ref: 007f5ba773955966c74f531cdf70901e43cb195d
parent: bf9a78085eabddbfe7ec9b975245fb96a26a9f7e
parent: 0a14d4c79b88098eebfaa44dbbd78ae98accf08a
author: Ethan Hugg <ethanhugg@gmail.com>
date: Mon Feb 10 05:36:05 EST 2014
Merge pull request #258 from mstorsjo/endian-cleanup Avoid endian-specific code
--- a/codec/common/deblocking_common.h
+++ b/codec/common/deblocking_common.h
@@ -15,12 +15,6 @@
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/common/macros.h
+++ b/codec/common/macros.h
@@ -234,34 +234,6 @@
#define CALC_BI_STRIDE(width,bitcount) ((((width * bitcount) + 31) & ~31) >> 3)
-#ifdef WORDS_BIGENDIAN
-static inline uint32_t ENDIAN_FIX (uint32_t x) {-return x;
-}
-#else //!WORDS_BIGENDIAN
-
-#if defined(_MSC_VER) && defined(_M_IX86)
-static inline uint32_t ENDIAN_FIX (uint32_t x) {-__asm {- mov eax, x
- bswap eax
- mov x, eax
-}
-return x;
-}
-#else // GCC
-static inline uint32_t ENDIAN_FIX (uint32_t x) {-#ifdef X86_ARCH
-__asm__ __volatile__ ("bswap %0":"+r" (x));-#else
-x = ((x & 0xff000000) >> 24) | ((x & 0xff0000) >> 8) |
- ((x & 0xff00) << 8) | ((x & 0xff) << 24);
-#endif
-return x;
-}
-#endif//GCC
-
-#endif//!WORDS_BIGENDIAN
#ifndef BUTTERFLY1x2
--- a/codec/decoder/core/src/deblocking.cpp
+++ b/codec/decoder/core/src/deblocking.cpp
@@ -145,7 +145,6 @@
void_t inline DeblockingBSInsideMBAvsbase (int8_t* pNnzTab, uint8_t nBS[2][4][4], int32_t iLShiftFactor) {uint32_t uiNnz32b0, uiNnz32b1, uiNnz32b2, uiNnz32b3;
- ENFORCE_STACK_ALIGN_1D (uint8_t, uiBsx3, 4, 4);
uiNnz32b0 = * (uint32_t*) (pNnzTab + 0);
uiNnz32b1 = * (uint32_t*) (pNnzTab + 4);
@@ -152,27 +151,23 @@
uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);
uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);
- * (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];
+ nBS[0][1][0] = (pNnzTab[0] | pNnzTab[1]) << iLShiftFactor;
+ nBS[0][2][0] = (pNnzTab[1] | pNnzTab[2]) << iLShiftFactor;
+ nBS[0][3][0] = (pNnzTab[2] | pNnzTab[3]) << 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];
+ nBS[0][1][1] = (pNnzTab[4] | pNnzTab[5]) << iLShiftFactor;
+ nBS[0][2][1] = (pNnzTab[5] | pNnzTab[6]) << iLShiftFactor;
+ nBS[0][3][1] = (pNnzTab[6] | pNnzTab[7]) << iLShiftFactor;
* (uint32_t*)nBS[1][1] = (uiNnz32b0 | uiNnz32b1) << 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];
+ nBS[0][1][2] = (pNnzTab[8] | pNnzTab[9]) << iLShiftFactor;
+ nBS[0][2][2] = (pNnzTab[9] | pNnzTab[10]) << iLShiftFactor;
+ nBS[0][3][2] = (pNnzTab[10] | pNnzTab[11]) << iLShiftFactor;
* (uint32_t*)nBS[1][2] = (uiNnz32b1 | uiNnz32b2) << 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];
+ nBS[0][1][3] = (pNnzTab[12] | pNnzTab[13]) << iLShiftFactor;
+ nBS[0][2][3] = (pNnzTab[13] | pNnzTab[14]) << iLShiftFactor;
+ nBS[0][3][3] = (pNnzTab[14] | pNnzTab[15]) << iLShiftFactor;
* (uint32_t*)nBS[1][3] = (uiNnz32b2 | uiNnz32b3) << iLShiftFactor;
}
@@ -188,22 +183,26 @@
uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);
uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);
- * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b0);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[i] | pNnzTab[i + 1];
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 = DEBLOCK_BS_SHIFTED (uiNnz32b1);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[4 + i] | pNnzTab[4 + i + 1];
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 = DEBLOCK_BS_SHIFTED (uiNnz32b2);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[8 + i] | pNnzTab[8 + i + 1];
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 = DEBLOCK_BS_SHIFTED (uiNnz32b3);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[12 + i] | pNnzTab[12 + i + 1];
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);
--- a/codec/encoder/core/inc/svc_enc_golomb.h
+++ b/codec/encoder/core/inc/svc_enc_golomb.h
