⭐ 欢迎来到虫虫下载站! | 📦 资源下载 📁 资源专辑 ℹ️ 关于我们
⭐ 虫虫下载站

📄 blit_arm_fix.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.
💻 C
📖 第 1 页 / 共 2 页
字号:
		int i,Shift;
		
		for (i=0;i<3;++i)
			Mask |= 1 << (p->DstPos[i] + p->DstSize[i] -1);

		Mask2 = Mask;
		Shift = 0;
		if (p->DstBPP==8 && p->DstDoubleX)
			Mask2 |= Mask << 16;
		else
		if (p->DstBPP <= 16)
		{
			if (p->DirX<0) Shift = p->DstBPP;
			Mask2 |= Mask << p->DstBPP;
		}
		
		Invert ^= RotateRight(Mask2,Shift+p->DstPos[0]);
	}

	p->YMul = InstCreate32(p->_YMul,NONE,NONE,NONE,0,0);
	p->RVMul = InstCreate32(p->_RVMul,NONE,NONE,NONE,0,0);
	p->RAdd = InstCreate32(p->_RAdd,NONE,NONE,NONE,0,0);
	p->GUMul = InstCreate32(p->_GUMul,NONE,NONE,NONE,0,0);
	p->GVMul = InstCreate32(p->_GVMul,NONE,NONE,NONE,0,0);
	p->GAdd = InstCreate32(p->_GAdd,NONE,NONE,NONE,0,0);
	p->BUMul = InstCreate32(p->_BUMul,NONE,NONE,NONE,0,0);
	p->BAdd = InstCreate32(p->_BAdd,NONE,NONE,NONE,0,0);

	if (Invert)
		p->InvertMask = InstCreate32(Invert,NONE,NONE,NONE,0,0);

	if (p->OnlyDiff)
		p->DiffMask = InstCreate32(0xFCFCFCFC,NONE,NONE,NONE,0,0);

	CodeBegin();
	I2C(SUB,SP,SP,OFS(stack,StackFrame));

	I2C(LDR,R9,R1,0); //Dst[0] RGB
	I2C(LDR,R10,R2,4); //Src[1] U
	I2C(LDR,R11,R2,8); //Src[2] V
	I2C(LDR,R12,R2,0); //Src[0] Y
	I2C(STR,R3,SP,OFS(stack,DstPitch));

	I3(MOV,R6,NONE,R3); //DstPitch
	I2C(LDR,R7,SP,OFS(stack,SrcPitch));
	I2C(LDR,R0,SP,OFS(stack,Height));
	I2C(LDR,R4,SP,OFS(stack,Width));

	if (!p->ColorLookup)
		I1P(LDR,R8,p->YMul,0);
	else
		I1P(MOV,R8,p->LookUp,0);

	//YNext = 2*Src->Pitch - (Width >> SrcDoubleX)
	I3S(MOV,R1,NONE,R7,LSL,1);
	I3S(SUB,R1,R1,R4,LSR,p->SrcDoubleX); 
	I2C(STR,R1,SP,OFS(stack,YNext));

	//UVNext = (Src->Pitch >> 1) - (Width >> SrcDoubleX >> 1);
	I3S(MOV,R2,NONE,R7,ASR,1);
	I3S(SUB,R2,R2,R4,LSR,p->SrcDoubleX+1); 
	I2C(STR,R2,SP,OFS(stack,UVNext));

	if (p->DirX<0 && p->DstBPP==16) //adjust reversed destination for block size
		I2C(SUB,R9,R9,-p->DstStepX-(p->DstBPP >> 3));
	if (p->DstBPP==32)
		I2C(ADD,R9,R9,p->DstStepX/2);

	if (p->SwapXY)
	{
		// EndOfRect = Dst + ((Height * DstBPP * DirX) >> 3) - (DstPitch << DstDoubleY)
		I3S(SUB,R9,R9,R6,LSL,p->DstDoubleY);
		I2C(MOV,R1,NONE,p->DstBPP * p->DirX);
		I3(MUL,R0,R1,R0);
		I3S(ADD,R0,R9,R0,ASR,3);
		I2C(STR,R0,SP,OFS(stack,EndOfRect));

