Home
last modified time | relevance | path

Searched refs:_mm256_adds_epi16 (Results 1 – 25 of 50) sorted by relevance

12

/external/libaom/libaom/aom_dsp/x86/
Daom_subpixel_8t_intrin_avx2.c115 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); in aom_filter_block1d4_h4_avx2()
205 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2); in aom_filter_block1d4_h8_avx2()
211 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); in aom_filter_block1d4_h8_avx2()
308 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); in aom_filter_block1d8_h4_avx2()
311 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); in aom_filter_block1d8_h4_avx2()
411 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2); in aom_filter_block1d8_h8_avx2()
421 __m256i sum23 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); in aom_filter_block1d8_h8_avx2()
422 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, sum23); in aom_filter_block1d8_h8_avx2()
425 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32); in aom_filter_block1d8_h8_avx2()
534 srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2); in aom_filter_block1d16_h4_avx2()
[all …]
Dtxfm_common_avx2.h55 *in0 = _mm256_adds_epi16(_in0, _in1); in btf_16_adds_subs_avx2()
70 *out0 = _mm256_adds_epi16(_in0, _in1); in btf_16_adds_subs_out_avx2()
239 in[i] = _mm256_adds_epi16(in[i], round); in round_shift_16bit_w16_avx2()
/external/libaom/libaom/av1/encoder/x86/
Dav1_fwd_txfm2d_avx2.c2041 x1[0] = _mm256_adds_epi16(input[0], input[7]); in fdct8x8_new_avx2()
2043 x1[1] = _mm256_adds_epi16(input[1], input[6]); in fdct8x8_new_avx2()
2045 x1[2] = _mm256_adds_epi16(input[2], input[5]); in fdct8x8_new_avx2()
2047 x1[3] = _mm256_adds_epi16(input[3], input[4]); in fdct8x8_new_avx2()
2052 x2[0] = _mm256_adds_epi16(x1[0], x1[3]); in fdct8x8_new_avx2()
2054 x2[1] = _mm256_adds_epi16(x1[1], x1[2]); in fdct8x8_new_avx2()
2073 x3[4] = _mm256_adds_epi16(x2[4], x2[5]); in fdct8x8_new_avx2()
2076 x3[7] = _mm256_adds_epi16(x2[7], x2[6]); in fdct8x8_new_avx2()
2151 x3[0] = _mm256_adds_epi16(x2[0], x2[2]); in fadst8x8_new_avx2()
2153 x3[1] = _mm256_adds_epi16(x2[1], x2[3]); in fadst8x8_new_avx2()
[all …]
Dav1_quantize_avx2.c116 __m256i q = _mm256_adds_epi16(abs_coeff, qp[0]); in quantize()
200 qcoeff256 = _mm256_adds_epi16(qcoeff256, round256); in av1_quantize_lp_avx2()
229 qcoeff256 = _mm256_adds_epi16(qcoeff256, round256); in av1_quantize_lp_avx2()
303 __m256i q = _mm256_adds_epi16(abs_coeff, qp[0]); in quantize_32x32()
378 __m256i q = _mm256_adds_epi16(abs_coeff, qp[0]); in quantize_64x64()
/external/libvpx/libvpx/vp9/encoder/x86/
Dvp9_quantize_avx2.c92 qcoeff256 = _mm256_adds_epi16(qcoeff256, round256); in vp9_quantize_fp_avx2()
120 qcoeff256 = _mm256_adds_epi16(qcoeff256, round256); in vp9_quantize_fp_avx2()
/external/libvpx/libvpx/vpx_dsp/x86/
Dconvolve_avx2.h70 sum1 = _mm256_adds_epi16(sum1, sum2); in convolve8_16_avx2()
147 const __m256i nearest_src = _mm256_adds_epi16(*src, *half_depth); in mm256_round_epi16()
Dvpx_subpixel_8t_intrin_avx2.c388 dst_first = _mm256_adds_epi16(tmp_0, tmp_1); in vpx_filter_block1d16_h4_avx2()
399 dst_second = _mm256_adds_epi16(tmp_0, tmp_1); in vpx_filter_block1d16_h4_avx2()
425 dst_first = _mm256_adds_epi16(tmp_0, tmp_1); in vpx_filter_block1d16_h4_avx2()
510 res_reg_lo = _mm256_adds_epi16(res_reg_m1001_lo, res_reg_1223_lo); in vpx_filter_block1d16_v4_avx2()
515 res_reg_hi = _mm256_adds_epi16(res_reg_m1001_hi, res_reg_1223_hi); in vpx_filter_block1d16_v4_avx2()
593 dst_reg = _mm256_adds_epi16(tmp_0, tmp_1); in vpx_filter_block1d8_h4_avx2()
703 res_reg = _mm256_adds_epi16(res_reg_m1001, res_reg_1223); in vpx_filter_block1d8_v4_avx2()
/external/XNNPACK/src/qs8-vaddc/gen/
Dminmax-avx2-mul32-ld64-x32.c58 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), v… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x32()
59 …__m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), v… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x32()
