blit_arm_stretch.c

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.

		p->PalPtr = InstCreate32((int)p->LookUp_Data,NONE,NONE,NONE,0,0);
		MB(); I1P(LDR,R3,p->PalPtr,0);
		if (p->Dither)
		{
			I2C(MOV,R1,NONE,0);
			I2C(MOV,R2,NONE,0);
		}
	}
	else
	{
		if (p->Dither)
		{
			I2C(MVN,R1,NONE,0x80000000);
			I2C(MVN,R2,NONE,0x80000000);
			I2C(MVN,R3,NONE,0x80000000);
		}
	}

	I2C(MOV,R5,NONE,0); //Pos

	LoopY = Label(0);
	I0P(B,AL,LoopY);

	if (p->InvertMask) InstPost(p->InvertMask);
	if (p->DiffMask) InstPost(p->DiffMask);
	if (p->PalPtr) InstPost(p->PalPtr);

	ColCount = 32;
	for (i=1;i<16 && !(p->RScaleX & i);i<<=1)
		ColCount >>= 1;

	InstPost(LoopY);

	EndOfLine = Label(0);

	//R4 Width
	//R5 Pos
	//R7 SrcPitch

	if (p->SwapXY)
	{
		I3(MUL,R4,R6,R4); //R6=DstPitch
		I3(ADD,R4,R9,R4);
	}
	else
	{
		if (p->DirX > 0)
			I3S(ADD,R4,R9,R4,LSL,p->DstBPP2);
		else
			I3S(SUB,R4,R9,R4,LSL,p->DstBPP2);
	}
	I2C(STR,R4,SP,OFS(stack,EndOfLine));

	if (p->SwapXY)
	{
		//we need next row in R14,R7,R8 (YUV)

		IMul(R6,R5,p->RScaleY); //R6 = RScaleY*Pos
		I2C(ADD,R8,R6,p->RScaleY);//R8 = RScaleY*(Pos+1)

		// dY  = ((RScaleY * (Pos+1)) >> 4) - ((RScaleY * Pos) >> 4);
		I3S(MOV,R14,NONE,R8,LSR,4);
		I3S(SUB,R14,R14,R6,LSR,4);
		I3(MUL,R14,R7,R14);

		// dUV = ((RScaleY * (Pos+1)) >> (4+SrcUVY2)) - ((RScaleY * Pos) >> (4+SrcUVY2));
		I3S(MOV,R8,NONE,R8,LSR,4+p->SrcUVY2);
		I3S(SUB,R8,R8,R6,LSR,4+p->SrcUVY2);
		I3(MUL,R8,R7,R8);

		I3(ADD,R14,R12,R14);
		I3S(ADD,R7,R10,R8,ASR,p->SrcUVX2);
		I3S(ADD,R8,R11,R8,ASR,p->SrcUVX2);
	}

	I2C(STR,R5,SP,OFS(stack,Pos));

	// special case, when R7,R8 must be saved/restored and incremented separatly
	Special = p->OnlyDiff && (p->Dither) && (p->SwapXY);
	if (Special)
	{
		I2C(STR,R7,SP,OFS(stack,SaveR7));
		I2C(STR,R8,SP,OFS(stack,SaveR8));
	}

	if (p->OnlyDiff)
	{
		MB(); I2C(LDR,R5,SP,OFS(stack,DiffMask));
		MB(); I2C(LDR,R6,SP,OFS(stack,Src2SrcLast));
	}

	LoopX = Label(1);
	RegDirty = 1;
	NoInc = 0;

	for (Col=0;Col<ColCount;)
	{
		int x = (Col * p->RScaleX) >> 4;

		int IncY;
		int IncUV;
		int IncY2 = 0;
		int IncUV2 = 0;
		int Col1 = Col;
		int Col2 = Col;

		IncY = (((Col+1) * p->RScaleX) >> 4) - ((Col * p->RScaleX) >> 4);
		IncUV = (((Col+1) * p->RScaleX) >> (4+p->SrcUVX2)) - ((Col * p->RScaleX) >> (4+p->SrcUVX2));
		Col++;

		if (!(p->SwapXY))
		{
			IncY2 = (((Col+1) * p->RScaleX) >> 4) - ((Col * p->RScaleX) >> 4);
			IncUV2 = (((Col+1) * p->RScaleX) >> (4+p->SrcUVX2)) - ((Col * p->RScaleX) >> (4+p->SrcUVX2));

			Col2++;
			Col++;
		}
		else
		if (!p->OnlyDiff)
		{
			IncY2 = IncY;
			IncUV2 = IncUV;
		}

		if (p->OnlyDiff)
		{
			reg RC = (reg)(p->SwapXY ? R0 : R14);
			reg RA = (reg)(p->Dither ? R7 : R1);
			reg RB = (reg)(p->Dither ? R8 : R2);

