shithub: openh264

Info  •  Files  •  Log  •  Branches

refine pNzc set access

--- a/codec/decoder/core/src/decode_slice.cpp

+++ b/codec/decoder/core/src/decode_slice.cpp

@@ -440,7 +440,8 @@

   int32_t iMbX = pCurLayer->iMbX;

   int32_t iMbY = pCurLayer->iMbY;

-  int32_t iMbXy = pCurLayer->iMbXyIndex;

+  const int32_t iMbXy = pCurLayer->iMbXyIndex;

+  int8_t *pNzc = pCurLayer->pNzc[iMbXy];

   int32_t i;

   uint32_t uiMbType = 0, uiCbp = 0, uiCbpL = 0, uiCbpC = 0;

   uint32_t uiCode;

@@ -505,7 +506,7 @@

     //step 3: update QP and pNonZeroCount

     pCurLayer->pLumaQp[iMbXy] = 0;

     pCurLayer->pChromaQp[iMbXy] = 0;

-    memset (pCurLayer->pNzc[iMbXy], 16, sizeof (pCurLayer->pNzc[iMbXy]));   //Rec. 9.2.1 for PCM, nzc=16

+    memset (pNzc, 16, sizeof (pCurLayer->pNzc[iMbXy]));   //Rec. 9.2.1 for PCM, nzc=16

     return 0;

   } else if (0 == uiMbType) { //reference to JM

     ENFORCE_STACK_ALIGN_1D (int8_t, pIntraPredMode, 48, 16);

@@ -540,12 +541,12 @@

   }

   memset (pCurLayer->pScaledTCoeff[iMbXy], 0, 384 * sizeof (pCurLayer->pScaledTCoeff[iMbXy][0]));

-  ST32A4 (&pCurLayer->pNzc[iMbXy][0], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][4], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][8], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][12], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][16], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][20], 0);

+  ST32A4 (&pNzc[0], 0);

+  ST32A4 (&pNzc[4], 0);

+  ST32A4 (&pNzc[8], 0);

+  ST32A4 (&pNzc[12], 0);

+  ST32A4 (&pNzc[16], 0);

+  ST32A4 (&pNzc[20], 0);

   if (pCurLayer->pCbp[iMbXy] == 0 && IS_INTRA4x4 (pCurLayer->pMbType[iMbXy])) {

     pCurLayer->pLumaQp[iMbXy] = pSlice->iLastMbQp;

@@ -599,10 +600,10 @@

             return -1;//abnormal

           }

         }

-        ST32A4 (&pCurLayer->pNzc[iMbXy][0], LD32 (&pNonZeroCount[1 + 8 * 1]));

-        ST32A4 (&pCurLayer->pNzc[iMbXy][4], LD32 (&pNonZeroCount[1 + 8 * 2]));

-        ST32A4 (&pCurLayer->pNzc[iMbXy][8], LD32 (&pNonZeroCount[1 + 8 * 3]));

-        ST32A4 (&pCurLayer->pNzc[iMbXy][12], LD32 (&pNonZeroCount[1 + 8 * 4]));

+        ST32A4 (&pNzc[0], LD32 (&pNonZeroCount[1 + 8 * 1]));

+        ST32A4 (&pNzc[4], LD32 (&pNonZeroCount[1 + 8 * 2]));

+        ST32A4 (&pNzc[8], LD32 (&pNonZeroCount[1 + 8 * 3]));

+        ST32A4 (&pNzc[12], LD32 (&pNonZeroCount[1 + 8 * 4]));

       }

     } else { //non-MB_TYPE_INTRA16x16

       for (iId8x8 = 0; iId8x8 < 4; iId8x8++) {

@@ -622,10 +623,10 @@

           ST16 (&pNonZeroCount[g_kuiCacheNzcScanIdx[ (iId8x8 << 2) + 2]], 0);

         }

       }

-      ST32A4 (&pCurLayer->pNzc[iMbXy][0], LD32 (&pNonZeroCount[1 + 8 * 1]));

-      ST32A4 (&pCurLayer->pNzc[iMbXy][4], LD32 (&pNonZeroCount[1 + 8 * 2]));

-      ST32A4 (&pCurLayer->pNzc[iMbXy][8], LD32 (&pNonZeroCount[1 + 8 * 3]));

-      ST32A4 (&pCurLayer->pNzc[iMbXy][12], LD32 (&pNonZeroCount[1 + 8 * 4]));

+      ST32A4 (&pNzc[0], LD32 (&pNonZeroCount[1 + 8 * 1]));

+      ST32A4 (&pNzc[4], LD32 (&pNonZeroCount[1 + 8 * 2]));

+      ST32A4 (&pNzc[8], LD32 (&pNonZeroCount[1 + 8 * 3]));

+      ST32A4 (&pNzc[12], LD32 (&pNonZeroCount[1 + 8 * 4]));

