Home
last modified time | relevance | path

Searched refs:_mm512_loadu_ps (Results 1 – 25 of 259) sorted by relevance

1234567891011

/external/XNNPACK/src/f32-dwconv/gen/
Dup32x25-minmax-avx512f-acc2.c170 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
171 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
179 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
180 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
188 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
189 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
197 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
198 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
206 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
207 const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f_acc2()
[all …]
Dup32x25-minmax-avx512f.c170 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
171 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
179 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
180 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
188 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
189 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
197 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
198 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
206 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
207 const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x25__avx512f()
[all …]
Dup32x9-minmax-avx512f-acc2.c90 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
91 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
99 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
100 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
108 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
109 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
117 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
118 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
126 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
127 const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f_acc2()
[all …]
Dup32x9-minmax-avx512f.c90 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
91 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
99 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
100 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
108 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
109 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
117 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
118 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
126 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
127 const __m512 vi4xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i4 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x9__avx512f()
[all …]
Dup16x25-minmax-avx512f.c169 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
175 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
181 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
187 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
193 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
199 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
205 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
211 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
217 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
223 const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f()
[all …]
Dup16x25-minmax-avx512f-acc2.c169 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
175 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
181 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
187 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
193 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
199 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
205 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
211 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
217 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
223 const __m512 vi9x0123456789ABCDEF = _mm512_loadu_ps(i9); in xnn_f32_dwconv_minmax_ukernel_up16x25__avx512f_acc2()
[all …]
Dup32x4-minmax-avx512f.c65 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
66 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
74 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
75 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
83 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
84 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
92 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
93 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
116 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
122 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f()
[all …]
Dup32x4-minmax-avx512f-acc2.c65 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
66 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
74 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
75 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
83 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
84 const __m512 vi2xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i2 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
92 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
93 const __m512 vi3xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i3 + 16); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
119 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
125 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up32x4__avx512f_acc2()
[all …]
Dup16x9-minmax-avx512f-acc2.c89 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
95 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
101 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
107 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
113 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