Dminmax-avx2-mul32-ld64-x16.c50 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), v… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x16()
Dminmax-avx2-mul32-ld64-x24.c54 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), v… in xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24()
/external/XNNPACK/src/qs8-vadd/gen/
Dminmax-avx2-mul32-ld64-x32.c67 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), v… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32()
68 …__m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(vaccGHIJKLMN, vaccOPQRSTUV), v… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x32()
Dminmax-avx2-mul32-ld64-x16.c55 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), v… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x16()
Dminmax-avx2-mul32-ld64-x24.c61 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), v… in xnn_qs8_vadd_minmax_ukernel__avx2_mul32_ld64_x24()
/external/libaom/libaom/av1/common/x86/
Dav1_inv_txfm_avx2.h48 __m256i u = _mm256_adds_epi16(_mm256_cvtepu8_epi16(pred), res); in write_recon_w16_avx2()
/external/XNNPACK/src/qs8-dwconv/
Dunipass-avx512skx-mul32.c.in109 …vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs…
111 …vout${ABC[C:C+4]}${ABC[C+8:C+12]}${ABC[C+4:C+8]}${ABC[C+12:C+16]} = _mm256_adds_epi16(_mm256_packs…
186 …__m256i vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]} = _mm256_adds_epi16(_mm256_packs_epi32…
188 …__m256i vout${ABC[0:4]}${ABC[8:12]}${ABC[4:8]}${ABC[12:16]} = _mm256_adds_epi16(_mm256_packs_epi32…
/external/XNNPACK/src/qs8-igemm/gen/
D3x8c8-minmax-avx2.c197 …__m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), vo… in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
198 …__m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), vo… in xnn_qs8_igemm_minmax_ukernel_3x8c8__avx2()
D1x16c8-minmax-avx512skx.c119 …__m256i vacc0x084C2A6E195D3B7F = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0… in xnn_qs8_igemm_minmax_ukernel_1x16c8__avx512skx()
D1x8c8-minmax-avx2.c131 …__m256i vacc00x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc0x01234567), vo… in xnn_qs8_igemm_minmax_ukernel_1x8c8__avx2()
/external/XNNPACK/src/qs8-gemm/gen/
D3x8c8-minmax-avx2.c180 …__m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), vo… in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
181 …__m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), vo… in xnn_qs8_gemm_minmax_ukernel_3x8c8__avx2()
D3x8c8-xw-minmax-avx2.c176 …__m256i vacc01x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc1x01234567), vo… in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
177 …__m256i vacc22x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc2x01234567, vacc2x01234567), vo… in xnn_qs8_gemm_xw_minmax_ukernel_3x8c8__avx2()
D1x16c8-minmax-avx512skx.c108 …__m256i vacc0x084C2A6E195D3B7F = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0… in xnn_qs8_gemm_minmax_ukernel_1x16c8__avx512skx()
D1x8c8-minmax-avx2.c118 …__m256i vacc00x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc0x01234567), vo… in xnn_qs8_gemm_minmax_ukernel_1x8c8__avx2()
D1x8c8-xw-minmax-avx2.c114 …__m256i vacc00x01234567 = _mm256_adds_epi16(_mm256_packs_epi32(vacc0x01234567, vacc0x01234567), vo… in xnn_qs8_gemm_xw_minmax_ukernel_1x8c8__avx2()
/external/XNNPACK/src/qs8-dwconv/gen/
Dup16x9-minmax-avx512skx-mul32.c171 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc012… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
250 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc012… in xnn_qs8_dwconv_minmax_ukernel_up16x9__avx512skx_mul32()
Dup32x9-minmax-avx512skx-mul32.c211 …__m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vaccGHI… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()
308 …__m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc012… in xnn_qs8_dwconv_minmax_ukernel_up32x9__avx512skx_mul32()

12