Home
last modified time | relevance | path

Searched refs:vi20x01234567 (Results 1 – 12 of 12) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-minmax-fma3-acc2.c290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
294 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
418 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2() local
420 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3_acc2()
Dup8x25-minmax-fma3.c290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
294 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
416 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3() local
418 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up8x25__fma3()
Dup8x25-minmax-avx.c290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
294 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
416 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx() local
418 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx()
Dup16x25-minmax-fma3.c351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
357 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
531 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
535 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
657 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3() local
659 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3()
Dup16x25-minmax-fma3-acc2.c351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
357 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
534 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
538 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
662 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2() local
664 vacc01234567p0 = _mm256_fmadd_ps(vi20x01234567, vk20x01234567, vacc01234567p0); in xnn_f32_dwconv_minmax_ukernel_up16x25__fma3_acc2()
Dup16x25-minmax-avx.c351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
357 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
531 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
535 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
657 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx() local
659 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx()
Dup8x25-minmax-avx-acc2.c290 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
294 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
418 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2() local
420 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up8x25__avx_acc2()
Dup16x25-minmax-avx-acc2.c351 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
357 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
534 const __m256 vi20x01234567 = _mm256_loadu_ps(i20); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
538 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
662 const __m256 vi20x01234567 = _mm256_maskload_ps(i20, vmask); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2() local
664 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi20x01234567, vk20x01234567)); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx_acc2()
/external/XNNPACK/src/f16-dwconv/gen/
Dup8x25-minmax-neonfp16arith-acc2.c250 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
252 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
362 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2() local
364 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith_acc2()
Dup8x25-minmax-neonfp16arith.c250 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
252 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
360 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith() local
362 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up8x25__neonfp16arith()
Dup16x25-minmax-neonfp16arith-acc2.c311 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
315 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
442 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
444 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
554 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2() local
556 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith_acc2()
Dup16x25-minmax-neonfp16arith.c311 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
315 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
439 const float16x8_t vi20x01234567 = vld1q_f16(i20); i20 += 8; in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
441 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()
549 const float16x8_t vi20x01234567 = vld1q_f16(i20); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith() local
551 vacc01234567p0 = vfmaq_f16(vacc01234567p0, vi20x01234567, vk20x01234567); in xnn_f16_dwconv_minmax_ukernel_up16x25__neonfp16arith()