
}
}
}
void xnn_f16_vrdivc_minmax_ukernel__aarch64_neonfp16arith_x8(
size_t batch,
const void* restrict input_a,
const void* restrict input_b,
void* restrict output,
const union xnn_f16_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(uint16_t) == 0);
assert(input_a != NULL);
assert(input_b != NULL);
assert(output != NULL);
const uint16_t* a = (const uint16_t*) input_a;
const uint16_t* b = (const uint16_t*) input_b;
uint16_t* o = (uint16_t*) output;
const float16x8_t vy_min = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.min));
const float16x8_t vy_max = vreinterpretq_f16_u16(vld1q_dup_u16(¶ms->fp16arith.max));
const float16x8_t vb = vreinterpretq_f16_u16(vld1q_dup_u16(b));
for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(a)); a += 8;
float16x8_t vy01234567 = vdivq_f16(vb, va01234567);
vy01234567 = vmaxq_f16(vy01234567, vy_min);
vy01234567 = vminq_f16(vy01234567, vy_max);
vst1q_u16(o, vreinterpretq_u16_f16(vy01234567)); o += 8;
}
if XNN_UNLIKELY(batch != 0) {
const float16x8_t va01234567 = vreinterpretq_f16_u16(vld1q_u16(a));
float16x8_t vy01234567 = vdivq_f16(vb, va01234567);
vy01234567 = vmaxq_f16(vy01234567, vy_min);
vy01234567 = vminq_f16(vy01234567, vy_max);
float16x4_t vy0123 = vget_low_f16(vy01234567);
if (batch & (4 * sizeof(uint16_t))) {
vst1_u16(o, vreinterpret_u16_f16(vy0123)); o += 4;
vy0123 = vget_high_f16(vy01234567);
}
if (batch & (2 * sizeof(uint16_t))) {
vst1_lane_u32((void*) o, vreinterpret_u32_f16(vy0123), 0); o += 2;
vy0123 = vext_f16(vy0123, vy0123, 2);
}
if (batch & (1 * sizeof(uint16_t))) {
vst1_lane_u16(o, vreinterpret_u16_f16(vy0123), 0);
}
}
}
void xnn_f16_vsqrt_ukernel__aarch64_neonfp16arith_sqrt_x8(
size_t batch,
const void* input,
void* output,
const union xnn_f16_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(batch != 0);
assert(batch % sizeof(uint16_t) == 0);
assert(input != NULL);
assert(output != NULL);
const uint16_t* i = (const uint16_t*) input;
uint16_t* o = (uint16_t*) output;
for (; batch >= 8 * sizeof(uint16_t); batch -= 8 * sizeof(uint16_t)) {
float16x8_t vacc = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
vacc = vsqrtq_f16(vacc);
vst1q_u16(o, vreinterpretq_u16_f16(vacc)); o += 8;
}
if XNN_UNLIKELY(batch != 0) {
const float16x8_t vacc = vreinterpretq_f16_u16(vld1q_u16(i));
float16x4_t vacc_lo = vsqrt_f16(vget_low_f16(vacc));
if (batch & (4 * sizeof(uint16_t))) {
vst1_u16(o, vreinterpret_u16_f16(vacc_lo)); o += 4;
vacc_lo = vsqrt_f16(vget_high_f16(vacc));
}
if (batch & (2 * sizeof(uint16_t))) {
vst1_lane_u32((void*) o, vreinterpret_u32_f16(vacc_lo), 0); o += 2;
vacc_lo = vext_f16(vacc_lo, vacc_lo, 2);
}
if (batch & (1 * sizeof(uint16_t))) {
vst1_lane_u16(o, vreinterpret_u16_f16(vacc_lo), 0);
}
}
}
void xnn_f16_vtanh_ukernel__aarch64_neonfp16arith_expm1minus_rr1_p3h2ts_div_x32(
size_t n,
const void* input,
void* output,
const union xnn_f16_tanh_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
assert(n != 0);
assert(n % sizeof(uint16_t) == 0);
