Home
last modified time | relevance | path

Searched refs:_mm_add_epi16 (Results 1 – 25 of 219) sorted by relevance

123456789

/external/libvpx/libvpx/vpx_dsp/x86/
Dfwd_txfm_sse2.c29 tmp = _mm_add_epi16(in0, in1); in vpx_fdct4x4_1_sse2()
54 u0 = _mm_add_epi16(in0, in1); in vpx_fdct8x8_1_sse2()
55 u1 = _mm_add_epi16(in2, in3); in vpx_fdct8x8_1_sse2()
62 sum = _mm_add_epi16(u0, u1); in vpx_fdct8x8_1_sse2()
64 in0 = _mm_add_epi16(in0, in1); in vpx_fdct8x8_1_sse2()
65 in2 = _mm_add_epi16(in2, in3); in vpx_fdct8x8_1_sse2()
66 sum = _mm_add_epi16(sum, in0); in vpx_fdct8x8_1_sse2()
69 sum = _mm_add_epi16(sum, in2); in vpx_fdct8x8_1_sse2()
100 u0 = _mm_add_epi16(in0, in1); in vpx_fdct16x16_1_sse2()
101 u1 = _mm_add_epi16(in2, in3); in vpx_fdct16x16_1_sse2()
[all …]
Dinv_txfm_sse2.h139 d0 = _mm_add_epi16(in_x, d0); in recon_and_store()
148 out[0] = _mm_add_epi16(in[0], final_rounding); in round_shift_8x8()
149 out[1] = _mm_add_epi16(in[1], final_rounding); in round_shift_8x8()
150 out[2] = _mm_add_epi16(in[2], final_rounding); in round_shift_8x8()
151 out[3] = _mm_add_epi16(in[3], final_rounding); in round_shift_8x8()
152 out[4] = _mm_add_epi16(in[4], final_rounding); in round_shift_8x8()
153 out[5] = _mm_add_epi16(in[5], final_rounding); in round_shift_8x8()
154 out[6] = _mm_add_epi16(in[6], final_rounding); in round_shift_8x8()
155 out[7] = _mm_add_epi16(in[7], final_rounding); in round_shift_8x8()
198 d[0] = _mm_add_epi16(d[0], in[0]); in recon_and_store4x4_sse2()
[all …]
Dhighbd_loopfilter_sse2.c254 pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5), _mm_add_epi16(p4, p3)); in vpx_highbd_lpf_horizontal_16_sse2()
255 pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5), _mm_add_epi16(q4, q3)); in vpx_highbd_lpf_horizontal_16_sse2()
257 pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1)); in vpx_highbd_lpf_horizontal_16_sse2()
258 pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); in vpx_highbd_lpf_horizontal_16_sse2()
260 pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1)); in vpx_highbd_lpf_horizontal_16_sse2()
261 pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); in vpx_highbd_lpf_horizontal_16_sse2()
263 _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); in vpx_highbd_lpf_horizontal_16_sse2()
264 pixetFilter_p2p1p0 = _mm_add_epi16( in vpx_highbd_lpf_horizontal_16_sse2()
265 four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); in vpx_highbd_lpf_horizontal_16_sse2()
267 _mm_srli_epi16(_mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7, p0)), 4); in vpx_highbd_lpf_horizontal_16_sse2()
[all …]
Dloopfilter_sse2.c392 pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16), in vpx_lpf_horizontal_16_sse2()
393 _mm_add_epi16(p4_16, p3_16)); in vpx_lpf_horizontal_16_sse2()
394 pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16), in vpx_lpf_horizontal_16_sse2()
395 _mm_add_epi16(q4_16, q3_16)); in vpx_lpf_horizontal_16_sse2()
397 pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16)); in vpx_lpf_horizontal_16_sse2()
398 pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); in vpx_lpf_horizontal_16_sse2()
400 pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16)); in vpx_lpf_horizontal_16_sse2()
401 pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); in vpx_lpf_horizontal_16_sse2()
403 _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); in vpx_lpf_horizontal_16_sse2()
404 pixetFilter_p2p1p0 = _mm_add_epi16( in vpx_lpf_horizontal_16_sse2()
[all …]
Dloopfilter_avx2.c194 pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16), in vpx_lpf_horizontal_16_avx2()
195 _mm_add_epi16(p4_16, p3_16)); in vpx_lpf_horizontal_16_avx2()
196 pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16), in vpx_lpf_horizontal_16_avx2()
197 _mm_add_epi16(q4_16, q3_16)); in vpx_lpf_horizontal_16_avx2()
199 pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16)); in vpx_lpf_horizontal_16_avx2()
