/third_party/python/Modules/_blake2/impl/ |
H A D | blake2s-round.h | 44 #define G1(row1,row2,row3,row4,buf) \ 45 row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \ 49 row2 = _mm_xor_si128( row2, row3 ); \ 50 row2 = _mm_roti_epi32(row2, -12); 52 #define G2(row1,row2,row3,row4,buf) \ 53 row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \ 57 row2 = _mm_xor_si128( row2, row [all...] |
H A D | blake2s.c | 268 __m128i row1, row2, row3, row4; in blake2s_compress() local 305 row2 = ff1 = LOADU( &S->h[4] ); in blake2s_compress() 319 STOREU( &S->h[4], _mm_xor_si128( ff1, _mm_xor_si128( row2, row4 ) ) ); in blake2s_compress()
|
/third_party/skia/third_party/externals/libjpeg-turbo/simd/arm/ |
H A D | jquanti-neon.c | 61 int16x8_t row2 = in jsimd_convsamp_neon() local 76 vst1q_s16(workspace + 2 * DCTSIZE, row2); in jsimd_convsamp_neon() 107 int16x8_t row2 = vld1q_s16(workspace + (i + 2) * DCTSIZE); in jsimd_quantize_neon() local 126 int16x8_t sign_row2 = vshrq_n_s16(row2, 15); in jsimd_quantize_neon() 131 uint16x8_t abs_row2 = vreinterpretq_u16_s16(vabsq_s16(row2)); in jsimd_quantize_neon() 159 row2 = vcombine_s16(vshrn_n_s32(row2_l, 16), vshrn_n_s32(row2_h, 16)); in jsimd_quantize_neon() 169 row2 = vreinterpretq_s16_u16(vshlq_u16(vreinterpretq_u16_s16(row2), in jsimd_quantize_neon() 179 row2 = veorq_s16(row2, sign_row in jsimd_quantize_neon() [all...] |
H A D | jidctint-neon.c | 99 int16x4_t row2, 118 int16x4_t row2, 203 int16x4_t row2 = vld1_s16(coef_block + 2 * DCTSIZE); in jsimd_idct_islow_neon() local 228 bitmap = vorr_s16(bitmap, row2); in jsimd_idct_islow_neon() 239 jsimd_idct_islow_pass1_sparse(row0, row1, row2, row3, quant_row0, in jsimd_idct_islow_neon() 244 jsimd_idct_islow_pass1_regular(row0, row1, row2, row3, row4, row5, in jsimd_idct_islow_neon() 256 row2 = vld1_s16(coef_block + 2 * DCTSIZE + 4); in jsimd_idct_islow_neon() 279 bitmap = vorr_s16(bitmap, row2); in jsimd_idct_islow_neon() 299 jsimd_idct_islow_pass1_sparse(row0, row1, row2, row3, quant_row0, in jsimd_idct_islow_neon() 304 jsimd_idct_islow_pass1_regular(row0, row1, row2, row in jsimd_idct_islow_neon() 338 jsimd_idct_islow_pass1_regular(int16x4_t row0, int16x4_t row1, int16x4_t row2, int16x4_t row3, int16x4_t row4, int16x4_t row5, int16x4_t row6, int16x4_t row7, int16x4_t quant_row0, int16x4_t quant_row1, int16x4_t quant_row2, int16x4_t quant_row3, int16x4_t quant_row4, int16x4_t quant_row5, int16x4_t quant_row6, int16x4_t quant_row7, int16_t *workspace_1, int16_t *workspace_2) jsimd_idct_islow_pass1_regular() argument 473 jsimd_idct_islow_pass1_sparse(int16x4_t row0, int16x4_t row1, int16x4_t row2, int16x4_t row3, int16x4_t quant_row0, int16x4_t quant_row1, int16x4_t quant_row2, int16x4_t quant_row3, int16_t *workspace_1, int16_t *workspace_2) jsimd_idct_islow_pass1_sparse() argument [all...] |
H A D | jidctred-neon.c | 204 int16x8_t row2 = vld1q_s16(coef_block + 2 * DCTSIZE); in jsimd_idct_4x4_neon() local 216 int16x8_t bitmap = vorrq_s16(row1, row2); in jsimd_idct_4x4_neon() 243 row2 = dcval; in jsimd_idct_4x4_neon() 264 int16x4_t z2 = vmul_s16(vget_high_s16(row2), quant_row2); in jsimd_idct_4x4_neon() 296 row2 = vcombine_s16(dcval, vrshrn_n_s32(vsubq_s32(tmp12, tmp0), in jsimd_idct_4x4_neon() 317 int16x4_t z2 = vmul_s16(vget_low_s16(row2), quant_row2); in jsimd_idct_4x4_neon() 349 row2 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp12, tmp0), in jsimd_idct_4x4_neon() 364 int16x8_t z2 = vmulq_s16(row2, quant_row2); in jsimd_idct_4x4_neon() 414 row2 = vcombine_s16(vrshrn_n_s32(vsubq_s32(tmp12_l, tmp0_l), in jsimd_idct_4x4_neon() 422 int16x8x2_t row_23 = vtrnq_s16(row2, row in jsimd_idct_4x4_neon() [all...] |
