softidct.h

大名鼎鼎的CE下播放软件,TCPPMP的源代码!!!2410下可以流畅的解QVGA的H264,MPEG4等格式.

#endif

#if defined(FREESCALE_MX1)
void MX1PopNone(softidct* p);
void MX1Intra8x8(softidct* p,idct_block_t* Data,int Length,int ScanType);
void MX1Inter8x8Back(softidct* p,idct_block_t* Block,int Length);
void MX1Inter8x8BackFwd(softidct* p,idct_block_t* Block,int Length);
void MX1Inter8x8QPEL(softidct* p,idct_block_t* Block,int Length);
void MX1Inter8x8GMC(softidct* p,idct_block_t* Block,int Length);
void MX1Intra8x8Swap(softidct* p,idct_block_t* Data,int Length,int ScanType);
void MX1Inter8x8BackSwap(softidct* p,idct_block_t* Block,int Length);
void MX1Inter8x8BackFwdSwap(softidct* p,idct_block_t* Block,int Length);
#endif

#if defined(ARM)
void WMMXIntra8x8(softidct* p,idct_block_t* Data,int Length,int ScanType);
void WMMXInter8x8Back(softidct* p,idct_block_t* Block,int Length);
void WMMXInter8x8BackFwd(softidct* p,idct_block_t* Block,int Length);
void WMMXIntra8x8Swap(softidct* p,idct_block_t* Data,int Length,int ScanType);
void WMMXInter8x8QPEL(softidct* p,idct_block_t* Block,int Length);
void WMMXInter8x8GMC(softidct* p,idct_block_t* Block,int Length);
void WMMXInter8x8BackSwap(softidct* p,idct_block_t* Block,int Length);
void WMMXInter8x8BackFwdSwap(softidct* p,idct_block_t* Block,int Length);

void STDCALL WMMXIDCT_Block8x4(idct_block_t *Block, uint8_t *Dest, int DestStride, const uint8_t *Src);
void STDCALL WMMXIDCT_Block8x8(idct_block_t *Block, uint8_t *Dest, int DestStride, const uint8_t *Src);
void STDCALL WMMXIDCT_Const8x8(int v,uint8_t * Dst,int DstStride,uint8_t * Src);
void STDCALL WMMXIDCT_Const4x4(int v,uint8_t * Dst,int DstStride,uint8_t * Src);

void STDCALL WMMXCopyBlockM(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL WMMXCopyBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch); 
void STDCALL WMMXCopyBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch); 
void STDCALL WMMXCopyBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL WMMXCopyBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL WMMXCopyBlockHorRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL WMMXCopyBlockVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL WMMXCopyBlockHorVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL WMMXAddBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch);
void STDCALL WMMXAddBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch); 
void STDCALL WMMXAddBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);
void STDCALL WMMXAddBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);

void STDCALL FastCopyBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch); 
void STDCALL FastCopyBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch); 
void STDCALL FastCopyBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL FastCopyBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL FastCopyBlockHorRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL FastCopyBlockVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL FastCopyBlockHorVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL FastAddBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch);
void STDCALL FastAddBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch); 
void STDCALL FastAddBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);
void STDCALL FastAddBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);

void STDCALL PreLoadCopyBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch); 
void STDCALL PreLoadCopyBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch); 
void STDCALL PreLoadCopyBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL PreLoadCopyBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL PreLoadCopyBlockHorRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL PreLoadCopyBlockVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL PreLoadCopyBlockHorVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
void STDCALL PreLoadAddBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch);
void STDCALL PreLoadAddBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch); 
void STDCALL PreLoadAddBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);
void STDCALL PreLoadAddBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);

#endif

#if defined(MIPS64)

#define CopyBlock8x8 CopyBlock
#define CopyBlock16x16 CopyMBlock

extern void STDCALL CopyMBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
extern void STDCALL CopyMBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch); 
extern void STDCALL CopyMBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
extern void STDCALL CopyMBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
extern void STDCALL CopyMBlockHorRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
extern void STDCALL CopyMBlockVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);
extern void STDCALL CopyMBlockHorVerRound(uint8_t * Src, uint8_t * Dst, int SrcPitch, int DstPitch);

extern void Inter8x8Add(softidct* p,idct_block_t *Block,int Length);

// Src must be aligned
extern void AddBlock8x8(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);
extern void AddBlock16x16(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);

#if defined(TARGET_WINCE) || defined(TARGET_WIN32)
#define MIPS64_WIN32
#endif

#ifdef MIPS64_WIN32
static INLINE void DisableInterrupts() {
	__asm(	".set noreorder;"
			"mfc0	$4,$12;"
			"li		$5,-1;"
			"sll	$5,$5,1;"
			"and	$4,$4,$5;"
			"mtc0	$4,$12;"
			"nop;"
			".set reorder;"
			);
}

static INLINE void EnableInterrupts() {
	__asm(	".set noreorder;"
			"mfc0	$4,$12;"
			"nop;"
			"nop;"
			"ori	$4,$4,0x001;"
			"mtc0	$4,$12;"
			"nop;"
			".set reorder;"
			);
}
#endif