@@ -46,6 +46,12 @@
namespace WelsSVCEnc {+#define WRITE_BE_32(ptr, val) do { \+ (ptr)[0] = (val) >> 24; \
+ (ptr)[1] = (val) >> 16; \
+ (ptr)[2] = (val) >> 8; \
+ (ptr)[3] = (val) >> 0; \
+ } while (0)
/************************************************************************/
/* GOLOMB CODIMG FOR WELS ENCODER */
/************************************************************************/
@@ -73,7 +79,7 @@
else {\(n) -= iLeftBits;\
uiCurBits = (uiCurBits<<iLeftBits) | ((v)>>(n));\
- *((uint32_t*)pBufPtr) = ENDIAN_FIX(uiCurBits);\
+ WRITE_BE_32(pBufPtr, uiCurBits);\
pBufPtr += 4;\
uiCurBits = (v) & ((1<<(n))-1);\
iLeftBits = 32 - (n);\
@@ -141,7 +147,7 @@
} else {n -= pBs->iLeftBits;
pBs->uiCurBits = (pBs->uiCurBits << pBs->iLeftBits) | (kuiValue >> n);
- * ((uint32_t*)pBs->pBufPtr) = ENDIAN_FIX (pBs->uiCurBits);
+ WRITE_BE_32(pBs->pBufPtr, pBs->uiCurBits);
pBs->pBufPtr += 4;
pBs->uiCurBits = kuiValue & ((1 << n) - 1);
pBs->iLeftBits = 32 - n;
@@ -160,7 +166,7 @@
static inline void BsFlush (SBitStringAux* pBs) {-* (uint32_t*)pBs->pBufPtr = ENDIAN_FIX (pBs->uiCurBits << pBs->iLeftBits);
+WRITE_BE_32(pBs->pBufPtr, pBs->uiCurBits << pBs->iLeftBits);
pBs->pBufPtr += 4 - pBs->iLeftBits / 8;
pBs->iLeftBits = 32;
pBs->uiCurBits = 0; // for future writing safe, 5/19/2010
--- a/codec/encoder/core/src/deblocking.cpp
+++ b/codec/encoder/core/src/deblocking.cpp
@@ -156,7 +156,6 @@
void inline DeblockingBSInsideMBAvsbase (int8_t* pNnzTab, uint8_t uiBS[2][4][4], int32_t iLShiftFactor) {uint32_t uiNnz32b0, uiNnz32b1, uiNnz32b2, uiNnz32b3;
- ENFORCE_STACK_ALIGN_1D (uint8_t, uiBsx3, 4, 4);
uiNnz32b0 = * (uint32_t*) (pNnzTab + 0);
uiNnz32b1 = * (uint32_t*) (pNnzTab + 4);
@@ -163,27 +162,23 @@
uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);
uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);
- * (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];
+ uiBS[0][1][0] = (pNnzTab[0] | pNnzTab[1]) << iLShiftFactor;
+ uiBS[0][2][0] = (pNnzTab[1] | pNnzTab[2]) << iLShiftFactor;
+ uiBS[0][3][0] = (pNnzTab[2] | pNnzTab[3]) << 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];
+ uiBS[0][1][1] = (pNnzTab[4] | pNnzTab[5]) << iLShiftFactor;
+ uiBS[0][2][1] = (pNnzTab[5] | pNnzTab[6]) << iLShiftFactor;
+ uiBS[0][3][1] = (pNnzTab[6] | pNnzTab[7]) << iLShiftFactor;
* (uint32_t*)uiBS[1][1] = (uiNnz32b0 | uiNnz32b1) << 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];
+ uiBS[0][1][2] = (pNnzTab[8] | pNnzTab[9]) << iLShiftFactor;
+ uiBS[0][2][2] = (pNnzTab[9] | pNnzTab[10]) << iLShiftFactor;
+ uiBS[0][3][2] = (pNnzTab[10] | pNnzTab[11]) << iLShiftFactor;
* (uint32_t*)uiBS[1][2] = (uiNnz32b1 | uiNnz32b2) << 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];
+ uiBS[0][1][3] = (pNnzTab[12] | pNnzTab[13]) << iLShiftFactor;
+ uiBS[0][2][3] = (pNnzTab[13] | pNnzTab[14]) << iLShiftFactor;
+ uiBS[0][3][3] = (pNnzTab[14] | pNnzTab[15]) << iLShiftFactor;
* (uint32_t*)uiBS[1][3] = (uiNnz32b2 | uiNnz32b3) << iLShiftFactor;
}
@@ -197,22 +192,26 @@
uiNnz32b2 = * (uint32_t*) (pNnzTab + 8);
uiNnz32b3 = * (uint32_t*) (pNnzTab + 12);
- * (uint32_t*)uiBsx4 = DEBLOCK_BS_SHIFTED (uiNnz32b0);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[i] | pNnzTab[i + 1];
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 = DEBLOCK_BS_SHIFTED (uiNnz32b1);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[4 + i] | pNnzTab[4 + i + 1];
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 = DEBLOCK_BS_SHIFTED (uiNnz32b2);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[8 + i] | pNnzTab[8 + i + 1];
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 = DEBLOCK_BS_SHIFTED (uiNnz32b3);
+ for (int i = 0; i < 3; i++)
+ uiBsx4[i] = pNnzTab[12 + i] | pNnzTab[12 + i + 1];
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);