~1.21x faster (68 vs. 56 cycles) compared with mmxext function
Unified · Split
| @@ -1163,6 +1163,33 @@ cglobal deblock_h_chroma_8, 5, 7, 8, 0-16, pix_, stride_, alpha_, beta_, tc0_ · 1162 unmodified lines | ||
| 1163 | 1163 | STORE_8_ROWS PASS8ROWS(pix_q - 2, r5 - 2, stride_q, r6) |
| 1164 | 1164 | RET |
| 1165 | 1165 | |
| 1166 | +cglobal deblock_h_chroma422_8, 5, 7, 8, 0-16, pix_, stride_, alpha_, beta_, tc0_, | |
| 1167 | + CHROMA_H_START_XMM r5, r6 | |
| 1168 | + LOAD_8_ROWS PASS8ROWS(pix_q - 2, r5 - 2, stride_q, r6) | |
| 1169 | + TRANSPOSE_8x4B_XMM | |
| 1170 | + movq [rsp], m0 | |
| 1171 | + movq [rsp + 8], m3 | |
| 1172 | + CHROMA_INTER_BODY_XMM 2 | |
| 1173 | + movq m0, [rsp] | |
| 1174 | + movq m3, [rsp + 8] | |
| 1175 | + TRANSPOSE_4x8B_XMM | |
| 1176 | + STORE_8_ROWS PASS8ROWS(pix_q - 2, r5 - 2, stride_q, r6) | |
| 1177 | + | |
| 1178 | + lea pix_q, [pix_q + 8*stride_q] | |
| 1179 | + lea r5, [r5 + 8*stride_q] | |
| 1180 | + add tc0_q, 2 | |
| 1181 | + | |
| 1182 | + LOAD_8_ROWS PASS8ROWS(pix_q - 2, r5 - 2, stride_q, r6) | |
| 1183 | + TRANSPOSE_8x4B_XMM | |
| 1184 | + movq [rsp], m0 | |
| 1185 | + movq [rsp + 8], m3 | |
| 1186 | + CHROMA_INTER_BODY_XMM 2 | |
| 1187 | + movq m0, [rsp] | |
| 1188 | + movq m3, [rsp + 8] | |
| 1189 | + TRANSPOSE_4x8B_XMM | |
| 1190 | + STORE_8_ROWS PASS8ROWS(pix_q - 2, r5 - 2, stride_q, r6) | |
| 1191 | +RET | |
| 1192 | + | |
| 1166 | 1193 | %endmacro ; DEBLOCK_CHROMA_XMM |
| 1167 | 1194 | |
| 1168 | 1195 | DEBLOCK_CHROMA_XMM avx |
| @@ -321,6 +321,8 @@ av_cold void ff_h264dsp_init_x86(H264DSPContext *c, const int bit_depth, · 320 unmodified lines | ||
| 321 | 321 | c->h264_v_loop_filter_chroma = ff_deblock_v_chroma_8_avx; |
| 322 | 322 | if (chroma_format_idc <= 1) { |
| 323 | 323 | c->h264_h_loop_filter_chroma = ff_deblock_h_chroma_8_avx; |
| 324 | + } else { | |
| 325 | + c->h264_h_loop_filter_chroma = ff_deblock_h_chroma422_8_avx; | |
| 324 | 326 | } |
| 325 | 327 | } |
| 326 | 328 | } else if (bit_depth == 10) { |