		//DstNext = DstStepX - Width*DstPitch;
		MB(); I3(MUL,R2,R6,R4);
		I2C(MOV,R0,NONE,p->DstStepX); 
		I3(SUB,R0,R0,R2); 
		I2C(STR,R0,SP,OFS(stack,DstNext));
	}
	else
	{
		// EndOfRect = Dst + DstPitch * Height
		I3(MUL,R0,R6,R0);
		I3(ADD,R0,R9,R0);
		I2C(STR,R0,SP,OFS(stack,EndOfRect));

		//DstNext = ((DstPitch*2 << DstDoubleY) - DirX * Width << DstBPP2;
		I3S(MOV,R2,NONE,R6,LSL,p->DstDoubleY+1);
		I3S(p->DirX>0?SUB:ADD,R2,R2,R4,LSL,p->DstBPP2); 
		I2C(STR,R2,SP,OFS(stack,DstNext));
	}
	I3(ADD,R14,R12,R7);

	if (p->Dither)
	{
		if (!p->ColorLookup)
		{
			I2C(MVN,R1,NONE,0x80000000);
			I2C(MVN,R2,NONE,0x80000000);
			I2C(MVN,R3,NONE,0x80000000);
		}
		else
		{
			I2C(MOV,R1,NONE,0x80);
			I2C(MOV,R2,NONE,0x80);
			I2C(MOV,R3,NONE,0x80);
		}
	}

	LoopY = Label(1);

	if (p->SwapXY)
	{
		I3(MUL,R0,R6,R4); //R6=dstpitch
		I3(ADD,R7,R9,R0);
	}
	else
	{
		if (p->DirX > 0)
			I3S(ADD,R7,R9,R4,LSL,p->DstBPP2);
		else
			I3S(SUB,R7,R9,R4,LSL,p->DstBPP2);
	}
	I2C(STR,R7,SP,OFS(stack,EndOfLine));

	// preload
	if (p->ARM5) // not needed for non slices mode, but testing show it didn't help on ARM4
	{
		if (!p->Slices)
		{
			//R4 width
			//R0,R7,R5,R6 tmp

			dyninst* PreLoadEnd = Label(0);
			dyninst* PreLoad1;
			dyninst* PreLoad2;
			dyninst* PreLoad3;
			dyninst* PreLoad4;

			I3S(ADD,R0,R12,R4,ASR,(p->SrcDoubleX?1:0)-(p->SrcHalfX?1:0));
			I2C(ADD,R5,R12,32);
			I3(CMP,NONE,R5,R0);
			I0P(B,CS,PreLoadEnd);

			//y0
			PreLoad1 = Label(1);
			Byte(); I2C(LDR,R6,R5,-32);
			I2C(ADD,R5,R5,64);
			I3(CMP,NONE,R5,R0);
			Byte(); I2C(LDR,R7,R5,-64);
			I0P(B,CC,PreLoad1);

			I3S(ADD,R0,R14,R4,ASR,(p->SrcDoubleX?1:0)-(p->SrcHalfX?1:0));
			I2C(ADD,R5,R14,32);

			//y1
			PreLoad2 = Label(1);
			Byte(); I2C(LDR,R6,R5,-32);
			I2C(ADD,R5,R5,64);
			I3(CMP,NONE,R5,R0);
			Byte(); I2C(LDR,R7,R5,-64);
			I0P(B,CC,PreLoad2);

			I3S(ADD,R0,R10,R4,ASR,(p->SrcDoubleX?1:0)-(p->SrcHalfX?1:0)+p->SrcUVX2);
			I2C(ADD,R5,R10,32);
			I3(CMP,NONE,R5,R0);
			I0P(B,CS,PreLoadEnd);

