Showing the changes to codec/decoder/core/src/deblocking.cpp (1 of 25 files). Show all
Unified · Split
| @@ -136,6 +136,19 @@ static const uint8_t g_kuiTableBIdx[2][8] = { · 135 unmodified lines | |||
| 136 | }, | 136 | }, |
| 137 | }; | 137 | }; |
| 138 | 138 | ||
| 139 | static const uint8_t g_kuiTableB8x8Idx[2][16] = { | ||
| 140 | { | ||
| 141 | 0, 1, 4, 5, 8, 9, 12, 13, // 0 1 | 2 3 | ||
| 142 | 2, 3, 6, 7, 10, 11, 14, 15 // 4 5 | 6 7 | ||
| 143 | }, // ------------ | ||
| 144 | // 8 9 | 10 11 | ||
| 145 | { | ||
| 146 | // 12 13 | 14 15 | ||
| 147 | 0, 1, 4, 5, 2, 3, 6, 7, | ||
| 148 | 8, 9, 12, 13, 10, 11, 14, 15 | ||
| 149 | }, | ||
| 150 | }; | ||
| 151 | |||
| 139 | #define TC0_TBL_LOOKUP(tc, iIndexA, pBS, bChroma) \ | 152 | #define TC0_TBL_LOOKUP(tc, iIndexA, pBS, bChroma) \ |
| 140 | {\ | 153 | {\ |
| 141 | tc[0] = g_kiTc0Table(iIndexA)[pBS[0]] + bChroma;\ | 154 | tc[0] = g_kiTc0Table(iIndexA)[pBS[0]] + bChroma;\ |
| @@ -170,7 +183,22 @@ void inline DeblockingBSInsideMBAvsbase (int8_t* pNnzTab, uint8_t nBS[2][4][4], · 28 unmodified lines | |||
| 170 | nBS[0][2][3] = (pNnzTab[13] | pNnzTab[14]) << iLShiftFactor; | 183 | nBS[0][2][3] = (pNnzTab[13] | pNnzTab[14]) << iLShiftFactor; |
| 171 | nBS[0][3][3] = (pNnzTab[14] | pNnzTab[15]) << iLShiftFactor; | 184 | nBS[0][3][3] = (pNnzTab[14] | pNnzTab[15]) << iLShiftFactor; |
| 172 | * (uint32_t*)nBS[1][3] = (uiNnz32b2 | uiNnz32b3) << iLShiftFactor; | 185 | * (uint32_t*)nBS[1][3] = (uiNnz32b2 | uiNnz32b3) << iLShiftFactor; |
| 186 | } | ||
| 187 | |||
| 188 | void inline DeblockingBSInsideMBAvsbase8x8 (int8_t* pNnzTab, uint8_t nBS[2][4][4], int32_t iLShiftFactor) { | ||
| 189 | int8_t i8x8NnzTab[4]; | ||
| 190 | for (int32_t i = 0; i < 4; i++) { | ||
| 191 | int32_t iBlkIdx = i << 2; | ||
| 192 | i8x8NnzTab[i] = (pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx]] | pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx + 1]] | | ||
| 193 | pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx + 2]] | pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx + 3]]); | ||
| 194 | } | ||
| 173 | 195 | ||
| 196 | //vertical | ||
| 197 | nBS[0][2][0] = nBS[0][2][1] = (i8x8NnzTab[0] | i8x8NnzTab[1]) << iLShiftFactor; | ||
| 198 | nBS[0][2][2] = nBS[0][2][3] = (i8x8NnzTab[2] | i8x8NnzTab[3]) << iLShiftFactor; | ||
| 199 | //horizontal | ||
| 200 | nBS[1][2][0] = nBS[1][2][1] = (i8x8NnzTab[0] | i8x8NnzTab[2]) << iLShiftFactor; | ||
| 201 | nBS[1][2][2] = nBS[1][2][3] = (i8x8NnzTab[1] | i8x8NnzTab[3]) << iLShiftFactor; | ||
| 174 | } | 202 | } |
| 175 | 203 | ||
| 176 | void static inline DeblockingBSInsideMBNormal (PDqLayer pCurDqLayer, uint8_t nBS[2][4][4], int8_t* pNnzTab, | 204 | void static inline DeblockingBSInsideMBNormal (PDqLayer pCurDqLayer, uint8_t nBS[2][4][4], int8_t* pNnzTab, |
