shithub: openh264

Info  •  Files  •  Log  •  Branches

Merge pull request #256 from mstorsjo/bigendian

Make the library work on big endian

--- a/codec/common/deblocking_common.h

+++ b/codec/common/deblocking_common.h

@@ -15,6 +15,12 @@

                             int8_t* pTc);

 void_t DeblockChromaEq4H_c (uint8_t* pPixCb, uint8_t* pPixCr, int32_t iStride, int32_t iAlpha, int32_t iBeta);

+#ifdef WORDS_BIGENDIAN

+#define DEBLOCK_BS_SHIFTED(x) ((x) | ((x) << 8))

+#else

+#define DEBLOCK_BS_SHIFTED(x) ((x) | ((x) >> 8))

+#endif

+

 #if defined(__cplusplus)

 extern "C" {

 #endif//__cplusplus

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

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

@@ -152,24 +152,24 @@

   uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);

   uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);

-  * (uint32_t*)uiBsx3 = (uiNnz32b0 | (uiNnz32b0 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b0) << iLShiftFactor;

   nBS[0][1][0] = uiBsx3[0];

   nBS[0][2][0] = uiBsx3[1];

   nBS[0][3][0] = uiBsx3[2];

-  * (uint32_t*)uiBsx3 = (uiNnz32b1 | (uiNnz32b1 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b1) << iLShiftFactor;

   nBS[0][1][1] = uiBsx3[0];

   nBS[0][2][1] = uiBsx3[1];

   nBS[0][3][1] = uiBsx3[2];

   * (uint32_t*)nBS[1][1] = (uiNnz32b0 | uiNnz32b1) << iLShiftFactor;

-  * (uint32_t*)uiBsx3 = (uiNnz32b2 | (uiNnz32b2 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b2) << iLShiftFactor;

   nBS[0][1][2] = uiBsx3[0];

   nBS[0][2][2] = uiBsx3[1];

   nBS[0][3][2] = uiBsx3[2];

   * (uint32_t*)nBS[1][2] = (uiNnz32b1 | uiNnz32b2) << iLShiftFactor;

-  * (uint32_t*)uiBsx3 = (uiNnz32b3 | (uiNnz32b3 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b3) << iLShiftFactor;

   nBS[0][1][3] = uiBsx3[0];

   nBS[0][2][3] = uiBsx3[1];

   nBS[0][3][3] = uiBsx3[2];

@@ -188,22 +188,22 @@

   uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);

   uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);

-  * (uint32_t*)uiBsx4 = (uiNnz32b0 | (uiNnz32b0 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b0);

   nBS[0][1][0] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 1, 0);

   nBS[0][2][0] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 2, 1);

   nBS[0][3][0] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 3, 2);

-  * (uint32_t*)uiBsx4 = (uiNnz32b1 | (uiNnz32b1 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b1);

   nBS[0][1][1] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 5, 4);

   nBS[0][2][1] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 6, 5);

   nBS[0][3][1] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 7, 6);

-  * (uint32_t*)uiBsx4 = (uiNnz32b2 | (uiNnz32b2 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b2);

   nBS[0][1][2] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 9, 8);

   nBS[0][2][2] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 10, 9);

   nBS[0][3][2] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 11, 10);

-  * (uint32_t*)uiBsx4 = (uiNnz32b3 | (uiNnz32b3 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b3);

   nBS[0][1][3] = BS_EDGE (uiBsx4[0], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 13, 12);

   nBS[0][2][3] = BS_EDGE (uiBsx4[1], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 14, 13);

   nBS[0][3][3] = BS_EDGE (uiBsx4[2], iRefIndex, pCurDqLayer->pMv[LIST_0][iMbXy], 15, 14);

@@ -233,18 +233,18 @@

   uint32_t uiBSx4;

   //uint8_t* bS = static_cast<uint8_t*>(&uiBSx4);

   uint8_t* pBS = (uint8_t*) (&uiBSx4);

-  uint32_t uiBIdx  = * (uint32_t*) (&g_kuiTableBIdx[iEdge][0]);

-  uint32_t uiBnIdx = * (uint32_t*) (&g_kuiTableBIdx[iEdge][4]);

