| /external/XNNPACK/src/x8-zip/ |
| D | x3-sse2.c | 54 …_mm_shuffle_ps(_mm_castsi128_ps(vzexoyozo), _mm_castsi128_ps(vxeyezexo), _MM_SHUFFLE(3, 1, 2, 0))); in xnn_x8_zip_x3_ukernel__sse2()
57 …_mm_shuffle_ps(_mm_castsi128_ps(vxeyezexo), _mm_castsi128_ps(vyozoxeye), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_x8_zip_x3_ukernel__sse2()
60 …_mm_shuffle_ps(_mm_castsi128_ps(vyozoxeye), _mm_castsi128_ps(vzexoyozo), _MM_SHUFFLE(3, 1, 3, 1))); in xnn_x8_zip_x3_ukernel__sse2()
64 _mm_shuffle_ps(_mm_castsi128_ps(vtemp1), _mm_castsi128_ps(vtemp0), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_x8_zip_x3_ukernel__sse2()
67 _mm_shuffle_ps(_mm_castsi128_ps(vtemp2), _mm_castsi128_ps(vtemp1), _MM_SHUFFLE(3, 1, 2, 0))); in xnn_x8_zip_x3_ukernel__sse2()
70 _mm_shuffle_ps(_mm_castsi128_ps(vtemp0), _mm_castsi128_ps(vtemp2), _MM_SHUFFLE(3, 1, 3, 1))); in xnn_x8_zip_x3_ukernel__sse2()
103 …_mm_shuffle_ps(_mm_castsi128_ps(vzexoyozo), _mm_castsi128_ps(vxeyezexo), _MM_SHUFFLE(3, 1, 2, 0))); in xnn_x8_zip_x3_ukernel__sse2()
106 …_mm_shuffle_ps(_mm_castsi128_ps(vxeyezexo), _mm_castsi128_ps(vyozoxeye), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_x8_zip_x3_ukernel__sse2()
109 …_mm_shuffle_ps(_mm_castsi128_ps(vyozoxeye), _mm_castsi128_ps(vzexoyozo), _MM_SHUFFLE(3, 1, 3, 1))); in xnn_x8_zip_x3_ukernel__sse2()
113 _mm_shuffle_ps(_mm_castsi128_ps(vtemp1), _mm_castsi128_ps(vtemp0), _MM_SHUFFLE(2, 0, 2, 0))); in xnn_x8_zip_x3_ukernel__sse2()
[all …]
|
| /external/angle/third_party/zlib/ |
| D | crc_folding.c | 74 ps_crc0 = _mm_castsi128_ps(*xmm_crc0); in fold_1()
75 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_1()
101 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_2()
102 ps_crc1 = _mm_castsi128_ps(*xmm_crc1); in fold_2()
108 ps_crc0 = _mm_castsi128_ps(*xmm_crc0); in fold_2()
109 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_2()
134 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_3()
135 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_3()
141 ps_crc1 = _mm_castsi128_ps(*xmm_crc1); in fold_3()
142 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_3()
[all …]
|
| /external/rust/crates/libz-sys/src/zlib-ng/arch/x86/ |
| D | crc_folding.c | 49 ps_crc0 = _mm_castsi128_ps(*xmm_crc0); in fold_1()
50 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_1()
71 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_2()
72 ps_crc1 = _mm_castsi128_ps(*xmm_crc1); in fold_2()
78 ps_crc0 = _mm_castsi128_ps(*xmm_crc0); in fold_2()
79 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_2()
99 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_3()
100 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_3()
106 ps_crc1 = _mm_castsi128_ps(*xmm_crc1); in fold_3()
107 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_3()
[all …]
|
| /external/zlib/ |
| D | crc_folding.c | 74 ps_crc0 = _mm_castsi128_ps(*xmm_crc0); in fold_1()
75 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_1()
101 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_2()
102 ps_crc1 = _mm_castsi128_ps(*xmm_crc1); in fold_2()
108 ps_crc0 = _mm_castsi128_ps(*xmm_crc0); in fold_2()
109 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_2()
134 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_3()
135 ps_crc3 = _mm_castsi128_ps(*xmm_crc3); in fold_3()
141 ps_crc1 = _mm_castsi128_ps(*xmm_crc1); in fold_3()
142 ps_crc2 = _mm_castsi128_ps(*xmm_crc2); in fold_3()
[all …]
|
| /external/XNNPACK/src/f32-dwconv2d-chw/gen/ |
| D | 3x3p1-minmax-ssse3-2x4.c | 100 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
101 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
102 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
103 …const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
117 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
118 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
119 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
120 …const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
162 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
163 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4()
[all …]
|
| D | 3x3p1-minmax-ssse3-6x4.c | 156 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
157 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
158 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
159 …const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
160 …const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
161 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
162 …const __m128 vi6x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
163 …const __m128 vi7x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi7x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
193 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
194 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_6x4()
