📄 blit_arm_yuv.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:
上一页 12
						I3S(ADD,R6,R6,R6,LSL,1);
						I3S(ADD,R4,R6,R4,LSL,3);
						I3S(ADD,R4,R4,R5,LSL,4);
						I3S(MOV,R4,NONE,R4,ASR,5);
						S(); I2C(ADD,Val,R4,16+p->FX.Brightness);
						if (p->FX.Brightness+16<0)
						{ 
							C(MI);I2C(MOV,Val,NONE,0x00); 
						}
						if (p->FX.Brightness+16+215>=255)
						{
							I2C(CMP,NONE,Val,0xFF);
							C(GT);I2C(MOV,Val,NONE,0xFF);
						}
						break;
					case 1:
						// U = (-5*R - 9*G + 14*B)/32 + 128;
						I3S(RSB,R6,R6,R6,LSL,3);
						I3(ADD,R6,R6,R6);
						I3S(ADD,R4,R4,R4,LSL,2);
						I3S(ADD,R5,R5,R5,LSL,3);
						I3(SUB,R4,R6,R4);
						I3(SUB,R4,R4,R5);
						I3S(MOV,R4,NONE,R4,ASR,5);
						I2C(ADD,Val,R4,128);
						break;
					case 2:
						// V = (14*R - 12*G - 2*B)/32 + 128;
						I3(ADD,R6,R6,R6);
						I3S(RSB,R4,R4,R4,LSL,3);
						I3S(ADD,R5,R5,R5,LSL,1);
						I3S(RSB,R4,R6,R4,LSL,1);
						I3S(SUB,R4,R4,R5,LSL,2);
						I3S(MOV,R4,NONE,R4,ASR,5);
						I2C(ADD,Val,R4,128);
						break;
					}

					if (x>0)
						I3S(ORR,Out,Out,Val,LSL,x*8);
				}
				I3S(ADD,R7,R7,R14,ASR,SrcUVPitch-HalfY);
			}
			I3S(SUB,R7,R7,R14,ASR,SrcUVPitch-2-HalfY);
		}
		else
		{
			int mode = HalfX + 2*HalfY;
			int x = 2;

			if (!p->SrcYUV && mode==0)
			{
				//possible alignment problems...
				mode = 1;
				x = 1;
			}

			switch (mode)
			{
			case 0:
				I3S(LDR_POST,R0,R7,R14,ASR,SrcUVPitch-1);
				I3S(LDR_POST,R2,R7,R14,ASR,SrcUVPitch);
				I3S(LDR_POSTSUB,R3,R7,R14,ASR,SrcUVPitch-1);
				I3S(LDR_POSTSUB,R1,R7,R14,ASR,SrcUVPitch);
				break;
			case 1: // halfx (not average...)
				Byte(); I2C(LDR,R0,R7,3*x);
				Byte(); I2C(LDR,R4,R7,2*x);
				Byte(); I2C(LDR,R5,R7,1*x);
				Byte(); I3S(LDR_POST,R6,R7,R14,ASR,SrcUVPitch-1);
				I3S(ORR,R0,R4,R0,LSL,8);
				I3S(ORR,R0,R5,R0,LSL,8);
				I3S(ORR,R0,R6,R0,LSL,8);

				Byte(); I2C(LDR,R2,R7,3*x);
				Byte(); I2C(LDR,R4,R7,2*x);
				Byte(); I2C(LDR,R5,R7,1*x);
				Byte(); I3S(LDR_POST,R6,R7,R14,ASR,SrcUVPitch);
				I3S(ORR,R2,R4,R2,LSL,8);
				I3S(ORR,R2,R5,R2,LSL,8);
				I3S(ORR,R2,R6,R2,LSL,8);

				Byte(); I2C(LDR,R3,R7,3*x);
				Byte(); I2C(LDR,R4,R7,2*x);
				Byte(); I2C(LDR,R5,R7,1*x);
				Byte(); I3S(LDR_POSTSUB,R6,R7,R14,ASR,SrcUVPitch-1);
				I3S(ORR,R3,R4,R3,LSL,8);
				I3S(ORR,R3,R5,R3,LSL,8);
				I3S(ORR,R3,R6,R3,LSL,8);

