| /external/python/cpython3/Modules/_blake2/impl/ |
| D | blake2b-load-sse41.h | 28 b0 = _mm_unpackhi_epi64(m0, m1); \
29 b1 = _mm_unpackhi_epi64(m2, m3); \
44 b0 = _mm_unpackhi_epi64(m4, m5); \
45 b1 = _mm_unpackhi_epi64(m6, m7); \
53 b1 = _mm_unpackhi_epi64(m4, m6); \
69 b1 = _mm_unpackhi_epi64(m5, m2); \
77 b1 = _mm_unpackhi_epi64(m3, m1); \
85 b1 = _mm_unpackhi_epi64(m2, m7); \
101 b1 = _mm_unpackhi_epi64(m3, m4); \
116 b0 = _mm_unpackhi_epi64(m3, m1); \
[all …]
|
| D | blake2b-round.h | 123 row4l = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t0, t0)); \
124 row4h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row4h, row4h)); \
125 row2l = _mm_unpackhi_epi64(row2l, _mm_unpacklo_epi64(row2h, row2h)); \
126 row2h = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(t1, t1))
134 row2l = _mm_unpackhi_epi64(row2h, _mm_unpacklo_epi64(row2l, row2l)); \
135 row2h = _mm_unpackhi_epi64(t0, _mm_unpacklo_epi64(row2h, row2h)); \
136 row4l = _mm_unpackhi_epi64(row4l, _mm_unpacklo_epi64(row4h, row4h)); \
137 row4h = _mm_unpackhi_epi64(row4h, _mm_unpacklo_epi64(t1, t1))
|
| D | blake2s-load-sse41.h | 101 t1 = _mm_unpackhi_epi64(m0,m2); \
106 t0 = _mm_unpackhi_epi64(m1,m3); \
111 t0 = _mm_unpackhi_epi64(m3,m1); \
112 t1 = _mm_unpackhi_epi64(m2,m0); \
160 t1 = _mm_unpackhi_epi64(m0,t0); \
175 t0 = _mm_unpackhi_epi64(m0,m3); \
219 t2 = _mm_unpackhi_epi64(t0,t1); \
|
| /external/libaom/libaom/aom_dsp/x86/ |
| D | transpose_sse2.h | 87 out[1] = _mm_unpackhi_epi64(c0, c0); in transpose_8bit_8x8()
89 out[3] = _mm_unpackhi_epi64(c1, c1); in transpose_8bit_8x8()
91 out[5] = _mm_unpackhi_epi64(c2, c2); in transpose_8bit_8x8()
93 out[7] = _mm_unpackhi_epi64(c3, c3); in transpose_8bit_8x8()
157 out[1] = _mm_unpackhi_epi64(b0, b1); in transpose_16bit_4x8()
159 out[3] = _mm_unpackhi_epi64(b2, b3); in transpose_16bit_4x8()
201 out[1] = _mm_unpackhi_epi64(b0, zeros); in transpose_16bit_8x4()
203 out[3] = _mm_unpackhi_epi64(b4, zeros); in transpose_16bit_8x4()
205 out[5] = _mm_unpackhi_epi64(b2, zeros); in transpose_16bit_8x4()
207 out[7] = _mm_unpackhi_epi64(b6, zeros); in transpose_16bit_8x4()
[all …]
|
| D | adaptive_quantize_sse2.c | 48 threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]); in aom_quantize_b_adaptive_sse2()
70 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC in aom_quantize_b_adaptive_sse2()
86 round = _mm_unpackhi_epi64(round, round); in aom_quantize_b_adaptive_sse2()
87 quant = _mm_unpackhi_epi64(quant, quant); in aom_quantize_b_adaptive_sse2()
88 shift = _mm_unpackhi_epi64(shift, shift); in aom_quantize_b_adaptive_sse2()
89 dequant = _mm_unpackhi_epi64(dequant, dequant); in aom_quantize_b_adaptive_sse2()
93 round = _mm_unpackhi_epi64(round, round); in aom_quantize_b_adaptive_sse2()
94 quant = _mm_unpackhi_epi64(quant, quant); in aom_quantize_b_adaptive_sse2()
95 shift = _mm_unpackhi_epi64(shift, shift); in aom_quantize_b_adaptive_sse2()
111 dequant = _mm_unpackhi_epi64(dequant, dequant); in aom_quantize_b_adaptive_sse2()
[all …]
|
| D | highbd_adaptive_quantize_sse2.c | 125 threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]); in aom_highbd_quantize_b_adaptive_sse2()