+  const uint8_t *pBIdx  = &g_kuiTableBIdx[iEdge][0];

+  const uint8_t *pBnIdx = &g_kuiTableBIdx[iEdge][4];

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

-    if (pCurDqLayer->pNzc[iMbXy][uiBIdx & 0xff] | pCurDqLayer->pNzc[iNeighMb][uiBnIdx & 0xff]) {

+    if (pCurDqLayer->pNzc[iMbXy][*pBIdx] | pCurDqLayer->pNzc[iNeighMb][*pBnIdx]) {

       pBS[i] = 2;

     } else {

-      pBS[i] = MB_BS_MV (pCurDqLayer->pRefIndex[LIST_0], pCurDqLayer->pMv[LIST_0], iMbXy, iNeighMb, (uiBIdx & 0xff),

-                         (uiBnIdx & 0xff));

+      pBS[i] = MB_BS_MV (pCurDqLayer->pRefIndex[LIST_0], pCurDqLayer->pMv[LIST_0], iMbXy, iNeighMb, *pBIdx,

+                         *pBnIdx);

     }

-    uiBIdx  = uiBIdx  >> 8;

-    uiBnIdx = uiBnIdx >> 8;

+    pBIdx++;

+    pBnIdx++;

   }

   return uiBSx4;

 }

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

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

@@ -326,11 +326,11 @@

   int32_t i, j;

   int32_t iA, iB, iC, iD;

   uint8_t* pSrcNext = pSrc + iSrcStride;

-  const uint32_t kuiABCD = * ((uint32_t*)g_kuiABCD[iMvY & 0x07][iMvX & 0x07]);

-  iA = (kuiABCD) & 0xff;

-  iB = (kuiABCD >>  8) & 0xff;

-  iC = (kuiABCD >> 16) & 0xff;

-  iD = (kuiABCD >> 24) & 0xff;

+  const uint8_t *pABCD = g_kuiABCD[iMvY & 0x07][iMvX & 0x07];

+  iA = pABCD[0];

+  iB = pABCD[1];

+  iC = pABCD[2];

+  iD = pABCD[3];

   for (i = 0; i < iHeight; i++) {

     for (j = 0; j < iWidth; j++) {

       pDst[j] = (iA * pSrc[j] + iB * pSrc[j + 1] + iC * pSrcNext[j] + iD * pSrcNext[j + 1] + 32) >> 6;

--- a/codec/encoder/core/src/deblocking.cpp

+++ b/codec/encoder/core/src/deblocking.cpp

@@ -163,24 +163,24 @@

   uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);

   uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);

-  * (uint32_t*)uiBsx3 = (uiNnz32b0 | (uiNnz32b0 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b0) << iLShiftFactor;

   uiBS[0][1][0] = uiBsx3[0];

   uiBS[0][2][0] = uiBsx3[1];

   uiBS[0][3][0] = uiBsx3[2];

-  * (uint32_t*)uiBsx3 = (uiNnz32b1 | (uiNnz32b1 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b1) << iLShiftFactor;

   uiBS[0][1][1] = uiBsx3[0];

   uiBS[0][2][1] = uiBsx3[1];

   uiBS[0][3][1] = uiBsx3[2];

   * (uint32_t*)uiBS[1][1] = (uiNnz32b0 | uiNnz32b1) << iLShiftFactor;

-  * (uint32_t*)uiBsx3 = (uiNnz32b2 | (uiNnz32b2 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b2) << iLShiftFactor;

   uiBS[0][1][2] = uiBsx3[0];

   uiBS[0][2][2] = uiBsx3[1];

   uiBS[0][3][2] = uiBsx3[2];

   * (uint32_t*)uiBS[1][2] = (uiNnz32b1 | uiNnz32b2) << iLShiftFactor;

-  * (uint32_t*)uiBsx3 = (uiNnz32b3 | (uiNnz32b3 >> 8)) << iLShiftFactor;