200 pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); in vpx_lpf_horizontal_16_avx2()
202 pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16)); in vpx_lpf_horizontal_16_avx2()
203 pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); in vpx_lpf_horizontal_16_avx2()
205 _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); in vpx_lpf_horizontal_16_avx2()
206 pixetFilter_p2p1p0 = _mm_add_epi16( in vpx_lpf_horizontal_16_avx2()
[all …]
Dinv_txfm_ssse3.c79 out[0] = _mm_add_epi16(step1[0], step1[7]); in idct32_34_8x32_quarter_1()
80 out[1] = _mm_add_epi16(step1[1], step1[6]); in idct32_34_8x32_quarter_1()
81 out[2] = _mm_add_epi16(step1[2], step1[5]); in idct32_34_8x32_quarter_1()
82 out[3] = _mm_add_epi16(step1[3], step1[4]); in idct32_34_8x32_quarter_1()
202 step2[4] = _mm_add_epi16(step1[4], step1[5]); in idct32_135_8x32_quarter_1()
205 step2[7] = _mm_add_epi16(step1[7], step1[6]); in idct32_135_8x32_quarter_1()
208 step1[0] = _mm_add_epi16(step2[0], step2[3]); in idct32_135_8x32_quarter_1()
209 step1[1] = _mm_add_epi16(step2[0], step2[2]); in idct32_135_8x32_quarter_1()
217 out[0] = _mm_add_epi16(step1[0], step1[7]); in idct32_135_8x32_quarter_1()
218 out[1] = _mm_add_epi16(step1[1], step1[6]); in idct32_135_8x32_quarter_1()
[all …]
Dinv_txfm_ssse3.h53 step2[4] = _mm_add_epi16(step1[4], step1[5]); // step2 4&7 in idct8x8_12_add_kernel_ssse3()
60 tmp[0] = _mm_add_epi16(step2[0], step2[2]); // step1 0&1 in idct8x8_12_add_kernel_ssse3()
66 tmp[0] = _mm_add_epi16(step1[3], step2[4]); // output 3&0 in idct8x8_12_add_kernel_ssse3()
67 tmp[1] = _mm_add_epi16(step1[2], step1[5]); // output 2&1 in idct8x8_12_add_kernel_ssse3()
85 step2[4] = _mm_add_epi16(step1[4], step1[5]); in idct8x8_12_add_kernel_ssse3()
88 step2[7] = _mm_add_epi16(step1[7], step1[6]); in idct8x8_12_add_kernel_ssse3()
91 step1[0] = _mm_add_epi16(step2[0], step2[3]); in idct8x8_12_add_kernel_ssse3()
92 step1[1] = _mm_add_epi16(step2[0], step2[2]); in idct8x8_12_add_kernel_ssse3()
98 io[0] = _mm_add_epi16(step1[0], step2[7]); in idct8x8_12_add_kernel_ssse3()
99 io[1] = _mm_add_epi16(step1[1], step1[6]); in idct8x8_12_add_kernel_ssse3()
[all …]
/external/libaom/libaom/aom_dsp/x86/
Dhighbd_loopfilter_sse2.c371 __m128i sum_p = _mm_add_epi16(pq[5], _mm_add_epi16(pq[4], pq[3])); in highbd_lpf_internal_14_sse2()
372 __m128i sum_lp = _mm_add_epi16(pq[0], _mm_add_epi16(pq[2], pq[1])); in highbd_lpf_internal_14_sse2()
373 sum_p = _mm_add_epi16(sum_p, sum_lp); in highbd_lpf_internal_14_sse2()
378 sum_p_0 = _mm_add_epi16(eight, _mm_add_epi16(sum_p, sum_q)); in highbd_lpf_internal_14_sse2()
379 sum_lp = _mm_add_epi16(four, _mm_add_epi16(sum_lp, sum_lq)); in highbd_lpf_internal_14_sse2()
381 flat_p[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(pq[3], pq[0])); in highbd_lpf_internal_14_sse2()
382 flat_q[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(q[3], q[0])); in highbd_lpf_internal_14_sse2()
384 sum_p6 = _mm_add_epi16(pq[6], pq[6]); in highbd_lpf_internal_14_sse2()
385 sum_p3 = _mm_add_epi16(pq[3], pq[3]); in highbd_lpf_internal_14_sse2()
390 work0_0 = _mm_add_epi16(_mm_add_epi16(pq[6], pq[0]), pq[1]); in highbd_lpf_internal_14_sse2()
[all …]
Dloopfilter_sse2.c487 pixelFilter_p = _mm_add_epi16(p5_16, _mm_add_epi16(p4_16, p3_16)); in lpf_internal_14_dual_sse2()
488 pixelFilter_q = _mm_add_epi16(q5_16, _mm_add_epi16(q4_16, q3_16)); in lpf_internal_14_dual_sse2()
490 pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16)); in lpf_internal_14_dual_sse2()
491 pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0); in lpf_internal_14_dual_sse2()
493 pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16)); in lpf_internal_14_dual_sse2()
494 pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0); in lpf_internal_14_dual_sse2()
496 _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q)); in lpf_internal_14_dual_sse2()