assert(input != NULL);
assert(output != NULL);
const float16x8_t vsat_cutoff = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x4482)));
const float16x8_t vmagic_bias = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x620F)));
const float16x8_t vminus_log2e = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xBDC5)));
const float16x8_t vln2 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x398C)));
const float16x8_t vc3 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xBD5B)));
const float16x8_t vc2 = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x4008)));
const float16x8_t vtwo = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0x4000)));
const float16x8_t vminus_one = vreinterpretq_f16_u16(vmovq_n_u16(UINT16_C(0xBC00)));
const uint16x8_t vsign_mask = vmovq_n_u16(UINT16_C(0x8000));
const uint16_t* i = (const uint16_t*) input;
uint16_t* o = (uint16_t*) output;
for (; n >= 4 * sizeof(float16x8_t); n -= 4 * sizeof(float16x8_t)) {
const float16x8_t vx0 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
const float16x8_t vx1 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
const float16x8_t vx2 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
const float16x8_t vx3 = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
float16x8_t vz0 = vabsq_f16(vx0);
float16x8_t vz1 = vabsq_f16(vx1);
float16x8_t vz2 = vabsq_f16(vx2);
float16x8_t vz3 = vabsq_f16(vx3);
vz0 = vminq_f16(vz0, vsat_cutoff);
vz1 = vminq_f16(vz1, vsat_cutoff);
vz2 = vminq_f16(vz2, vsat_cutoff);
vz3 = vminq_f16(vz3, vsat_cutoff);
float16x8_t vn0 = vfmaq_f16(vmagic_bias, vz0, vminus_log2e);
float16x8_t vn1 = vfmaq_f16(vmagic_bias, vz1, vminus_log2e);
float16x8_t vn2 = vfmaq_f16(vmagic_bias, vz2, vminus_log2e);
float16x8_t vn3 = vfmaq_f16(vmagic_bias, vz3, vminus_log2e);
const float16x8_t vs0 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn0), 10));
vn0 = vsubq_f16(vn0, vmagic_bias);
const float16x8_t vs1 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn1), 10));
vn1 = vsubq_f16(vn1, vmagic_bias);
const float16x8_t vs2 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn2), 10));
vn2 = vsubq_f16(vn2, vmagic_bias);
const float16x8_t vs3 = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn3), 10));
vn3 = vsubq_f16(vn3, vmagic_bias);
const float16x8_t vt0 = vfmaq_f16(vz0, vn0, vln2);
const float16x8_t vt1 = vfmaq_f16(vz1, vn1, vln2);
const float16x8_t vt2 = vfmaq_f16(vz2, vn2, vln2);
const float16x8_t vt3 = vfmaq_f16(vz3, vn3, vln2);
float16x8_t vp0 = vfmaq_f16(vc2, vc3, vt0);
float16x8_t vp1 = vfmaq_f16(vc2, vc3, vt1);
float16x8_t vp2 = vfmaq_f16(vc2, vc3, vt2);
float16x8_t vp3 = vfmaq_f16(vc2, vc3, vt3);
vp0 = vfmsq_f16(vtwo, vp0, vt0);
vp1 = vfmsq_f16(vtwo, vp1, vt1);
vp2 = vfmsq_f16(vtwo, vp2, vt2);
vp3 = vfmsq_f16(vtwo, vp3, vt3);
const float16x8_t vts0 = vmulq_f16(vt0, vs0);
const float16x8_t vsmo0 = vaddq_f16(vs0, vminus_one);
const float16x8_t vts1 = vmulq_f16(vt1, vs1);
const float16x8_t vsmo1 = vaddq_f16(vs1, vminus_one);
