ShuriZma
/
desmume
mirror of https://github.com/TASEmulators/desmume.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Colorspace Handler: Colorspace conversions no longer do big-endian byte swapping by default. Instead, byte swapping must be explicitly requested. 

- Byte swapping can now be independently controlled for both input and output data.
- As an application to this new API, VRAM display mode now shows the correct colors on big-endian systems.
This commit is contained in:
rogerman 2021-09-11 20:59:41 -07:00
parent 329eade565
commit ad64e73bb2
22 changed files with 506 additions and 313 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
desmume/src/FIFO.cpp
View File

@ -436,11 +436,11 @@ void _DISP_FIFOrecv_LineOpaque32_vec(u32 *__restrict dst)
if (OUTPUTFORMAT == NDSColorFormat_BGR666_Rev)
{
ColorspaceConvert555To6665Opaque_AltiVec<false>(fifoColor, dstLo, dstHi);
ColorspaceConvert555To6665Opaque_AltiVec<false, BESwapDst>(fifoColor, dstLo, dstHi);
}
else if (OUTPUTFORMAT == NDSColorFormat_BGR888_Rev)
{
ColorspaceConvert555To8888Opaque_AltiVec<false>(fifoColor, dstLo, dstHi);
ColorspaceConvert555To8888Opaque_AltiVec<false, BESwapDst>(fifoColor, dstLo, dstHi);
}
vec_st(dstLo, d + 0, dst);
@ -449,11 +449,11 @@ void _DISP_FIFOrecv_LineOpaque32_vec(u32 *__restrict dst)
#else
if (OUTPUTFORMAT == NDSColorFormat_BGR666_Rev)
{
ColorspaceConvertBuffer555To6665Opaque<false, false>((u16 *)(disp_fifo.buf + disp_fifo.head), dst, GPU_FRAMEBUFFER_NATIVE_WIDTH * sizeof(u16));
ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapDst>((u16 *)(disp_fifo.buf + disp_fifo.head), dst, GPU_FRAMEBUFFER_NATIVE_WIDTH * sizeof(u16));
}
else if (OUTPUTFORMAT == NDSColorFormat_BGR888_Rev)
{
ColorspaceConvertBuffer555To8888Opaque<false, false>((u16 *)(disp_fifo.buf + disp_fifo.head), dst, GPU_FRAMEBUFFER_NATIVE_WIDTH * sizeof(u16));
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>((u16 *)(disp_fifo.buf + disp_fifo.head), dst, GPU_FRAMEBUFFER_NATIVE_WIDTH * sizeof(u16));
}
_DISP_FIFOrecv_LineAdvance();
#endif
@ -499,13 +499,13 @@ void DISP_FIFOrecv_LineOpaque(u32 *__restrict dst)
if (OUTPUTFORMAT == NDSColorFormat_BGR666_Rev)
{
dst[i+0] = COLOR555TO6665_OPAQUE((src >> 0) & 0x7FFF);
dst[i+1] = COLOR555TO6665_OPAQUE((src >> 16) & 0x7FFF);
dst[i+0] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>((src >> 0) & 0x7FFF) );
dst[i+1] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>((src >> 16) & 0x7FFF) );
}
else if (OUTPUTFORMAT == NDSColorFormat_BGR888_Rev)
{
dst[i+0] = COLOR555TO8888_OPAQUE((src >> 0) & 0x7FFF);
dst[i+1] = COLOR555TO8888_OPAQUE((src >> 16) & 0x7FFF);
dst[i+0] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>((src >> 0) & 0x7FFF) );
dst[i+1] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>((src >> 16) & 0x7FFF) );
}
}
}

28
desmume/src/GPU.cpp
View File

@ -365,7 +365,7 @@ void GPUEngineBase::_Reset_Base()
renderState.selectedBGLayer = &this->_BGLayer[GPULayerID_BG0];
renderState.backdropColor16 = LE_TO_LOCAL_16(this->_paletteBG[0]) & 0x7FFF;
renderState.workingBackdropColor16 = renderState.backdropColor16;
renderState.workingBackdropColor32.color = (dispInfo.colorFormat == NDSColorFormat_BGR666_Rev) ? COLOR555TO666(LOCAL_TO_LE_16(renderState.workingBackdropColor16)) : COLOR555TO888(LOCAL_TO_LE_16(renderState.workingBackdropColor16));
renderState.workingBackdropColor32.color = LOCAL_TO_LE_32( (dispInfo.colorFormat == NDSColorFormat_BGR666_Rev) ? COLOR555TO666(LOCAL_TO_LE_16(renderState.workingBackdropColor16)) : COLOR555TO888(LOCAL_TO_LE_16(renderState.workingBackdropColor16)) );
renderState.colorEffect = (ColorEffect)this->_IORegisterMap->BLDCNT.ColorEffect;
renderState.blendEVA = 0;
renderState.blendEVB = 0;
@ -864,7 +864,7 @@ void GPUEngineBase::UpdateRenderStates(const size_t l)
{
currRenderState.workingBackdropColor16 = currRenderState.backdropColor16;
}
currRenderState.workingBackdropColor32.color = (OUTPUTFORMAT == NDSColorFormat_BGR666_Rev) ? COLOR555TO666(LOCAL_TO_LE_16(currRenderState.workingBackdropColor16)) : COLOR555TO888(LOCAL_TO_LE_16(currRenderState.workingBackdropColor16));
currRenderState.workingBackdropColor32.color = LOCAL_TO_LE_32( (OUTPUTFORMAT == NDSColorFormat_BGR666_Rev) ? COLOR555TO666(LOCAL_TO_LE_16(currRenderState.workingBackdropColor16)) : COLOR555TO888(LOCAL_TO_LE_16(currRenderState.workingBackdropColor16)) );
// Save the current render states to this line's compositor info.
compInfo.renderState = currRenderState;
@ -3949,7 +3949,7 @@ void GPUEngineA::_RenderLine_DisplayCaptureCustom(const IOREG_DISPCAPCNT &DISPCA
{
if (OUTPUTFORMAT == NDSColorFormat_BGR888_Rev)
{
ColorspaceConvertBuffer555To8888Opaque<false, false>(this->_fifoLine16, (u32 *)srcBPtr, GPU_FRAMEBUFFER_NATIVE_WIDTH);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>(this->_fifoLine16, (u32 *)srcBPtr, GPU_FRAMEBUFFER_NATIVE_WIDTH);
}
this->_RenderLine_DispCapture_Copy<OUTPUTFORMAT, 1, CAPTURELENGTH, true, false>(lineInfo, srcBPtr, dstCustomPtr, captureLengthExt);
@ -3990,7 +3990,7 @@ void GPUEngineA::_RenderLine_DisplayCaptureCustom(const IOREG_DISPCAPCNT &DISPCA
{
if ((OUTPUTFORMAT == NDSColorFormat_BGR888_Rev) && (DISPCAPCNT.SrcB != 0))
{
ColorspaceConvertBuffer555To8888Opaque<false, false>(this->_fifoLine16, (u32 *)srcBPtr, GPU_FRAMEBUFFER_NATIVE_WIDTH);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>(this->_fifoLine16, (u32 *)srcBPtr, GPU_FRAMEBUFFER_NATIVE_WIDTH);
}
CopyLineExpandHinted<0xFFFF, true, false, false, 4>(lineInfo, srcBPtr, this->_captureWorkingB32);
@ -4139,7 +4139,7 @@ void GPUEngineA::_RenderLine_DisplayCapture(const GPUEngineCompositorInfo &compI
{
if (willReadNativeVRAM)
{
ColorspaceConvertBuffer555To8888Opaque<false, false>(vramNative16, (u32 *)vramCustom32, GPU_FRAMEBUFFER_NATIVE_WIDTH);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>(vramNative16, (u32 *)vramCustom32, GPU_FRAMEBUFFER_NATIVE_WIDTH);
}
}
@ -4669,7 +4669,7 @@ void GPUEngineA::_HandleDisplayModeVRAM(const GPUEngineLineInfo &lineInfo)
{
const u16 *src = this->_VRAMNativeBlockPtr[DISPCNT.VRAM_Block] + lineInfo.blockOffsetNative;
u32 *dst = (u32 *)this->_nativeBuffer + lineInfo.blockOffsetNative;
ColorspaceConvertBuffer555To6665Opaque<false, false>(src, dst, GPU_FRAMEBUFFER_NATIVE_WIDTH);
ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapSrcDst>(src, dst, GPU_FRAMEBUFFER_NATIVE_WIDTH);
break;
}
@ -4677,7 +4677,7 @@ void GPUEngineA::_HandleDisplayModeVRAM(const GPUEngineLineInfo &lineInfo)
{
const u16 *src = this->_VRAMNativeBlockPtr[DISPCNT.VRAM_Block] + lineInfo.blockOffsetNative;
u32 *dst = (u32 *)this->_nativeBuffer + lineInfo.blockOffsetNative;
ColorspaceConvertBuffer555To8888Opaque<false, false>(src, dst, GPU_FRAMEBUFFER_NATIVE_WIDTH);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapSrcDst>(src, dst, GPU_FRAMEBUFFER_NATIVE_WIDTH);
break;
}
}
@ -4694,7 +4694,7 @@ void GPUEngineA::_HandleDisplayModeVRAM(const GPUEngineLineInfo &lineInfo)
{
const u16 *src = (u16 *)this->_VRAMCustomBlockPtr[DISPCNT.VRAM_Block] + lineInfo.blockOffsetCustom;
u32 *dst = (u32 *)this->_customBuffer + lineInfo.blockOffsetCustom;
ColorspaceConvertBuffer555To6665Opaque<false, false>(src, dst, lineInfo.pixelCount);
ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapSrcDst>(src, dst, lineInfo.pixelCount);
break;
}
@ -5937,11 +5937,11 @@ void GPUSubsystem::ClearWithColor(const u16 colorBGRA5551)
break;
case NDSColorFormat_BGR666_Rev:
color32.color = COLOR555TO6665_OPAQUE(colorBGRA5551 & 0x7FFF);
color32.color = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>(colorBGRA5551 & 0x7FFF) );
break;
case NDSColorFormat_BGR888_Rev:
color32.color = COLOR555TO8888_OPAQUE(colorBGRA5551 & 0x7FFF);
color32.color = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(colorBGRA5551 & 0x7FFF) );
break;
default:
@ -6139,7 +6139,7 @@ bool GPUSubsystem::LoadState(EMUFILE &is, int size)
{
if (this->_displayInfo.isDisplayEnabled[NDSDisplayID_Main])
{
ColorspaceConvertBuffer555To6665Opaque<false, false>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Main], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Main], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapDst>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Main], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Main], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
}
else
{
@ -6148,7 +6148,7 @@ bool GPUSubsystem::LoadState(EMUFILE &is, int size)
if (this->_displayInfo.isDisplayEnabled[NDSDisplayID_Touch])
{
ColorspaceConvertBuffer555To6665Opaque<false, false>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Touch], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Touch], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapDst>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Touch], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Touch], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
}
else
{
@ -6161,7 +6161,7 @@ bool GPUSubsystem::LoadState(EMUFILE &is, int size)
{
if (this->_displayInfo.isDisplayEnabled[NDSDisplayID_Main])
{
ColorspaceConvertBuffer555To8888Opaque<false, false>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Main], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Main], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Main], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Main], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
}
else
{
@ -6170,7 +6170,7 @@ bool GPUSubsystem::LoadState(EMUFILE &is, int size)
if (this->_displayInfo.isDisplayEnabled[NDSDisplayID_Touch])
{
ColorspaceConvertBuffer555To8888Opaque<false, false>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Touch], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Touch], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>((u16 *)this->_displayInfo.customBuffer[NDSDisplayID_Touch], (u32 *)this->_displayInfo.nativeBuffer[NDSDisplayID_Touch], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
}
else
{

