Home
last modified time | relevance | path

Searched refs:vi21x01234567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-fma3-acc2.c219 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
223 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
345 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
347 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
Dup8x25-fma3.c219 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
223 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
343 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
345 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
Dup16x25-fma3.c283 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
289 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
460 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
464 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
584 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
586 vacc01234567p0 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
Dup16x25-fma3-acc2.c283 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
289 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
463 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
467 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
589 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
591 vacc01234567p1 = _mm256_fmadd_ps(vi21x01234567, vk21x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
Dup8x25-avx.c219 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up8x25__avx() local
223 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
343 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx() local
345 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
Dup8x25-avx-acc2.c219 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
223 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
345 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
347 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
Dup16x25-avx-acc2.c283 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
289 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
463 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
467 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
589 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
591 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
Dup16x25-avx.c283 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__avx() local
289 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
460 const __m256 vi21x01234567 = _mm256_loadu_ps(i21); in xnn_f32_dwconv_ukernel_up16x25__avx() local
464 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
584 const __m256 vi21x01234567 = _mm256_maskload_ps(i21, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx() local
586 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi21x01234567, vk21x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()