     }

     //chroma

@@ -653,10 +654,10 @@

           iIndex++;

         }

       }

-      ST16 (&pCurLayer->pNzc[iMbXy][16], LD16 (&pNonZeroCount[6 + 8 * 1]));

-      ST16 (&pCurLayer->pNzc[iMbXy][20], LD16 (&pNonZeroCount[6 + 8 * 2]));

-      ST16 (&pCurLayer->pNzc[iMbXy][18], LD16 (&pNonZeroCount[6 + 8 * 4]));

-      ST16 (&pCurLayer->pNzc[iMbXy][22], LD16 (&pNonZeroCount[6 + 8 * 5]));

+      ST16A2 (&pNzc[16], LD16A2 (&pNonZeroCount[6 + 8 * 1]));

+      ST16A2 (&pNzc[20], LD16A2 (&pNonZeroCount[6 + 8 * 2]));

+      ST16A2 (&pNzc[18], LD16A2 (&pNonZeroCount[6 + 8 * 4]));

+      ST16A2 (&pNzc[22], LD16A2 (&pNonZeroCount[6 + 8 * 5]));

     }

     BsEndCavlc (pBs);

   }

@@ -705,8 +706,8 @@

   int32_t iMbX = pCurLayer->iMbX;

   int32_t iMbY = pCurLayer->iMbY;

-  int32_t iMbXy = pCurLayer->iMbXyIndex;

-

+  const int32_t iMbXy = pCurLayer->iMbXyIndex;

+  int8_t* pNzc = pCurLayer->pNzc[iMbXy];

   int32_t i;

   uint32_t uiMbType = 0, uiCbp = 0, uiCbpL = 0, uiCbpC = 0;

   uint32_t uiCode;

@@ -796,12 +797,12 @@

       pCurLayer->pLumaQp[iMbXy] = 0;

       pCurLayer->pChromaQp[iMbXy] = 0;

       //Rec. 9.2.1 for PCM, nzc=16

-      ST32A4 (&pCurLayer->pNzc[iMbXy][0], 0x10101010);

-      ST32A4 (&pCurLayer->pNzc[iMbXy][4], 0x10101010);

-      ST32A4 (&pCurLayer->pNzc[iMbXy][8], 0x10101010);

-      ST32A4 (&pCurLayer->pNzc[iMbXy][12], 0x10101010);

-      ST32A4 (&pCurLayer->pNzc[iMbXy][16], 0x10101010);

-      ST32A4 (&pCurLayer->pNzc[iMbXy][20], 0x10101010);

+      ST32A4 (&pNzc[0], 0x10101010);

+      ST32A4 (&pNzc[4], 0x10101010);

+      ST32A4 (&pNzc[8], 0x10101010);

+      ST32A4 (&pNzc[12], 0x10101010);

+      ST32A4 (&pNzc[16], 0x10101010);

+      ST32A4 (&pNzc[20], 0x10101010);

       return 0;

     } else {

       if (0 == uiMbType) {

@@ -845,12 +846,12 @@

   memset(pCurLayer->pScaledTCoeff[iMbXy], 0, MB_COEFF_LIST_SIZE * sizeof(int16_t));

-  ST32A4 (&pCurLayer->pNzc[iMbXy][0], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][4], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][8], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][12], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][16], 0);

-  ST32A4 (&pCurLayer->pNzc[iMbXy][20], 0);

+  ST32A4 (&pNzc[0], 0);

+  ST32A4 (&pNzc[4], 0);

+  ST32A4 (&pNzc[8], 0);

+  ST32A4 (&pNzc[12], 0);

+  ST32A4 (&pNzc[16], 0);

+  ST32A4 (&pNzc[20], 0);

   if (pCurLayer->pCbp[iMbXy] == 0 && !IS_INTRA16x16 (pCurLayer->pMbType[iMbXy]) && !IS_I_BL (pCurLayer->pMbType[iMbXy])) {

     pCurLayer->pLumaQp[iMbXy] = pSlice->iLastMbQp;

     pCurLayer->pChromaQp[iMbXy] = g_kuiChromaQp[WELS_CLIP3 (pCurLayer->pLumaQp[iMbXy] +

@@ -901,10 +902,10 @@

             return -1;//abnormal

           }

         }

-        ST32A4 (&pCurLayer->pNzc[iMbXy][0], LD32 (&pNonZeroCount[1 + 8 * 1]));

-        ST32A4 (&pCurLayer->pNzc[iMbXy][4], LD32 (&pNonZeroCount[1 + 8 * 2]));

-        ST32A4 (&pCurLayer->pNzc[iMbXy][8], LD32 (&pNonZeroCount[1 + 8 * 3]));

-        ST32A4 (&pCurLayer->pNzc[iMbXy][12], LD32 (&pNonZeroCount[1 + 8 * 4]));

