Home
last modified time | relevance | path

Searched refs:_mm_unpackhi_epi32 (Results 1 – 25 of 133) sorted by relevance

123456

/external/libaom/libaom/aom_dsp/x86/
Dtranspose_sse2.h73 const __m128i c1 = _mm_unpackhi_epi32(b0, b2); in transpose_8bit_8x8()
75 const __m128i c3 = _mm_unpackhi_epi32(b1, b3); in transpose_8bit_8x8()
116 out[2] = _mm_unpackhi_epi32(a0, a1); in transpose_16bit_4x4()
148 const __m128i b2 = _mm_unpackhi_epi32(a0, a1); in transpose_16bit_4x8()
149 const __m128i b3 = _mm_unpackhi_epi32(a2, a3); in transpose_16bit_4x8()
187 const __m128i b4 = _mm_unpackhi_epi32(a0, a1); in transpose_16bit_8x4()
188 const __m128i b6 = _mm_unpackhi_epi32(a4, a5); in transpose_16bit_8x4()
252 const __m128i b4 = _mm_unpackhi_epi32(a0, a1); in transpose_16bit_8x8()
253 const __m128i b5 = _mm_unpackhi_epi32(a2, a3); in transpose_16bit_8x8()
254 const __m128i b6 = _mm_unpackhi_epi32(a4, a5); in transpose_16bit_8x8()
[all …]
Dlpf_common_sse2.h43 ww0 = _mm_unpackhi_epi32(w0, w1); // 02 12 22 32 03 13 23 33 in highbd_transpose6x6_sse2()
70 ww1 = _mm_unpackhi_epi32(w0, w1); // 02 12 22 32 03 13 23 33 in highbd_transpose4x8_8x4_low_sse2()
89 ww3 = _mm_unpackhi_epi32(w0, w1); // 06 16 26 36 07 17 27 37 in highbd_transpose4x8_8x4_high_sse2()
150 ww0 = _mm_unpackhi_epi32(w0, w1); // 02 12 22 32 03 13 23 33 in highbd_transpose8x8_low_sse2()
151 ww1 = _mm_unpackhi_epi32(w2, w3); // 42 52 62 72 43 53 63 73 in highbd_transpose8x8_low_sse2()
183 ww0 = _mm_unpackhi_epi32(w0, w1); // 06 16 26 36 07 17 27 37 in highbd_transpose8x8_high_sse2()
184 ww1 = _mm_unpackhi_epi32(w2, w3); // 46 56 66 76 47 57 67 77 in highbd_transpose8x8_high_sse2()
340 *d2 = _mm_unpackhi_epi32( in transpose8x8_low_sse2()
378 *d2d3 = _mm_unpackhi_epi32( in transpose8x8_sse2()
388 *d6d7 = _mm_unpackhi_epi32( in transpose8x8_sse2()
[all …]
Daom_subpixel_8t_intrin_sse2.c38 const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg); in aom_filter_block1d16_h4_sse2()
63 __m128i res_hi = _mm_unpackhi_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2); in aom_filter_block1d16_h4_sse2()
87 res_hi = _mm_unpackhi_epi32(srcRegFilt32b2_1, srcRegFilt32b2_2); in aom_filter_block1d16_h4_sse2()
132 const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg); in aom_filter_block1d16_v4_sse2()
276 const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg); in aom_filter_block1d8_h4_sse2()
301 __m128i res_hi = _mm_unpackhi_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2); in aom_filter_block1d8_h4_sse2()
343 const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg); in aom_filter_block1d8_v4_sse2()
443 const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg); in aom_filter_block1d4_h4_sse2()
507 const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg); in aom_filter_block1d4_v4_sse2()
/external/libvpx/libvpx/vpx_dsp/x86/
Dtranspose_sse2.h72 const __m128i c1 = _mm_unpackhi_epi32(b0, b2); in transpose_8bit_8x8()
74 const __m128i c3 = _mm_unpackhi_epi32(b1, b3); in transpose_8bit_8x8()
112 out[1] = _mm_unpackhi_epi32(a0, a1); in transpose_16bit_4x4()
143 const __m128i b2 = _mm_unpackhi_epi32(a0, a1); in transpose_16bit_4x8()
144 const __m128i b3 = _mm_unpackhi_epi32(a2, a3); in transpose_16bit_4x8()
199 const __m128i b4 = _mm_unpackhi_epi32(a0, a1); in transpose_16bit_8x8()
200 const __m128i b5 = _mm_unpackhi_epi32(a2, a3); in transpose_16bit_8x8()
201 const __m128i b6 = _mm_unpackhi_epi32(a4, a5); in transpose_16bit_8x8()
202 const __m128i b7 = _mm_unpackhi_epi32(a6, a7); in transpose_16bit_8x8()
257 const __m128i a2 = _mm_unpackhi_epi32(in[0], in[1]); in transpose_32bit_4x4()
[all …]
Dfwd_dct32x32_impl_sse2.h1680 u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]); in FDCT32x32_2D()
1682 u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]); in FDCT32x32_2D()
1723 u[1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]); in FDCT32x32_2D()
1725 u[3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]); in FDCT32x32_2D()
1727 u[5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]); in FDCT32x32_2D()