			RegDirty = 1;
			p->Skip = Label(0);

			//RC,RA,RB,R4 for temporary
			//R5(DiffMask),R6(Src2SrcLast)

			if (p->SwapXY)
			{
				Byte(); I3(LDR,RC,R12,R6);
				Byte(); I2C(LDR_POST,RA,R12,IncY);
				Byte(); I3(LDR,R4,R14,R6);	
				Byte(); I2C(LDR_POST,RB,R14,IncY);

				IncY = IncY2 = -IncY;
			}
			else
			{
				Byte(); I3(LDR,RC,R12,R6);
				Byte(); I2C(LDR_POST,RA,R12,IncY);
				Byte(); I3(LDR,R4,R12,R6);	
				Byte(); I2C(LDR_POST,RB,R12,IncY2);

				IncY = -IncY-IncY2;
				IncY2 = -IncY2;
			}

			I3(EOR,RC,RC,RA);
			S(); I3(TST,NONE,RC,R5);
			I3(EOR,R4,R4,RB);
			C(EQ); S(); I3(TST,NONE,R4,R5);

			C(EQ); Byte(); I3(LDR,RC,R10,R6); //last u
			Byte(); I2C(LDR_POST,RA,R10,IncUV+IncUV2); //u
			C(EQ); Byte(); I3(LDR,R4,R11,R6); //last v
			Byte(); I2C(LDR_POST,RB,R11,IncUV+IncUV2); //v

			IncUV = -IncUV-IncUV2;

			if (Special) //R7,R8 not incremented, using direct position
				IncUV2 = x >> p->SrcUVX2;
			else
			if (p->SwapXY) //IncUV2 will increment R7,R8
			{
				I2C(ADD,R7,R7,-IncUV);
				I2C(ADD,R8,R8,-IncUV);
				IncUV2 = IncUV;
			}
			else
				IncUV2 = -IncUV2;

			C(EQ); I3(EOR,RC,RC,RA);
			C(EQ); S(); I3(TST,NONE,RC,R5);
			C(EQ); I3(EOR,R4,R4,RB);
			C(EQ); S(); I3(TST,NONE,R4,R5);

			C(EQ); I2C(LDR,R4,SP,OFS(stack,EndOfLine));

#ifdef SHOWDIFF
			I2C(MVN,RC,NONE,0);
			if (DstBPP==8) Half();
			C(EQ); I2(STR,RC,R9);
#endif

			if (p->SwapXY)
			{
				MB(); I2C(LDR,RC,SP,OFS(stack,DstPitch));
				C(EQ); I3(ADD,R9,R9,RC);
			}
			else
			{
				C(EQ); I2C(ADD,R9,R9,p->DstStepX);
			}
			I0P(B,EQ,p->Skip);

			if (Special)
			{
				I2C(LDR,R7,SP,OFS(stack,SaveR7));
				I2C(LDR,R8,SP,OFS(stack,SaveR8));
			}
		}

		p->Upper = p->DirX < 0;
		if (ReUse && NoInc)
		{
			p->Pos = 0;
			if (!p->SwapXY && p->DstBPP==16)
				I3S(MOV,R0,NONE,R0,LSL,p->Upper?16:-16);
			if (!p->OnlyDiff)
				Inc_RGB_UV_Pixel(p, IncY, IncUV);
		}
		else
		{
			p->Pos = -1;
			Stretch_RGB_UV_Pixel(p, Col1, IncY, IncUV);
			RegDirty = 1;
		}

		if (!p->SwapXY) 
			NoInc = !IncY && !IncUV;

		if (p->DstPalette)
			Byte();

		if (p->DstPalette || p->DstBPP==32)
		{
			if (p->DstBPP==32 && p->Pos)
				I3S(MOV,R0,NONE,R0,ROR,-p->Pos);
			if (p->SwapXY)
				I2C(STR,R0,R9,-p->DstStepX/2);
			else
				I2C(STR_POST,R0,R9,p->DstStepX/2);
		}

		p->Upper = p->DirX > 0;
		if (ReUse && NoInc)
		{
			if (!p->SwapXY && p->DstBPP==16)
				I3S(ORR,R0,R0,R0,ROR,16);
			if (!p->OnlyDiff)
				Inc_RGB_UV_Pixel(p, IncY2, IncUV2);
		}
		else
		{
			if (p->DstBPP!=16)
				p->Pos = -1;
			Stretch_RGB_UV_Pixel(p, Col2, IncY2, IncUV2);
			RegDirty = 1;
		}