			//u
			PreLoad3 = Label(1);
			Byte(); I2C(LDR,R6,R5,-32);
			I2C(ADD,R5,R5,64);
			I3(CMP,NONE,R5,R0);
			Byte(); I2C(LDR,R7,R5,-64);
			I0P(B,CC,PreLoad3);

			I3S(ADD,R0,R11,R4,ASR,(p->SrcDoubleX?1:0)-(p->SrcHalfX?1:0)+p->SrcUVX2);
			I2C(ADD,R5,R11,32);

			//v
			PreLoad4 = Label(1);
			Byte(); I2C(LDR,R6,R5,-32);
			I2C(ADD,R5,R5,64);
			I3(CMP,NONE,R5,R0);
			Byte(); I2C(LDR,R7,R5,-64);
			I0P(B,CC,PreLoad4);

			if (p->OnlyDiff) //restore R7
				I2C(LDR,R7,SP,OFS(stack,EndOfLine));

			InstPost(PreLoadEnd);
		}
		else
		{
			//preload next
			MB(); I2C(LDR,R6,SP,OFS(stack,SrcPitch));
			I3S(PLD,NONE,R12,R6,LSL,1);
			I3S(PLD,NONE,R14,R6,LSL,1);
			I3S(PLD,NONE,R10,R6,ASR,p->SrcUVPitch2);
			I3S(PLD,NONE,R11,R6,ASR,p->SrcUVPitch2);
		}
	}

    if (p->OnlyDiff)
	{
		MB(); I1P(LDR,R5,p->DiffMask,0);
		MB(); I2C(LDR,R6,SP,OFS(stack,Src2SrcLast));
		p->Skip = Label(0);
	}

	LoopX = Label(1);
	{
		int PitchDouble;
		reg Pitch;

		Fix_RGB_UV_LoadUV(p);

		p->Pos = -1;

		Fix_RGB_UV_Pixel(p,0,0);
		if (p->DstBPP==32)
		{
			if (p->Pos)
				I3S(MOV,R0,NONE,R0,ROR,-p->Pos);

			assert(!p->DstDoubleX && !p->DstDoubleY);
			MB(); I2C(LDR,R1,SP,OFS(stack,DstPitch));
			I2C(ADD,R1,R1,-p->DstStepX/2);
			I3(STR,R0,R9,R1);
			p->Pos = -1;
		}

		Fix_RGB_UV_Pixel(p,1,0);
		if (p->Pos) 
			I3S(MOV,R0,NONE,R0,ROR,-p->Pos);

		Pitch = (reg)(p->Dither ? R7:R1);
		MB(); I2C(LDR,Pitch,SP,OFS(stack,DstPitch));

		if (p->DstBPP==8 && p->DstDoubleX) 
			I3S(ORR,R0,R0,R0,LSL,8);

		if (p->DstBPP==16 && p->DstDoubleX) 
		{
			I2C(ADD,R9,R9,4);
			I3S(MOV,R3,NONE,R0,LSR,16);
			I3S(MOV,R0,NONE,R0,LSL,16);
			I3S(ORR,R3,R3,R3,LSL,16);
			I3S(ORR,R0,R0,R0,LSR,16);

			if (p->DstDoubleY)
			{
				I3S(ADD,R2,Pitch,Pitch,LSL,1); //R2=3*DstPitch
				I3(STR,R3,R9,R2);
			}
			I3S(STR,R3,R9,Pitch,LSL,p->DstDoubleY);
			I2C(SUB,R9,R9,4);
		}

		if (p->DstDoubleY)
		{
			I3S(ADD,R2,Pitch,Pitch,LSL,1); //R2=3*DstPitch
			if (p->DstBPP==8 && !p->DstDoubleX) Half();
			I3(STR,R0,R9,R2);
		}
		