#elif defined(MIPS32)

// Src must be aligned
extern void STDCALL AddBlock4x4(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);
extern void STDCALL AddBlock8x8(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);
extern void STDCALL CopyBlock4x4(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);
extern void STDCALL CopyBlock16x16(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);
extern void STDCALL CopyBlock8x8(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);

#else

// AddBlock DstPitch=8
extern void STDCALL AddBlock(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlockHor(uint8_t * Src, uint8_t * Dst, int SrcPitch); 
extern void STDCALL AddBlockVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlockHorVer(uint8_t * Src, uint8_t * Dst, int SrcPitch);

extern void STDCALL AddBlock4x4_00(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_01(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_02(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_03(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_10(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_11(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_12(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_13(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_20(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_21(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_22(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_23(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_30(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_31(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_32(uint8_t * Src, uint8_t * Dst, int SrcPitch);
extern void STDCALL AddBlock4x4_33(uint8_t * Src, uint8_t * Dst, int SrcPitch);

// Src must be aligned
extern void STDCALL CopyBlock4x4(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);
extern void STDCALL CopyBlock8x8(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);
extern void STDCALL CopyBlock16x16(uint8_t * Src, uint8_t * Dst, int SrcPitch,int DstPitch);

#endif

#ifdef MMX
  #ifdef _MSC_VER
    #if defined(TARGET_WINCE) || defined(_M_AMD64)
      extern void STDCALL EMMS();
    #else
      static INLINE void EMMS() { __asm { emms } }
    #endif
  #else
	#define	EMMS() __asm__ __volatile__ ("emms")
  #endif
#else
  static INLINE void EMMS() {}
#endif

#ifdef ARM
  #define SAT(Value) Value = Value < 0 ? 0: (Value > 255 ? 255: Value);
#else
  // upper bits are trashed! use only lower 8 bits afterward
  #define SAT(Value) Value <<= Value >> 16; Value |= (Value << 23) >> 31;
#endif


static INLINE void ProcessHead(softidct* p,int x,int y)
{
#ifdef MIPS64_WIN32
	if (y != p->NextIRQ)
		EnableInterrupts();
#endif
}

static INLINE void ProcessTail(softidct* p,int x,int y)
{
#ifdef MIPS64_WIN32
	if (y != p->NextIRQ)
	{
		p->NextIRQ = y;
		DisableInterrupts();
	}
#endif
}

static INLINE void SetPtr420(softidct* p,int x,int y,int shift)
{
	x <<= 4-shift;
	y <<= 4-shift;
	x += EDGE;
	y += EDGE;
	y *= p->BufWidth;
	p->CurrPitch = p->BufWidth;
	p->DstPtr = p->Dst+x+y;
	p->RefPtr[0] = p->Ref[0]+x+y;
	p->RefPtr[1] = p->Ref[1]+x+y;
	p->Ptr = p->Tab;
	p->Tab[3] = 1 | 2*(-y-x-(8>>shift)*p->BufWidth-(8>>shift) + 
		p->YToU + (x>>1) + (y>>2));	//Y[1;1] -> U
}

static INLINE void SetPtr422(softidct* p,int x,int y,int shift)
{
	x <<= 4-shift;
	y <<= 3-shift;
	x += EDGE;
	y += EDGE;
	y *= p->BufWidth;
	p->CurrPitch = p->BufWidth;
	p->DstPtr = p->Dst+x+y;
	p->RefPtr[0] = p->Ref[0]+x+y;
	p->RefPtr[1] = p->Ref[1]+x+y;
	p->Ptr = p->Tab;
	p->Tab[1] = 1 | 2*(-y-x-(8>>shift) + 
		p->YToU + (x>>1) + (y>>1));	//Y[0;1] -> U
}

static INLINE void SetPtr444(softidct* p,int x,int y,int shift)
{
	x <<= 3-shift;
	y <<= 3-shift;
	x += EDGE;
	y += EDGE;
	y *= p->BufWidth;
	p->CurrPitch = p->BufWidth;
	p->DstPtr = p->Dst+x+y;
	p->RefPtr[0] = p->Ref[0]+x+y;
	p->RefPtr[1] = p->Ref[1]+x+y;
	p->Ptr = p->Tab;
}

static INLINE void IncPtr(softidct* p,bool_t Ref0,bool_t Ref1)
{
	uint8_t* DstPtr = p->DstPtr;
	int v = *(p->Ptr++);
	if (v & 1) p->CurrPitch >>= 1;
	v >>= 1;
	p->DstPtr = DstPtr + v;
	if (Ref0) p->RefPtr[0] += v;
	if (Ref1) p->RefPtr[1] += v;
}

static INLINE void IncPtrLum(softidct* p)
{
	int v = (p->Tab[0]+p->Tab[1]+p->Tab[2]+p->Tab[3]) >> 1;
	p->CurrPitch >>= 1;
	p->DstPtr += v;
	p->RefPtr[0] += v;
	p->RefPtr[1] += v;
	p->Ptr = &p->Tab[4];
}

#endif