[all …]
|
| D | 3x3p1-minmax-ssse3-1x4.c | 86 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
87 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
88 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
98 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
99 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
100 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
130 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
131 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
132 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
139 … const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4()
[all …]
|
| D | 3x3p1-minmax-ssse3-4x4.c | 128 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
129 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
130 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
131 …const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
132 …const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
133 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
155 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
156 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
157 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
158 …const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_4x4()
[all …]
|
| D | 3x3p1-minmax-ssse3-3x4.c | 114 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
115 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
116 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
117 …const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
118 …const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
136 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
137 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
138 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
139 …const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
140 …const __m128 vi4x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_3x4()
[all …]
|
| D | 3x3p1-minmax-ssse3-5x4.c | 142 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
143 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
144 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
145 …const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
146 …const __m128 vi4x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi4x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
147 …const __m128 vi5x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi5x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
148 …const __m128 vi6x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi6x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
174 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
175 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
176 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_5x4()
[all …]
|
| D | 3x3p1-minmax-ssse3-1x4-acc2.c | 86 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
87 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
88 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
98 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
99 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
100 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
131 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
132 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
133 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
140 … const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc2()
[all …]
|
| D | 3x3p1-minmax-ssse3-1x4-acc3.c | 86 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
87 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
88 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
98 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
99 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
100 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
132 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
133 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
134 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
141 … const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc3()
[all …]
|
| D | 3x3p1-minmax-ssse3-1x4-acc4.c | 86 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
87 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
88 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
98 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
99 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
100 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
133 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
134 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
135 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
142 … const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(vzero, _mm_castps_si128(vi0x4567), 4)); in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_1x4_acc4()