+        ST32A4 (&pNzc[0], LD32 (&pNonZeroCount[1 + 8 * 1]));

+        ST32A4 (&pNzc[4], LD32 (&pNonZeroCount[1 + 8 * 2]));

+        ST32A4 (&pNzc[8], LD32 (&pNonZeroCount[1 + 8 * 3]));

+        ST32A4 (&pNzc[12], LD32 (&pNonZeroCount[1 + 8 * 4]));

       }

     } else { //non-MB_TYPE_INTRA16x16

       for (iId8x8 = 0; iId8x8 < 4; iId8x8++) {

@@ -924,10 +925,10 @@

           ST16 (&pNonZeroCount[g_kuiCacheNzcScanIdx[ (iId8x8 << 2) + 2]], 0);

         }

       }

-      ST32A4 (&pCurLayer->pNzc[iMbXy][0], LD32 (&pNonZeroCount[1 + 8 * 1]));

-      ST32A4 (&pCurLayer->pNzc[iMbXy][4], LD32 (&pNonZeroCount[1 + 8 * 2]));

-      ST32A4 (&pCurLayer->pNzc[iMbXy][8], LD32 (&pNonZeroCount[1 + 8 * 3]));

-      ST32A4 (&pCurLayer->pNzc[iMbXy][12], LD32 (&pNonZeroCount[1 + 8 * 4]));

+      ST32A4 (&pNzc[0], LD32 (&pNonZeroCount[1 + 8 * 1]));

+      ST32A4 (&pNzc[4], LD32 (&pNonZeroCount[1 + 8 * 2]));

+      ST32A4 (&pNzc[8], LD32 (&pNonZeroCount[1 + 8 * 3]));

+      ST32A4 (&pNzc[12], LD32 (&pNonZeroCount[1 + 8 * 4]));

     }

@@ -956,10 +957,10 @@

           iIndex++;

         }

       }

-      ST16 (&pCurLayer->pNzc[iMbXy][16], LD16 (&pNonZeroCount[6 + 8 * 1]));

-      ST16 (&pCurLayer->pNzc[iMbXy][20], LD16 (&pNonZeroCount[6 + 8 * 2]));

-      ST16 (&pCurLayer->pNzc[iMbXy][18], LD16 (&pNonZeroCount[6 + 8 * 4]));

-      ST16 (&pCurLayer->pNzc[iMbXy][22], LD16 (&pNonZeroCount[6 + 8 * 5]));

+      ST16A2 (&pNzc[16], LD16A2 (&pNonZeroCount[6 + 8 * 1]));

+      ST16A2 (&pNzc[20], LD16A2 (&pNonZeroCount[6 + 8 * 2]));

+      ST16A2 (&pNzc[18], LD16A2 (&pNonZeroCount[6 + 8 * 4]));

+      ST16A2 (&pNzc[22], LD16A2 (&pNonZeroCount[6 + 8 * 5]));

     }

     BsEndCavlc (pBs);

   }

@@ -973,7 +974,8 @@

   PSlice pSlice			 = &pCurLayer->sLayerInfo.sSliceInLayer;

   PSliceHeader pSliceHeader		    = &pSlice->sSliceHeaderExt.sSliceHeader;

-  int32_t iMbXy = pCurLayer->iMbXyIndex;

+  const int32_t iMbXy = pCurLayer->iMbXyIndex;

+  int8_t *pNzc = pCurLayer->pNzc[iMbXy];

   int32_t iBaseModeFlag, i;

   int32_t iRet = 0; //should have the return value to indicate decoding error or not, It's NECESSARY--2010.4.15

   uint32_t uiCode;

@@ -990,12 +992,12 @@

     int16_t iMv[2];

     pCurLayer->pMbType[iMbXy] = MB_TYPE_SKIP;

-    ST32A4 (&pCurLayer->pNzc[iMbXy][0], 0);

-    ST32A4 (&pCurLayer->pNzc[iMbXy][4], 0);

-    ST32A4 (&pCurLayer->pNzc[iMbXy][8], 0);

-    ST32A4 (&pCurLayer->pNzc[iMbXy][12], 0);

-    ST32A4 (&pCurLayer->pNzc[iMbXy][16], 0);

-    ST32A4 (&pCurLayer->pNzc[iMbXy][20], 0);

+    ST32A4 (&pNzc[0], 0);

+    ST32A4 (&pNzc[4], 0);

+    ST32A4 (&pNzc[8], 0);

+    ST32A4 (&pNzc[12], 0);

+    ST32A4 (&pNzc[16], 0);

+    ST32A4 (&pNzc[20], 0);

     pCurLayer->pInterPredictionDoneFlag[iMbXy] = 0;

     memset (pCurLayer->pRefIndex[0][iMbXy], 0, sizeof (int8_t) * 16);