Home
last modified time | relevance | path

Searched refs:vi5x01234567 (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup16x9-fma3-acc2.c107 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
113 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
191 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
195 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
253 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
255 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
Dup16x9-fma3.c107 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
113 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
188 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
192 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
248 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
250 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
Dup16x9-avx-acc2.c107 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
113 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
191 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
195 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
253 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
255 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
Dup16x9-avx.c107 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__avx() local
113 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
188 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x9__avx() local
192 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
248 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x9__avx() local
250 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
Dup8x9-fma3-acc2.c91 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
95 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2()
153 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
155 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2()
Dup8x9-fma3.c91 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
95 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3()
151 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
153 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3()
Dup8x9-avx.c91 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x9__avx() local
95 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx()
151 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x9__avx() local
153 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx()
Dup8x9-avx-acc2.c91 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
95 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2()
153 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
155 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2()
Dup8x25-fma3-acc2.c123 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
127 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
281 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
283 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
Dup8x25-fma3.c123 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
127 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
279 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
281 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
Dup16x25-fma3.c139 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
145 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
364 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
368 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
520 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
522 vacc01234567p0 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
Dup16x25-fma3-acc2.c139 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
145 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
367 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
371 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
525 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
527 vacc01234567p1 = _mm256_fmadd_ps(vi5x01234567, vk5x01234567, vacc01234567p1); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
Dup8x25-avx.c123 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x25__avx() local
127 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
279 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx() local
281 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
Dup8x25-avx-acc2.c123 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
127 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
281 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
283 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
Dup16x25-avx-acc2.c139 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
145 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
367 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
371 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
525 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
527 vacc01234567p1 = _mm256_add_ps(vacc01234567p1, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
Dup16x25-avx.c139 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__avx() local
145 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
364 const __m256 vi5x01234567 = _mm256_loadu_ps(i5); in xnn_f32_dwconv_ukernel_up16x25__avx() local
368 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
520 const __m256 vi5x01234567 = _mm256_maskload_ps(i5, vmask); in xnn_f32_dwconv_ukernel_up16x25__avx() local
522 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi5x01234567, vk5x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()