				Byte(); I2C(LDR,R1,R7,3*x);
				Byte(); I2C(LDR,R4,R7,2*x);
				Byte(); I2C(LDR,R5,R7,1*x);
				Byte(); I3S(LDR_POSTSUB,R6,R7,R14,ASR,SrcUVPitch);
				I3S(ORR,R1,R4,R1,LSL,8);
				I3S(ORR,R1,R5,R1,LSL,8);
				I3S(ORR,R1,R6,R1,LSL,8);
				break;
			case 2: // halfy
				I3S(LDR,R4,R7,R14,ASR,SrcUVPitch);
				I3S(LDR_POST,R0,R7,R14,ASR,SrcUVPitch-1-1);

				I3S(LDR,R6,R7,R14,ASR,SrcUVPitch);
				I3S(LDR_POST,R2,R7,R14,ASR,SrcUVPitch-1);

				I3S(AND,R4,R5,R4,LSR,1);
				I3S(AND,R0,R5,R0,LSR,1);
				I3(ADD,R0,R0,R4);

				I3S(AND,R6,R5,R6,LSR,1);
				I3S(AND,R2,R5,R2,LSR,1);
				I3(ADD,R2,R2,R6);

				I3S(LDR,R4,R7,R14,ASR,SrcUVPitch);
				I3S(LDR_POSTSUB,R3,R7,R14,ASR,SrcUVPitch-1-1);

				I3S(LDR,R6,R7,R14,ASR,SrcUVPitch);
				I3S(LDR_POSTSUB,R1,R7,R14,ASR,SrcUVPitch-1);

				I3S(AND,R4,R5,R4,LSR,1);
				I3S(AND,R3,R5,R3,LSR,1);
				I3(ADD,R3,R3,R4);

				I3S(AND,R6,R5,R6,LSR,1);
				I3S(AND,R1,R5,R1,LSR,1);
				I3(ADD,R1,R1,R4);
				break;
			case 3: // halfx + halfy (no average..., used by palette mode as well)
				Byte(); I2C(LDR,R0,R7,6);
				Byte(); I2C(LDR,R4,R7,4);
				Byte(); I2C(LDR,R5,R7,2);
				Byte(); I3S(LDR_POST,R6,R7,R14,ASR,SrcUVPitch-1-1);
				I3S(ORR,R0,R4,R0,LSL,8);
				I3S(ORR,R0,R5,R0,LSL,8);
				I3S(ORR,R0,R6,R0,LSL,8);

				Byte(); I2C(LDR,R2,R7,6);
				Byte(); I2C(LDR,R4,R7,4);
				Byte(); I2C(LDR,R5,R7,2);
				Byte(); I3S(LDR_POST,R6,R7,R14,ASR,SrcUVPitch-1);
				I3S(ORR,R2,R4,R2,LSL,8);
				I3S(ORR,R2,R5,R2,LSL,8);
				I3S(ORR,R2,R6,R2,LSL,8);

				Byte(); I2C(LDR,R3,R7,6);
				Byte(); I2C(LDR,R4,R7,4);
				Byte(); I2C(LDR,R5,R7,2);
				Byte(); I3S(LDR_POSTSUB,R6,R7,R14,ASR,SrcUVPitch-1-1);
				I3S(ORR,R3,R4,R3,LSL,8);
				I3S(ORR,R3,R5,R3,LSL,8);
				I3S(ORR,R3,R6,R3,LSL,8);

				Byte(); I2C(LDR,R1,R7,6);
				Byte(); I2C(LDR,R4,R7,4);
				Byte(); I2C(LDR,R5,R7,2);
				Byte(); I3S(LDR_POSTSUB,R6,R7,R14,ASR,SrcUVPitch-1);
				I3S(ORR,R1,R4,R1,LSL,8);
				I3S(ORR,R1,R5,R1,LSL,8);
				I3S(ORR,R1,R6,R1,LSL,8);
				break;
			}
		}

		if (p->SwapXY && (HalfX || HalfY || !p->SrcYUV))
			I2C(STR,R0,SP,OFS(stack,SaveR0));