1729 u[7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]); in FDCT32x32_2D()
1731 u[9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]); in FDCT32x32_2D()
1733 u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]); in FDCT32x32_2D()
1735 u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]); in FDCT32x32_2D()
1737 u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]); in FDCT32x32_2D()
[all …]
Dconvolve_sse2.h28 __m128i tmp = _mm_unpackhi_epi32(*reg, *reg); in extract_quarter_3_epi16_sse2()
70 const __m128i tmp_2 = _mm_unpackhi_epi32(*src_1, *src_2); in mm_zip_epi32_sse2()
/external/libgav1/libgav1/src/dsp/x86/
Dtranspose_sse4.h115 out[1] = _mm_unpackhi_epi32(b0, b1); in Transpose8x8To4x16_U8()
117 out[3] = _mm_unpackhi_epi32(b2, b3); in Transpose8x8To4x16_U8()
135 const __m128i dcba_hi = _mm_unpackhi_epi32(ba, dc); in Transpose4x4_U16()
175 const __m128i b2 = _mm_unpackhi_epi32(a0, a1); in Transpose4x8To8x4_U16()
176 const __m128i b3 = _mm_unpackhi_epi32(a2, a3); in Transpose4x8To8x4_U16()
214 const __m128i b4 = _mm_unpackhi_epi32(a0, a1); in Transpose8x4To4x8_U16()
215 const __m128i b6 = _mm_unpackhi_epi32(a4, a5); in Transpose8x4To4x8_U16()
279 const __m128i b4 = _mm_unpackhi_epi32(a0, a1); in Transpose8x8_U16()
280 const __m128i b5 = _mm_unpackhi_epi32(a2, a3); in Transpose8x8_U16()
281 const __m128i b6 = _mm_unpackhi_epi32(a4, a5); in Transpose8x8_U16()
[all …]
/external/python/cpython3/Modules/_blake2/impl/
Dblake2s-load-sse41.h48 t0 = _mm_unpackhi_epi32(m0,m1); \
54 t0 = _mm_unpackhi_epi32(m2,m3); \
78 t0 = _mm_unpackhi_epi32(m0,m1); \
79 t1 = _mm_unpackhi_epi32(t0, m2); \
96 t1 = _mm_unpackhi_epi32(m1,m2); \
122 t0 = _mm_unpackhi_epi32(m0,m1); \
159 t0 = _mm_unpackhi_epi32(m1,m2); \
164 t0 = _mm_unpackhi_epi32(m0,m1); \
182 t1 = _mm_unpackhi_epi32(m1,m2); \
186 t0 = _mm_unpackhi_epi32(m1,m3); \
[all …]
/external/XNNPACK/src/qs8-gemm/gen/
D3x4c8-minmax-sse2-ld64.c111 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
112 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
113 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
114 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
115 …const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
116 …const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
118 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
119 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
120 …__m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld64()
D3x4c8-minmax-sse2-ld128.c109 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
110 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
111 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
112 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
113 …const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
114 …const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
116 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
117 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
118 …__m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_3x4c8__sse2_ld128()
D3x4c8-xw-minmax-sse2.c107 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
108 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
109 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
110 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
111 …const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
112 …const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
114 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
115 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
116 …__m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_xw_minmax_ukernel_3x4c8__sse2()
D2x4c8-xw-minmax-sse2.c90 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
91 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
92 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
93 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
95 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
96 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_xw_minmax_ukernel_2x4c8__sse2()