H A D | jfdctfst-neon.c | 154 int16x8_t row2 = vreinterpretq_s16_s32(rows_26.val[0]); in jsimd_fdct_ifast_neon() local 167 tmp2 = vaddq_s16(row2, row5); in jsimd_fdct_ifast_neon() 168 tmp5 = vsubq_s16(row2, row5); in jsimd_fdct_ifast_neon() 182 row2 = vaddq_s16(tmp13, z1); /* phase 5 */ in jsimd_fdct_ifast_neon() 208 vst1q_s16(data + 2 * DCTSIZE, row2); in jsimd_fdct_ifast_neon()
|
H A D | jidctfst-neon.c | 71 int16x8_t row2 = vld1q_s16(coef_block + 2 * DCTSIZE); in jsimd_idct_ifast_neon() local 84 int16x8_t bitmap = vorrq_s16(row1, row2); in jsimd_idct_ifast_neon() 103 row2 = dcval; in jsimd_idct_ifast_neon() 128 int16x4_t tmp1 = vmul_s16(vget_high_s16(row2), quant_row2); in jsimd_idct_ifast_neon() 180 row2 = vcombine_s16(dcval, vadd_s16(tmp2, tmp5)); in jsimd_idct_ifast_neon() 203 int16x4_t tmp1 = vmul_s16(vget_low_s16(row2), quant_row2); in jsimd_idct_ifast_neon() 255 row2 = vcombine_s16(vadd_s16(tmp2, tmp5), dcval); in jsimd_idct_ifast_neon() 273 int16x8_t tmp1 = vmulq_s16(row2, quant_row2); in jsimd_idct_ifast_neon() 325 row2 = vaddq_s16(tmp2, tmp5); in jsimd_idct_ifast_neon() 333 int16x8x2_t rows_23 = vtrnq_s16(row2, row in jsimd_idct_ifast_neon() [all...] |
H A D | jfdctint-neon.c | 253 int16x8_t row2 = vreinterpretq_s16_s32(rows_26.val[0]); in jsimd_fdct_islow_neon() local 266 tmp2 = vaddq_s16(row2, row5); in jsimd_fdct_islow_neon() 267 tmp5 = vsubq_s16(row2, row5); in jsimd_fdct_islow_neon() 288 row2 = vcombine_s16(vrshrn_n_s32(row2_scaled_l, DESCALE_P2), in jsimd_fdct_islow_neon() 370 vst1q_s16(data + 2 * DCTSIZE, row2); in jsimd_fdct_islow_neon()
|
/third_party/skia/third_party/externals/libjpeg-turbo/simd/arm/aarch32/ |
H A D | jchuff-neon.c | 70 int16x8_t row2 = vld1q_dup_s16(block + 12); in jsimd_huff_encode_one_block_neon() local 71 row2 = vld1q_lane_s16(block + 19, row2, 1); in jsimd_huff_encode_one_block_neon() 72 row2 = vld1q_lane_s16(block + 26, row2, 2); in jsimd_huff_encode_one_block_neon() 73 row2 = vld1q_lane_s16(block + 33, row2, 3); in jsimd_huff_encode_one_block_neon() 74 row2 = vld1q_lane_s16(block + 40, row2, 4); in jsimd_huff_encode_one_block_neon() 75 row2 in jsimd_huff_encode_one_block_neon() [all...] |
/third_party/skia/third_party/externals/swiftshader/src/Pipeline/ |
H A D | ShaderCore.cpp | 472 void transpose4x4(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3) in transpose4x4() argument 475 Int2 tmp1 = UnpackHigh(row2, row3); in transpose4x4() 477 Int2 tmp3 = UnpackLow(row2, row3); in transpose4x4() 481 row2 = UnpackLow(tmp0, tmp1); in transpose4x4() 485 void transpose4x3(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3) in transpose4x3() argument 488 Int2 tmp1 = UnpackHigh(row2, row3); in transpose4x3() 490 Int2 tmp3 = UnpackLow(row2, row3); in transpose4x3() 494 row2 = UnpackLow(tmp0, tmp1); in transpose4x3() 497 void transpose4x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) in transpose4x4() argument 500 Float4 tmp1 = UnpackLow(row2, row in transpose4x4() 510 transpose4x3(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose4x3() argument 522 transpose4x2(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose4x2() argument 531 transpose4x1(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose4x1() argument 539 transpose2x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose2x4() argument 550 transpose4xN(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3, int N) transpose4xN() argument [all...] |