		RA = (reg)YUV_4(p,R0,0,Plane);
		I3S(STR_POST,RA,R11,R12,ASR,DstUVPitch-1);
		RA = (reg)YUV_4(p,R2,16,Plane);
		I3S(STR_POST,RA,R11,R12,ASR,DstUVPitch);
		RA = (reg)YUV_4(p,R3,24,Plane);
		I3S(STR_POSTSUB,RA,R11,R12,ASR,DstUVPitch-1);
		RA = (reg)YUV_4(p,R1,8,Plane);
		I3S(STR_POSTSUB,RA,R11,R12,ASR,DstUVPitch);

		if (p->LookUp && p->LookUpOfs)
		{
			I2C(SUB,R8,R8,p->LookUpOfs);
			p->LookUpOfs = 0;
		}

		if (p->RealOnlyDiff)
			InstPost(p->Skip);

		I2C(ADD,R7,R7,(4 << HalfX)*(p->SrcBPP/8));
		if (p->SwapXY)
			I3S(ADD,R11,R11,R12,ASR,DstUVPitch-2);
		else
			I2C(ADD,R11,R11,4*p->DirX);
		I3(CMP,NONE,R7,R10);
		I0P(B,NE,LoopX);

		I2C(LDR,R0,SP,OFS(stack,SrcNext[UV]));
		I2C(LDR,R1,SP,OFS(stack,DstNext[UV]));
		I3(ADD,R7,R7,R0);
		I3(ADD,R11,R11,R1);
	}

	I2C(STR,R7,SP,OFS(stack,Src[Plane]));
	I2C(STR,R11,SP,OFS(stack,Dst[Plane]));
}

void Fix_Any_YUV(blit_soft* p)
{
	dyninst* Add32 = NULL;
	dyninst* MaskCarry = NULL;
	dyninst* LoopY;
	reg RSrcWidth;

	p->SrcAlignPos = p->DstAlignPos = p->DstAlignSize = 8;

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

	p->DiffMask = NULL;

	I2C(LDR,R6,R1,0);//Dst[0] U
	I2C(LDR,R7,R1,4);//Dst[1] V
	I2C(LDR,R8,R1,8);//Dst[2] Y

	I2C(LDR,R10,R2,0); //Src[0] U
	if (!p->SrcYUV)
	{
		I3(MOV,R11,NONE,R10);
		I3(MOV,R12,NONE,R10);
		p->SrcUVX2 = 0;
		p->SrcUVY2 = 0;
		p->SrcUVPitch2 = 0;
		p->Caps = VC_BRIGHTNESS;
	}	
	else
	{
		I2C(LDR,R11,R2,4); //Src[1] V
		I2C(LDR,R12,R2,8); //Src[2] Y
	}

	if (p->DirX<0)
	{
		//adjust reversed destination for block size
		I2C(SUB,R6,R6,3);
		I2C(SUB,R7,R7,3);
		I2C(SUB,R8,R8,3);
	}

	I2C(STR,R6,SP,OFS(stack,Dst[0]));
	I2C(STR,R7,SP,OFS(stack,Dst[1]));
	I2C(STR,R8,SP,OFS(stack,Dst[2]));

	I2C(STR,R10,SP,OFS(stack,Src[0]));
	I2C(STR,R11,SP,OFS(stack,Src[1]));
	I2C(STR,R12,SP,OFS(stack,Src[2]));

	I3(MOV,R12,NONE,R3);	 //DstPitch
	I2C(LDR,R14,SP,OFS(stack,SrcPitch));

	I2C(LDR,R2,SP,OFS(stack,Height));
	I2C(LDR,R1,SP,OFS(stack,Width));

	p->RealOnlyDiff = p->OnlyDiff;
	p->HalfX = (boolmem_t)(p->SrcUVX2+1 == p->DstUVX2);
	p->HalfY = (boolmem_t)(p->SrcUVY2+1 == p->DstUVY2);

	if (p->HalfX || p->HalfY)
		p->RealOnlyDiff = 0;

	if (p->RealOnlyDiff)
	{
		int Mask = 0x03030303;
		p->DiffMask = InstCreate32(Mask,NONE,NONE,NONE,0,0);
		I1P(LDR,R5,p->DiffMask,0);
		I2C(LDR,R6,SP,OFS(stack,Src2SrcLast));
	}
	else
	if ((p->HalfX + 2*p->HalfY)==2)
	{
		int Mask = 0x7F7F7F7F;
		p->DiffMask = InstCreate32(Mask,NONE,NONE,NONE,0,0);
		I1P(LDR,R5,p->DiffMask,0);
	}

	p->LookUp = NULL;
	if (p->SrcYUV)
		CalcYUVLookUp(p);
	else if (p->Src.Palette)
		CalcPalYUVLookUp(p);