| @@ -179,73 +207,148 @@ void static inline DeblockingBSInsideMBNormal (PDqLayer pCurDqLayer, uint8_t nBS · 2 unmodified lines | |||
| 179 | int8_t* iRefIndex = pCurDqLayer->pRefIndex[LIST_0][iMbXy]; | 207 | int8_t* iRefIndex = pCurDqLayer->pRefIndex[LIST_0][iMbXy]; |
| 180 | ENFORCE_STACK_ALIGN_1D (uint8_t, uiBsx4, 4, 4); | 208 | ENFORCE_STACK_ALIGN_1D (uint8_t, uiBsx4, 4, 4); |
| 181 | 209 | ||
| 182 | uiNnz32b0 = * (uint32_t*) (pNnzTab + 0); | 210 | int8_t i8x8NnzTab[4]; |
| 183 | uiNnz32b1 = * (uint32_t*) (pNnzTab + 4); | ||
| 184 | uiNnz32b2 = * (uint32_t*) (pNnzTab + 8); | ||
| 185 | uiNnz32b3 = * (uint32_t*) (pNnzTab + 12); | ||
| 186 | 211 | ||
| 187 | for (int i = 0; i < 3; i++) | 212 | if (pCurDqLayer->pTransformSize8x8Flag[iMbXy]) { |
| 188 | uiBsx4[i] = pNnzTab[i] | pNnzTab[i + 1]; | 213 | for (int32_t i = 0; i < 4; i++) { |
| 189 | nBS[0][1][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 1, 0); | 214 | int32_t iBlkIdx = i << 2; |
| 190 | nBS[0][2][0] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 2, 1); | 215 | i8x8NnzTab[i] = (pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx]] | pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx + 1]] | |
| 191 | nBS[0][3][0] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 3, 2); | 216 | pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx + 2]] | pNnzTab[g_kuiMbCountScan4Idx[iBlkIdx + 3]]); |
| 192 | 217 | } | |
| 193 | for (int i = 0; i < 3; i++) | 218 | //vertical |
| 194 | uiBsx4[i] = pNnzTab[4 + i] | pNnzTab[4 + i + 1]; | 219 | nBS[0][2][0] = nBS[0][2][1] = BS_EDGE ((i8x8NnzTab[0] | i8x8NnzTab[1]), iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], |
| 195 | nBS[0][1][1] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 5, 4); | 220 | g_kuiMbCountScan4Idx[1 << 2], g_kuiMbCountScan4Idx[0]); |
| 196 | nBS[0][2][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 6, 5); | 221 | nBS[0][2][2] = nBS[0][2][3] = BS_EDGE ((i8x8NnzTab[2] | i8x8NnzTab[3]), iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], |
| 197 | nBS[0][3][1] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 7, 6); | 222 | g_kuiMbCountScan4Idx[3 << 2], g_kuiMbCountScan4Idx[2 << 2]); |
| 198 | 223 | ||
| 199 | for (int i = 0; i < 3; i++) | 224 | //horizontal |
| 200 | uiBsx4[i] = pNnzTab[8 + i] | pNnzTab[8 + i + 1]; | 225 | nBS[1][2][0] = nBS[1][2][1] = BS_EDGE ((i8x8NnzTab[0] | i8x8NnzTab[2]), iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], |
| 201 | nBS[0][1][2] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 9, 8); | 226 | g_kuiMbCountScan4Idx[2 << 2], g_kuiMbCountScan4Idx[0]); |
| 202 | nBS[0][2][2] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 10, 9); | 227 | nBS[1][2][2] = nBS[1][2][3] = BS_EDGE ((i8x8NnzTab[1] | i8x8NnzTab[3]), iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], |