119 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
125 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
131 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
137 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f_acc2()
Dup16x9-minmax-avx512f.c89 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
95 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
101 const __m512 vi2x0123456789ABCDEF = _mm512_loadu_ps(i2); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
107 const __m512 vi3x0123456789ABCDEF = _mm512_loadu_ps(i3); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
113 const __m512 vi4x0123456789ABCDEF = _mm512_loadu_ps(i4); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
119 const __m512 vi5x0123456789ABCDEF = _mm512_loadu_ps(i5); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
125 const __m512 vi6x0123456789ABCDEF = _mm512_loadu_ps(i6); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
131 const __m512 vi7x0123456789ABCDEF = _mm512_loadu_ps(i7); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
137 const __m512 vi8x0123456789ABCDEF = _mm512_loadu_ps(i8); in xnn_f32_dwconv_minmax_ukernel_up16x9__avx512f()
/external/XNNPACK/src/f32-vbinary/gen/
Dvmin-avx512f-x32.c34 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); in xnn_f32_vmin_ukernel__avx512f_x32()
35 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16); in xnn_f32_vmin_ukernel__avx512f_x32()
38 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b); in xnn_f32_vmin_ukernel__avx512f_x32()
39 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16); in xnn_f32_vmin_ukernel__avx512f_x32()
52 const __m512 va = _mm512_loadu_ps(a); in xnn_f32_vmin_ukernel__avx512f_x32()
55 const __m512 vb = _mm512_loadu_ps(b); in xnn_f32_vmin_ukernel__avx512f_x32()
Dvmax-avx512f-x32.c34 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); in xnn_f32_vmax_ukernel__avx512f_x32()
35 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16); in xnn_f32_vmax_ukernel__avx512f_x32()
38 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b); in xnn_f32_vmax_ukernel__avx512f_x32()
39 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16); in xnn_f32_vmax_ukernel__avx512f_x32()
52 const __m512 va = _mm512_loadu_ps(a); in xnn_f32_vmax_ukernel__avx512f_x32()
55 const __m512 vb = _mm512_loadu_ps(b); in xnn_f32_vmax_ukernel__avx512f_x32()
Dvsqrdiff-avx512f-x32.c34 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
35 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
38 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
39 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
54 const __m512 va = _mm512_loadu_ps(a); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
57 const __m512 vb = _mm512_loadu_ps(b); in xnn_f32_vsqrdiff_ukernel__avx512f_x32()
Dvadd-minmax-avx512f-x32.c36 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
37 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
40 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
41 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
59 const __m512 va = _mm512_loadu_ps(a); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
62 const __m512 vb = _mm512_loadu_ps(b); in xnn_f32_vadd_minmax_ukernel__avx512f_x32()
Dvdiv-minmax-avx512f-x32.c36 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
37 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
40 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
41 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
59 const __m512 va = _mm512_loadu_ps(a); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
62 const __m512 vb = _mm512_loadu_ps(b); in xnn_f32_vdiv_minmax_ukernel__avx512f_x32()
Dvmul-minmax-avx512f-x32.c36 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
37 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
40 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
41 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
59 const __m512 va = _mm512_loadu_ps(a); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
62 const __m512 vb = _mm512_loadu_ps(b); in xnn_f32_vmul_minmax_ukernel__avx512f_x32()
Dvsub-minmax-avx512f-x32.c36 const __m512 va0123456789ABCDEF = _mm512_loadu_ps(a); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
37 const __m512 vaGHIJKLMNOPQRSTUV = _mm512_loadu_ps(a + 16); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
40 const __m512 vb0123456789ABCDEF = _mm512_loadu_ps(b); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
41 const __m512 vbGHIJKLMNOPQRSTUV = _mm512_loadu_ps(b + 16); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
59 const __m512 va = _mm512_loadu_ps(a); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
62 const __m512 vb = _mm512_loadu_ps(b); in xnn_f32_vsub_minmax_ukernel__avx512f_x32()
/external/XNNPACK/src/f32-vscale/
Davx512f-x64.c26 const __m512 vx0 = _mm512_loadu_ps(x); in xnn_f32_vscale_ukernel__avx512f_x64()
27 const __m512 vx1 = _mm512_loadu_ps(x + 16); in xnn_f32_vscale_ukernel__avx512f_x64()
28 const __m512 vx2 = _mm512_loadu_ps(x + 32); in xnn_f32_vscale_ukernel__avx512f_x64()
29 const __m512 vx3 = _mm512_loadu_ps(x + 48); in xnn_f32_vscale_ukernel__avx512f_x64()
44 const __m512 vx = _mm512_loadu_ps(x); in xnn_f32_vscale_ukernel__avx512f_x64()
/external/XNNPACK/src/f32-rmax/
Davx512f.c27 const __m512 vx0 = _mm512_loadu_ps(x); in xnn_f32_rmax_ukernel__avx512f()
28 const __m512 vx1 = _mm512_loadu_ps(x + 16); in xnn_f32_rmax_ukernel__avx512f()
29 const __m512 vx2 = _mm512_loadu_ps(x + 32); in xnn_f32_rmax_ukernel__avx512f()
30 const __m512 vx3 = _mm512_loadu_ps(x + 48); in xnn_f32_rmax_ukernel__avx512f()
40 const __m512 vx = _mm512_loadu_ps(x); in xnn_f32_rmax_ukernel__avx512f()
/external/XNNPACK/src/f32-vsqrt/gen/