497 pixetFilter_p2p1p0 = _mm_add_epi16( in lpf_internal_14_dual_sse2()
498 four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0)); in lpf_internal_14_dual_sse2()
500 _mm_add_epi16(pixelFilter_p, in lpf_internal_14_dual_sse2()
[all …]
/external/libgav1/libgav1/src/dsp/x86/
Dcdef_sse4.cc103 *partial_lo = _mm_add_epi16(*partial_lo, _mm_slli_si128(v_src_16[1], 2)); in AddPartial_D0_D4()
105 *partial_hi = _mm_add_epi16(*partial_hi, _mm_srli_si128(v_src_16[1], 14)); in AddPartial_D0_D4()
108 *partial_lo = _mm_add_epi16(*partial_lo, _mm_slli_si128(v_src_16[2], 4)); in AddPartial_D0_D4()
110 *partial_hi = _mm_add_epi16(*partial_hi, _mm_srli_si128(v_src_16[2], 12)); in AddPartial_D0_D4()
113 *partial_lo = _mm_add_epi16(*partial_lo, _mm_slli_si128(v_src_16[3], 6)); in AddPartial_D0_D4()
115 *partial_hi = _mm_add_epi16(*partial_hi, _mm_srli_si128(v_src_16[3], 10)); in AddPartial_D0_D4()
118 *partial_lo = _mm_add_epi16(*partial_lo, _mm_slli_si128(v_src_16[4], 8)); in AddPartial_D0_D4()
120 *partial_hi = _mm_add_epi16(*partial_hi, _mm_srli_si128(v_src_16[4], 8)); in AddPartial_D0_D4()
123 *partial_lo = _mm_add_epi16(*partial_lo, _mm_slli_si128(v_src_16[5], 10)); in AddPartial_D0_D4()
125 *partial_hi = _mm_add_epi16(*partial_hi, _mm_srli_si128(v_src_16[5], 6)); in AddPartial_D0_D4()
[all …]
Dintra_edge_sse4.cc61 _mm_add_epi16(outers4_lo, _mm_srli_si128(outers4_lo, 4)); in ComputeKernel1Store12()
63 _mm_add_epi16(outers4_hi, _mm_srli_si128(outers4_hi, 4)); in ComputeKernel1Store12()
66 _mm_add_epi16(partial_sums_lo, centers8_lo), 4); in ComputeKernel1Store12()
68 _mm_add_epi16(partial_sums_hi, centers8_hi), 4); in ComputeKernel1Store12()
89 _mm_add_epi16(outers_lo, _mm_slli_epi16(outers_lo, 2)); in ComputeKernel2Store12()
91 _mm_add_epi16(outers_hi, _mm_slli_epi16(outers_hi, 2)); in ComputeKernel2Store12()
93 const __m128i centers6_lo = _mm_add_epi16(_mm_slli_epi16(centers_lo, 1), in ComputeKernel2Store12()
95 const __m128i centers6_hi = _mm_add_epi16(_mm_slli_epi16(centers_hi, 1), in ComputeKernel2Store12()
99 _mm_add_epi16(outers5_lo, _mm_srli_si128(outers5_lo, 4)); in ComputeKernel2Store12()
102 _mm_add_epi16(centers6_lo, partial_sums_lo), 4); in ComputeKernel2Store12()
[all …]
/external/XNNPACK/src/qs8-gavgpool/gen/
D7p7x-minmax-sse41-c24-acc2.c74 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
75 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
76 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
77 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
78 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
79 __m128i vacc1xGHIJKLMN = _mm_add_epi16(vxi2xGHIJKLMN, vxi3xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
81 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
82 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
83 vacc0xGHIJKLMN = _mm_add_epi16(vacc0xGHIJKLMN, vxi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
84 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c24_acc2()
[all …]
D7p7x-minmax-sse41-c16-acc2.c67 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
68 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
69 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
70 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
72 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
73 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
74 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
75 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
76 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
77 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi6x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c16_acc2()
[all …]
D7p7x-minmax-sse2-c24-acc2.c95 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
96 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
97 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