H A D | ShaderCore.hpp | 214 void transpose4x4(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3); 215 void transpose4x3(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3); 216 void transpose4x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 217 void transpose4x3(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 218 void transpose4x2(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 219 void transpose4x1(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 220 void transpose2x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 221 void transpose4xN(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3, int N);
|
/third_party/skia/third_party/externals/libjpeg-turbo/simd/arm/aarch64/ |
H A D | jchuff-neon.c | 125 int16x8_t row2 = in jsimd_huff_encode_one_block_neon() local 144 row2 = in jsimd_huff_encode_one_block_neon() 146 4), row2, 0); in jsimd_huff_encode_one_block_neon() 147 row2 = in jsimd_huff_encode_one_block_neon() 149 0), row2, 5); in jsimd_huff_encode_one_block_neon() 163 int16x8_t abs_row2 = vabsq_s16(row2); in jsimd_huff_encode_one_block_neon() 176 vreinterpretq_u16_s16(veorq_s16(abs_row2, vshrq_n_s16(row2, 15))); in jsimd_huff_encode_one_block_neon()
|
/third_party/ffmpeg/libavcodec/mips/ |
H A D | vp8_lpf_msa.c | 380 v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; in ff_vp8_h_loop_filter16_msa() local 388 LD_UB8(temp_src, pitch, row0, row1, row2, row3, row4, row5, row6, row7); in ff_vp8_h_loop_filter16_msa() 392 TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, in ff_vp8_h_loop_filter16_msa() 445 v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; in ff_vp8_h_loop_filter8uv_msa() local 453 LD_UB8(src_u - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7); in ff_vp8_h_loop_filter8uv_msa() 456 TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, in ff_vp8_h_loop_filter8uv_msa() 525 v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; in ff_vp8_h_loop_filter_simple_msa() local 531 LD_UB8(temp_src, pitch, row0, row1, row2, row3, row4, row5, row6, row7); in ff_vp8_h_loop_filter_simple_msa() 535 TRANSPOSE16x4_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, in ff_vp8_h_loop_filter_simple_msa() 595 v16u8 row0, row1, row2, row in ff_vp8_h_loop_filter8uv_inner_msa() local 649 v16u8 row0, row1, row2, row3, row4, row5, row6, row7; ff_vp8_h_loop_filter16_inner_msa() local [all...] |
/third_party/ffmpeg/libavcodec/ |
H A D | ivi_dsp.c | 670 int i, row2, row4, row8; in ff_ivi_col_slant8() local 673 row2 = pitch << 1; in ff_ivi_col_slant8() 681 out[0], out[pitch], out[row2], out[row2 + pitch], out[row4], in ff_ivi_col_slant8() 682 out[row4 + pitch], out[row4 + row2], out[row8 - pitch], in ff_ivi_col_slant8() 685 out[0] = out[pitch] = out[row2] = out[row2 + pitch] = out[row4] = in ff_ivi_col_slant8() 686 out[row4 + pitch] = out[row4 + row2] = out[row8 - pitch] = 0; in ff_ivi_col_slant8() 731 int i, row2; in ff_ivi_col_slant4() local 734 row2 in ff_ivi_col_slant4() [all...] |