	if (p->LookUp_Data)
	{
		p->LookUp = InstCreate(p->LookUp_Data,p->LookUp_Size,NONE,NONE,NONE,0,0);
		free(p->LookUp_Data);
		p->LookUp_Data = NULL;
		I1P(MOV,R8,p->LookUp,0);
		p->LookUpOfs = 0;
	}
	else
	if (p->FX.Brightness)
	{
		int i = p->FX.Brightness;
		if (i<0) i=-i;
		if (i>127) i=127;
		MaskCarry = InstCreate32(0x80808080U,NONE,NONE,NONE,0,0);
		Add32 = InstCreate32(0x01010101U * (uint8_t)i,NONE,NONE,NONE,0,0);
		I1P(LDR,R8,MaskCarry,0);
		I1P(LDR,R9,Add32,0);
	}

	//EndOfRect
	//DstNext[2] 
	//SrcNext[2]

	RSrcWidth = R1;
	if (p->SrcBPP2>0)
	{
		IMul(R3,R1,p->SrcBPP/8);
		RSrcWidth = R3;
	}

	//SrcYNext = 4*Src->Pitch - Width
	I3S(RSB,R0,RSrcWidth,R14,LSL,2); 
	I2C(STR,R0,SP,OFS(stack,SrcNext[0]));

	//SrcUVNext = (4 << HalfY)*(Src->Pitch >> SrcUVPitch2) - (Width >> SrcUVX)
	I3S(MOV,R0,NONE,R14,LSL,2+p->HalfY-p->SrcUVPitch2); 
	I3S(SUB,R0,R0,RSrcWidth,LSR,p->SrcUVX2); 
	I2C(STR,R0,SP,OFS(stack,SrcNext[1]));

	//EndOfRect = Src + Src->Pitch * Height
	I3(MUL,R0,R14,R2);
	I3(ADD,R0,R0,R10);
	I2C(STR,R0,SP,OFS(stack,EndOfRect));

	if (p->SwapXY)
	{
		//DstYNext = 4*DirX - Width*Dst->Pitch
		I3(MUL,R3,R1,R12);
		I2C(MOV,R0,NONE,4*p->DirX);
		I3(SUB,R0,R0,R3);
		I2C(STR,R0,SP,OFS(stack,DstNext[0]));

		//DstUVNext = 4*DirX - (Width >> DstUVY2)*(Dst->Pitch >> DstUVPitch2)
		if (p->DstUVY2+p->DstUVPitch2) I3S(MOV,R3,NONE,R3,ASR,p->DstUVY2+p->DstUVPitch2);
		I2C(MOV,R0,NONE,4*p->DirX);
		I3(SUB,R0,R0,R3);
		I2C(STR,R0,SP,OFS(stack,DstNext[1]));
	}
	else
	{
		//DstYNext = 4*Dst->Pitch - DirX * Width
		I3S(p->DirX<0?ADD:RSB,R0,R1,R12,LSL,2); 
		I2C(STR,R0,SP,OFS(stack,DstNext[0]));

		//DstUVNext = 4*Dst->Pitch >> DstUVPitch2 - DirX * (Width >> DstUVX2)
		I3S(MOV,R0,NONE,R12,LSL,2-p->DstUVPitch2);
		I3S(p->DirX<0?ADD:SUB,R0,R0,R1,LSR,p->DstUVX2); 
		I2C(STR,R0,SP,OFS(stack,DstNext[1]));
	}

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

	if (p->DiffMask) InstPost(p->DiffMask);
	if (Add32) InstPost(Add32);
	if (MaskCarry) InstPost(MaskCarry);
	if (p->LookUp)
	{
		Align(16);
		InstPost(p->LookUp);
	}

	InstPost(LoopY);
	{
		if (p->LookUp)
			I2C(ADD,R8,R8,256+4);
		YUV_4X4(p,1);
		if (p->LookUp)
			I2C(ADD,R8,R8,256);
		YUV_4X4(p,2);
		if (p->LookUp)
			I2C(SUB,R8,R8,256+256+4);
		YUV_4X4(p,0);

		MB(); I2C(LDR,R2,SP,OFS(stack,EndOfRect));
		I3(CMP,NONE,R2,R7);
		I0P(B,NE,LoopY);
	}

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

#endif
上一页 12
⌨️ 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 + -