| /external/XNNPACK/src/f32-raddextexp/ |
| D | avx2-p5.c.in | 132 const __m256 vmax_acce${ABC[A:A+2]} = _mm256_max_ps(vacce${A}, vacce${A+1});
134 const __m256 vmax_acce${ABC[A]} = vacce${A};
139 …const __m256 vmax_acce${ABC[A:min(A+ACC_SLICE*2, ACCUMULATORS)]} = _mm256_max_ps(vmax_acce${ABC[A:…
143 …const __m256 vdelta_acce${K} = _mm256_max_ps(_mm256_sub_ps(vacce${K}, vmax_acce${ABC[0:ACCUMULATOR…
151 __m256 vacce = vmax_acce${ABC[0:ACCUMULATORS]};
239 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1));
240 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, …
241 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, …
242 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent);
251 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce));
|
| D | avx512f-p5-scalef.c.in | 114 const __m512 vmax_acce${ABC[A:A+2]} = _mm512_max_ps(vacce${A}, vacce${A+1});
116 const __m512 vmax_acce${ABC[A]} = vacce${A};
121 …const __m512 vmax_acce${ABC[A:min(A+ACC_SLICE*2, ACCUMULATORS)]} = _mm512_max_ps(vmax_acce${ABC[A:…
125 const __m512 vdelta_acce${K} = _mm512_sub_ps(vacce${K}, vmax_acce${ABC[0:ACCUMULATORS]});
130 __m512 vacce = vmax_acce${ABC[0:ACCUMULATORS]};
197 const float vmax_acce = _mm512_reduce_max_ps(vacce);
198 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce));
201 sum[1] = vmax_acce;
|
| /external/XNNPACK/src/f32-raddextexp/gen/ |
| D | avx2-p5-x64.c | 283 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x64() local
284 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x64()
285 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x64()
286 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64()
295 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x64()
|
| D | avx2-p5-x72.c | 296 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x72() local
297 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x72()
298 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x72()
299 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72()
308 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x72()
|
| D | avx2-p5-x64-acc2.c | 298 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2() local
299 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
300 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
301 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
310 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc2()
|
| D | avx2-p5-x64-acc4.c | 318 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4() local
319 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
320 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
321 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
330 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x64_acc4()
|
| D | avx2-p5-x80-acc2.c | 324 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2() local
325 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
326 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
327 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
336 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc2()
|
| D | avx2-p5-x80.c | 309 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x80() local
310 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x80()
311 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x80()
312 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80()
321 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x80()
|
| D | avx2-p5-x72-acc3.c | 322 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3() local
323 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
324 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
325 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
334 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x72_acc3()
|
| D | avx2-p5-x96.c | 335 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96() local
336 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
337 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
338 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
347 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96()
|
| D | avx2-p5-x96-acc3.c | 361 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3() local
362 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
363 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
364 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
373 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc3()
|
| D | avx2-p5-x96-acc2.c | 350 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2() local
351 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
352 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
353 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
362 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc2()
|
| D | avx2-p5-x80-acc5.c | 355 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5() local
356 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
357 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
358 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
367 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x80_acc5()
|
| D | avx2-p5-x96-acc6.c | 390 __m256 vmax_acce = _mm256_max_ps(vacce, _mm256_permute2f128_ps(vacce, vacce, 1)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6() local
391 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(1, 0, 3, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
392 …vmax_acce = _mm256_max_ps(vmax_acce, _mm256_shuffle_ps(vmax_acce, vmax_acce, _MM_SHUFFLE(2, 3, 0, … in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
393 const __m256 vdelta_acce = _mm256_max_ps(_mm256_sub_ps(vacce, vmax_acce), vmin_exponent); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
402 _mm_store_ss(&sum[1], _mm256_castps256_ps128(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx2_p5_x96_acc6()
|
| D | avx512f-p5-scalef-x128.c | 239 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128() local
240 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
243 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128()
|
| D | avx512f-p5-scalef-x128-acc2.c | 250 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2() local
251 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
254 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc2()
|
| D | avx512f-p5-scalef-x144.c | 251 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144() local
252 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
255 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144()
|
| D | avx512f-p5-scalef-x128-acc4.c | 266 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4() local
267 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
270 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x128_acc4()
|
| D | avx512f-p5-scalef-x160-acc2.c | 274 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2() local
275 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
278 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc2()
|
| D | avx512f-p5-scalef-x160.c | 263 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160() local
264 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
267 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160()
|
| D | avx512f-p5-scalef-x144-acc3.c | 271 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3() local
272 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
275 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x144_acc3()
|
| D | avx512f-p5-scalef-x192-acc2.c | 298 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2() local
299 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
302 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc2()
|
| D | avx512f-p5-scalef-x160-acc5.c | 299 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5() local
300 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
303 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x160_acc5()
|
| D | avx512f-p5-scalef-x192.c | 287 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192() local
288 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
291 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192()
|
| D | avx512f-p5-scalef-x192-acc3.c | 307 const float vmax_acce = _mm512_reduce_max_ps(vacce); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3() local
308 const __m512 vdelta_acce = _mm512_sub_ps(vacce, _mm512_set1_ps(vmax_acce)); in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
311 sum[1] = vmax_acce; in xnn_f32_raddextexp_ukernel__avx512f_p5_scalef_x192_acc3()
|