Home
last modified time | relevance | path

Searched refs:vi12x01234567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-fma3-acc2.c165 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
169 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
309 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
311 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
Dup8x25-fma3.c165 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
169 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
307 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
309 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
Dup16x25-fma3.c202 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
208 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
406 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
410 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
548 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
550 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
Dup16x25-fma3-acc2.c202 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
208 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
409 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
413 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
553 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
555 vacc01234567p0 = _mm256_fmadd_ps(vi12x01234567, vk12x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
Dup8x25-avx.c165 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up8x25__avx() local
169 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
307 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx() local
309 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
Dup8x25-avx-acc2.c165 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
169 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
309 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
311 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
Dup16x25-avx-acc2.c202 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
208 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
409 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
413 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
553 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
555 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
Dup16x25-avx.c202 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__avx() local
208 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
406 const __m256 vi12x01234567 = _mm256_loadu_ps(i12); in xnn_f32_dwconv_ukernel_up16x25__avx() local
410 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
548 const __m256 vi12x01234567 = _mm256_maskload_ps(i12, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx() local
550 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi12x01234567, vk12x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()