162 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC in aom_highbd_quantize_b_adaptive_sse2()
175 round = _mm_unpackhi_epi64(round, round); in aom_highbd_quantize_b_adaptive_sse2()
176 quant = _mm_unpackhi_epi64(quant, quant); in aom_highbd_quantize_b_adaptive_sse2()
177 shift = _mm_unpackhi_epi64(shift, shift); in aom_highbd_quantize_b_adaptive_sse2()
178 dequant = _mm_unpackhi_epi64(dequant, dequant); in aom_highbd_quantize_b_adaptive_sse2()
182 round = _mm_unpackhi_epi64(round, round); in aom_highbd_quantize_b_adaptive_sse2()
183 quant = _mm_unpackhi_epi64(quant, quant); in aom_highbd_quantize_b_adaptive_sse2()
184 shift = _mm_unpackhi_epi64(shift, shift); in aom_highbd_quantize_b_adaptive_sse2()
199 dequant = _mm_unpackhi_epi64(dequant, dequant); in aom_highbd_quantize_b_adaptive_sse2()
[all …]
|
| D | lpf_common_sse2.h | 40 *d1 = _mm_unpackhi_epi64(ww0, in highbd_transpose6x6_sse2()
51 *d3 = _mm_unpackhi_epi64(ww0, _mm_srli_si128(w2, 4)); // 03 13 23 33 43 53 in highbd_transpose6x6_sse2()
55 *d5 = _mm_unpackhi_epi64(ww0, in highbd_transpose6x6_sse2()
73 *d1 = _mm_unpackhi_epi64(ww0, zero); // 01 11 21 31 xx xx xx xx in highbd_transpose4x8_8x4_low_sse2()
75 *d3 = _mm_unpackhi_epi64(ww1, zero); // 03 13 23 33 xx xx xx xx in highbd_transpose4x8_8x4_low_sse2()
92 *d5 = _mm_unpackhi_epi64(ww2, zero); // 05 15 25 35 xx xx xx xx in highbd_transpose4x8_8x4_high_sse2()
94 *d7 = _mm_unpackhi_epi64(ww3, zero); // 07 17 27 37 xx xx xx xx in highbd_transpose4x8_8x4_high_sse2()
148 *d1 = _mm_unpackhi_epi64(ww0, ww1); // 01 11 21 31 41 51 61 71 in highbd_transpose8x8_low_sse2()
154 *d3 = _mm_unpackhi_epi64(ww0, ww1); // 03 13 23 33 43 53 63 73 in highbd_transpose8x8_low_sse2()
181 *d5 = _mm_unpackhi_epi64(ww0, ww1); // 05 15 25 35 45 55 65 75 in highbd_transpose8x8_high_sse2()
[all …]
|
| D | quantize_ssse3.c | 104 zbin = _mm_unpackhi_epi64(zbin, zbin); in aom_quantize_b_64x64_ssse3()
116 round = _mm_unpackhi_epi64(round, round); in aom_quantize_b_64x64_ssse3()
117 quant = _mm_unpackhi_epi64(quant, quant); in aom_quantize_b_64x64_ssse3()
118 shift = _mm_unpackhi_epi64(shift, shift); in aom_quantize_b_64x64_ssse3()
119 dequant = _mm_unpackhi_epi64(dequant, dequant); in aom_quantize_b_64x64_ssse3()
122 round = _mm_unpackhi_epi64(round, round); in aom_quantize_b_64x64_ssse3()
123 quant = _mm_unpackhi_epi64(quant, quant); in aom_quantize_b_64x64_ssse3()
124 shift = _mm_unpackhi_epi64(shift, shift); in aom_quantize_b_64x64_ssse3()
139 dequant = _mm_unpackhi_epi64(dequant, dequant); in aom_quantize_b_64x64_ssse3()
|
| D | quantize_sse2.c | 54 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC in aom_quantize_b_sse2()
59 round = _mm_unpackhi_epi64(round, round); in aom_quantize_b_sse2()
60 quant = _mm_unpackhi_epi64(quant, quant); in aom_quantize_b_sse2()
61 shift = _mm_unpackhi_epi64(shift, shift); in aom_quantize_b_sse2()
77 dequant = _mm_unpackhi_epi64(dequant, dequant); in aom_quantize_b_sse2()
|
| D | intrapred_x86.h | 23 const __m128i high = _mm_unpackhi_epi64(x, x); in dc_sum_16_sse2()