H A D | a64multienc.c | 107 uint8_t row1, row2; in render_charset() local 142 row1 = 0; row2 = 0; in render_charset() 155 row2 <<= 2; in render_charset() 162 row2 |= 3-(index2[pix] & 3); in render_charset() 164 row2 |= 3-(index1[pix] & 3); in render_charset() 174 if (INTERLACED) charset[y+0x800] = row2; in render_charset()
|
/third_party/skia/tools/ |
H A D | image_diff_metric.cpp | 46 const uint8_t* row2 = reinterpret_cast<const uint8_t*>(bm[1].pixmap().addr32(0, y)); in main() local 48 totalDiffs += std::abs((int)row1[i] - (int)row2[i]); in main()
|
/third_party/skia/third_party/externals/swiftshader/src/Shader/ |
H A D | ShaderCore.hpp | 87 void transpose4x4(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3); 88 void transpose4x3(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3); 89 void transpose4x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 90 void transpose4x3(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 91 void transpose4x2(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 92 void transpose4x1(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 93 void transpose2x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3); 94 void transpose4xN(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3, int N);
|
H A D | ShaderCore.cpp | 488 void transpose4x4(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3) in transpose4x4() argument 491 Int2 tmp1 = UnpackHigh(row2, row3); in transpose4x4() 493 Int2 tmp3 = UnpackLow(row2, row3); in transpose4x4() 497 row2 = UnpackLow(tmp0, tmp1); in transpose4x4() 501 void transpose4x3(Short4 &row0, Short4 &row1, Short4 &row2, Short4 &row3) in transpose4x3() argument 504 Int2 tmp1 = UnpackHigh(row2, row3); in transpose4x3() 506 Int2 tmp3 = UnpackLow(row2, row3); in transpose4x3() 510 row2 = UnpackLow(tmp0, tmp1); in transpose4x3() 513 void transpose4x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) in transpose4x4() argument 516 Float4 tmp1 = UnpackLow(row2, row in transpose4x4() 526 transpose4x3(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose4x3() argument 538 transpose4x2(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose4x2() argument 547 transpose4x1(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose4x1() argument 555 transpose2x4(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3) transpose2x4() argument 566 transpose4xN(Float4 &row0, Float4 &row1, Float4 &row2, Float4 &row3, int N) transpose4xN() argument [all...] |
H A D | VertexProgram.cpp | 1029 Vector4f row2 = fetchRegister(src1, 2); in M3X3() local 1033 dst.z = dot3(src0, row2); in M3X3() 1040 Vector4f row2 = fetchRegister(src1, 2); in M3X4() local 1045 dst.z = dot3(src0, row2); in M3X4() 1053 Vector4f row2 = fetchRegister(src1, 2); in M4X3() local 1057 dst.z = dot4(src0, row2); in M4X3() 1064 Vector4f row2 = fetchRegister(src1, 2); in M4X4() local 1069 dst.z = dot4(src0, row2); in M4X4()
|
/third_party/ffmpeg/libavcodec/loongarch/ |
