Home
last modified time | relevance | path

Searched refs:vk6x01234567 (Results 1 – 16 of 16) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup16x9-fma3-acc2.c120 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
122 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
200 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
201 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
258 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
259 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
Dup16x9-fma3.c120 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
122 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
197 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
198 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
253 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
254 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
Dup16x9-avx-acc2.c120 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
122 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
200 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
201 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
258 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
259 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
Dup16x9-avx.c120 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__avx() local
122 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
197 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__avx() local
198 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
253 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x9__avx() local
254 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
Dup8x9-fma3-acc2.c100 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
101 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2()
158 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
159 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2()
Dup8x9-fma3.c100 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
101 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3()
156 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
157 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3()
Dup8x9-avx.c100 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__avx() local
101 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx()
156 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__avx() local
157 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx()
Dup8x9-avx-acc2.c100 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
101 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2()
158 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
159 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2()
Dup8x25-fma3-acc2.c132 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
133 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
286 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
287 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
Dup8x25-fma3.c132 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
133 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
284 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
285 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
Dup16x25-fma3.c152 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
154 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
373 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
374 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
525 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
526 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
Dup16x25-fma3-acc2.c152 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
154 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
376 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
377 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
530 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
531 vacc01234567p0 = _mm256_fmadd_ps(vi6x01234567, vk6x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
Dup8x25-avx.c132 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__avx() local
133 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
284 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__avx() local
285 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
Dup8x25-avx-acc2.c132 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
133 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
286 const __m256 vk6x01234567 = _mm256_load_ps(w + 56); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
287 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
Dup16x25-avx-acc2.c152 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
154 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
376 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
377 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
530 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
531 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
Dup16x25-avx.c152 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__avx() local
154 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
373 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__avx() local
374 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
525 const __m256 vk6x01234567 = _mm256_load_ps(w + 112); in xnn_f32_dwconv_ukernel_up16x25__avx() local
526 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi6x01234567, vk6x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()