Home
last modified time | relevance | path

Searched refs:vk24x01234567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-fma3-acc2.c240 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
241 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
358 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
359 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
Dup8x25-fma3.c240 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
241 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
356 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
357 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
Dup16x25-fma3.c314 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
316 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
481 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
482 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
597 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
598 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
Dup16x25-fma3-acc2.c314 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
316 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
484 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
485 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
602 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
603 vacc01234567p0 = _mm256_fmadd_ps(vi24x01234567, vk24x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
Dup8x25-avx.c240 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__avx() local
241 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
356 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__avx() local
357 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
Dup8x25-avx-acc2.c240 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
241 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
358 const __m256 vk24x01234567 = _mm256_load_ps(w + 200); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
359 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
Dup16x25-avx-acc2.c314 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
316 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
484 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
485 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
602 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
603 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
Dup16x25-avx.c314 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__avx() local
316 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
481 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__avx() local
482 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
597 const __m256 vk24x01234567 = _mm256_load_ps(w + 400); in xnn_f32_dwconv_ukernel_up16x25__avx() local
598 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi24x01234567, vk24x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()