+  * (uint32_t*)uiBsx3 = DEBLOCK_BS_SHIFTED (uiNnz32b3) << iLShiftFactor;

   uiBS[0][1][3] = uiBsx3[0];

   uiBS[0][2][3] = uiBsx3[1];

   uiBS[0][3][3] = uiBsx3[2];

@@ -197,22 +197,22 @@

   uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);

   uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);

-  * (uint32_t*)uiBsx4 = (uiNnz32b0 | (uiNnz32b0 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b0);

   uiBS[0][1][0] = BS_EDGE (uiBsx4[0], iRefIdx, pCurMb->sMv, 1, 0);

   uiBS[0][2][0] = BS_EDGE (uiBsx4[1], iRefIdx, pCurMb->sMv, 2, 1);

   uiBS[0][3][0] = BS_EDGE (uiBsx4[2], iRefIdx, pCurMb->sMv, 3, 2);

-  * (uint32_t*)uiBsx4 = (uiNnz32b1 | (uiNnz32b1 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b1);

   uiBS[0][1][1] = BS_EDGE (uiBsx4[0], iRefIdx, pCurMb->sMv, 5, 4);

   uiBS[0][2][1] = BS_EDGE (uiBsx4[1], iRefIdx, pCurMb->sMv, 6, 5);

   uiBS[0][3][1] = BS_EDGE (uiBsx4[2], iRefIdx, pCurMb->sMv, 7, 6);

-  * (uint32_t*)uiBsx4 = (uiNnz32b2 | (uiNnz32b2 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b2);

   uiBS[0][1][2] = BS_EDGE (uiBsx4[0], iRefIdx, pCurMb->sMv, 9, 8);

   uiBS[0][2][2] = BS_EDGE (uiBsx4[1], iRefIdx, pCurMb->sMv, 10, 9);

   uiBS[0][3][2] = BS_EDGE (uiBsx4[2], iRefIdx, pCurMb->sMv, 11, 10);

-  * (uint32_t*)uiBsx4 = (uiNnz32b3 | (uiNnz32b3 >> 8));

+  * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b3);

   uiBS[0][1][3] = BS_EDGE (uiBsx4[0], iRefIdx, pCurMb->sMv, 13, 12);

   uiBS[0][2][3] = BS_EDGE (uiBsx4[1], iRefIdx, pCurMb->sMv, 14, 13);

   uiBS[0][3][3] = BS_EDGE (uiBsx4[2], iRefIdx, pCurMb->sMv, 15, 14);

@@ -241,11 +241,12 @@

   int32_t i;

   uint32_t uiBSx4;

   uint8_t* pBS = (uint8_t*) (&uiBSx4);

-  uint32_t uiBIdx  = * (uint32_t*) (&g_kuiTableBIdx[iEdge][0]);

-  uint32_t uiBnIdx = * (uint32_t*) (&g_kuiTableBIdx[iEdge][4]);

+  const uint8_t *pBIdx  = &g_kuiTableBIdx[iEdge][0];

+  const uint8_t *pBnIdx = &g_kuiTableBIdx[iEdge][4];

+

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

-    if (pCurMb->pNonZeroCount[uiBIdx & 0xff] | pNeighMb->pNonZeroCount[uiBnIdx & 0xff]) {

+    if (pCurMb->pNonZeroCount[*pBIdx] | pNeighMb->pNonZeroCount[*pBnIdx]) {

       pBS[i] = 2;

     } else {

       pBS[i] =

@@ -252,10 +253,10 @@

 #ifndef SINGLE_REF_FRAME

         (pCurMb->uiRefIndex[g_kiTableBlock8x8Idx[1][iEdge][i]] - pNeighMb->uiRefIndex[g_kiTableBlock8x8NIdx[1][iEdge][i]]) ||

 #endif

-        MB_BS_MV (pCurMb->sMv, pNeighMb->sMv, (uiBIdx & 0xff), (uiBnIdx & 0xff));

+        MB_BS_MV (pCurMb->sMv, pNeighMb->sMv, *pBIdx, *pBnIdx);

     }

-    uiBIdx  = uiBIdx  >> 8;

-    uiBnIdx = uiBnIdx >> 8;

+    pBIdx++;

+    pBnIdx++;

   }

   return uiBSx4;

 }