[all …]
|
| D | 3x3p1-minmax-ssse3-2x4-acc2.c | 100 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
101 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
102 …const __m128 vi2x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
103 …const __m128 vi3x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
117 …const __m128 vi0x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
118 …const __m128 vi1x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
119 …const __m128 vi2x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi2x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
120 …const __m128 vi3x5678 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi3x89AB), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
164 …const __m128 vi0x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi0x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
165 …const __m128 vi1x3456 = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(vi1x4567), _mm_castps_si… in xnn_f32_dwconv2d_chw_ukernel_3x3p1__ssse3_2x4_acc2()
[all …]
|
| /external/XNNPACK/src/f32-vrnd/gen/ |
| D | vrndu-sse2-x8.c | 38 …const __m128 vrndmask0123 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx0123, vmagi… in xnn_f32_vrndu_ukernel__sse2_x8()
39 …const __m128 vrndmask4567 = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx4567, vmagi… in xnn_f32_vrndu_ukernel__sse2_x8()
47 … const __m128 vadjmask0123 = _mm_or_ps(_mm_cmpge_ps(vrndx0123, vx0123), _mm_castsi128_ps(vmagic)); in xnn_f32_vrndu_ukernel__sse2_x8()
48 … const __m128 vadjmask4567 = _mm_or_ps(_mm_cmpge_ps(vrndx4567, vx4567), _mm_castsi128_ps(vmagic)); in xnn_f32_vrndu_ukernel__sse2_x8()
65 const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); in xnn_f32_vrndu_ukernel__sse2_x8()
68 const __m128 vadjmask = _mm_or_ps(_mm_cmpge_ps(vrndx, vx), _mm_castsi128_ps(vmagic)); in xnn_f32_vrndu_ukernel__sse2_x8()
78 const __m128 vrndmask = _mm_castsi128_ps(_mm_or_si128(vmagic, _mm_cmpeq_epi32(vintx, vmagic))); in xnn_f32_vrndu_ukernel__sse2_x8()
81 const __m128 vadjmask = _mm_or_ps(_mm_cmpge_ps(vrndx, vx), _mm_castsi128_ps(vmagic)); in xnn_f32_vrndu_ukernel__sse2_x8()
|
| /external/webrtc/common_audio/third_party/ooura/fft_size_128/ |
| D | ooura_fft_sse2.cc | 27 static __inline __m128 _mm_castsi128_ps(__m128i a) { in _mm_castsi128_ps() function
112 _mm_shuffle_ps(_mm_castsi128_ps(a_00), _mm_castsi128_ps(a_32), in cftmdl_128_SSE2()
115 _mm_shuffle_ps(_mm_castsi128_ps(a_08), _mm_castsi128_ps(a_40), in cftmdl_128_SSE2()
125 _mm_shuffle_ps(_mm_castsi128_ps(a_16), _mm_castsi128_ps(a_48), in cftmdl_128_SSE2()
128 _mm_shuffle_ps(_mm_castsi128_ps(a_24), _mm_castsi128_ps(a_56), in cftmdl_128_SSE2()
136 const __m128 x3i0_3r0_3i1_x3r1 = _mm_castsi128_ps(_mm_shuffle_epi32( in cftmdl_128_SSE2()
186 _mm_shuffle_ps(_mm_castsi128_ps(a_00), _mm_castsi128_ps(a_32), in cftmdl_128_SSE2()
189 _mm_shuffle_ps(_mm_castsi128_ps(a_08), _mm_castsi128_ps(a_40), in cftmdl_128_SSE2()
199 _mm_shuffle_ps(_mm_castsi128_ps(a_16), _mm_castsi128_ps(a_48), in cftmdl_128_SSE2()
202 _mm_shuffle_ps(_mm_castsi128_ps(a_24), _mm_castsi128_ps(a_56), in cftmdl_128_SSE2()
[all …]
|
| /external/XNNPACK/src/qu8-requantization/ |
| D | precise-ssse3.c | 83 …_mm_shuffle_ps(_mm_castsi128_ps(x_abs_scaled02), _mm_castsi128_ps(x_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__ssse3()
85 …_mm_shuffle_ps(_mm_castsi128_ps(y_abs_scaled02), _mm_castsi128_ps(y_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__ssse3()
87 …_mm_shuffle_ps(_mm_castsi128_ps(z_abs_scaled02), _mm_castsi128_ps(z_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__ssse3()
89 …_mm_shuffle_ps(_mm_castsi128_ps(w_abs_scaled02), _mm_castsi128_ps(w_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__ssse3()
|
| D | precise-sse2.c | 88 …_mm_shuffle_ps(_mm_castsi128_ps(x_abs_scaled02), _mm_castsi128_ps(x_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__sse2()
90 …_mm_shuffle_ps(_mm_castsi128_ps(y_abs_scaled02), _mm_castsi128_ps(y_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__sse2()
92 …_mm_shuffle_ps(_mm_castsi128_ps(z_abs_scaled02), _mm_castsi128_ps(z_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__sse2()
94 …_mm_shuffle_ps(_mm_castsi128_ps(w_abs_scaled02), _mm_castsi128_ps(w_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qu8_requantize_precise__sse2()
|
| /external/XNNPACK/src/qs8-requantization/ |
| D | precise-ssse3.c | 83 …_mm_shuffle_ps(_mm_castsi128_ps(x_abs_scaled02), _mm_castsi128_ps(x_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__ssse3()
85 …_mm_shuffle_ps(_mm_castsi128_ps(y_abs_scaled02), _mm_castsi128_ps(y_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__ssse3()
87 …_mm_shuffle_ps(_mm_castsi128_ps(z_abs_scaled02), _mm_castsi128_ps(z_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__ssse3()