20
desmume/src/GPU_Operations.cpp
View File

@ -242,8 +242,8 @@ void PixelOperation::InitLUTs()
cur.bits.blue = (cur.bits.blue + ((31 - cur.bits.blue) * i / 16));
cur.bits.alpha = 0;
PixelOperation::BrightnessUpTable555[i][j] = cur.val;
PixelOperation::BrightnessUpTable666[i][j].color = COLOR555TO666(cur.val);
PixelOperation::BrightnessUpTable888[i][j].color = COLOR555TO888(cur.val);
PixelOperation::BrightnessUpTable666[i][j].color = LOCAL_TO_LE_32( COLOR555TO666(cur.val) );
PixelOperation::BrightnessUpTable888[i][j].color = LOCAL_TO_LE_32( COLOR555TO888(cur.val) );
cur.val = j;
cur.bits.red = (cur.bits.red - (cur.bits.red * i / 16));
@ -251,8 +251,8 @@ void PixelOperation::InitLUTs()
cur.bits.blue = (cur.bits.blue - (cur.bits.blue * i / 16));
cur.bits.alpha = 0;
PixelOperation::BrightnessDownTable555[i][j] = cur.val;
PixelOperation::BrightnessDownTable666[i][j].color = COLOR555TO666(cur.val);
PixelOperation::BrightnessDownTable888[i][j].color = COLOR555TO888(cur.val);
PixelOperation::BrightnessDownTable666[i][j].color = LOCAL_TO_LE_32( COLOR555TO666(cur.val) );
PixelOperation::BrightnessDownTable888[i][j].color = LOCAL_TO_LE_32( COLOR555TO888(cur.val) );
}
}
@ -289,11 +289,11 @@ FORCEINLINE void PixelOperation::_copy16(GPUEngineCompositorInfo &compInfo, cons
break;
case NDSColorFormat_BGR666_Rev:
dstColor32.color = ColorspaceConvert555To6665Opaque<false>(srcColor16);
dstColor32.color = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>(srcColor16) );
break;
case NDSColorFormat_BGR888_Rev:
dstColor32.color = ColorspaceConvert555To8888Opaque<false>(srcColor16);
dstColor32.color = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(srcColor16) );
break;
}
@ -562,7 +562,7 @@ FORCEINLINE void PixelOperation::_unknownEffect16(GPUEngineCompositorInfo &compI
switch (selectedEffect)
{
case ColorEffect_Disable:
dstColor32.color = ColorspaceConvert555To6665Opaque<false>(srcColor16);
dstColor32.color = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>(srcColor16) );
break;
case ColorEffect_IncreaseBrightness:
@ -576,7 +576,7 @@ FORCEINLINE void PixelOperation::_unknownEffect16(GPUEngineCompositorInfo &compI
case ColorEffect_Blend:
{
FragmentColor srcColor32;
srcColor32.color = ColorspaceConvert555To6665Opaque<false>(srcColor16);
srcColor32.color = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>(srcColor16) );
dstColor32 = (LAYERTYPE == GPULayerType_3D) ? colorop.blend3D<OUTPUTFORMAT>(srcColor32, dstColor32) : colorop.blend<OUTPUTFORMAT>(srcColor32, dstColor32, blendEVA, blendEVB);
break;
}
@ -587,7 +587,7 @@ FORCEINLINE void PixelOperation::_unknownEffect16(GPUEngineCompositorInfo &compI
switch (selectedEffect)
{
case ColorEffect_Disable:
dstColor32.color = ColorspaceConvert555To8888Opaque<false>(srcColor16);
dstColor32.color = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(srcColor16) );
break;
case ColorEffect_IncreaseBrightness:
@ -601,7 +601,7 @@ FORCEINLINE void PixelOperation::_unknownEffect16(GPUEngineCompositorInfo &compI
case ColorEffect_Blend:
{
FragmentColor srcColor32;
srcColor32.color = ColorspaceConvert555To8888Opaque<false>(srcColor16);
srcColor32.color = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(srcColor16) );
dstColor32 = (LAYERTYPE == GPULayerType_3D) ? colorop.blend3D<OUTPUTFORMAT>(srcColor32, dstColor32) : colorop.blend<OUTPUTFORMAT>(srcColor32, dstColor32, blendEVA, blendEVB);
break;
}

8
desmume/src/frontend/cocoa/OGLDisplayOutput.cpp
View File

@ -4663,7 +4663,7 @@ void OGLClientFetchObject::FetchNativeDisplayToSrcClone(const NDSDisplayID displ
if (this->_fetchColorFormatOGL == GL_UNSIGNED_SHORT_1_5_5_5_REV)
{
ColorspaceConvertBuffer555To8888Opaque<false, false>((const uint16_t *)this->_fetchDisplayInfo[bufferIndex].nativeBuffer[displayID], this->_srcNativeClone[displayID][bufferIndex], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>((const uint16_t *)this->_fetchDisplayInfo[bufferIndex].nativeBuffer[displayID], this->_srcNativeClone[displayID][bufferIndex], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
}
else
{
@ -5500,9 +5500,9 @@ OGLImage::OGLImage(OGLContextInfo *contextInfo, GLsizei imageWidth, GLsizei imag
const GLint vtxBuffer[8] = {
(GLint)(-_normalWidth/2.0), (GLint)( _normalHeight/2.0),
(GLint)( _normalWidth/2.0), (GLint)( _normalHeight/2.0),
(GLint)(-_normalWidth/2.0), (GLint)(-_normalHeight/2.0),
(GLint)( _normalWidth/2.0), (GLint)(-_normalHeight/2.0)
(GLint)( _normalWidth/2.0), (GLint)( _normalHeight/2.0),
(GLint)(-_normalWidth/2.0), (GLint)(-_normalHeight/2.0),
(GLint)( _normalWidth/2.0), (GLint)(-_normalHeight/2.0)
};
// Set up VBOs