D2x4c8-minmax-sse2-ld128.c92 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
93 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
94 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
95 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
97 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
98 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld128()
D2x4c8-minmax-sse2-ld64.c94 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
95 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
96 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
97 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
99 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
100 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_gemm_minmax_ukernel_2x4c8__sse2_ld64()
/external/XNNPACK/src/qs8-igemm/gen/
D3x4c8-minmax-sse2-ld128.c126 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
127 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
128 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
129 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
130 …const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
131 …const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
133 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
134 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
135 …__m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld128()
D3x4c8-minmax-sse2-ld64.c128 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
129 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
130 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
131 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
132 …const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
133 …const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
135 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
136 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
137 …__m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_3x4c8__sse2_ld64()
D2x4c8-minmax-sse2-ld128.c107 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
108 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
109 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
110 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
112 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
113 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld128()
D2x4c8-minmax-sse2-ld64.c109 …const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64()
110 …const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64()
111 …const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64()
112 …const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(va… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64()
114 …__m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64()
115 …__m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc… in xnn_qs8_igemm_minmax_ukernel_2x4c8__sse2_ld64()
/external/libavc/common/x86/
Dih264_ihadamard_scaling_sse42.c127 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3 in ih264_ihadamard_scaling_4x4_sse42()
128 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3 in ih264_ihadamard_scaling_4x4_sse42()
156 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d1 in ih264_ihadamard_scaling_4x4_sse42()
157 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d3 in ih264_ihadamard_scaling_4x4_sse42()
226 …temp1 = _mm_unpackhi_epi32(plane_0, plane_1); //a0-a1+a2-a3 a0-a1-a2+a3 b0-b1+b2-b3 b0-b1-b2… in ih264_ihadamard_scaling_2x2_uv_sse42()
Dih264_ihadamard_scaling_ssse3.c129 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3 in ih264_ihadamard_scaling_4x4_ssse3()
130 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3 in ih264_ihadamard_scaling_4x4_ssse3()
158 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d1 in ih264_ihadamard_scaling_4x4_ssse3()
159 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d3 in ih264_ihadamard_scaling_4x4_ssse3()
Dih264_iquant_itrans_recon_sse42.c185 temp2 = _mm_unpackhi_epi32(resq_r0, resq_r1); //a2 b2 a3 b3 in ih264_iquant_itrans_recon_4x4_sse42()
