Home
last modified time | relevance | path

Searched refs:vk16x01234567 (Results 1 – 8 of 8) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup8x25-fma3-acc2.c192 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
193 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
326 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
327 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
Dup8x25-fma3.c192 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
193 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
324 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
325 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
Dup16x25-fma3.c242 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
244 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
433 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
434 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
565 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
566 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
Dup16x25-fma3-acc2.c242 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
244 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
436 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
437 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
570 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
571 vacc01234567p0 = _mm256_fmadd_ps(vi16x01234567, vk16x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
Dup8x25-avx.c192 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__avx() local
193 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
324 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__avx() local
325 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
Dup8x25-avx-acc2.c192 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
193 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
326 const __m256 vk16x01234567 = _mm256_load_ps(w + 136); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
327 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
Dup16x25-avx-acc2.c242 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
244 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
436 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
437 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
570 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
571 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
Dup16x25-avx.c242 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__avx() local
244 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
433 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__avx() local
434 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
565 const __m256 vk16x01234567 = _mm256_load_ps(w + 272); in xnn_f32_dwconv_ukernel_up16x25__avx() local
566 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi16x01234567, vk16x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()