Davx512f-nr1fma1adj-x128.c30 const __m512 vx0 = _mm512_loadu_ps(x); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
31 const __m512 vx1 = _mm512_loadu_ps(x + 16); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
32 const __m512 vx2 = _mm512_loadu_ps(x + 32); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
33 const __m512 vx3 = _mm512_loadu_ps(x + 48); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
34 const __m512 vx4 = _mm512_loadu_ps(x + 64); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
35 const __m512 vx5 = _mm512_loadu_ps(x + 80); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
36 const __m512 vx6 = _mm512_loadu_ps(x + 96); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
37 const __m512 vx7 = _mm512_loadu_ps(x + 112); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
121 const __m512 vx = _mm512_loadu_ps(x); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x128()
Davx512f-nr1fma1adj-x112.c30 const __m512 vx0 = _mm512_loadu_ps(x); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
31 const __m512 vx1 = _mm512_loadu_ps(x + 16); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
32 const __m512 vx2 = _mm512_loadu_ps(x + 32); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
33 const __m512 vx3 = _mm512_loadu_ps(x + 48); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
34 const __m512 vx4 = _mm512_loadu_ps(x + 64); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
35 const __m512 vx5 = _mm512_loadu_ps(x + 80); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
36 const __m512 vx6 = _mm512_loadu_ps(x + 96); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
111 const __m512 vx = _mm512_loadu_ps(x); in xnn_f32_vsqrt_ukernel__avx512f_nr1fma1adj_x112()
/external/XNNPACK/src/f32-prelu/gen/
Davx512f-2x32.c53 const __m512 vi0x0123456789ABCDEF = _mm512_loadu_ps(i0); in xnn_f32_prelu_ukernel__avx512f_2x32()
54 const __m512 vi0xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i0 + 16); in xnn_f32_prelu_ukernel__avx512f_2x32()
56 const __m512 vi1x0123456789ABCDEF = _mm512_loadu_ps(i1); in xnn_f32_prelu_ukernel__avx512f_2x32()
57 const __m512 vi1xGHIJKLMNOPQRSTUV = _mm512_loadu_ps(i1 + 16); in xnn_f32_prelu_ukernel__avx512f_2x32()
80 const __m512 vi0 = _mm512_loadu_ps(i0); in xnn_f32_prelu_ukernel__avx512f_2x32()
82 const __m512 vi1 = _mm512_loadu_ps(i1); in xnn_f32_prelu_ukernel__avx512f_2x32()
/external/XNNPACK/src/f32-raddstoreexpminusmax/gen/
Davx512f-p5-scalef-x192.c43 const __m512 vi0 = _mm512_loadu_ps(input); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
44 const __m512 vi1 = _mm512_loadu_ps(input + 16); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
45 const __m512 vi2 = _mm512_loadu_ps(input + 32); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
46 const __m512 vi3 = _mm512_loadu_ps(input + 48); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
47 const __m512 vi4 = _mm512_loadu_ps(input + 64); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
48 const __m512 vi5 = _mm512_loadu_ps(input + 80); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
49 const __m512 vi6 = _mm512_loadu_ps(input + 96); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
50 const __m512 vi7 = _mm512_loadu_ps(input + 112); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
51 const __m512 vi8 = _mm512_loadu_ps(input + 128); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
52 const __m512 vi9 = _mm512_loadu_ps(input + 144); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]
Davx512f-p5-scalef-x192-acc2.c44 const __m512 vi0 = _mm512_loadu_ps(input); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
45 const __m512 vi1 = _mm512_loadu_ps(input + 16); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
46 const __m512 vi2 = _mm512_loadu_ps(input + 32); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
47 const __m512 vi3 = _mm512_loadu_ps(input + 48); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
48 const __m512 vi4 = _mm512_loadu_ps(input + 64); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
49 const __m512 vi5 = _mm512_loadu_ps(input + 80); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
50 const __m512 vi6 = _mm512_loadu_ps(input + 96); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
51 const __m512 vi7 = _mm512_loadu_ps(input + 112); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
52 const __m512 vi8 = _mm512_loadu_ps(input + 128); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
53 const __m512 vi9 = _mm512_loadu_ps(input + 144); in xnn_f32_raddstoreexpminusmax_ukernel__avx512f_p5_scalef_x192_acc2()
[all …]
/external/XNNPACK/src/f32-raddexpminusmax/gen/
Davx512f-p5-scalef-x192.c42 const __m512 vi0 = _mm512_loadu_ps(input); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
43 const __m512 vi1 = _mm512_loadu_ps(input + 16); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
44 const __m512 vi2 = _mm512_loadu_ps(input + 32); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
45 const __m512 vi3 = _mm512_loadu_ps(input + 48); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
46 const __m512 vi4 = _mm512_loadu_ps(input + 64); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
47 const __m512 vi5 = _mm512_loadu_ps(input + 80); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
48 const __m512 vi6 = _mm512_loadu_ps(input + 96); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
49 const __m512 vi7 = _mm512_loadu_ps(input + 112); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
50 const __m512 vi8 = _mm512_loadu_ps(input + 128); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
51 const __m512 vi9 = _mm512_loadu_ps(input + 144); in xnn_f32_raddexpminusmax_ukernel__avx512f_p5_scalef_x192()
[all …]

1234567891011