| /external/XNNPACK/src/f32-velu/gen/ |
| D | velu-avx-rr2-lut16-p3-x48.c | 84 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
86 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
87 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
88 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
89 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
90 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
113 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
115 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
116 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
117 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x48()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x32.c | 78 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
80 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
81 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
82 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
83 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
84 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
107 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
109 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
110 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
111 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x32()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x40.c | 81 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
83 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
84 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
85 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
86 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
87 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
110 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
112 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
113 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
114 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x40()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x24.c | 75 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
77 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
78 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
79 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
80 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
81 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
104 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
106 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
107 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
108 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x24()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x16.c | 72 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
74 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
75 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
76 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
77 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
78 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx0_hi, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
101 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
103 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_hi, 2)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
104 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 1)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
105 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx1_lo, 3)… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x16()
[all …]
|
| D | velu-avx-rr2-lut16-p3-x8.c | 69 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
71 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
72 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
73 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
74 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
75 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
135 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
137 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
138 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
139 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))… in xnn_f32_velu_ukernel__avx_rr2_lut16_p3_x8()
[all …]
|
| /external/mesa3d/src/gallium/drivers/swr/rasterizer/common/ |
| D | simdlib_256_avx.inl | 248 aHi = _mm_extract_epi32(vAHi, 0);
249 countHi = _mm_extract_epi32(vCountHi, 0);
253 aLow = _mm_extract_epi32(vALow, 0);
254 countLow = _mm_extract_epi32(vCountLow, 0);
258 aHi = _mm_extract_epi32(vAHi, 1);
259 countHi = _mm_extract_epi32(vCountHi, 1);
263 aLow = _mm_extract_epi32(vALow, 1);
264 countLow = _mm_extract_epi32(vCountLow, 1);
268 aHi = _mm_extract_epi32(vAHi, 2);
269 countHi = _mm_extract_epi32(vCountHi, 2);
[all …]
|
| D | simdlib_128_avx.inl | 166 a = _mm_extract_epi32(vA, 0);
167 count = _mm_extract_epi32(vB, 0);
171 a = _mm_extract_epi32(vA, 1);
172 count = _mm_extract_epi32(vB, 1);
176 a = _mm_extract_epi32(vA, 2);
177 count = _mm_extract_epi32(vB, 2);
181 a = _mm_extract_epi32(vA, 3);
182 count = _mm_extract_epi32(vB, 3);
207 a = _mm_extract_epi32(vA, 0);
208 count = _mm_extract_epi32(vB, 0);
[all …]
|
| /external/XNNPACK/src/math/ |
| D | expm1minus-f32-avx-rr2-lut16-p3.c | 90 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2))… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
92 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2))… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
93 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
94 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
95 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
96 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))… in xnn_math_f32_expm1minus__avx_rr2_lut16_p3()
|
| D | sigmoid-f32-avx-rr2-lut64-p2-div.c | 94 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_lo, 2))… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
96 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_hi, 2))… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
97 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
98 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
99 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_hi, 1))… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
100 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_64 + (uint32_t) _mm_extract_epi32(vidx_hi, 3))… in xnn_math_f32_sigmoid__avx_rr2_lut64_p2_div()
|
| /external/XNNPACK/src/f32-velu/ |
| D | avx-rr2-lut16-p3.c.in | 74 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_lo,…