89 …_mm_shuffle_ps(_mm_castsi128_ps(w_abs_scaled02), _mm_castsi128_ps(w_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__ssse3()
|
| D | precise-sse2.c | 88 …_mm_shuffle_ps(_mm_castsi128_ps(x_abs_scaled02), _mm_castsi128_ps(x_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__sse2()
90 …_mm_shuffle_ps(_mm_castsi128_ps(y_abs_scaled02), _mm_castsi128_ps(y_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__sse2()
92 …_mm_shuffle_ps(_mm_castsi128_ps(z_abs_scaled02), _mm_castsi128_ps(z_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__sse2()
94 …_mm_shuffle_ps(_mm_castsi128_ps(w_abs_scaled02), _mm_castsi128_ps(w_abs_scaled13), _MM_SHUFFLE(2, … in xnn_qs8_requantize_precise__sse2()
|
| /external/XNNPACK/src/f32-sigmoid/gen/ |
| D | avx-rr2-p5-div-x80.c | 76 …const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
77 …const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
79 …const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
80 …const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
82 …const __m128 vs_lo2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
83 …const __m128 vs_hi2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
85 …const __m128 vs_lo3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
86 …const __m128 vs_hi3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
88 …const __m128 vs_lo4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
89 …const __m128 vs_hi4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn4, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x80()
[all …]
|
| D | avx-rr2-p5-div-x72.c | 73 …const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
74 …const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
76 …const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
77 …const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
79 …const __m128 vs_lo2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
80 …const __m128 vs_hi2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
82 …const __m128 vs_lo3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
83 …const __m128 vs_hi3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
85 …const __m128 vs_lo4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
86 …const __m128 vs_hi4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn4, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x72()
[all …]
|
| D | avx-rr2-p5-div-x64.c | 70 …const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
71 …const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
73 …const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
74 …const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
76 …const __m128 vs_lo2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
77 …const __m128 vs_hi2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
79 …const __m128 vs_lo3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
80 …const __m128 vs_hi3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
82 …const __m128 vs_lo4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
83 …const __m128 vs_hi4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn4, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x64()
[all …]
|
| D | avx-rr2-p5-div-x56.c | 67 …const __m128 vs_lo0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
68 …const __m128 vs_hi0 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
70 …const __m128 vs_lo1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
71 …const __m128 vs_hi1 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
73 …const __m128 vs_lo2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
74 …const __m128 vs_hi2 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
76 …const __m128 vs_lo3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
77 …const __m128 vs_hi3 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
79 …const __m128 vs_lo4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)… in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
80 …const __m128 vs_hi4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn4, … in xnn_f32_sigmoid_ukernel__avx_rr2_p5_div_x56()
[all …]
|
| /external/XNNPACK/src/f32-prelu/gen/ |
| D | sse2-2x8.c | 59 …const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
61 …const __m128 vmask0x4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
63 …const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
65 …const __m128 vmask1x4567 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
89 …const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
91 …const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
111 …const __m128 vmask0x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
113 …const __m128 vmask1x0123 = _mm_castsi128_ps(_mm_cmpgt_epi32(_mm_setzero_si128(), _mm_castps_si128(… in xnn_f32_prelu_ukernel__sse2_2x8()
|