Home
last modified time | relevance | path

Searched refs:vk0x01234567 (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup16x4-fma3.c56 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__fma3() local
58 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x4__fma3()
106 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__fma3() local
107 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x4__fma3()
144 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__fma3() local
145 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x4__fma3()
Dup16x4-avx-acc2.c56 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local
58 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2()
109 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local
110 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2()
149 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2() local
150 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x4__avx_acc2()
Dup16x4-avx.c56 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__avx() local
58 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x4__avx()
106 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__avx() local
107 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x4__avx()
144 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__avx() local
145 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x4__avx()
Dup16x4-fma3-acc2.c56 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local
58 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2()
109 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local
110 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2()
149 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2() local
150 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x4__fma3_acc2()
Dup8x4-fma3.c54 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__fma3() local
55 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x4__fma3()
92 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__fma3() local
93 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x4__fma3()
Dup8x4-avx.c54 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__avx() local
55 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x4__avx()
92 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__avx() local
93 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x4__avx()
Dup8x4-fma3-acc2.c54 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local
55 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2()
94 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2() local
95 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x4__fma3_acc2()
Dup8x4-avx-acc2.c54 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local
55 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2()
94 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2() local
95 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x4__avx_acc2()
Dup16x9-fma3-acc2.c66 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
68 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
164 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
165 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
234 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2() local
235 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3_acc2()
Dup16x9-fma3.c66 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
68 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
161 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
162 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
229 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__fma3() local
230 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x9__fma3()
Dup16x9-avx-acc2.c66 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
68 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
164 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
165 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
234 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2() local
235 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx_acc2()
Dup16x9-avx.c66 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__avx() local
68 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
161 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__avx() local
162 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
229 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x9__avx() local
230 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x9__avx()
Dup8x9-fma3-acc2.c64 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
65 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2()
134 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2() local
135 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3_acc2()
Dup8x9-fma3.c64 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
65 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3()
132 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__fma3() local
133 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x9__fma3()
Dup8x9-avx.c64 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__avx() local
65 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx()
132 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__avx() local
133 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx()
Dup8x9-avx-acc2.c64 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
65 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2()
134 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2() local
135 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x9__avx_acc2()
Dup8x25-fma3-acc2.c96 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
97 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
262 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2() local
263 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3_acc2()
Dup8x25-fma3.c96 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
97 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
260 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__fma3() local
261 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up8x25__fma3()
Dup16x25-fma3.c98 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
100 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
337 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
338 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
501 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__fma3() local
502 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3()
Dup16x25-fma3-acc2.c98 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
100 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
340 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
341 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
506 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2() local
507 vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0); in xnn_f32_dwconv_ukernel_up16x25__fma3_acc2()
Dup8x25-avx.c96 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__avx() local
97 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
260 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__avx() local
261 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx()
Dup8x25-avx-acc2.c96 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
97 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
262 const __m256 vk0x01234567 = _mm256_load_ps(w + 8); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2() local
263 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up8x25__avx_acc2()
Dup16x25-avx-acc2.c98 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
100 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
340 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
341 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
506 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2() local
507 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx_acc2()
Dup16x25-avx.c98 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__avx() local
100 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
337 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__avx() local
338 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()
501 const __m256 vk0x01234567 = _mm256_load_ps(w + 16); in xnn_f32_dwconv_ukernel_up16x25__avx() local
502 vacc01234567p0 = _mm256_add_ps(vacc01234567p0, _mm256_mul_ps(vi0x01234567, vk0x01234567)); in xnn_f32_dwconv_ukernel_up16x25__avx()