Home
last modified time | relevance | path

Searched refs:_mm512_load_ps (Results 1 – 25 of 45) sorted by relevance

12

/external/XNNPACK/src/f32-dwconv/gen/
Dup32x25-minmax-avx512f-acc2.c166 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
167 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
174 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
175 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
183 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
184 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
192 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
193 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
201 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
202 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
[all …]
Dup32x25-minmax-avx512f.c166 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
167 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
174 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
175 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
183 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
184 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
192 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
193 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
201 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
202 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
[all …]
Dup32x9-minmax-avx512f-acc2.c86 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
87 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
94 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
95 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
103 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
104 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
112 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
113 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
121 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
122 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
[all …]
Dup32x9-minmax-avx512f.c86 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
87 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
94 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
95 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
103 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
104 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
112 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
113 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
121 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
122 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
[all …]
Dup32x4-minmax-avx512f.c61 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
62 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
69 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
70 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
78 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
79 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
87 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
88 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
96 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
97 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
[all …]
Dup32x4-minmax-avx512f-acc2.c61 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
62 __m512 vaccGHIJKLMNOPQRSTUVp0 = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
69 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
70 const __m512 vk0xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
78 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
79 const __m512 vk1xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
87 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
88 const __m512 vk2xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
96 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
97 const __m512 vk3xGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
[all …]
Dup16x25-minmax-avx512f.c166 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
172 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
178 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
184 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
190 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
196 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
202 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
208 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
214 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
220 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
[all …]
Dup16x25-minmax-avx512f-acc2.c166 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
172 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
178 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
184 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
190 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
196 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
202 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
208 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
214 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
220 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
[all …]
Dup16x9-minmax-avx512f-acc2.c86 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
92 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
98 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
104 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
110 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
116 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
122 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
128 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
134 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
140 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
Dup16x9-minmax-avx512f.c86 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
92 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
98 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
104 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
110 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
116 const __m512 vk4x0123456789ABCDEF = _mm512_load_ps(w + 80); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
122 const __m512 vk5x0123456789ABCDEF = _mm512_load_ps(w + 96); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
128 const __m512 vk6x0123456789ABCDEF = _mm512_load_ps(w + 112); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
134 const __m512 vk7x0123456789ABCDEF = _mm512_load_ps(w + 128); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
140 const __m512 vk8x0123456789ABCDEF = _mm512_load_ps(w + 144); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