| 203 | nBS[0][3][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 11, 10); | 228 | g_kuiMbCountScan4Idx[3 << 2], g_kuiMbCountScan4Idx[1 << 2]); |
| 204 | 229 | } else { | |
| 205 | for (int i = 0; i < 3; i++) | 230 | uiNnz32b0 = * (uint32_t*) (pNnzTab + 0); |
| 206 | uiBsx4[i] = pNnzTab[12 + i] | pNnzTab[12 + i + 1]; | 231 | uiNnz32b1 = * (uint32_t*) (pNnzTab + 4); |
| 207 | nBS[0][1][3] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 13, 12); | 232 | uiNnz32b2 = * (uint32_t*) (pNnzTab + 8); |
| 208 | nBS[0][2][3] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 14, 13); | 233 | uiNnz32b3 = * (uint32_t*) (pNnzTab + 12); |
| 209 | nBS[0][3][3] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 15, 14); | 234 | |
| 210 | 235 | for (int i = 0; i < 3; i++) | |
| 211 | // horizontal | 236 | uiBsx4[i] = pNnzTab[i] | pNnzTab[i + 1]; |
| 212 | * (uint32_t*)uiBsx4 = (uiNnz32b0 | uiNnz32b1); | 237 | nBS[0][1][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 1, 0); |
| 213 | nBS[1][1][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 4, 0); | 238 | nBS[0][2][0] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 2, 1); |
| 214 | nBS[1][1][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 5, 1); | 239 | nBS[0][3][0] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 3, 2); |
| 215 | nBS[1][1][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 6, 2); | 240 | |
| 216 | nBS[1][1][3] = BS_EDGE (uiBsx4[3], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 7, 3); | 241 | for (int i = 0; i < 3; i++) |
| 217 | 242 | uiBsx4[i] = pNnzTab[4 + i] | pNnzTab[4 + i + 1]; | |
| 218 | * (uint32_t*)uiBsx4 = (uiNnz32b1 | uiNnz32b2); | 243 | nBS[0][1][1] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 5, 4); |
| 219 | nBS[1][2][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 8, 4); | 244 | nBS[0][2][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 6, 5); |
| 220 | nBS[1][2][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 9, 5); | 245 | nBS[0][3][1] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 7, 6); |
| 221 | nBS[1][2][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 10, 6); | 246 | |
| 222 | nBS[1][2][3] = BS_EDGE (uiBsx4[3], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 11, 7); | 247 | for (int i = 0; i < 3; i++) |
| 223 | 248 | uiBsx4[i] = pNnzTab[8 + i] | pNnzTab[8 + i + 1]; | |
| 224 | * (uint32_t*)uiBsx4 = (uiNnz32b2 | uiNnz32b3); | 249 | nBS[0][1][2] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 9, 8); |
| 225 | nBS[1][3][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 12, 8); | 250 | nBS[0][2][2] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 10, 9); |
| 226 | nBS[1][3][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 13, 9); | 251 | nBS[0][3][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 11, 10); |