116
desmume/src/frontend/cocoa/cocoa_rom.mm
View File

@ -1,6 +1,6 @@
/*
Copyright (C) 2011 Roger Manuel
Copyright (C) 2011-2017 DeSmuME team
Copyright (C) 2011-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -640,17 +640,15 @@ static NSMutableDictionary *saveTypeValues = nil;
@end
/********************************************************************************************
RomIconToRGBA8888()
@function RomIconToRGBA8888()
Reads the icon image data from a ROM and converts it to an RGBA8888 formatted bitmap.
@brief Reads the icon image data from a ROM and converts it to an RGBA8888 formatted bitmap.
Takes:
bitmapData - Write pointer for the icon's pixel data.
@param bitmapData Write pointer for the icon's pixel data.
Returns:
Nothing.
@return Nothing
Details:
@discussion
- If bitmapData is NULL, then this function immediately returns and does nothing.
- If no ROM is loaded, then bitmapData will have a black square icon.
- The caller is responsible for ensuring that bitmapData points to a valid
@ -660,110 +658,68 @@ static NSMutableDictionary *saveTypeValues = nil;
********************************************************************************************/
void RomIconToRGBA8888(uint32_t *bitmapData)
{
const RomBanner &ndsRomBanner = gameInfo.getRomBanner(); // Contains the memory addresses we need to get our read pointer locations.
const uint16_t *iconClutPtr; // Read pointer for the icon's CLUT.
const uint32_t *iconPixPtr; // Read pointer for the icon's pixel data.
uint32_t clut[16]; // 4-bit indexed CLUT, storing RGBA8888 values for each color.
uint32_t pixRowColors; // Temp location for storing an 8 pixel row of 4-bit indexed color values from the icon's pixel data.
unsigned int pixRowIndex; // Temp location for tracking which pixel row of an 8x8 square that we are reading.
unsigned int x; // Temp location for tracking which of the 8x8 pixel squares that we are reading (x-dimension).
unsigned int y; // Temp location for tracking which of the 8x8 pixel squares that we are reading (y-dimension).
uint32_t *bitmapPixPtr; // Write pointer for the RGBA8888 bitmap pixel data, relative to the passed in *bitmapData pointer.
if (bitmapData == NULL)
{
return;
}
// Set all of our icon read pointers.
iconClutPtr = (uint16_t *)ndsRomBanner.palette + 1;
iconPixPtr = (uint32_t *)ndsRomBanner.bitmap;
const RomBanner &ndsRomBanner = gameInfo.getRomBanner(); // Contains the memory addresses we need to get our read pointer locations.
const uint32_t *iconPixPtr = (uint32_t *)ndsRomBanner.bitmap; // Read pointer for the icon's pixel data.
// Setup the 4-bit CLUT.
//
// The actual color values are stored with the ROM icon data in RGB555 format.
// We convert these color values and store them in the CLUT as RGBA8888 values.
//
// The first entry always represents the alpha, so we can just ignore it.
clut[0] = 0x00000000;
ColorspaceConvertBuffer555To8888Opaque<false, true>((u16 *)iconClutPtr, &clut[1], 15);
// The first entry always represents the alpha, so just set it to 0.
const uint16_t *clut4 = (uint16_t *)ndsRomBanner.palette;
CACHE_ALIGN uint32_t clut32[16];
ColorspaceConvertBuffer555To8888Opaque<false, true, BESwapNone>(clut4, clut32, 16);
clut32[0] = 0x00000000;
// Load the image from the icon pixel data.
//
// ROM icons are stored in 4-bit indexed color and have dimensions of 32x32 pixels.
// Also, ROM icons are split into 16 separate 8x8 pixel squares arranged in a 4x4
// Also, ROM icons are split into 16 separate 8x8 pixel tiles arranged in a 4x4
// array. Here, we sequentially read from the ROM data, and adjust our write
// location appropriately within the bitmap memory block.
for(y = 0; y < 4; y++)
for (size_t y = 0; y < 4; y++)
{
for(x = 0; x < 4; x++)
for (size_t x = 0; x < 4; x++)
{
for(pixRowIndex = 0; pixRowIndex < 8; pixRowIndex++, iconPixPtr++)
for (size_t p = 0; p < 8; p++, iconPixPtr++)
{
// Load the entire row of pixels as a single 32-bit chunk.
pixRowColors = *iconPixPtr;
// Load an entire row of palette color indices as a single 32-bit chunk.
const uint32_t palIdx = LE_TO_LOCAL_32(*iconPixPtr);
// Set the write location. The formula below calculates the proper write
// location depending on the position of the read pointer. We use a more
// optimized version of this formula in practice.
//
// bitmapPixPtr = bitmapData + ( ((y * 8) + pixRowIndex) * 32 ) + (x * 8);
bitmapPixPtr = bitmapData + ( ((y << 3) + pixRowIndex) << 5 ) + (x << 3);
// bitmapOutPtr = bitmapData + ( ((y * 8) + palIdx) * 32 ) + (x * 8);
uint32_t *bitmapOutPtr = bitmapData + ( ((y << 3) + p) << 5 ) + (x << 3);
*bitmapOutPtr = clut32[(palIdx & 0x0000000F) >> 0];
// Set the RGBA8888 bitmap pixels using our CLUT from earlier.
bitmapOutPtr++;
*bitmapOutPtr = clut32[(palIdx & 0x000000F0) >> 4];
#ifdef MSB_FIRST
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0x0F000000) >> 24]);
bitmapOutPtr++;
*bitmapOutPtr = clut32[(palIdx & 0x00000F00) >> 8];
bitmapPixPtr++;
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0xF0000000) >> 28]);
bitmapOutPtr++;
*bitmapOutPtr = clut32[(palIdx & 0x0000F000) >> 12];
bitmapPixPtr++;
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0x000F0000) >> 16]);
bitmapOutPtr++;
*bitmapOutPtr = clut32[(palIdx & 0x000F0000) >> 16];
bitmapPixPtr++;
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0x00F00000) >> 20]);
bitmapOutPtr++;
*bitmapOutPtr = clut32[(palIdx & 0x00F00000) >> 20];
bitmapPixPtr++;
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0x00000F00) >> 8]);
bitmapOutPtr++;
*bitmapOutPtr = clut32[(palIdx & 0x0F000000) >> 24];
bitmapPixPtr++;
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0x0000F000) >> 12]);
bitmapPixPtr++;
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0x0000000F)]);
bitmapPixPtr++;
*bitmapPixPtr = LOCAL_TO_LE_32(clut[(pixRowColors & 0x000000F0) >> 4]);
#else
*bitmapPixPtr = clut[(pixRowColors & 0x0000000F)];
bitmapPixPtr++;
*bitmapPixPtr = clut[(pixRowColors & 0x000000F0) >> 4];
bitmapPixPtr++;
*bitmapPixPtr = clut[(pixRowColors & 0x00000F00) >> 8];
bitmapPixPtr++;
*bitmapPixPtr = clut[(pixRowColors & 0x0000F000) >> 12];
bitmapPixPtr++;
*bitmapPixPtr = clut[(pixRowColors & 0x000F0000) >> 16];
bitmapPixPtr++;
*bitmapPixPtr = clut[(pixRowColors & 0x00F00000) >> 20];
bitmapPixPtr++;
*bitmapPixPtr = clut[(pixRowColors & 0x0F000000) >> 24];
bitmapPixPtr++;
*bitmapPixPtr = clut[(pixRowColors & 0xF0000000) >> 28];
#endif
bitmapOutPtr++;
*bitmapOutPtr = clut32[(palIdx & 0xF0000000) >> 28];
}
}
}

2
desmume/src/frontend/cocoa/userinterface/MacMetalDisplayView.mm
View File

@ -2456,7 +2456,7 @@ void MacMetalFetchObject::_FetchNativeDisplayByID(const NDSDisplayID displayID,
if (this->_fetchDisplayInfo[bufferIndex].pixelBytes == 2)
{
ColorspaceConvertBuffer555To8888Opaque<false, false>((const uint16_t *)this->_fetchDisplayInfo[bufferIndex].nativeBuffer[displayID], this->_srcNativeClone[displayID][bufferIndex], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>((const uint16_t *)this->_fetchDisplayInfo[bufferIndex].nativeBuffer[displayID], this->_srcNativeClone[displayID][bufferIndex], GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT);
}
else
{