Dup16x4-minmax-avx512f-acc2.c61 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f_acc2()
67 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f_acc2()
73 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f_acc2()
79 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f_acc2()
85 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f_acc2()
Dup16x4-minmax-avx512f.c61 __m512 vacc0123456789ABCDEFp0 = _mm512_load_ps(w); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f()
67 const __m512 vk0x0123456789ABCDEF = _mm512_load_ps(w + 16); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f()
73 const __m512 vk1x0123456789ABCDEF = _mm512_load_ps(w + 32); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f()
79 const __m512 vk2x0123456789ABCDEF = _mm512_load_ps(w + 48); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f()
85 const __m512 vk3x0123456789ABCDEF = _mm512_load_ps(w + 64); in xnn_f32_dwconv_minmax_ukernel_up16x4__avx512f()
/external/XNNPACK/src/f32-gemm/gen-inc/
D8x16inc-minmax-avx512f-broadcast.c87 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
88 __m512 vacc1x0123456789ABCDEF = _mm512_load_ps(acc + 16); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
89 __m512 vacc2x0123456789ABCDEF = _mm512_load_ps(acc + 32); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
90 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
91 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
92 __m512 vacc5x0123456789ABCDEF = _mm512_load_ps(acc + 80); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
93 __m512 vacc6x0123456789ABCDEF = _mm512_load_ps(acc + 96); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
94 __m512 vacc7x0123456789ABCDEF = _mm512_load_ps(acc + 112); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
99 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemminc_minmax_ukernel_8x16__avx512f_broadcast()
D7x16inc-minmax-avx512f-broadcast.c81 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
82 __m512 vacc1x0123456789ABCDEF = _mm512_load_ps(acc + 16); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
83 __m512 vacc2x0123456789ABCDEF = _mm512_load_ps(acc + 32); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
84 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
85 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
86 __m512 vacc5x0123456789ABCDEF = _mm512_load_ps(acc + 80); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
87 __m512 vacc6x0123456789ABCDEF = _mm512_load_ps(acc + 96); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
92 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemminc_minmax_ukernel_7x16__avx512f_broadcast()
D6x16inc-minmax-avx512f-broadcast.c75 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
76 __m512 vacc1x0123456789ABCDEF = _mm512_load_ps(acc + 16); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
77 __m512 vacc2x0123456789ABCDEF = _mm512_load_ps(acc + 32); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
78 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
79 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
80 __m512 vacc5x0123456789ABCDEF = _mm512_load_ps(acc + 80); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
85 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemminc_minmax_ukernel_6x16__avx512f_broadcast()
D5x16inc-minmax-avx512f-broadcast.c69 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
70 __m512 vacc1x0123456789ABCDEF = _mm512_load_ps(acc + 16); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
71 __m512 vacc2x0123456789ABCDEF = _mm512_load_ps(acc + 32); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
72 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
73 __m512 vacc4x0123456789ABCDEF = _mm512_load_ps(acc + 64); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
78 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemminc_minmax_ukernel_5x16__avx512f_broadcast()
D4x16inc-minmax-avx512f-broadcast.c63 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
64 __m512 vacc1x0123456789ABCDEF = _mm512_load_ps(acc + 16); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
65 __m512 vacc2x0123456789ABCDEF = _mm512_load_ps(acc + 32); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
66 __m512 vacc3x0123456789ABCDEF = _mm512_load_ps(acc + 48); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
71 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemminc_minmax_ukernel_4x16__avx512f_broadcast()
D1x16inc-minmax-avx512f-broadcast.c45 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(acc + 0); in xnn_f32_gemminc_minmax_ukernel_1x16__avx512f_broadcast()
50 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemminc_minmax_ukernel_1x16__avx512f_broadcast()
/external/XNNPACK/src/f32-gemm/
Davx512-broadcast.c.in66 __m512 vacc${M}x${ABC[N:N+16]} = _mm512_load_ps(acc + ${M*NR+N});
69 __m512 vacc0x${ABC[0:16]} = _mm512_load_ps(w);
71 __m512 vacc0x${ABC[N:N+16]} = _mm512_load_ps(w + ${N});
79 const __m512 vb${ABC[0:16]} = _mm512_load_ps(w);
81 const __m512 vb${ABC[N:N+16]} = _mm512_load_ps(w + ${N});
/external/XNNPACK/src/f32-gemm/gen/
D1x16-minmax-avx512f-broadcast.c43 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast()
48 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_minmax_ukernel_1x16__avx512f_broadcast()
D4x16-minmax-avx512f-broadcast.c61 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast()
69 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_gemm_minmax_ukernel_4x16__avx512f_broadcast()
/external/XNNPACK/src/f32-prelu/gen/
Davx512f-2x32.c49 const __m512 vw0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_prelu_ukernel__avx512f_2x32()
50 const __m512 vwGHIJKLMNOPQRSTUV = _mm512_load_ps(w + 16); in xnn_f32_prelu_ukernel__avx512f_2x32()
77 const __m512 vw = _mm512_load_ps(w); in xnn_f32_prelu_ukernel__avx512f_2x32()
/external/XNNPACK/src/f32-dwconv/
Dup-avx512.c.in47 __m512 vacc${ABC[0:16]}p0 = _mm512_load_ps(w);
49 __m512 vacc${ABC[C:C+16]}p0 = _mm512_load_ps(w + ${C});
59 const __m512 vk${K}x${ABC[C:C+16]} = _mm512_load_ps(w + ${(K + 1) * CHANNEL_TILE + C});
90 __m512 vacc${ABC[0:16]}p0 = _mm512_load_ps(w);
96 const __m512 vk${K}x${ABC[0:16]} = _mm512_load_ps(w + ${(K + 1) * CHANNEL_TILE});
/external/XNNPACK/src/f32-igemm/gen/
D1x16-minmax-avx512f-broadcast.c47 __m512 vacc0x0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast()
61 const __m512 vb0123456789ABCDEF = _mm512_load_ps(w); in xnn_f32_igemm_minmax_ukernel_1x16__avx512f_broadcast()
/external/XNNPACK/src/f32-igemm/
Davx512-broadcast.c.in59 __m512 vacc0x${ABC[0:16]} = _mm512_load_ps(w);
61 __m512 vacc0x${ABC[N:N+16]} = _mm512_load_ps(w + ${N});
79 const __m512 vb${ABC[0:16]} = _mm512_load_ps(w);
81 const __m512 vb${ABC[N:N+16]} = _mm512_load_ps(w + ${N});

12