| 227 | nBS[1][3][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 14, 10); | 252 | |
| 228 | nBS[1][3][3] = BS_EDGE (uiBsx4[3], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 15, 11); | 253 | for (int i = 0; i < 3; i++) |
| 254 | uiBsx4[i] = pNnzTab[12 + i] | pNnzTab[12 + i + 1]; | ||
| 255 | nBS[0][1][3] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 13, 12); | ||
| 256 | nBS[0][2][3] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 14, 13); | ||
| 257 | nBS[0][3][3] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 15, 14); | ||
| 258 | |||
| 259 | // horizontal | ||
| 260 | * (uint32_t*)uiBsx4 = (uiNnz32b0 | uiNnz32b1); | ||
| 261 | nBS[1][1][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 4, 0); | ||
| 262 | nBS[1][1][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 5, 1); | ||
| 263 | nBS[1][1][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 6, 2); | ||
| 264 | nBS[1][1][3] = BS_EDGE (uiBsx4[3], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 7, 3); | ||
| 265 | |||
| 266 | * (uint32_t*)uiBsx4 = (uiNnz32b1 | uiNnz32b2); | ||
| 267 | nBS[1][2][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 8, 4); | ||
| 268 | nBS[1][2][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 9, 5); | ||
| 269 | nBS[1][2][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 10, 6); | ||
| 270 | nBS[1][2][3] = BS_EDGE (uiBsx4[3], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 11, 7); | ||
| 271 | |||
| 272 | * (uint32_t*)uiBsx4 = (uiNnz32b2 | uiNnz32b3); | ||
| 273 | nBS[1][3][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 12, 8); | ||
| 274 | nBS[1][3][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 13, 9); | ||
| 275 | nBS[1][3][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 14, 10); | ||
| 276 | nBS[1][3][3] = BS_EDGE (uiBsx4[3], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 15, 11); | ||
| 277 | } | ||
| 229 | } | 278 | } |
| 230 | 279 | ||
| 231 | uint32_t DeblockingBsMarginalMBAvcbase (PDqLayer pCurDqLayer, int32_t iEdge, int32_t iNeighMb, int32_t iMbXy) { | 280 | uint32_t DeblockingBsMarginalMBAvcbase (PDqLayer pCurDqLayer, int32_t iEdge, int32_t iNeighMb, int32_t iMbXy) { |
| 232 | int32_t i; | 281 | int32_t i, j; |
| 233 | uint32_t uiBSx4; | 282 | uint32_t uiBSx4; |
| 234 | //uint8_t* bS = static_cast<uint8_t*>(&uiBSx4); | ||
| 235 | uint8_t* pBS = (uint8_t*) (&uiBSx4); | 283 | uint8_t* pBS = (uint8_t*) (&uiBSx4); |
| 236 | const uint8_t* pBIdx = &g_kuiTableBIdx[iEdge][0]; | 284 | const uint8_t* pBIdx = &g_kuiTableBIdx[iEdge][0]; |
| 237 | const uint8_t* pBnIdx = &g_kuiTableBIdx[iEdge][4]; | 285 | const uint8_t* pBnIdx = &g_kuiTableBIdx[iEdge][4]; |
| 238 | 286 | const uint8_t* pB8x8Idx = &g_kuiTableB8x8Idx[iEdge][0]; | |