76 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_hi,…
77 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_lo,…
78 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_lo,…
79 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_hi,…
80 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx${N}_hi,…
151 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 2))…
153 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_hi, 2))…
154 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 1))…
155 …const int*) ((uintptr_t) xnn_table_exp2minus_k_over_16 + (uint32_t) _mm_extract_epi32(vidx_lo, 3))…
[all …]
|
| /external/libaom/aom_dsp/flow_estimation/x86/ |
| D | corner_match_sse4.c | 95 int sum1 = _mm_extract_epi32(res, 0); in av1_compute_cross_correlation_sse4_1()
96 int sum2 = _mm_extract_epi32(res, 1); in av1_compute_cross_correlation_sse4_1()
97 int sumsq2 = _mm_extract_epi32(res, 2); in av1_compute_cross_correlation_sse4_1()
98 int cross = _mm_extract_epi32(res, 3); in av1_compute_cross_correlation_sse4_1()
|
| /external/libaom/aom_dsp/x86/ |
| D | highbd_variance_avx2.c | 83 sum_long += _mm_extract_epi32(v_d, 0); in aom_highbd_var_filter_block2d_bil_avx2()
84 sse_long += _mm_extract_epi32(v_d, 1); in aom_highbd_var_filter_block2d_bil_avx2()
132 sum_long += _mm_extract_epi32(v_d, 0); in aom_highbd_var_filter_block2d_bil_avx2()
133 sse_long += _mm_extract_epi32(v_d, 1); in aom_highbd_var_filter_block2d_bil_avx2()
196 sum_long += _mm_extract_epi32(v_d, 0); in aom_highbd_var_filter_block2d_bil_avx2()
197 sse_long += _mm_extract_epi32(v_d, 1); in aom_highbd_var_filter_block2d_bil_avx2()
267 sum_long += _mm_extract_epi32(v_d, 0); in aom_highbd_var_filter_block2d_bil_avx2()
268 sse_long += _mm_extract_epi32(v_d, 1); in aom_highbd_var_filter_block2d_bil_avx2()
319 sum_long += _mm_extract_epi32(v_d, 0); in aom_highbd_var_filter_block2d_bil_avx2()
320 sse_long += _mm_extract_epi32(v_d, 1); in aom_highbd_var_filter_block2d_bil_avx2()
[all …]
|
| D | avg_intrin_sse4.c | 54 const int mean_int = _mm_extract_epi32(result, 0); in aom_vector_var_sse4_1()
55 const int sse_int = _mm_extract_epi32(result, 2); in aom_vector_var_sse4_1()
|
| /external/libavc/encoder/x86/ |
| D | ime_distortion_metrics_sse42.c | 205 val1 = _mm_extract_epi32(sad_val,0); in ime_compute_sad_16x16_sse42()
206 val2 = _mm_extract_epi32(sad_val, 2); in ime_compute_sad_16x16_sse42()
303 val1 = _mm_extract_epi32(sad_val,0); in ime_compute_sad_16x8_sse42()
304 val2 = _mm_extract_epi32(sad_val, 2); in ime_compute_sad_16x8_sse42()
404 val1 = _mm_extract_epi32(sad_val, 0); in ime_compute_sad_16x16_ea8_sse42()
405 val2 = _mm_extract_epi32(sad_val, 2); in ime_compute_sad_16x16_ea8_sse42()
457 val1 = _mm_extract_epi32(sad_val, 0); in ime_compute_sad_16x16_ea8_sse42()
458 val2 = _mm_extract_epi32(sad_val, 2); in ime_compute_sad_16x16_ea8_sse42()
560 val1 = _mm_extract_epi32(sad_val, 0); in ime_compute_sad_16x16_fast_sse42()
561 val2 = _mm_extract_epi32(sad_val, 2); in ime_compute_sad_16x16_fast_sse42()
[all …]
|
| /external/libaom/av1/common/x86/ |
| D | selfguided_sse4.c | 220 _mm_set_epi32(av1_x_by_xplus1[_mm_extract_epi32(z, 3)], in calc_ab()
221 av1_x_by_xplus1[_mm_extract_epi32(z, 2)], in calc_ab()
222 av1_x_by_xplus1[_mm_extract_epi32(z, 1)], in calc_ab()
223 av1_x_by_xplus1[_mm_extract_epi32(z, 0)]); in calc_ab()
361 _mm_set_epi32(av1_x_by_xplus1[_mm_extract_epi32(z, 3)], in calc_ab_fast()
362 av1_x_by_xplus1[_mm_extract_epi32(z, 2)], in calc_ab_fast()
363 av1_x_by_xplus1[_mm_extract_epi32(z, 1)], in calc_ab_fast()
364 av1_x_by_xplus1[_mm_extract_epi32(z, 0)]); in calc_ab_fast()
|
| /external/tensorflow/tensorflow/compiler/xla/pjrt/ |
| D | transpose_kernels.h | 65 *reinterpret_cast<uint32_t*>(b + ldb * 0) = _mm_extract_epi32(x, 0);
66 *reinterpret_cast<uint32_t*>(b + ldb * 1) = _mm_extract_epi32(x, 1);
67 *reinterpret_cast<uint32_t*>(b + ldb * 2) = _mm_extract_epi32(x, 2);
68 *reinterpret_cast<uint32_t*>(b + ldb * 3) = _mm_extract_epi32(x, 3);
|
| /external/XNNPACK/src/qc8-igemm/gen/ |
| D | 4x4c2s4-minmax-fp32-sse41-ld128.c | 179 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
181 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
183 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
|
| D | 4x4c2s4-minmax-fp32-sse41-ld64.c | 181 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld64()
183 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld64()
185 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld64()
|
| D | 4x4c2s4-minmax-fp32-xop-ld128.c | 184 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__xop_ld128()
186 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__xop_ld128()
188 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qc8_igemm_minmax_fp32_ukernel_4x4c2s4__xop_ld128()
|
| /external/XNNPACK/src/qs8-gemm/gen/ |
| D | 4x4c2s4-minmax-fp32-sse41-ld128.c | 161 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
162 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
163 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
|
| D | 4x4c2s4-minmax-fp32-xop-ld64.c | 168 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__xop_ld64()
169 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__xop_ld64()
170 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qs8_gemm_minmax_fp32_ukernel_4x4c2s4__xop_ld64()
|
| /external/XNNPACK/src/qu8-igemm/gen/ |
| D | 4x4c2s4-minmax-fp32-avx-ld128.c | 179 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2s4__avx_ld128()
181 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2s4__avx_ld128()
183 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qu8_igemm_minmax_fp32_ukernel_4x4c2s4__avx_ld128()
|
| /external/XNNPACK/src/qc8-gemm/gen/ |
| D | 4x4c2s4-minmax-fp32-sse41-ld128.c | 162 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
163 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
164 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2s4__sse41_ld128()
|
| /external/XNNPACK/src/qu8-gemm/gen/ |
| D | 4x4c2s4-minmax-fp32-xop-ld64.c | 168 unaligned_store_u32(c1, (uint32_t) _mm_extract_epi32(vout, 1)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__xop_ld64()
169 unaligned_store_u32(c2, (uint32_t) _mm_extract_epi32(vout, 2)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__xop_ld64()
170 unaligned_store_u32(c3, (uint32_t) _mm_extract_epi32(vout, 3)); in xnn_qu8_gemm_minmax_fp32_ukernel_4x4c2s4__xop_ld64()
|