34 const __m128i high = _mm_unpackhi_epi64(x0, x0); in dc_sum_32_sse2()
|
| /external/libvpx/libvpx/vpx_dsp/x86/ |
| D | transpose_sse2.h | 86 out[1] = _mm_unpackhi_epi64(c0, c0); in transpose_8bit_8x8()
88 out[3] = _mm_unpackhi_epi64(c1, c1); in transpose_8bit_8x8()
90 out[5] = _mm_unpackhi_epi64(c2, c2); in transpose_8bit_8x8()
92 out[7] = _mm_unpackhi_epi64(c3, c3); in transpose_8bit_8x8()
152 out[1] = _mm_unpackhi_epi64(b0, b1); in transpose_16bit_4x8()
154 out[3] = _mm_unpackhi_epi64(b2, b3); in transpose_16bit_4x8()
214 out[1] = _mm_unpackhi_epi64(b0, b1); in transpose_16bit_8x8()
216 out[3] = _mm_unpackhi_epi64(b4, b5); in transpose_16bit_8x8()
218 out[5] = _mm_unpackhi_epi64(b2, b3); in transpose_16bit_8x8()
220 out[7] = _mm_unpackhi_epi64(b6, b7); in transpose_16bit_8x8()
[all …]
|
| D | quantize_avx.c | 58 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC in vpx_quantize_b_avx()
72 round = _mm_unpackhi_epi64(round, round); in vpx_quantize_b_avx()
73 quant = _mm_unpackhi_epi64(quant, quant); in vpx_quantize_b_avx()
74 shift = _mm_unpackhi_epi64(shift, shift); in vpx_quantize_b_avx()
75 dequant = _mm_unpackhi_epi64(dequant, dequant); in vpx_quantize_b_avx()
78 round = _mm_unpackhi_epi64(round, round); in vpx_quantize_b_avx()
79 quant = _mm_unpackhi_epi64(quant, quant); in vpx_quantize_b_avx()
80 shift = _mm_unpackhi_epi64(shift, shift); in vpx_quantize_b_avx()
95 dequant = _mm_unpackhi_epi64(dequant, dequant); in vpx_quantize_b_avx()
199 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC. in vpx_quantize_b_32x32_avx()
[all …]
|
| D | quantize_ssse3.c | 51 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC in vpx_quantize_b_ssse3()
55 round = _mm_unpackhi_epi64(round, round); in vpx_quantize_b_ssse3()
56 quant = _mm_unpackhi_epi64(quant, quant); in vpx_quantize_b_ssse3()
57 shift = _mm_unpackhi_epi64(shift, shift); in vpx_quantize_b_ssse3()
72 dequant = _mm_unpackhi_epi64(dequant, dequant); in vpx_quantize_b_ssse3()
167 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC. in vpx_quantize_b_32x32_ssse3()
183 round = _mm_unpackhi_epi64(round, round); in vpx_quantize_b_32x32_ssse3()
184 quant = _mm_unpackhi_epi64(quant, quant); in vpx_quantize_b_32x32_ssse3()
185 shift = _mm_unpackhi_epi64(shift, shift); in vpx_quantize_b_32x32_ssse3()
186 dequant = _mm_unpackhi_epi64(dequant, dequant); in vpx_quantize_b_32x32_ssse3()
[all …]
|
| D | quantize_sse2.c | 55 zbin = _mm_unpackhi_epi64(zbin, zbin); // Switch DC to AC in vpx_quantize_b_sse2()
60 round = _mm_unpackhi_epi64(round, round); in vpx_quantize_b_sse2()
61 quant = _mm_unpackhi_epi64(quant, quant); in vpx_quantize_b_sse2()
62 shift = _mm_unpackhi_epi64(shift, shift); in vpx_quantize_b_sse2()
78 dequant = _mm_unpackhi_epi64(dequant, dequant); in vpx_quantize_b_sse2()
|
| D | inv_txfm_ssse3.h | 44 tmp[1] = _mm_unpackhi_epi64(io[0], io[0]); in idct8x8_12_add_kernel_ssse3()
46 tmp[3] = _mm_unpackhi_epi64(io[1], io[1]); in idct8x8_12_add_kernel_ssse3()
55 step2[6] = _mm_unpackhi_epi64(step2[5], step2[5]); // step2 6 in idct8x8_12_add_kernel_ssse3()