186 temp4 = _mm_unpackhi_epi32(resq_r2, resq_r3); //c2 d2 c3 d3 in ih264_iquant_itrans_recon_4x4_sse42()
221 temp2 = _mm_unpackhi_epi32(resq_r0, resq_r1); //c0 c1 d0 d1 in ih264_iquant_itrans_recon_4x4_sse42()
222 temp4 = _mm_unpackhi_epi32(resq_r2, resq_r3); //c2 c3 d2 d3 in ih264_iquant_itrans_recon_4x4_sse42()
435 temp2 = _mm_unpackhi_epi32(resq_r0, resq_r1); //a2 b2 a3 b3 in ih264_iquant_itrans_recon_chroma_4x4_sse42()
436 temp4 = _mm_unpackhi_epi32(resq_r2, resq_r3); //c2 d2 c3 d3 in ih264_iquant_itrans_recon_chroma_4x4_sse42()
471 temp2 = _mm_unpackhi_epi32(resq_r0, resq_r1); //c0 c1 d0 d1 in ih264_iquant_itrans_recon_chroma_4x4_sse42()
472 temp4 = _mm_unpackhi_epi32(resq_r2, resq_r3); //c2 c3 d2 d3 in ih264_iquant_itrans_recon_chroma_4x4_sse42()
Dih264_resi_trans_quant_sse42.c168 temp3 = _mm_unpackhi_epi32(temp0, temp2); //a2 b2 c2 d2 a3 b3 c3 d3 in ih264_resi_trans_quant_4x4_sse42()
206 temp3 = _mm_unpackhi_epi32(temp0, temp2); //c0 c1 c2 c3 d0 d1 d2 d3 in ih264_resi_trans_quant_4x4_sse42()
451 temp3 = _mm_unpackhi_epi32(temp0, temp2); //a2 b2 c2 d2 a3 b3 c3 d3 in ih264_resi_trans_quant_chroma_4x4_sse42()
489 temp3 = _mm_unpackhi_epi32(temp0, temp2); //c0 c1 c2 c3 d0 d1 d2 d3 in ih264_resi_trans_quant_chroma_4x4_sse42()
708 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //a2 b2 a3 b3 in ih264_hadamard_quant_4x4_sse42()
709 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 d2 c3 d3 in ih264_hadamard_quant_4x4_sse42()
737 temp1 = _mm_unpackhi_epi32(src_r0, src_r1); //c0 c1 d0 d1 in ih264_hadamard_quant_4x4_sse42()
738 temp3 = _mm_unpackhi_epi32(src_r2, src_r3); //c2 c3 d2 d3 in ih264_hadamard_quant_4x4_sse42()
925 …temp1 = _mm_unpackhi_epi32(plane_0, plane_1); //a0-a1+a2-a3 a0-a1-a2+a3 b0-b1+b2-b3 b0-b1-b2… in ih264_hadamard_quant_2x2_uv_sse42()
/external/libaom/libaom/av1/encoder/x86/
Dav1_fwd_txfm2d_sse4.c168 __m128i temp1 = _mm_unpackhi_epi32(inputA[0], inputA[2]); in transpose_32_4x4x2()
170 __m128i temp3 = _mm_unpackhi_epi32(inputA[1], inputA[3]); in transpose_32_4x4x2()
173 output[1 * stride] = _mm_unpackhi_epi32(temp0, temp2); in transpose_32_4x4x2()
175 output[3 * stride] = _mm_unpackhi_epi32(temp1, temp3); in transpose_32_4x4x2()
178 temp1 = _mm_unpackhi_epi32(inputB[0], inputB[2]); in transpose_32_4x4x2()
180 temp3 = _mm_unpackhi_epi32(inputB[1], inputB[3]); in transpose_32_4x4x2()
183 output[5 * stride] = _mm_unpackhi_epi32(temp0, temp2); in transpose_32_4x4x2()
185 output[7 * stride] = _mm_unpackhi_epi32(temp1, temp3); in transpose_32_4x4x2()
Dav1_txfm1d_sse4.h64 __m128i temp1 = _mm_unpackhi_epi32(input[0 * stride], input[2 * stride]); in transpose_32_4x4()
66 __m128i temp3 = _mm_unpackhi_epi32(input[1 * stride], input[3 * stride]); in transpose_32_4x4()
69 output[1 * stride] = _mm_unpackhi_epi32(temp0, temp2); in transpose_32_4x4()
71 output[3 * stride] = _mm_unpackhi_epi32(temp1, temp3); in transpose_32_4x4()
/external/libhevc/common/x86/
Dihevc_itrans_recon_sse42_intr.c250 m_temp_reg_21 = _mm_unpackhi_epi32(m_temp_reg_24, m_temp_reg_25); in ihevc_itrans_recon_4x4_ttype1_sse42()
341 m_temp_reg_21 = _mm_unpackhi_epi32(m_temp_reg_24, m_temp_reg_25); in ihevc_itrans_recon_4x4_ttype1_sse42()
576 m_temp_reg_21 = _mm_unpackhi_epi32(m_temp_reg_24, m_temp_reg_25); in ihevc_itrans_recon_4x4_sse42()
668 m_temp_reg_21 = _mm_unpackhi_epi32(m_temp_reg_24, m_temp_reg_25); in ihevc_itrans_recon_4x4_sse42()
1064 m_temp_reg_1 = _mm_unpackhi_epi32(m_temp_reg_10, m_temp_reg_11); in ihevc_itrans_recon_8x8_sse42()
1070 m_temp_reg_5 = _mm_unpackhi_epi32(m_temp_reg_12, m_temp_reg_13); in ihevc_itrans_recon_8x8_sse42()
1282 m_temp_reg_1 = _mm_unpackhi_epi32(m_temp_reg_10, m_temp_reg_11); in ihevc_itrans_recon_8x8_sse42()
1287 m_temp_reg_5 = _mm_unpackhi_epi32(m_temp_reg_12, m_temp_reg_13); in ihevc_itrans_recon_8x8_sse42()
1524 m_temp_reg_1 = _mm_unpackhi_epi32(m_temp_reg_10, m_temp_reg_11); in ihevc_itrans_recon_8x8_sse42()
1526 m_temp_reg_3 = _mm_unpackhi_epi32(m_temp_reg_14, m_temp_reg_15); in ihevc_itrans_recon_8x8_sse42()
[all …]

123456