| 239 | for (i = 0; i < 4; i++) { | 287 | const uint8_t* pBn8x8Idx = &g_kuiTableB8x8Idx[iEdge][8]; |
| 240 | if (pCurDqLayer->pNzc[iMbXy][*pBIdx] | pCurDqLayer->pNzc[iNeighMb][*pBnIdx]) { | 288 | |
| 241 | pBS[i] = 2; | 289 | if (pCurDqLayer->pTransformSize8x8Flag[iMbXy] && pCurDqLayer->pTransformSize8x8Flag[iNeighMb]) { |
| 242 | } else { | 290 | for (i = 0; i < 2; i++) { |
| 243 | pBS[i] = MB_BS_MV (pCurDqLayer->pRefIndex[LIST_0], pCurDqLayer->pMv[LIST_0], iMbXy, iNeighMb, *pBIdx, | 291 | uint8_t uiNzc = 0; |
| 244 | *pBnIdx); | 292 | for (j = 0; uiNzc == 0 && j < 4; j++) { |
| 293 | uiNzc |= (pCurDqLayer->pNzc[iMbXy][* (pB8x8Idx + j)] | pCurDqLayer->pNzc[iNeighMb][* (pBn8x8Idx + j)]); | ||
| 294 | } | ||
| 295 | if (uiNzc) { | ||
| 296 | pBS[i << 1] = pBS[1 + (i << 1)] = 2; | ||
| 297 | } else { | ||
| 298 | pBS[i << 1] = pBS[1 + (i << 1)] = MB_BS_MV (pCurDqLayer->pRefIndex[LIST_0], pCurDqLayer->pMv[LIST_0], iMbXy, iNeighMb, | ||
| 299 | *pB8x8Idx, *pBn8x8Idx); | ||
| 300 | } | ||
| 301 | pB8x8Idx += 4; | ||
| 302 | pBn8x8Idx += 4; | ||
| 303 | } | ||
| 304 | } else if (pCurDqLayer->pTransformSize8x8Flag[iMbXy]) { | ||
| 305 | for (i = 0; i < 2; i++) { | ||
| 306 | uint8_t uiNzc = 0; | ||
| 307 | for (j = 0; uiNzc == 0 && j < 4; j++) { | ||
| 308 | uiNzc |= pCurDqLayer->pNzc[iMbXy][* (pB8x8Idx + j)]; | ||
| 309 | } | ||
| 310 | for (j = 0; j < 2; j++) { | ||
| 311 | if (uiNzc | pCurDqLayer->pNzc[iNeighMb][*pBnIdx]) { | ||
| 312 | pBS[j + (i << 1)] = 2; | ||
| 313 | } else { | ||
| 314 | pBS[j + (i << 1)] = MB_BS_MV (pCurDqLayer->pRefIndex[LIST_0], pCurDqLayer->pMv[LIST_0], iMbXy, iNeighMb, *pB8x8Idx, | ||
| 315 | *pBnIdx); | ||
| 316 | } | ||
| 317 | pBnIdx++; | ||
| 318 | } | ||
| 319 | pB8x8Idx += 4; | ||
| 320 | } | ||
| 321 | } else if (pCurDqLayer->pTransformSize8x8Flag[iNeighMb]) { | ||
| 322 | for (i = 0; i < 2; i++) { | ||
| 323 | uint8_t uiNzc = 0; | ||
| 324 | for (j = 0; uiNzc == 0 && j < 4; j++) { | ||
| 325 | uiNzc |= pCurDqLayer->pNzc[iNeighMb][* (pBn8x8Idx + j)]; | ||
| 326 | } | ||
| 327 | for (j = 0; j < 2; j++) { | ||
| 328 | if (uiNzc | pCurDqLayer->pNzc[iMbXy][*pBIdx]) { | ||
| 329 | pBS[j + (i << 1)] = 2; | ||
| 330 | } else { | ||
| 331 | pBS[j + (i << 1)] = MB_BS_MV (pCurDqLayer->pRefIndex[LIST_0], pCurDqLayer->pMv[LIST_0], iMbXy, iNeighMb, *pBIdx, | ||
| 332 | *pBn8x8Idx); | ||
| 333 | } | ||
| 334 | pBIdx++; | ||
| 335 | } | ||
| 336 | pBn8x8Idx += 4; | ||
| 337 | } | ||
| 338 | } else { | ||
| 339 | // only 4x4 transform | ||
| 340 | for (i = 0; i < 4; i++) { | ||
| 341 | if (pCurDqLayer->pNzc[iMbXy][*pBIdx] | pCurDqLayer->pNzc[iNeighMb][*pBnIdx]) { | ||
| 342 | pBS[i] = 2; | ||
| 343 | } else { | ||
| 344 | pBS[i] = MB_BS_MV (pCurDqLayer->pRefIndex[LIST_0], pCurDqLayer->pMv[LIST_0], iMbXy, iNeighMb, *pBIdx, | ||
| 345 | *pBnIdx); | ||
| 346 | } | ||
| 347 | pBIdx++; | ||