62 step1[2] = _mm_unpackhi_epi64(tmp[1], tmp[0]); // step1 2&1 in idct8x8_12_add_kernel_ssse3()
|
| D | inv_txfm_sse2.h | 55 out[1] = _mm_unpackhi_epi64(tr1_0, tr1_1); in idct8x8_12_transpose_16bit_4x8()
57 out[3] = _mm_unpackhi_epi64(tr1_2, tr1_3); in idct8x8_12_transpose_16bit_4x8()
318 step2[6] = _mm_unpackhi_epi64(step2[5], zero); // step2 6 in idct8x8_12_add_kernel_sse2()
326 step1[2] = _mm_unpackhi_epi64(tmp[1], tmp[0]); // step1 2&1 in idct8x8_12_add_kernel_sse2()
459 step1[13] = _mm_unpackhi_epi64(step2[11], zero); in idct16x16_10_pass1()
460 step1[14] = _mm_unpackhi_epi64(step2[8], zero); in idct16x16_10_pass1()
477 step2[6] = _mm_unpackhi_epi64(step1[4], zero); in idct16x16_10_pass1()
489 step1[12] = _mm_unpackhi_epi64(step1[11], zero); in idct16x16_10_pass1()
490 step1[13] = _mm_unpackhi_epi64(step1[10], zero); in idct16x16_10_pass1()
491 step1[14] = _mm_unpackhi_epi64(step1[9], zero); in idct16x16_10_pass1()
[all …]
|
| /external/libgav1/libgav1/src/dsp/x86/ |
| D | transpose_sse4.h | 184 out[1] = _mm_unpackhi_epi64(b0, b1); in Transpose4x8To8x4_U16()
186 out[3] = _mm_unpackhi_epi64(b2, b3); in Transpose4x8To8x4_U16()
228 out[1] = _mm_unpackhi_epi64(b0, zeros); in Transpose8x4To4x8_U16()
230 out[3] = _mm_unpackhi_epi64(b4, zeros); in Transpose8x4To4x8_U16()
232 out[5] = _mm_unpackhi_epi64(b2, zeros); in Transpose8x4To4x8_U16()
234 out[7] = _mm_unpackhi_epi64(b6, zeros); in Transpose8x4To4x8_U16()
294 out[1] = _mm_unpackhi_epi64(b0, b1); in Transpose8x8_U16()
296 out[3] = _mm_unpackhi_epi64(b4, b5); in Transpose8x8_U16()
298 out[5] = _mm_unpackhi_epi64(b2, b3); in Transpose8x8_U16()
300 out[7] = _mm_unpackhi_epi64(b6, b7); in Transpose8x8_U16()
|
| /external/XNNPACK/src/x8-zip/ |
| D | xm-sse2.c | 59 vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
71 vxyzw1 = _mm_unpackhi_epi64(vxyzw1, vxyzw1); in xnn_x8_zip_xm_ukernel__sse2()
83 vxyzw2 = _mm_unpackhi_epi64(vxyzw2, vxyzw2); in xnn_x8_zip_xm_ukernel__sse2()
95 vxyzw3 = _mm_unpackhi_epi64(vxyzw3, vxyzw3); in xnn_x8_zip_xm_ukernel__sse2()
122 vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
134 vxyzw1 = _mm_unpackhi_epi64(vxyzw1, vxyzw1); in xnn_x8_zip_xm_ukernel__sse2()
166 vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
181 vxyzw0 = _mm_unpackhi_epi64(vxyzw0, vxyzw0); in xnn_x8_zip_xm_ukernel__sse2()
|
| /external/libavc/common/x86/ |
| D | ih264_ihadamard_scaling_sse42.c | 130 src_r1 = _mm_unpackhi_epi64(temp0, temp2); //a1 b1 c1 d1 in ih264_ihadamard_scaling_4x4_sse42()
132 src_r3 = _mm_unpackhi_epi64(temp1, temp3); //a3 b3 c3 d3 in ih264_ihadamard_scaling_4x4_sse42()
159 src_r1 = _mm_unpackhi_epi64(temp0, temp2); //b0 b1 b2 b3 in ih264_ihadamard_scaling_4x4_sse42()
161 src_r3 = _mm_unpackhi_epi64(temp1, temp3); //d0 d1 d2 d3 in ih264_ihadamard_scaling_4x4_sse42()
229 …plane_1 = _mm_unpackhi_epi64(temp0, temp1); //b0+b1+b2+b3 b0+b1-b2-b3 b0-b1+b2-b3 b0-b1-b2… in ih264_ihadamard_scaling_2x2_uv_sse42()
|
| D | ih264_ihadamard_scaling_ssse3.c | 132 src_r1 = _mm_unpackhi_epi64(temp0, temp2); //a1 b1 c1 d1 in ih264_ihadamard_scaling_4x4_ssse3()