98 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
99 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
100 __m128i vacc1xGHIJKLMN = _mm_add_epi16(vxi2xGHIJKLMN, vxi3xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
102 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
103 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
104 vacc0xGHIJKLMN = _mm_add_epi16(vacc0xGHIJKLMN, vxi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
105 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c24_acc2()
[all …]
D7p7x-minmax-ssse3-c24-acc2.c95 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
96 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
97 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
98 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
99 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
100 __m128i vacc1xGHIJKLMN = _mm_add_epi16(vxi2xGHIJKLMN, vxi3xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
102 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
103 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
104 vacc0xGHIJKLMN = _mm_add_epi16(vacc0xGHIJKLMN, vxi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
105 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c24_acc2()
[all …]
D7p7x-minmax-ssse3-c16-acc2.c81 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
82 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
83 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
84 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
86 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
87 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
88 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
89 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
90 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
91 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi6x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__ssse3_c16_acc2()
[all …]
D7p7x-minmax-sse41-c8-acc2.c60 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
61 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
63 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
64 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
65 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
68 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vacc1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
106 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
107 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
109 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
110 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse41_c8_acc2()
[all …]
D7p7x-minmax-sse2-c16-acc2.c81 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
82 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
83 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
84 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
86 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
87 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
88 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
89 vacc1x89ABCDEF = _mm_add_epi16(vacc1x89ABCDEF, vxi5x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
90 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi6x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
91 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi6x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7p7x__sse2_c16_acc2()
[all …]
D7x-minmax-sse41-c24-acc2.c91 __m128i vacc0x01234567 = _mm_add_epi16(vxi0x01234567, vxi1x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
92 __m128i vacc0x89ABCDEF = _mm_add_epi16(vxi0x89ABCDEF, vxi1x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
93 __m128i vacc0xGHIJKLMN = _mm_add_epi16(vxi0xGHIJKLMN, vxi1xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
94 __m128i vacc1x01234567 = _mm_add_epi16(vxi2x01234567, vxi3x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