		PitchDouble = p->DstDoubleY;
		if (p->DstBPP==8 && !p->DstDoubleX) 
		{
			if (PitchDouble) // can't use STR with Half() and LSL,#1 at the same time
			{
				PitchDouble = 0;
				I3(ADD,Pitch,Pitch,Pitch);
			}
			Half();
		}
		I3S(STR,R0,R9,Pitch,LSL,PitchDouble);
		if (p->SwapXY)
			I3S(ADD,R9,R9,Pitch,LSL,1+PitchDouble);

		p->Pos = -1;
		Fix_RGB_UV_Pixel(p,0,1);
		if (p->DstBPP==32)
		{
			if (p->Pos)
				I3S(MOV,R0,NONE,R0,ROR,-p->Pos);

			assert(!p->DstDoubleX && !p->DstDoubleY);
			I2C(STR,R0,R9,-p->DstStepX/2);
			p->Pos = -1;
		}

		Fix_RGB_UV_Pixel(p,1,1);
		if (p->Pos) 
			I3S(MOV,R0,NONE,R0,ROR,-p->Pos);

		if (p->DstBPP==8 && p->DstDoubleX) 
			I3S(ORR,R0,R0,R0,LSL,8);

		if (p->DstDoubleY)
		{
			MB(); I2C(LDR,R1,SP,OFS(stack,DstPitch));
		}

		if (p->DstBPP==16 && p->DstDoubleX) 
		{
			I3S(MOV,R3,NONE,R0,LSR,16);
			I3S(MOV,R0,NONE,R0,LSL,16);
			I3S(ORR,R3,R3,R3,LSL,16);
			I3S(ORR,R0,R0,R0,LSR,16);

			if (p->DstDoubleY)
			{
				I2C(ADD,R2,R1,4); //DstPitch+4
				I3(STR,R3,R9,R2);
			}
			I2C(STR,R3,R9,4);
		}

		if (p->DstDoubleY)
		{
			if (p->DstBPP==8 && !p->DstDoubleX) Half();
			I3(STR,R0,R9,R1);
		}

		if (p->DstBPP==8 && !p->DstDoubleX) Half();
		if (p->SwapXY)
			I2(STR,R0,R9);
		else
			I2C(STR_POST,R0,R9,p->DstStepX);

		MB(); I2C(LDR,R7,SP,OFS(stack,EndOfLine));
		if (p->OnlyDiff)
		{
			MB(); I1P(LDR,R5,p->DiffMask,0);
			MB(); I2C(LDR,R6,SP,OFS(stack,Src2SrcLast));
			InstPost(p->Skip);
		}

		I3(CMP,NONE,R9,R7);
		I0P(B,NE,LoopX);
	}

	I2C(LDR,R0,SP,OFS(stack,YNext));
	I2C(LDR,R4,SP,OFS(stack,DstNext));
	I2C(LDR,R6,SP,OFS(stack,UVNext));
	I2C(LDR,R5,SP,OFS(stack,EndOfRect));
	
	//increment pointers
	I3(ADD,R12,R12,R0);
	I3(ADD,R14,R14,R0);
	I3(ADD,R9,R9,R4);
	I3(ADD,R10,R10,R6);
	I3(ADD,R11,R11,R6);

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

	I3(CMP,NONE,R9,R5);
	I0P(B,NE,LoopY);

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

	InstPost(p->YMul);
	InstPost(p->RVMul);
	InstPost(p->RAdd);
	InstPost(p->GUMul);
	InstPost(p->GVMul);
	InstPost(p->GAdd);
	InstPost(p->BUMul);
	InstPost(p->BAdd);
	if (p->InvertMask) InstPost(p->InvertMask);
	if (p->DiffMask) InstPost(p->DiffMask);
	if (p->PalPtr) InstPost(p->PalPtr);
	if (p->LookUp)
	{
		Align(16);
		InstPost(p->LookUp);
	}
}

#endif

⌨️ 快捷键说明

复制代码 Ctrl + C
搜索代码 Ctrl + F
全屏模式 F11
切换主题 Ctrl + Shift + D
显示快捷键 ?
增大字号 Ctrl + =
减小字号 Ctrl + -