134 src_r3 = _mm_unpackhi_epi64(temp1, temp3); //a3 b3 c3 d3 in ih264_ihadamard_scaling_4x4_ssse3()
161 src_r1 = _mm_unpackhi_epi64(temp0, temp2); //b0 b1 b2 b3 in ih264_ihadamard_scaling_4x4_ssse3()
163 src_r3 = _mm_unpackhi_epi64(temp1, temp3); //d0 d1 d2 d3 in ih264_ihadamard_scaling_4x4_ssse3()
|
| D | ih264_resi_trans_quant_sse42.c | 171 src_r1 = _mm_unpackhi_epi64(temp1, zero_8x16b); //a1 b1 c1 d1 in ih264_resi_trans_quant_4x4_sse42()
173 src_r3 = _mm_unpackhi_epi64(temp3, zero_8x16b); //a3 b3 c3 d3 in ih264_resi_trans_quant_4x4_sse42()
209 src_r1 = _mm_unpackhi_epi64(temp1, zero_8x16b); //b0 b1 b2 b3 in ih264_resi_trans_quant_4x4_sse42()
211 src_r3 = _mm_unpackhi_epi64(temp3, zero_8x16b); //d0 d1 d2 d3 in ih264_resi_trans_quant_4x4_sse42()
454 src_r1 = _mm_unpackhi_epi64(temp1, zero_8x16b); //a1 b1 c1 d1 in ih264_resi_trans_quant_chroma_4x4_sse42()
456 src_r3 = _mm_unpackhi_epi64(temp3, zero_8x16b); //a3 b3 c3 d3 in ih264_resi_trans_quant_chroma_4x4_sse42()
492 src_r1 = _mm_unpackhi_epi64(temp1, zero_8x16b); //b0 b1 b2 b3 in ih264_resi_trans_quant_chroma_4x4_sse42()
494 src_r3 = _mm_unpackhi_epi64(temp3, zero_8x16b); //d0 d1 d2 d3 in ih264_resi_trans_quant_chroma_4x4_sse42()
711 src_r1 = _mm_unpackhi_epi64(temp0, temp2); //a1 b1 c1 d1 in ih264_hadamard_quant_4x4_sse42()
713 src_r3 = _mm_unpackhi_epi64(temp1, temp3); //a3 b3 c3 d3 in ih264_hadamard_quant_4x4_sse42()
[all …]
|
| /external/libopus/celt/x86/ |
| D | pitch_sse4_1.c | 98 acc1 = _mm_add_epi32(acc1, _mm_unpackhi_epi64(acc1, acc1)); in celt_inner_prod_sse4_1()
141 sum0 = _mm_add_epi32(sum0, _mm_unpackhi_epi64( sum0, sum0)); in xcorr_kernel_sse4_1()
144 sum1 = _mm_add_epi32(sum1, _mm_unpackhi_epi64( sum1, sum1)); in xcorr_kernel_sse4_1()
147 sum2 = _mm_add_epi32(sum2, _mm_unpackhi_epi64( sum2, sum2)); in xcorr_kernel_sse4_1()
150 sum3 = _mm_add_epi32(sum3, _mm_unpackhi_epi64( sum3, sum3)); in xcorr_kernel_sse4_1()
|
| /external/libhevc/decoder/x86/ |
| D | ihevcd_fmt_conv_ssse3_intr.c | 154 src_v_8x16b = _mm_unpackhi_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
167 src_v_8x16b = _mm_unpackhi_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
180 src_v_8x16b = _mm_unpackhi_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
193 src_v_8x16b = _mm_unpackhi_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
229 src_v_8x16b = _mm_unpackhi_epi64(temp0_8x16b, temp1_8x16b); in ihevcd_fmt_conv_420sp_to_420p_ssse3()
|
| /external/webp/src/dsp/ |
| D | common_sse2.h | 94 *out1 = _mm_unpackhi_epi64(transpose1_0, transpose1_1); in VP8Transpose_2_4x4_16b()
96 *out3 = _mm_unpackhi_epi64(transpose1_2, transpose1_3); in VP8Transpose_2_4x4_16b()
182 *in0 = _mm_unpackhi_epi64(C1, C3); in VP8L32bToPlanar_SSE2()
184 *in2 = _mm_unpackhi_epi64(C0, C2); in VP8L32bToPlanar_SSE2()
|
| /external/XNNPACK/src/x32-zip/ |
| D | x4-sse2.c | 43 const __m128i vxyzw1 = _mm_unpackhi_epi64(vxy_lo, vzw_lo); in xnn_x32_zip_x4_ukernel__sse2()
45 const __m128i vxyzw3 = _mm_unpackhi_epi64(vxy_hi, vzw_hi); in xnn_x32_zip_x4_ukernel__sse2()
69 const __m128i vxyzw_hi = _mm_unpackhi_epi64(vxy, vzw); in xnn_x32_zip_x4_ukernel__sse2()
|