| 348 | pBnIdx++; | ||
| 245 | } | 349 | } |
| 246 | pBIdx++; | ||
| 247 | pBnIdx++; | ||
| 248 | } | 350 | } |
| 351 | |||
| 249 | return uiBSx4; | 352 | return uiBSx4; |
| 250 | } | 353 | } |
| 251 | int32_t DeblockingAvailableNoInterlayer (PDqLayer pCurDqLayer, int32_t iFilterIdc) { | 354 | int32_t DeblockingAvailableNoInterlayer (PDqLayer pCurDqLayer, int32_t iFilterIdc) { |
| @@ -501,7 +604,7 @@ void DeblockingInterMb (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, uint8_ · 249 unmodified lines | |||
| 501 | pFilter->iChromaQP[0] = pCurChromaQp[0]; | 604 | pFilter->iChromaQP[0] = pCurChromaQp[0]; |
| 502 | pFilter->iChromaQP[1] = pCurChromaQp[1]; | 605 | pFilter->iChromaQP[1] = pCurChromaQp[1]; |
| 503 | 606 | ||
| 504 | if (* (uint32_t*)nBS[0][1] != 0) { | 607 | if (* (uint32_t*)nBS[0][1] != 0 && !pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { |
| 505 | FilteringEdgeLumaV (pFilter, &pDestY[1 << 2], iLineSize, nBS[0][1]); | 608 | FilteringEdgeLumaV (pFilter, &pDestY[1 << 2], iLineSize, nBS[0][1]); |
| 506 | } | 609 | } |
| 507 | 610 | ||
| @@ -510,7 +613,7 @@ void DeblockingInterMb (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, uint8_ · 2 unmodified lines | |||
| 510 | FilteringEdgeChromaV (pFilter, &pDestCb[2 << 1], &pDestCr[2 << 1], iLineSizeUV, nBS[0][2]); | 613 | FilteringEdgeChromaV (pFilter, &pDestCb[2 << 1], &pDestCr[2 << 1], iLineSizeUV, nBS[0][2]); |
| 511 | } | 614 | } |
| 512 | 615 | ||
| 513 | if (* (uint32_t*)nBS[0][3] != 0) { | 616 | if (* (uint32_t*)nBS[0][3] != 0 && !pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { |
| 514 | FilteringEdgeLumaV (pFilter, &pDestY[3 << 2], iLineSize, nBS[0][3]); | 617 | FilteringEdgeLumaV (pFilter, &pDestY[3 << 2], iLineSize, nBS[0][3]); |
| 515 | } | 618 | } |
| 516 | 619 | ||
| @@ -536,7 +639,7 @@ void DeblockingInterMb (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, uint8_ · 19 unmodified lines | |||
| 536 | pFilter->iChromaQP[0] = pCurChromaQp[0]; | 639 | pFilter->iChromaQP[0] = pCurChromaQp[0]; |
| 537 | pFilter->iChromaQP[1] = pCurChromaQp[1]; | 640 | pFilter->iChromaQP[1] = pCurChromaQp[1]; |
| 538 | 641 | ||
| 539 | if (* (uint32_t*)nBS[1][1] != 0) { | 642 | if (* (uint32_t*)nBS[1][1] != 0 && !pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { |
| 540 | FilteringEdgeLumaH (pFilter, &pDestY[ (1 << 2)*iLineSize], iLineSize, nBS[1][1]); | 643 | FilteringEdgeLumaH (pFilter, &pDestY[ (1 << 2)*iLineSize], iLineSize, nBS[1][1]); |
| 541 | } | 644 | } |
| 542 | 645 | ||
| @@ -546,7 +649,7 @@ void DeblockingInterMb (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, uint8_ · 3 unmodified lines | |||
| 546 | nBS[1][2]); | 649 | nBS[1][2]); |
| 547 | } | 650 | } |
| 548 | 651 | ||
| 549 | if (* (uint32_t*)nBS[1][3] != 0) { | 652 | if (* (uint32_t*)nBS[1][3] != 0 && !pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { |
| 550 | FilteringEdgeLumaH (pFilter, &pDestY[ (3 << 2)*iLineSize], iLineSize, nBS[1][3]); | 653 | FilteringEdgeLumaH (pFilter, &pDestY[ (3 << 2)*iLineSize], iLineSize, nBS[1][3]); |
| 551 | } | 654 | } |
| 552 | } | 655 | } |
| @@ -581,9 +684,16 @@ void FilteringEdgeLumaHV (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, int3 · 28 unmodified lines | |||
| 581 | iBeta); | 684 | iBeta); |
| 582 | if (iAlpha | iBeta) { | 685 | if (iAlpha | iBeta) { |
| 583 | TC0_TBL_LOOKUP (iTc, iIndexA, uiBSx4, 0); | 686 | TC0_TBL_LOOKUP (iTc, iIndexA, uiBSx4, 0); |
| 584 | pFilter->pLoopf->pfLumaDeblockingLT4Hor (&pDestY[1 << 2], iLineSize, iAlpha, iBeta, iTc); | 687 | |
| 688 | if (!pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { | ||
| 689 | pFilter->pLoopf->pfLumaDeblockingLT4Hor (&pDestY[1 << 2], iLineSize, iAlpha, iBeta, iTc); | ||
| 690 | } | ||
| 691 | |||
| 585 | pFilter->pLoopf->pfLumaDeblockingLT4Hor (&pDestY[2 << 2], iLineSize, iAlpha, iBeta, iTc); | 692 | pFilter->pLoopf->pfLumaDeblockingLT4Hor (&pDestY[2 << 2], iLineSize, iAlpha, iBeta, iTc); |
| 586 | pFilter->pLoopf->pfLumaDeblockingLT4Hor (&pDestY[3 << 2], iLineSize, iAlpha, iBeta, iTc); | 693 | |
| 694 | if (!pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { | ||
| 695 | pFilter->pLoopf->pfLumaDeblockingLT4Hor (&pDestY[3 << 2], iLineSize, iAlpha, iBeta, iTc); | ||
| 696 | } | ||
| 587 | } | 697 | } |
| 588 | 698 | ||
| 589 | // luma h | 699 | // luma h |
| @@ -594,9 +704,15 @@ void FilteringEdgeLumaHV (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, int3 · 4 unmodified lines | |||
| 594 | 704 | ||
| 595 | pFilter->iLumaQP = iCurQp; | 705 | pFilter->iLumaQP = iCurQp; |
| 596 | if (iAlpha | iBeta) { | 706 | if (iAlpha | iBeta) { |
| 597 | pFilter->pLoopf->pfLumaDeblockingLT4Ver (&pDestY[ (1 << 2)*iLineSize], iLineSize, iAlpha, iBeta, iTc); | 707 | if (!pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { |
| 708 | pFilter->pLoopf->pfLumaDeblockingLT4Ver (&pDestY[ (1 << 2)*iLineSize], iLineSize, iAlpha, iBeta, iTc); | ||
| 709 | } | ||
| 710 | |||
| 598 | pFilter->pLoopf->pfLumaDeblockingLT4Ver (&pDestY[ (2 << 2)*iLineSize], iLineSize, iAlpha, iBeta, iTc); | 711 | pFilter->pLoopf->pfLumaDeblockingLT4Ver (&pDestY[ (2 << 2)*iLineSize], iLineSize, iAlpha, iBeta, iTc); |
| 599 | pFilter->pLoopf->pfLumaDeblockingLT4Ver (&pDestY[ (3 << 2)*iLineSize], iLineSize, iAlpha, iBeta, iTc); | 712 | |
| 713 | if (!pCurDqLayer->pTransformSize8x8Flag[iMbXyIndex]) { | ||
| 714 | pFilter->pLoopf->pfLumaDeblockingLT4Ver (&pDestY[ (3 << 2)*iLineSize], iLineSize, iAlpha, iBeta, iTc); | ||
| 715 | } | ||
| 600 | } | 716 | } |
| 601 | } | 717 | } |