H A D | h264dsp_lasx.c | 96 __m256i row0, row1, row2, row3, row4, row5, row6, row7; in ff_h264_h_lpf_luma_8_lasx() local 100 src, img_width_3x, row0, row1, row2, row3); in ff_h264_h_lpf_luma_8_lasx() 114 LASX_TRANSPOSE16x8_B(row0, row1, row2, row3, row4, row5, row6, in ff_h264_h_lpf_luma_8_lasx() 201 __m256i row0, row1, row2, row3, row4, row5, row6, row7; in ff_h264_h_lpf_luma_8_lasx() local 209 row0, row2); in ff_h264_h_lpf_luma_8_lasx() 212 DUP2_ARG2(__lasx_xvilvl_b, row2, row0, row3, row1, row4, row6); in ff_h264_h_lpf_luma_8_lasx() 213 DUP2_ARG2(__lasx_xvilvh_b, row2, row0, row3, row1, row5, row7); in ff_h264_h_lpf_luma_8_lasx() 393 __m256i row0, row1, row2, row3, row4, row5, row6, row7; in ff_h264_h_lpf_chroma_8_lasx() local 396 src, img_width_3x, row0, row1, row2, row3); in ff_h264_h_lpf_chroma_8_lasx() 402 DUP4_ARG2(__lasx_xvilvl_b, row2, row in ff_h264_h_lpf_chroma_8_lasx() 597 __m256i row0, row1, row2, row3, row4, row5, row6, row7; ff_h264_h_lpf_luma_intra_8_lasx() local 717 __m256i row0, row1, row2, row3, row4, row5, row6, row7; ff_h264_h_lpf_luma_intra_8_lasx() local 892 __m256i row0, row1, row2, row3, row4, row5, row6, row7; ff_h264_h_lpf_chroma_intra_8_lasx() local [all...] |
H A D | vp8_lpf_lsx.c | 335 __m128i row0, row1, row2, row3, row4, row5, row6, row7, row8; in ff_vp8_h_loop_filter16_lsx() local 349 temp_src + stride3, 0, row0, row1, row2, row3); in ff_vp8_h_loop_filter16_lsx() 360 LSX_TRANSPOSE16x8_B(row0, row1, row2, row3, row4, row5, row6, row7, row8, row9, row10, in ff_vp8_h_loop_filter16_lsx() 422 __m128i row0, row1, row2, row3, row4, row5, row6, row7, row8; in ff_vp8_h_loop_filter8uv_lsx() local 436 temp_src + stride3, 0, row0, row1, row2, row3); in ff_vp8_h_loop_filter8uv_lsx() 448 LSX_TRANSPOSE16x8_B(row0, row1, row2, row3, row4, row5, row6, row7, in ff_vp8_h_loop_filter8uv_lsx()
|
/third_party/skia/third_party/externals/libwebp/src/dsp/ |
H A D | dec_msa.c | 364 v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; in HFilter16() local 371 LD_UB8(ptmp, stride, row0, row1, row2, row3, row4, row5, row6, row7); in HFilter16() 374 TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, in HFilter16() 446 v16u8 row0, row1, row2, row3, row4, row5, row6, row7; in HFilterVertEdge16i() local 453 LD_UB8(src - 4, stride, row0, row1, row2, row3, row4, row5, row6, row7); in HFilterVertEdge16i() 456 TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, in HFilterVertEdge16i() 530 v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8; in HFilter8() local 537 LD_UB8(ptmp_src_u, stride, row0, row1, row2, row3, row4, row5, row6, row7); in HFilter8() 540 TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7, in HFilter8() 595 v16u8 row0, row1, row2, row in HFilter8i() local 636 v16u8 p1, p0, q1, q0, mask, row0, row1, row2, row3, row4, row5, row6, row7; SimpleHFilter16() local [all...] |
/third_party/skia/third_party/externals/d3d12allocator/src/ |
H A D | Common.h | 185 const vec4& row2,
in mat4() 189 _21(row2.x), _22(row2.y), _23(row2.z), _24(row2.w),
in mat4()
|
/third_party/skia/third_party/externals/swiftshader/third_party/llvm-10.0/llvm/lib/MC/ |
H A D | MCWin64EH.cpp | 597 uint32_t row2 = 0x0; in ARM64EmitUnwindInfo() local 598 row2 |= (CodeWords & 0xFF) << 16; in ARM64EmitUnwindInfo() 599 row2 |= (EpilogCount & 0xFFFF); in ARM64EmitUnwindInfo() 600 streamer.EmitIntValue(row2, 4); in ARM64EmitUnwindInfo()
|
/third_party/backends/backend/genesys/ |
H A D | image_pipeline.cpp | 308 const auto* row2 = buffer_.get_row_ptr(2); in get_next_row_data() local 315 std::uint16_t ch2 = get_raw_channel_from_row(row2, x, 0, format); in get_next_row_data() 515 const auto* row2 = buffer_.get_row_ptr(channel_shifts_[2]); in get_next_row_data() local 520 std::uint16_t ch2 = get_raw_channel_from_row(row2, x, 2, format); in get_next_row_data()
|