95 __m128i vacc1x89ABCDEF = _mm_add_epi16(vxi2x89ABCDEF, vxi3x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
96 __m128i vacc1xGHIJKLMN = _mm_add_epi16(vxi2xGHIJKLMN, vxi3xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
98 vacc0x01234567 = _mm_add_epi16(vacc0x01234567, vxi4x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
99 vacc0x89ABCDEF = _mm_add_epi16(vacc0x89ABCDEF, vxi4x89ABCDEF); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
100 vacc0xGHIJKLMN = _mm_add_epi16(vacc0xGHIJKLMN, vxi4xGHIJKLMN); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
101 vacc1x01234567 = _mm_add_epi16(vacc1x01234567, vxi5x01234567); in xnn_qs8_gavgpool_minmax_ukernel_7x__sse41_c24_acc2()
[all …]
/external/libavc/common/x86/
Dih264_deblk_luma_ssse3.c219 Alpha_8x16 = _mm_add_epi16(Alpha_8x16, const_val2_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
266 temp1 = _mm_add_epi16(p0_8x16, q1_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
267 temp2 = _mm_add_epi16(p1_8x16, q0_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
268 temp5 = _mm_add_epi16(temp1, const_val2_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
269 temp6 = _mm_add_epi16(temp2, const_val2_16x8); in ih264_deblk_luma_vert_bs4_ssse3()
272 temp1 = _mm_add_epi16(temp5, temp3); in ih264_deblk_luma_vert_bs4_ssse3()
273 temp2 = _mm_add_epi16(temp6, temp4); in ih264_deblk_luma_vert_bs4_ssse3()
278 temp6 = _mm_add_epi16(temp6, p0_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
279 temp5 = _mm_add_epi16(temp5, q0_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
280 temp1 = _mm_add_epi16(temp6, p2_8x16); in ih264_deblk_luma_vert_bs4_ssse3()
[all …]
Dih264_luma_intra_pred_filters_ssse3.c346 res1_8x16b = _mm_add_epi16(top_8x16b, top_sh_8x16b); in ih264_intra_pred_luma_4x4_mode_diag_dl_ssse3()
349 res1_8x16b = _mm_add_epi16(res1_8x16b, const_2_8x16b); in ih264_intra_pred_luma_4x4_mode_diag_dl_ssse3()
350 res1_8x16b = _mm_add_epi16(res2_8x16b, res1_8x16b); in ih264_intra_pred_luma_4x4_mode_diag_dl_ssse3()
431 …res1_8x16b = _mm_add_epi16(top_left_8x16b, top_left_sh_8x16b); //l3+l2 l2+l1 l1+l0 l0+tl… in ih264_intra_pred_luma_4x4_mode_diag_dr_ssse3()
435 res1_8x16b = _mm_add_epi16(res1_8x16b, const_2_8x16b); in ih264_intra_pred_luma_4x4_mode_diag_dr_ssse3()
436 …res1_8x16b = _mm_add_epi16(res2_8x16b, res1_8x16b); //l3+2*l2+l1+2 l2+2*l1+l0… in ih264_intra_pred_luma_4x4_mode_diag_dr_ssse3()
519 …w121_a1_8x16b = _mm_add_epi16(w121_a1_8x16b, w121_a2_8x16b); //l2+l1 l1+l0 l0+tl tl+t0 t0+t1 … in ih264_intra_pred_luma_4x4_mode_vert_r_ssse3()
524 w121_a1_8x16b = _mm_add_epi16(w121_a1_8x16b, w121_a2_8x16b); //l2+2*l1+l0 l1+2*l0+tl ... in ih264_intra_pred_luma_4x4_mode_vert_r_ssse3()
525 w121_a1_8x16b = _mm_add_epi16(w121_a1_8x16b, const_2_8x16b); in ih264_intra_pred_luma_4x4_mode_vert_r_ssse3()
612 …w121_a1_8x16b = _mm_add_epi16(w121_a1_8x16b, w121_a2_8x16b); //l3+l2 l2+l1 l1+l0 l0+tl tl+t0 … in ih264_intra_pred_luma_4x4_mode_horz_d_ssse3()
[all …]
Dih264_inter_pred_filters_ssse3.c279 res_r0r1_t1_8x16b = _mm_add_epi16(res_r0r1_t1_8x16b, res_r0r1_t2_8x16b); in ih264_inter_pred_luma_horz_ssse3()
280 res_r0r1_t3_8x16b = _mm_add_epi16(const_val16_8x16b, res_r0r1_t3_8x16b); in ih264_inter_pred_luma_horz_ssse3()
281 …res_r0r1_t1_8x16b = _mm_add_epi16(res_r0r1_t1_8x16b, res_r0r1_t3_8x16b); //a0*c0+a1*c1+a2*c2+a3*c3… in ih264_inter_pred_luma_horz_ssse3()
358 res_r0_t1_8x16b = _mm_add_epi16(res_r0_t1_8x16b, res_r0_t2_8x16b); in ih264_inter_pred_luma_horz_ssse3()
359 res_r1_t1_8x16b = _mm_add_epi16(res_r1_t1_8x16b, res_r1_t2_8x16b); in ih264_inter_pred_luma_horz_ssse3()
360 res_r0_t3_8x16b = _mm_add_epi16(res_r0_t3_8x16b, const_val16_8x16b); in ih264_inter_pred_luma_horz_ssse3()
361 res_r1_t3_8x16b = _mm_add_epi16(res_r1_t3_8x16b, const_val16_8x16b); in ih264_inter_pred_luma_horz_ssse3()