		if (!p->SwapXY) 
			NoInc = !IncY2 && !IncUV2;
		else
			NoInc = !IncY && !IncUV;

		if (p->Pos) 
			I3S(MOV,R0,NONE,R0,ROR,-p->Pos);

		if (p->DstPalette || p->DstBPP==32)
		{
			if (p->SwapXY)
			{
				if (RegDirty) { MB(); I2C(LDR,R5,SP,OFS(stack,DstPitch)); }
				if (p->DstPalette) Byte(); 
				I3(STR_POST,R0,R9,R5);
			}
			else
			{
				if (p->DstPalette) Byte(); 
				I2C(STR_POST,R0,R9,p->DstStepX/2);
			}
		}
		else
		{
			if (p->SwapXY)
			{
				if (RegDirty) { MB(); I2C(LDR,R5,SP,OFS(stack,DstPitch)); }
				I3(STR_POST,R0,R9,R5);
			}
			else
				I2C(STR_POST,R0,R9,p->DstStepX);
		}

		if (RegDirty) { MB(); I2C(LDR,R4,SP,OFS(stack,EndOfLine)); }

		RegDirty = 0;

		if (p->OnlyDiff)
		{
			MB(); I2C(LDR,R5,SP,OFS(stack,DiffMask));
			MB(); I2C(LDR,R6,SP,OFS(stack,Src2SrcLast));
			InstPost(p->Skip);
		}

		I3(CMP,NONE,R9,R4);
		if (Col == ColCount) 
		{
			//last item in block
			if (Special) 
			{
				I2C(LDR,R7,SP,OFS(stack,SaveR7));
				I2C(LDR,R8,SP,OFS(stack,SaveR8));

				I2C(ADD,R7,R7,(p->RScaleX * ColCount) >> (4+p->SrcUVX2));
				I2C(ADD,R8,R8,(p->RScaleX * ColCount) >> (4+p->SrcUVX2));

				I2C(STR,R7,SP,OFS(stack,SaveR7));
				I2C(STR,R8,SP,OFS(stack,SaveR8));
			}
			I0P(B,NE,LoopX);
		}
		else
			I0P(B,EQ,EndOfLine);
	}

	InstPost(EndOfLine);

	I2C(LDR,R6,SP,OFS(stack,DstNext));
	I2C(LDR,R7,SP,OFS(stack,SrcPitch));
	I2C(LDR,R5,SP,OFS(stack,Pos));

	I2C(LDR,R12,SP,OFS(stack,Y));
	I2C(LDR,R10,SP,OFS(stack,U));
	I2C(LDR,R11,SP,OFS(stack,V));

	I3(ADD,R9,R9,R6);
	
	//increment row of Y,U,V according to Pos -> Pos+RowStep
	RowStep = (p->SwapXY) ? 2:1;

	//R4,R6,R8 for temporary
	IMul(R6,R5,p->RScaleY); //R6 = RScaleY*Pos
	I2C(ADD,R5,R5,RowStep);
	IMul(R8,R5,p->RScaleY); //R8 = RScaleY*(Pos+RowStep)

	// dY  = ((RScaleY * (Pos+RowStep)) >> 4) - ((RScaleY * Pos) >> 4);
	I3S(MOV,R4,NONE,R8,LSR,4);
	I3S(SUB,R4,R4,R6,LSR,4);
	I3(MUL,R4,R7,R4);

	// dUV = ((RScaleY * (Pos+RowStep)) >> (4+SrcUVY2)) - ((RScaleY * Pos) >> (4+SrcUVY2));
	I3S(MOV,R8,NONE,R8,LSR,4+p->SrcUVY2);
	I3S(SUB,R8,R8,R6,LSR,4+p->SrcUVY2);
	I3(MUL,R8,R7,R8);

	I3(ADD,R12,R12,R4);
	I3S(ADD,R10,R10,R8,ASR,p->SrcUVPitch2); 
	I3S(ADD,R11,R11,R8,ASR,p->SrcUVPitch2);

	I2C(STR,R12,SP,OFS(stack,Y));
	I2C(STR,R10,SP,OFS(stack,U));
	I2C(STR,R11,SP,OFS(stack,V));

	//prepare registers for next row
	if (p->SwapXY) I2C(LDR,R6,SP,OFS(stack,DstPitch));
	I2C(LDR,R4,SP,OFS(stack,Width));

	MB(); I2C(LDR,R0,SP,OFS(stack,EndOfRect));
	I3(CMP,NONE,R9,R0);
	I0P(B,NE,LoopY);

	I2C(ADD,SP,SP,OFS(stack,StackFrame));
	CodeEnd();
}

#endif