4
desmume/src/frontend/posix/gtk/main.cpp
View File

@ -1301,7 +1301,7 @@ static int ConfigureDrawingArea(GtkWidget *widget, GdkEventConfigure *event, gpo
static inline void gpu_screen_to_rgb(u32* dst)
{
ColorspaceConvertBuffer555To8888Opaque<false, false>((const uint16_t *)GPU->GetDisplayInfo().masterNativeBuffer, dst, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT * 2);
ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapDst>((const uint16_t *)GPU->GetDisplayInfo().masterNativeBuffer, dst, GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT * 2);
}
static inline void drawScreen(cairo_t* cr, u32* buf, gint w, gint h) {
@ -1427,7 +1427,7 @@ static gboolean ExposeDrawingArea (GtkWidget *widget, GdkEventExpose *event, gpo
}
static void RedrawScreen() {
ColorspaceConvertBuffer555To8888Opaque<true, false>((const uint16_t *)GPU->GetDisplayInfo().masterNativeBuffer, (uint32_t *)video->GetSrcBufferPtr(), GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT * 2);
ColorspaceConvertBuffer555To8888Opaque<true, false, BESwapDst>((const uint16_t *)GPU->GetDisplayInfo().masterNativeBuffer, (uint32_t *)video->GetSrcBufferPtr(), GPU_FRAMEBUFFER_NATIVE_WIDTH * GPU_FRAMEBUFFER_NATIVE_HEIGHT * 2);
#ifdef HAVE_LIBAGG
aggDraw.hud->attach((u8*)video->GetSrcBufferPtr(), 256, 384, 1024);
osd->update();

2
desmume/src/frontend/windows/display.cpp
View File

@ -683,7 +683,7 @@ void DoDisplay()
//we have to do a copy here because we're about to draw the OSD onto it. bummer.
if (gpu_bpp == 15)
ColorspaceConvertBuffer555To8888Opaque<true, false>((u16 *)video.srcBuffer, video.buffer, video.srcBufferSize / 2);
ColorspaceConvertBuffer555To8888Opaque<true, false, BESwapNone>((u16 *)video.srcBuffer, video.buffer, video.srcBufferSize / 2);
else
ColorspaceConvertBuffer888XTo8888Opaque<true, false>((u32*)video.srcBuffer, video.buffer, video.srcBufferSize / 4);

8
desmume/src/rasterize.cpp
View File

@ -2034,7 +2034,7 @@ Render3DError SoftRasterizerRenderer::BeginRender(const GFX3D &engine)
}
// Convert the toon table colors
ColorspaceConvertBuffer555To6665Opaque<false, false>(engine.renderState.u16ToonTable, (u32 *)this->toonColor32LUT, 32);
ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapDst>(engine.renderState.u16ToonTable, (u32 *)this->toonColor32LUT, 32);
if (this->_enableEdgeMark)
{
@ -2103,7 +2103,7 @@ void SoftRasterizerRenderer::_UpdateEdgeMarkColorTable(const u16 *edgeMarkColorT
//we can do this by rendering a 3d frame and then freezing the system, but only changing the edge mark colors
for (size_t i = 0; i < 8; i++)
{
this->_edgeMarkTable[i].color = COLOR555TO6665(edgeMarkColorTable[i] & 0x7FFF, (this->currentRenderState->enableAntialiasing) ? 0x10 : 0x1F);
this->_edgeMarkTable[i].color = LE_TO_LOCAL_32( COLOR555TO6665(edgeMarkColorTable[i] & 0x7FFF, (this->currentRenderState->enableAntialiasing) ? 0x10 : 0x1F) );
//zero 20-jun-2013 - this doesnt make any sense. at least, it should be related to the 0x8000 bit. if this is undocumented behaviour, lets write about which scenario proves it here, or which scenario is requiring this code.
//// this seems to be the only thing that selectively disables edge marking
@ -2251,7 +2251,7 @@ Render3DError SoftRasterizerRenderer::RenderEdgeMarkingAndFog(const SoftRasteriz
if (param.enableFog)
{
FragmentColor fogColor;
fogColor.color = COLOR555TO6665( param.fogColor & 0x7FFF, (param.fogColor>>16) & 0x1F );
fogColor.color = LE_TO_LOCAL_32( COLOR555TO6665(param.fogColor & 0x7FFF, (param.fogColor>>16) & 0x1F) );
const size_t fogIndex = depth >> 9;
assert(fogIndex < 32768);
@ -2309,7 +2309,7 @@ Render3DError SoftRasterizerRenderer::ClearUsingImage(const u16 *__restrict colo
{
const size_t ir = readLine + ((x * xRatio) >> 16);
this->_framebufferColor[iw].color = COLOR555TO6665(colorBuffer[ir] & 0x7FFF, (colorBuffer[ir] >> 15) * 0x1F);
this->_framebufferColor[iw].color = LE_TO_LOCAL_32( COLOR555TO6665(colorBuffer[ir] & 0x7FFF, (colorBuffer[ir] >> 15) * 0x1F) );
this->_framebufferAttributes->depth[iw] = depthBuffer[ir];
this->_framebufferAttributes->isFogged[iw] = fogBuffer[ir];
this->_framebufferAttributes->opaquePolyID[iw] = opaquePolyID;

2
desmume/src/render3D.cpp
View File

@ -698,7 +698,7 @@ Render3DError Render3D::Render(const GFX3D &engine)
this->_isPoweredOn = true;
const u32 clearColorSwapped = LE_TO_LOCAL_32(engine.renderState.clearColor);
this->_clearColor6665.color = COLOR555TO6665(clearColorSwapped & 0x7FFF, (clearColorSwapped >> 16) & 0x1F);
this->_clearColor6665.color = LE_TO_LOCAL_32( COLOR555TO6665(clearColorSwapped & 0x7FFF, (clearColorSwapped >> 16) & 0x1F) );
this->_clearAttributes.opaquePolyID = (clearColorSwapped >> 24) & 0x3F;
//special value for uninitialized translucent polyid. without this, fires in spiderman2 dont display