| 602 | void FilteringEdgeChromaHV (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, int32_t iBoundryFlag) { | 718 | void FilteringEdgeChromaHV (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, int32_t iBoundryFlag) { |
| @@ -705,6 +821,7 @@ void WelsDeblockingMb (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, int32_t · 102 unmodified lines | |||
| 705 | 821 | ||
| 706 | switch (iCurMbType) { | 822 | switch (iCurMbType) { |
| 707 | case MB_TYPE_INTRA4x4: | 823 | case MB_TYPE_INTRA4x4: |
| 824 | case MB_TYPE_INTRA8x8: | ||
| 708 | case MB_TYPE_INTRA16x16: | 825 | case MB_TYPE_INTRA16x16: |
| 709 | case MB_TYPE_INTRA_PCM: | 826 | case MB_TYPE_INTRA_PCM: |
| 710 | DeblockingIntraMb (pCurDqLayer, pFilter, iBoundryFlag); | 827 | DeblockingIntraMb (pCurDqLayer, pFilter, iBoundryFlag); |
| @@ -728,7 +845,11 @@ void WelsDeblockingMb (PDqLayer pCurDqLayer, PDeblockingFilter pFilter, int32_t · 17 unmodified lines | |||
| 728 | //SKIP MB_16x16 or others | 845 | //SKIP MB_16x16 or others |
| 729 | if (iCurMbType != MB_TYPE_SKIP) { | 846 | if (iCurMbType != MB_TYPE_SKIP) { |
| 730 | if (iCurMbType == MB_TYPE_16x16) { | 847 | if (iCurMbType == MB_TYPE_16x16) { |
| 731 | DeblockingBSInsideMBAvsbase (pCurDqLayer->pNzc[iMbXyIndex], nBS, 1); | 848 | if (!pCurDqLayer->pTransformSize8x8Flag[pCurDqLayer->iMbXyIndex]) { |
| 849 | DeblockingBSInsideMBAvsbase (pCurDqLayer->pNzc[iMbXyIndex], nBS, 1); | ||
| 850 | } else { | ||
| 851 | DeblockingBSInsideMBAvsbase8x8 (pCurDqLayer->pNzc[iMbXyIndex], nBS, 1); | ||
| 852 | } | ||
| 732 | } else { | 853 | } else { |
| 733 | DeblockingBSInsideMBNormal (pCurDqLayer, nBS, pCurDqLayer->pNzc[iMbXyIndex], iMbXyIndex); | 854 | DeblockingBSInsideMBNormal (pCurDqLayer, nBS, pCurDqLayer->pNzc[iMbXyIndex], iMbXyIndex); |
| 734 | } | 855 | } |
| @@ -839,8 +960,8 @@ void DeblockingInit (SDeblockingFunc* pFunc, int32_t iCpu) { · 104 unmodified lines | |||
| 839 | if (iCpu & WELS_CPU_SSSE3) { | 960 | if (iCpu & WELS_CPU_SSSE3) { |
| 840 | pFunc->pfLumaDeblockingLT4Ver = DeblockLumaLt4V_ssse3; | 961 | pFunc->pfLumaDeblockingLT4Ver = DeblockLumaLt4V_ssse3; |
| 841 | pFunc->pfLumaDeblockingEQ4Ver = DeblockLumaEq4V_ssse3; | 962 | pFunc->pfLumaDeblockingEQ4Ver = DeblockLumaEq4V_ssse3; |
| 842 | pFunc->pfLumaDeblockingLT4Hor = DeblockLumaLt4H_ssse3; | 963 | pFunc->pfLumaDeblockingLT4Hor = DeblockLumaLt4H_ssse3; |
| 843 | pFunc->pfLumaDeblockingEQ4Hor = DeblockLumaEq4H_ssse3; | 964 | pFunc->pfLumaDeblockingEQ4Hor = DeblockLumaEq4H_ssse3; |
| 844 | pFunc->pfChromaDeblockingLT4Ver = DeblockChromaLt4V_ssse3; | 965 | pFunc->pfChromaDeblockingLT4Ver = DeblockChromaLt4V_ssse3; |
| 845 | pFunc->pfChromaDeblockingEQ4Ver = DeblockChromaEq4V_ssse3; | 966 | pFunc->pfChromaDeblockingEQ4Ver = DeblockChromaEq4V_ssse3; |
| 846 | pFunc->pfChromaDeblockingLT4Hor = DeblockChromaLt4H_ssse3; | 967 | pFunc->pfChromaDeblockingLT4Hor = DeblockChromaLt4H_ssse3; |