362 res_r0_t1_8x16b = _mm_add_epi16(res_r0_t1_8x16b, res_r0_t3_8x16b); in ih264_inter_pred_luma_horz_ssse3()
363 res_r1_t1_8x16b = _mm_add_epi16(res_r1_t1_8x16b, res_r1_t3_8x16b); in ih264_inter_pred_luma_horz_ssse3()
435 res_r0_t1_8x16b = _mm_add_epi16(res_r0_t1_8x16b, res_r0_t2_8x16b); in ih264_inter_pred_luma_horz_ssse3()
[all …]
/external/libhevc/common/x86/
Dihevc_inter_pred_filters_ssse3_intr.c356 res_temp5_8x16b = _mm_add_epi16(res_temp1_8x16b, res_temp2_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
357 res_temp6_8x16b = _mm_add_epi16(res_temp3_8x16b, res_temp4_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
358 res_temp5_8x16b = _mm_add_epi16(res_temp5_8x16b, res_temp6_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
389 res_temp15_8x16b = _mm_add_epi16(res_temp11_8x16b, res_temp12_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
390 res_temp16_8x16b = _mm_add_epi16(res_temp13_8x16b, res_temp14_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
391 res_temp15_8x16b = _mm_add_epi16(res_temp15_8x16b, res_temp16_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
445 res_temp5_8x16b = _mm_add_epi16(res_temp1_8x16b, res_temp2_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
446 res_temp6_8x16b = _mm_add_epi16(res_temp3_8x16b, res_temp4_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
447 res_temp5_8x16b = _mm_add_epi16(res_temp5_8x16b, res_temp6_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
483 res_temp15_8x16b = _mm_add_epi16(res_temp11_8x16b, res_temp12_8x16b); in ihevc_inter_pred_luma_horz_ssse3()
[all …]
Dihevc_intra_pred_filters_ssse3_intr.c533 src_temp1 = _mm_add_epi16(src_temp1, src_temp2); in ihevc_intra_pred_ref_filtering_ssse3()
534 src_temp1 = _mm_add_epi16(src_temp1, src_temp3); in ihevc_intra_pred_ref_filtering_ssse3()
535 src_temp1 = _mm_add_epi16(src_temp1, const_value_8x16); in ihevc_intra_pred_ref_filtering_ssse3()
549 src_temp4 = _mm_add_epi16(src_temp4, src_temp5); in ihevc_intra_pred_ref_filtering_ssse3()
550 src_temp4 = _mm_add_epi16(src_temp4, src_temp6); in ihevc_intra_pred_ref_filtering_ssse3()
551 src_temp4 = _mm_add_epi16(src_temp4, const_value_8x16); in ihevc_intra_pred_ref_filtering_ssse3()
720 row1_16x8b = _mm_add_epi16(row_16x8b, const_temp8_4x32b); in ihevc_intra_pred_luma_planar_ssse3()
722 row2_16x8b = _mm_add_epi16(row1_16x8b, const_temp8_4x32b); in ihevc_intra_pred_luma_planar_ssse3()
724 row3_16x8b = _mm_add_epi16(row2_16x8b, const_temp8_4x32b); in ihevc_intra_pred_luma_planar_ssse3()
745 res_temp_8x16b = _mm_add_epi16(res_temp_8x16b, res_temp4_8x16b); in ihevc_intra_pred_luma_planar_ssse3()
[all …]
/external/libavc/encoder/x86/
Dih264e_intra_modes_eval_ssse3.c229 sad_8x16b = _mm_add_epi16(sad_8x16b, sad1_8x16b); in ih264e_evaluate_intra16x16_modes_ssse3()
280 sad_8x16b = _mm_add_epi16(sad_8x16b, sad1_8x16b); in ih264e_evaluate_intra16x16_modes_ssse3()
353 sad_8x16b = _mm_add_epi16(sad_8x16b, sad1_8x16b); in ih264e_evaluate_intra16x16_modes_ssse3()
601 …temp1_8x16b = _mm_add_epi16(temp1_8x16b, temp2_8x16b); //l3+l3 l3+l2 l2+l1... … in ih264e_evaluate_intra_4x4_modes_ssse3()
604 …temp1_8x16b = _mm_add_epi16(temp1_8x16b, temp2_8x16b); //4*l3 l3+2*l3+l2 l3+2*l2+l1...… in ih264e_evaluate_intra_4x4_modes_ssse3()
606 …temp1_8x16b = _mm_add_epi16(const_2_8x16b, temp1_8x16b); //4*l3+2 3*l3+l2+2 l3+2*l2+l1+2.… in ih264e_evaluate_intra_4x4_modes_ssse3()
617 …temp2_8x16b = _mm_add_epi16(temp2_8x16b, temp3_8x16b); //t1+t2 t2+t3... t6+t7 … in ih264e_evaluate_intra_4x4_modes_ssse3()
619 …temp2_8x16b = _mm_add_epi16(temp2_8x16b, temp3_8x16b); //t1+2*t2+t3 t2+2*t3+t4.. t6+2*t7+… in ih264e_evaluate_intra_4x4_modes_ssse3()
621 …temp2_8x16b = _mm_add_epi16(const_2_8x16b, temp2_8x16b); //t1+2*t2+t3+2 t2+2*t3+t4+2 t3+2*t… in ih264e_evaluate_intra_4x4_modes_ssse3()
885 sad_8x16b = _mm_add_epi16(tmp1_8x16b, tmp2_8x16b); in ih264e_evaluate_intra_chroma_modes_ssse3()
[all …]

123456789