50
desmume/src/texcache.cpp
View File

@ -44,7 +44,7 @@ using std::max;
#define DO_DEBUG_DUMP_TEXTURE
#endif
#define CONVERT(color) ((TEXCACHEFORMAT == TexFormat_32bpp)?(COLOR555TO8888_OPAQUE(color)):COLOR555TO6665_OPAQUE(color))
#define CONVERT(color) ( (TEXCACHEFORMAT == TexFormat_32bpp) ? ColorspaceConvert555To8888Opaque<false>(color) : ColorspaceConvert555To6665Opaque<false>(color) )
//This class represents a number of regions of memory which should be viewed as contiguous
class MemSpan
@ -888,16 +888,16 @@ void NDSTextureUnpackI2(const size_t srcSize, const u8 *__restrict srcData, cons
u8 idx;
idx = *srcData & 0x03;
*dstBuffer++ = (idx == 0) ? 0 : CONVERT(srcPal[idx] & 0x7FFF);
*dstBuffer++ = (idx == 0) ? 0 : LE_TO_LOCAL_32( CONVERT(srcPal[idx] & 0x7FFF) );
idx = (*srcData >> 2) & 0x03;
*dstBuffer++ = (idx == 0) ? 0 : CONVERT(srcPal[idx] & 0x7FFF);
*dstBuffer++ = (idx == 0) ? 0 : LE_TO_LOCAL_32( CONVERT(srcPal[idx] & 0x7FFF) );
idx = (*srcData >> 4) & 0x03;
*dstBuffer++ = (idx == 0) ? 0 : CONVERT(srcPal[idx] & 0x7FFF);
*dstBuffer++ = (idx == 0) ? 0 : LE_TO_LOCAL_32( CONVERT(srcPal[idx] & 0x7FFF) );
idx = (*srcData >> 6) & 0x03;
*dstBuffer++ = (idx == 0) ? 0 : CONVERT(srcPal[idx] & 0x7FFF);
*dstBuffer++ = (idx == 0) ? 0 : LE_TO_LOCAL_32( CONVERT(srcPal[idx] & 0x7FFF) );
}
#endif
}
@ -939,10 +939,10 @@ void NDSTextureUnpackI2(const size_t srcSize, const u8 *__restrict srcData, cons
#else
for (size_t i = 0; i < srcSize; i++, srcData++)
{
*dstBuffer++ = CONVERT(srcPal[ *srcData & 0x03] & 0x7FFF);
*dstBuffer++ = CONVERT(srcPal[(*srcData >> 2) & 0x03] & 0x7FFF);
*dstBuffer++ = CONVERT(srcPal[(*srcData >> 4) & 0x03] & 0x7FFF);
*dstBuffer++ = CONVERT(srcPal[(*srcData >> 6) & 0x03] & 0x7FFF);
*dstBuffer++ = LE_TO_LOCAL_32( CONVERT(srcPal[ *srcData & 0x03] & 0x7FFF) );
*dstBuffer++ = LE_TO_LOCAL_32( CONVERT(srcPal[(*srcData >> 2) & 0x03] & 0x7FFF) );
*dstBuffer++ = LE_TO_LOCAL_32( CONVERT(srcPal[(*srcData >> 4) & 0x03] & 0x7FFF) );
*dstBuffer++ = LE_TO_LOCAL_32( CONVERT(srcPal[(*srcData >> 6) & 0x03] & 0x7FFF) );
}
#endif
}
@ -1009,10 +1009,10 @@ void NDSTextureUnpackI4(const size_t srcSize, const u8 *__restrict srcData, cons
u8 idx;
idx = *srcData & 0x0F;
*dstBuffer++ = (idx == 0) ? 0 : CONVERT(srcPal[idx] & 0x7FFF);
*dstBuffer++ = (idx == 0) ? 0 : LE_TO_LOCAL_32( CONVERT(srcPal[idx] & 0x7FFF) );
idx = *srcData >> 4;
*dstBuffer++ = (idx == 0) ? 0 : CONVERT(srcPal[idx] & 0x7FFF);
*dstBuffer++ = (idx == 0) ? 0 : LE_TO_LOCAL_32( CONVERT(srcPal[idx] & 0x7FFF) );
}
#endif
}
@ -1059,8 +1059,8 @@ void NDSTextureUnpackI4(const size_t srcSize, const u8 *__restrict srcData, cons
#else
for (size_t i = 0; i < srcSize; i++, srcData++)
{
*dstBuffer++ = CONVERT(srcPal[*srcData & 0x0F] & 0x7FFF);
*dstBuffer++ = CONVERT(srcPal[*srcData >> 4] & 0x7FFF);
*dstBuffer++ = LE_TO_LOCAL_32( CONVERT(srcPal[*srcData & 0x0F] & 0x7FFF) );
*dstBuffer++ = LE_TO_LOCAL_32( CONVERT(srcPal[*srcData >> 4] & 0x7FFF) );
}
#endif
}
@ -1074,14 +1074,14 @@ void NDSTextureUnpackI8(const size_t srcSize, const u8 *__restrict srcData, cons
for (size_t i = 0; i < srcSize; i++, srcData++)
{
const u8 idx = *srcData;
*dstBuffer++ = (idx == 0) ? 0 : CONVERT(srcPal[idx] & 0x7FFF);
*dstBuffer++ = (idx == 0) ? 0 : LE_TO_LOCAL_32( CONVERT(srcPal[idx] & 0x7FFF) );
}
}
else
{
for (size_t i = 0; i < srcSize; i++, srcData++)
{
*dstBuffer++ = CONVERT(srcPal[*srcData] & 0x7FFF);
*dstBuffer++ = LE_TO_LOCAL_32( CONVERT(srcPal[*srcData] & 0x7FFF) );
}
}
}
@ -1093,7 +1093,7 @@ void NDSTextureUnpackA3I5(const size_t srcSize, const u8 *__restrict srcData, co
{
const u16 c = srcPal[*srcData & 0x1F] & 0x7FFF;
const u8 alpha = *srcData >> 5;
*dstBuffer++ = (TEXCACHEFORMAT == TexFormat_15bpp) ? COLOR555TO6665(c, material_3bit_to_5bit[alpha]) : COLOR555TO8888(c, material_3bit_to_8bit[alpha]);
*dstBuffer++ = LE_TO_LOCAL_32( (TEXCACHEFORMAT == TexFormat_15bpp) ? COLOR555TO6665(c, material_3bit_to_5bit[alpha]) : COLOR555TO8888(c, material_3bit_to_8bit[alpha]) );
}
}
@ -1145,7 +1145,7 @@ void NDSTextureUnpackA5I3(const size_t srcSize, const u8 *__restrict srcData, co
{
const u16 c = srcPal[*srcData & 0x07] & 0x7FFF;
const u8 alpha = (*srcData >> 3);
*dstBuffer++ = (TEXCACHEFORMAT == TexFormat_15bpp) ? COLOR555TO6665(c, alpha) : COLOR555TO8888(c, material_5bit_to_8bit[alpha]);
*dstBuffer++ = LE_TO_LOCAL_32( (TEXCACHEFORMAT == TexFormat_15bpp) ? COLOR555TO6665(c, alpha) : COLOR555TO8888(c, material_5bit_to_8bit[alpha]) );
}
#endif
}
@ -1193,13 +1193,13 @@ void NDSTextureUnpack4x4(const size_t srcSize, const u32 *__restrict srcData, co
const u8 mode = pal1>>14;
CACHE_ALIGN u32 tmp_col[4];
tmp_col[0] = COLOR555TO8888_OPAQUE( PAL4X4(pal1offset) );
tmp_col[1] = COLOR555TO8888_OPAQUE( PAL4X4(pal1offset+1) );
tmp_col[0] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(PAL4X4(pal1offset+0)) );
tmp_col[1] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(PAL4X4(pal1offset+1)) );
switch (mode)
{
case 0:
tmp_col[2] = COLOR555TO8888_OPAQUE( PAL4X4(pal1offset+2) );
tmp_col[2] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(PAL4X4(pal1offset+2)) );
tmp_col[3] = 0x00000000;
break;
@ -1219,8 +1219,8 @@ void NDSTextureUnpack4x4(const size_t srcSize, const u32 *__restrict srcData, co
break;
case 2:
tmp_col[2] = COLOR555TO8888_OPAQUE( PAL4X4(pal1offset+2) );
tmp_col[3] = COLOR555TO8888_OPAQUE( PAL4X4(pal1offset+3) );
tmp_col[2] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(PAL4X4(pal1offset+2)) );
tmp_col[3] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(PAL4X4(pal1offset+3)) );
break;
case 3:
@ -1248,8 +1248,8 @@ void NDSTextureUnpack4x4(const size_t srcSize, const u32 *__restrict srcData, co
( ((g0*3 + g1*5)>>6) << 5 ) |
( ((b0*3 + b1*5)>>6) << 10 );
tmp_col[2] = COLOR555TO8888_OPAQUE(tmp1);
tmp_col[3] = COLOR555TO8888_OPAQUE(tmp2);
tmp_col[2] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(tmp1) );
tmp_col[3] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(tmp2) );
break;
}
@ -1319,7 +1319,7 @@ void NDSTextureUnpackDirect16Bit(const size_t srcSize, const u16 *__restrict src
for (; i < pixCount; i++, srcData++)
{
const u16 c = LOCAL_TO_LE_16(*srcData);
*dstBuffer++ = (c & 0x8000) ? CONVERT(c & 0x7FFF) : 0;
*dstBuffer++ = (c & 0x8000) ? LE_TO_LOCAL_32( CONVERT(c & 0x7FFF) ) : 0;
}
}

19
desmume/src/types.h
View File

@ -414,6 +414,25 @@ inline bool atomic_test_and_clear_barrier32(volatile s32 *V, s32 M) { return (s
#endif
// Flags used to determine how a conversion function swaps bytes for big-endian systems.
// These flags should be ignored on little-endian systems.
enum BESwapFlags
{
BESwapNone = 0x00, // No byte swapping for both incoming and outgoing data. All data is used as-is.
BESwapIn = 0x01, // All incoming data is byte swapped; outgoing data is used as-is.
BESwapPre = 0x01, // An alternate name for "BESwapIn"
BESwapSrc = 0x01, // An alternate name for "BESwapIn"
BESwapOut = 0x02, // All incoming data is used as-is; outgoing data is byte swapped.
BESwapPost = 0x02, // An alternate name for "BESwapOut"
BESwapDst = 0x02, // An alternate name for "BESwapOut"
BESwapInOut = 0x03, // Both incoming data and outgoing data are byte swapped.
BESwapPrePost = 0x03, // An alternate name for "BESwapInOut"
BESwapSrcDst = 0x03 // An alternate name for "BESwapInOut"
};
/* little endian (ds' endianess) to local endianess convert macros */
#ifdef MSB_FIRST /* local arch is big endian */
# define LE_TO_LOCAL_16(x) ((((x)&0xff)<<8)|(((x)>>8)&0xff))