const float16x8_t vts2 = vmulq_f16(vt2, vs2);
const float16x8_t vsmo2 = vaddq_f16(vs2, vminus_one);
const float16x8_t vts3 = vmulq_f16(vt3, vs3);
const float16x8_t vsmo3 = vaddq_f16(vs3, vminus_one);
const float16x8_t vemo0 = vfmsq_f16(vsmo0, vp0, vts0);
const float16x8_t vemo1 = vfmsq_f16(vsmo1, vp1, vts1);
const float16x8_t vemo2 = vfmsq_f16(vsmo2, vp2, vts2);
const float16x8_t vemo3 = vfmsq_f16(vsmo3, vp3, vts3);
const float16x8_t vepo0 = vaddq_f16(vemo0, vtwo);
const float16x8_t vepo1 = vaddq_f16(vemo1, vtwo);
const float16x8_t vepo2 = vaddq_f16(vemo2, vtwo);
const float16x8_t vepo3 = vaddq_f16(vemo3, vtwo);
float16x8_t vy0 = vdivq_f16(vemo0, vepo0);
float16x8_t vy1 = vdivq_f16(vemo1, vepo1);
float16x8_t vy2 = vdivq_f16(vemo2, vepo2);
float16x8_t vy3 = vdivq_f16(vemo3, vepo3);
vy0 = vbslq_f16(vsign_mask, vx0, vy0);
vy1 = vbslq_f16(vsign_mask, vx1, vy1);
vy2 = vbslq_f16(vsign_mask, vx2, vy2);
vy3 = vbslq_f16(vsign_mask, vx3, vy3);
vst1q_u16(o, vreinterpretq_u16_f16(vy0)); o += 8;
vst1q_u16(o, vreinterpretq_u16_f16(vy1)); o += 8;
vst1q_u16(o, vreinterpretq_u16_f16(vy2)); o += 8;
vst1q_u16(o, vreinterpretq_u16_f16(vy3)); o += 8;
}
for (; n >= 1 * sizeof(float16x8_t); n -= 1 * sizeof(float16x8_t)) {
const float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
float16x8_t vz = vabsq_f16(vx);
vz = vminq_f16(vz, vsat_cutoff);
float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
vn = vsubq_f16(vn, vmagic_bias);
const float16x8_t vt = vfmaq_f16(vz, vn, vln2);
float16x8_t vp = vfmaq_f16(vc2, vc3, vt);
vp = vfmsq_f16(vtwo, vp, vt);
const float16x8_t vts = vmulq_f16(vt, vs);
const float16x8_t vsmo = vaddq_f16(vs, vminus_one);
const float16x8_t vemo = vfmsq_f16(vsmo, vp, vts);
const float16x8_t vepo = vaddq_f16(vemo, vtwo);
float16x8_t vy = vdivq_f16(vemo, vepo);
vy = vbslq_f16(vsign_mask, vx, vy);
vst1q_u16(o, vreinterpretq_u16_f16(vy)); o += 8;
}
if (n != 0) {
const float16x8_t vx = vreinterpretq_f16_u16(vld1q_u16(i)); i += 8;
float16x8_t vz = vabsq_f16(vx);
vz = vminq_f16(vz, vsat_cutoff);
float16x8_t vn = vfmaq_f16(vmagic_bias, vz, vminus_log2e);
const float16x8_t vs = vreinterpretq_f16_s16(vshlq_n_s16(vreinterpretq_s16_f16(vn), 10));
vn = vsubq_f16(vn, vmagic_bias);
const float16x8_t vt = vfmaq_f16(vz, vn, vln2);
float16x8_t vp = vfmaq_f16(vc2, vc3, vt);
vp = vfmsq_f16(vtwo, vp, vt);
const float16x8_t vts = vmulq_f16(vt, vs);
const float16x8_t vsmo = vaddq_f16(vs, vminus_one);
const float16x8_t vemo = vfmsq_f16(vsmo, vp, vts);
const float16x8_t vepo = vaddq_f16(vemo, vtwo);
float16x8_t vy = vdivq_f16(vemo, vepo);
vy = vbslq_f16(vsign_mask, vx, vy);
float16x4_t vy_lo = vget_low_f16(vy);
if (n & 4 * sizeof(uint16_t)) {
vst1_u16(o, vreinterpret_u16_f16(vy_lo)); o += 4;
vy_lo = vget_high_f16(vy);
}
if (n & 2 * sizeof(uint16_t)) {
vst1_lane_u32((void*) o, vreinterpret_u32_f16(vy_lo), 0); o+= 2;
vy_lo = vext_f16(vy_lo, vy_lo, 2);
}
if (n & 1 * sizeof(uint16_t)) {
vst1_lane_u16(o, vreinterpret_u16_f16(vy_lo), 0);
}
}
}
|