226
desmume/src/utils/colorspacehandler/colorspacehandler.cpp
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2019 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -143,13 +143,13 @@ void ColorspaceHandlerInit()
for (size_t i = 0; i < 32768; i++)
{
color_555_to_666[i] = LE_TO_LOCAL_32( RGB15TO18_BITLOGIC(i) );
color_555_to_6665_opaque[i] = LE_TO_LOCAL_32( RGB15TO18_BITLOGIC(i) | 0x1F000000 );
color_555_to_6665_opaque_swap_rb[i] = LE_TO_LOCAL_32( RGB15TO18_SWAP_RB_BITLOGIC(i) | 0x1F000000 );
color_555_to_666[i] = RGB15TO18_BITLOGIC(i);
color_555_to_6665_opaque[i] = RGB15TO18_BITLOGIC(i) | 0x1F000000;
color_555_to_6665_opaque_swap_rb[i] = RGB15TO18_SWAP_RB_BITLOGIC(i) | 0x1F000000;
color_555_to_888[i] = LE_TO_LOCAL_32( RGB15TO24_BITLOGIC(i) );
color_555_to_8888_opaque[i] = LE_TO_LOCAL_32( RGB15TO24_BITLOGIC(i) | 0xFF000000 );
color_555_to_8888_opaque_swap_rb[i] = LE_TO_LOCAL_32( RGB15TO24_SWAP_RB_BITLOGIC(i) | 0xFF000000 );
color_555_to_888[i] = RGB15TO24_BITLOGIC(i);
color_555_to_8888_opaque[i] = RGB15TO24_BITLOGIC(i) | 0xFF000000;
color_555_to_8888_opaque_swap_rb[i] = RGB15TO24_SWAP_RB_BITLOGIC(i) | 0xFF000000;
}
#define RGB16_SWAP_RB_BITLOGIC(col) ( (((col)&0x001F)<<10) | ((col)&0x03E0) | (((col)&0x7C00)>>10) | ((col)&0x8000) )
@ -161,7 +161,7 @@ void ColorspaceHandlerInit()
}
}
template <bool SWAP_RB, bool IS_UNALIGNED>
template <bool SWAP_RB, bool IS_UNALIGNED, BESwapFlags BE_BYTESWAP>
void ColorspaceConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount)
{
size_t i = 0;
@ -173,36 +173,51 @@ void ColorspaceConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__re
{
if (IS_UNALIGNED)
{
i = csh.ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(src, dst, pixCountVector);
i = csh.ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned<BE_BYTESWAP>(src, dst, pixCountVector);
}
else
{
i = csh.ConvertBuffer555To8888Opaque_SwapRB(src, dst, pixCountVector);
i = csh.ConvertBuffer555To8888Opaque_SwapRB<BE_BYTESWAP>(src, dst, pixCountVector);
}
}
else
{
if (IS_UNALIGNED)
{
i = csh.ConvertBuffer555To8888Opaque_IsUnaligned(src, dst, pixCountVector);
i = csh.ConvertBuffer555To8888Opaque_IsUnaligned<BE_BYTESWAP>(src, dst, pixCountVector);
}
else
{
i = csh.ConvertBuffer555To8888Opaque(src, dst, pixCountVector);
i = csh.ConvertBuffer555To8888Opaque<BE_BYTESWAP>(src, dst, pixCountVector);
}
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert555To8888Opaque<SWAP_RB>(src[i]);
switch (BE_BYTESWAP)
{
case BESwapNone:
dst[i] = ColorspaceConvert555To8888Opaque<SWAP_RB>(src[i]);
break;
case BESwapIn:
dst[i] = ColorspaceConvert555To8888Opaque<SWAP_RB>(LE_TO_LOCAL_16(src[i]));
break;
case BESwapOut:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<SWAP_RB>(src[i]) );
break;
case BESwapInOut:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<SWAP_RB>(LE_TO_LOCAL_16(src[i])) );
break;
}
}
}
template <bool SWAP_RB, bool IS_UNALIGNED>
template <bool SWAP_RB, bool IS_UNALIGNED, BESwapFlags BE_BYTESWAP>
void ColorspaceConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount)
{
size_t i = 0;
@ -214,32 +229,47 @@ void ColorspaceConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__re
{
if (IS_UNALIGNED)
{
i = csh.ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(src, dst, pixCountVector);
i = csh.ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned<BE_BYTESWAP>(src, dst, pixCountVector);
}
else
{
i = csh.ConvertBuffer555To6665Opaque_SwapRB(src, dst, pixCountVector);
i = csh.ConvertBuffer555To6665Opaque_SwapRB<BE_BYTESWAP>(src, dst, pixCountVector);
}
}
else
{
if (IS_UNALIGNED)
{
i = csh.ConvertBuffer555To6665Opaque_IsUnaligned(src, dst, pixCountVector);
i = csh.ConvertBuffer555To6665Opaque_IsUnaligned<BE_BYTESWAP>(src, dst, pixCountVector);
}
else
{
i = csh.ConvertBuffer555To6665Opaque(src, dst, pixCountVector);
i = csh.ConvertBuffer555To6665Opaque<BE_BYTESWAP>(src, dst, pixCountVector);
}
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert555To6665Opaque<SWAP_RB>(src[i]);
switch (BE_BYTESWAP)
{
case BESwapNone:
dst[i] = ColorspaceConvert555To6665Opaque<SWAP_RB>(src[i]);
break;
case BESwapIn:
dst[i] = ColorspaceConvert555To6665Opaque<SWAP_RB>(LE_TO_LOCAL_16(src[i]));
break;
case BESwapOut:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<SWAP_RB>(src[i]) );
break;
case BESwapInOut:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<SWAP_RB>(LE_TO_LOCAL_16(src[i])) );
break;
}
}
}
@ -275,9 +305,7 @@ void ColorspaceConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert8888To6665<SWAP_RB>(src[i]);
@ -316,9 +344,7 @@ void ColorspaceConvertBuffer6665To8888(const u32 *src, u32 *dst, size_t pixCount
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert6665To8888<SWAP_RB>(src[i]);
@ -357,9 +383,7 @@ void ColorspaceConvertBuffer8888To5551(const u32 *__restrict src, u16 *__restric
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert8888To5551<SWAP_RB>(src[i]);
@ -398,9 +422,7 @@ void ColorspaceConvertBuffer6665To5551(const u32 *__restrict src, u16 *__restric
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert6665To5551<SWAP_RB>(src[i]);
@ -439,9 +461,7 @@ void ColorspaceConvertBuffer888XTo8888Opaque(const u32 *src, u32 *dst, size_t pi
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert888XTo8888Opaque<SWAP_RB>(src[i]);
@ -480,9 +500,7 @@ void ColorspaceConvertBuffer555XTo888(const u16 *__restrict src, u8 *__restrict
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
ColorspaceConvert555XTo888<SWAP_RB>(src[i], &dst[i*3]);
@ -521,9 +539,7 @@ void ColorspaceConvertBuffer888XTo888(const u32 *__restrict src, u8 *__restrict
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
ColorspaceConvert888XTo888<SWAP_RB>(src[i], &dst[i*3]);
@ -554,9 +570,7 @@ void ColorspaceCopyBuffer16(const u16 *src, u16 *dst, size_t pixCount)
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceCopy16<SWAP_RB>(src[i]);
@ -587,9 +601,7 @@ void ColorspaceCopyBuffer32(const u32 *src, u32 *dst, size_t pixCount)
}
#pragma LOOPVECTORIZE_DISABLE
#endif // USEMANUALVECTORIZATION
for (; i < pixCount; i++)
{
dst[i] = ColorspaceCopy32<SWAP_RB>(src[i]);
@ -773,72 +785,148 @@ void ColorspaceApplyIntensityToBuffer32(u32 *dst, size_t pixCount, float intensi
}
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
size_t i = 0;
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert555To8888Opaque<false>(src[i]);
switch (BE_BYTESWAP)
{
case BESwapNone:
dst[i] = ColorspaceConvert555To8888Opaque<false>(src[i]);
break;
case BESwapSrc:
dst[i] = ColorspaceConvert555To8888Opaque<false>(LE_TO_LOCAL_16(src[i]));
break;
case BESwapDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(src[i]) );
break;
case BESwapSrcDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<false>(LE_TO_LOCAL_16(src[i])) );
break;
}
}
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
size_t i = 0;
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert555To8888Opaque<true>(src[i]);
switch (BE_BYTESWAP)
{
case BESwapNone:
dst[i] = ColorspaceConvert555To8888Opaque<true>(src[i]);
break;
case BESwapSrc:
dst[i] = ColorspaceConvert555To8888Opaque<true>(LE_TO_LOCAL_16(src[i]));
break;
case BESwapDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<true>(src[i]) );
break;
case BESwapSrcDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To8888Opaque<true>(LE_TO_LOCAL_16(src[i])) );
break;
}
}
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return this->ColorspaceHandler::ConvertBuffer555To8888Opaque(src, dst, pixCount);
return this->ColorspaceHandler::ConvertBuffer555To8888Opaque<BE_BYTESWAP>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return this->ColorspaceHandler::ConvertBuffer555To8888Opaque_SwapRB(src, dst, pixCount);
return this->ColorspaceHandler::ConvertBuffer555To8888Opaque_SwapRB<BE_BYTESWAP>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
size_t i = 0;
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert555To6665Opaque<false>(src[i]);
switch (BE_BYTESWAP)
{
case BESwapNone:
dst[i] = ColorspaceConvert555To6665Opaque<false>(src[i]);
break;
case BESwapSrc:
dst[i] = ColorspaceConvert555To6665Opaque<false>(LE_TO_LOCAL_16(src[i]));
break;
case BESwapDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>(src[i]) );
break;
case BESwapSrcDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<false>(LE_TO_LOCAL_16(src[i])) );
break;
}
}
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
size_t i = 0;
for (; i < pixCount; i++)
{
dst[i] = ColorspaceConvert555To6665Opaque<true>(src[i]);
switch (BE_BYTESWAP)
{
case BESwapNone:
dst[i] = ColorspaceConvert555To6665Opaque<true>(src[i]);
break;
case BESwapSrc:
dst[i] = ColorspaceConvert555To6665Opaque<true>(LE_TO_LOCAL_16(src[i]));
break;
case BESwapDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<true>(src[i]) );
break;
case BESwapSrcDst:
dst[i] = LE_TO_LOCAL_32( ColorspaceConvert555To6665Opaque<true>(LE_TO_LOCAL_16(src[i])) );
break;
}
}
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return this->ColorspaceHandler::ConvertBuffer555To6665Opaque(src, dst, pixCount);
return this->ColorspaceHandler::ConvertBuffer555To6665Opaque<BE_BYTESWAP>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler::ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return this->ColorspaceHandler::ConvertBuffer555To6665Opaque_SwapRB(src, dst, pixCount);
return this->ColorspaceHandler::ConvertBuffer555To6665Opaque_SwapRB<BE_BYTESWAP>(src, dst, pixCount);
}
size_t ColorspaceHandler::ConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount) const
@ -1283,15 +1371,39 @@ size_t ColorspaceHandler::ApplyIntensityToBuffer32_SwapRB_IsUnaligned(u32 *dst,
return this->ApplyIntensityToBuffer32_SwapRB(dst, pixCount, intensity);
}
template void ColorspaceConvertBuffer555To8888Opaque<true, true>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, false>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, true>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, false>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, true, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, false, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, true, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, true, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, false, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, true, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, true, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, false, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, true, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, true, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<true, false, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, true, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To8888Opaque<false, false, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, true>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, false>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, true>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, false>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, true, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, false, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, true, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapNone>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, true, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, false, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, true, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapIn>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, true, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, false, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, true, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, true, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<true, false, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, true, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer555To6665Opaque<false, false, BESwapInOut>(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template void ColorspaceConvertBuffer8888To6665<true, true>(const u32 *src, u32 *dst, size_t pixCount);
template void ColorspaceConvertBuffer8888To6665<true, false>(const u32 *src, u32 *dst, size_t pixCount);

36
desmume/src/utils/colorspacehandler/colorspacehandler.h
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2017 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -108,22 +108,12 @@ extern CACHE_ALIGN u32 color_555_to_888[32768];
#define COLOR555TO6665_OPAQUE(col) (color_555_to_6665_opaque[(col)]) // Convert a 15-bit color to an opaque sparsely packed 32-bit color containing an RGBA6665 color
#define COLOR555TO6665_OPAQUE_SWAP_RB(col) (color_555_to_6665_opaque_swap_rb[(col)]) // Convert a 15-bit color to an opaque sparsely packed 32-bit color containing an RGBA6665 color with R and B components swapped
#define COLOR555TO666(col) (color_555_to_666[(col)]) // Convert a 15-bit color to a fully transparent sparsely packed 32-bit color containing an RGBA6665 color
#ifdef MSB_FIRST
#define COLOR555TO6665(col,alpha5) ((alpha5) | color_555_to_666[(col)]) // Convert a 15-bit color to a sparsely packed 32-bit color containing an RGBA6665 color with user-defined alpha, big-endian
#else
#define COLOR555TO6665(col,alpha5) (((alpha5)<<24) | color_555_to_666[(col)]) // Convert a 15-bit color to a sparsely packed 32-bit color containing an RGBA6665 color with user-defined alpha, little-endian
#endif
#define COLOR555TO6665(col,alpha5) (((alpha5)<<24) | color_555_to_666[(col)]) // Convert a 15-bit color to a sparsely packed 32-bit color containing an RGBA6665 color with user-defined alpha
#define COLOR555TO8888_OPAQUE(col) (color_555_to_8888_opaque[(col)]) // Convert a 15-bit color to an opaque 32-bit color
#define COLOR555TO8888_OPAQUE_SWAP_RB(col) (color_555_to_8888_opaque_swap_rb[(col)]) // Convert a 15-bit color to an opaque 32-bit color with R and B components swapped
#define COLOR555TO888(col) (color_555_to_888[(col)]) // Convert a 15-bit color to an opaque 24-bit color or a fully transparent 32-bit color
#ifdef MSB_FIRST
#define COLOR555TO8888(col,alpha8) ((alpha8) | color_555_to_888[(col)]) // Convert a 15-bit color to a 32-bit color with user-defined alpha, big-endian
#else
#define COLOR555TO8888(col,alpha8) (((alpha8)<<24) | color_555_to_888[(col)]) // Convert a 15-bit color to a 32-bit color with user-defined alpha, little-endian
#endif
#define COLOR555TO8888(col,alpha8) (((alpha8)<<24) | color_555_to_888[(col)]) // Convert a 15-bit color to a 32-bit color with user-defined alpha
//produce a 15bpp color from individual 5bit components
#define R5G5B5TORGB15(r,g,b) ( (r) | ((g)<<5) | ((b)<<10) )
@ -350,8 +340,8 @@ FORCEINLINE u32 ColorspaceApplyIntensity32(u32 srcColor, float intensity)
return ColorspaceApplyIntensity32<SWAP_RB>(srcColorComponent);
}
template<bool SWAP_RB, bool IS_UNALIGNED> void ColorspaceConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template<bool SWAP_RB, bool IS_UNALIGNED> void ColorspaceConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template<bool SWAP_RB, bool IS_UNALIGNED, BESwapFlags BE_BYTESWAP> void ColorspaceConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template<bool SWAP_RB, bool IS_UNALIGNED, BESwapFlags BE_BYTESWAP> void ColorspaceConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount);
template<bool SWAP_RB, bool IS_UNALIGNED> void ColorspaceConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount);
template<bool SWAP_RB, bool IS_UNALIGNED> void ColorspaceConvertBuffer6665To8888(const u32 *src, u32 *dst, size_t pixCount);
template<bool SWAP_RB, bool IS_UNALIGNED> void ColorspaceConvertBuffer8888To5551(const u32 *__restrict src, u16 *__restrict dst, size_t pixCount);
@ -372,15 +362,15 @@ class ColorspaceHandler
public:
ColorspaceHandler() {};
size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665_SwapRB(const u32 *src, u32 *dst, size_t pixCount) const;

10
desmume/src/utils/colorspacehandler/colorspacehandler_AVX2.cpp
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2019 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -904,41 +904,49 @@ size_t ColorspaceApplyIntensityToBuffer32_AVX2(u32 *dst, size_t pixCountVec256,
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX2<false, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX2<true, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX2<false, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX2<true, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX2<false, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX2<true, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX2<false, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX2::ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX2<true, true>(src, dst, pixCount);

18
desmume/src/utils/colorspacehandler/colorspacehandler_AVX2.h
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2017 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -47,15 +47,15 @@ class ColorspaceHandler_AVX2 : public ColorspaceHandler
public:
ColorspaceHandler_AVX2() {};
size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665_SwapRB(const u32 *src, u32 *dst, size_t pixCount) const;

10
desmume/src/utils/colorspacehandler/colorspacehandler_AVX512.cpp
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2019 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -857,41 +857,49 @@ size_t ColorspaceApplyIntensityToBuffer32_AVX512(u32 *dst, size_t pixCountVec512
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX512<false, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX512<true, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX512<false, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AVX512<true, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX512<false, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX512<true, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX512<false, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AVX512::ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AVX512<true, true>(src, dst, pixCount);

18
desmume/src/utils/colorspacehandler/colorspacehandler_AVX512.h
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2019 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -47,15 +47,15 @@ class ColorspaceHandler_AVX512 : public ColorspaceHandler
public:
ColorspaceHandler_AVX512() {};
size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665_SwapRB(const u32 *src, u32 *dst, size_t pixCount) const;

176
desmume/src/utils/colorspacehandler/colorspacehandler_AltiVec.cpp
View File

@ -23,62 +23,114 @@
#include <string.h>
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
FORCEINLINE void ColorspaceConvert555To8888_AltiVec(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi)
{
// Conversion algorithm:
// RGB 5-bit to 8-bit formula: dstRGB8 = (srcRGB5 << 3) | ((srcRGB5 >> 2) & 0x07)
dstLo = vec_unpackl((vector pixel)srcColor);
dstLo = vec_or( vec_sl((v128u8)dstLo, ((v128u8){0,3,3,3, 0,3,3,3, 0,3,3,3, 0,3,3,3})), vec_sr((v128u8)dstLo, ((v128u8){0,2,2,2, 0,2,2,2, 0,2,2,2, 0,2,2,2})) );
dstLo = vec_perm(dstLo, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x11, 0x05,0x06,0x07,0x13, 0x09,0x0A,0x0B,0x15, 0x0D,0x0E,0x0F,0x17}) : ((v128u8){0x03,0x02,0x01,0x11, 0x07,0x06,0x05,0x13, 0x0B,0x0A,0x09,0x15, 0x0F,0x0E,0x0D,0x17}));
v128u16 srcSwapped;
if ( (BE_BYTESWAP == BESwapSrc) || (BE_BYTESWAP == BESwapSrcDst) )
{
srcSwapped = vec_perm(srcColor, srcColor, ((v128u8){1,0, 3,2, 5,4, 7,6, 9,8, 11,10, 13,12, 15,14}));
}
else
{
srcSwapped = srcColor;
}
dstHi = vec_unpackh((vector pixel)srcColor);
dstLo = vec_unpackl((vector pixel)srcSwapped);
dstLo = vec_or( vec_sl((v128u8)dstLo, ((v128u8){0,3,3,3, 0,3,3,3, 0,3,3,3, 0,3,3,3})), vec_sr((v128u8)dstLo, ((v128u8){0,2,2,2, 0,2,2,2, 0,2,2,2, 0,2,2,2})) );
if ( (BE_BYTESWAP == BESwapDst) || (BE_BYTESWAP == BESwapSrcDst) )
{
dstLo = vec_perm(dstLo, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x11, 0x05,0x06,0x07,0x13, 0x09,0x0A,0x0B,0x15, 0x0D,0x0E,0x0F,0x17}) : ((v128u8){0x03,0x02,0x01,0x11, 0x07,0x06,0x05,0x13, 0x0B,0x0A,0x09,0x15, 0x0F,0x0E,0x0D,0x17}));
}
else
{
dstLo = vec_perm(dstLo, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x11,0x03,0x02,0x01, 0x13,0x07,0x06,0x05, 0x15,0x0B,0x0A,0x09, 0x17,0x0F,0x0E,0x0D}) : ((v128u8){0x11,0x01,0x02,0x03, 0x13,0x05,0x06,0x07, 0x15,0x09,0x0A,0x0B, 0x17,0x0D,0x0E,0x0F}));
}
dstHi = vec_unpackh((vector pixel)srcSwapped);
dstHi = vec_or( vec_sl((v128u8)dstHi, ((v128u8){0,3,3,3, 0,3,3,3, 0,3,3,3, 0,3,3,3})), vec_sr((v128u8)dstHi, ((v128u8){0,2,2,2, 0,2,2,2, 0,2,2,2, 0,2,2,2})) );
dstHi = vec_perm(dstHi, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x19, 0x05,0x06,0x07,0x1B, 0x09,0x0A,0x0B,0x1D, 0x0D,0x0E,0x0F,0x1F}) : ((v128u8){0x03,0x02,0x01,0x19, 0x07,0x06,0x05,0x1B, 0x0B,0x0A,0x09,0x1D, 0x0F,0x0E,0x0D,0x1F}));
if ( (BE_BYTESWAP == BESwapDst) || (BE_BYTESWAP == BESwapSrcDst) )
{
dstHi = vec_perm(dstHi, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x19, 0x05,0x06,0x07,0x1B, 0x09,0x0A,0x0B,0x1D, 0x0D,0x0E,0x0F,0x1F}) : ((v128u8){0x03,0x02,0x01,0x19, 0x07,0x06,0x05,0x1B, 0x0B,0x0A,0x09,0x1D, 0x0F,0x0E,0x0D,0x1F}));
}
else
{
dstHi = vec_perm(dstHi, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x19,0x03,0x02,0x01, 0x1B,0x07,0x06,0x05, 0x1D,0x0B,0x0A,0x09, 0x1F,0x0F,0x0E,0x0D}) : ((v128u8){0x19,0x01,0x02,0x03, 0x1B,0x05,0x06,0x07, 0x1D,0x09,0x0A,0x0B, 0x1F,0x0D,0x0E,0x0F}));
}
}
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
FORCEINLINE void ColorspaceConvert555XTo888X_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi)
{
const v128u16 srcAlphaBits16 = {0, 0, 0, 0, 0, 0, 0, 0};
ColorspaceConvert555To8888_AltiVec<SWAP_RB>(srcColor, srcAlphaBits16, dstLo, dstHi);
ColorspaceConvert555To8888_AltiVec<SWAP_RB, BE_BYTESWAP>(srcColor, srcAlphaBits16, dstLo, dstHi);
}
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
FORCEINLINE void ColorspaceConvert555To6665_AltiVec(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi)
{
// Conversion algorithm:
// RGB 5-bit to 6-bit formula: dstRGB6 = (srcRGB5 << 1) | ((srcRGB5 >> 4) & 0x01)
dstLo = vec_unpackl((vector pixel)srcColor);
dstLo = vec_or( vec_sl((v128u8)dstLo, ((v128u8){0,1,1,1, 0,1,1,1, 0,1,1,1, 0,1,1,1})), vec_sr((v128u8)dstLo, ((v128u8){0,4,4,4, 0,4,4,4, 0,4,4,4, 0,4,4,4})) );
dstLo = vec_perm(dstLo, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x11, 0x05,0x06,0x07,0x13, 0x09,0x0A,0x0B,0x15, 0x0D,0x0E,0x0F,0x17}) : ((v128u8){0x03,0x02,0x01,0x11, 0x07,0x06,0x05,0x13, 0x0B,0x0A,0x09,0x15, 0x0F,0x0E,0x0D,0x17}));
v128u16 srcSwapped;
if ( (BE_BYTESWAP == BESwapSrc) || (BE_BYTESWAP == BESwapSrcDst) )
{
srcSwapped = vec_perm(srcColor, srcColor, ((v128u8){1,0, 3,2, 5,4, 7,6, 9,8, 11,10, 13,12, 15,14}));
}
else
{
srcSwapped = srcColor;
}
dstHi = vec_unpackh((vector pixel)srcColor);
dstLo = vec_unpackl((vector pixel)srcSwapped);
dstLo = vec_or( vec_sl((v128u8)dstLo, ((v128u8){0,1,1,1, 0,1,1,1, 0,1,1,1, 0,1,1,1})), vec_sr((v128u8)dstLo, ((v128u8){0,4,4,4, 0,4,4,4, 0,4,4,4, 0,4,4,4})) );
if ( (BE_BYTESWAP == BESwapDst) || (BE_BYTESWAP == BESwapSrcDst) )
{
dstLo = vec_perm(dstLo, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x11, 0x05,0x06,0x07,0x13, 0x09,0x0A,0x0B,0x15, 0x0D,0x0E,0x0F,0x17}) : ((v128u8){0x03,0x02,0x01,0x11, 0x07,0x06,0x05,0x13, 0x0B,0x0A,0x09,0x15, 0x0F,0x0E,0x0D,0x17}));
}
else
{
dstLo = vec_perm(dstLo, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x11,0x03,0x02,0x01, 0x13,0x07,0x06,0x05, 0x15,0x0B,0x0A,0x09, 0x17,0x0F,0x0E,0x0D}) : ((v128u8){0x11,0x01,0x02,0x03, 0x13,0x05,0x06,0x07, 0x15,0x09,0x0A,0x0B, 0x17,0x0D,0x0E,0x0F}));
}
dstHi = vec_unpackh((vector pixel)srcSwapped);
dstHi = vec_or( vec_sl((v128u8)dstHi, ((v128u8){0,1,1,1, 0,1,1,1, 0,1,1,1, 0,1,1,1})), vec_sr((v128u8)dstHi, ((v128u8){0,4,4,4, 0,4,4,4, 0,4,4,4, 0,4,4,4})) );
dstHi = vec_perm(dstHi, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x19, 0x05,0x06,0x07,0x1B, 0x09,0x0A,0x0B,0x1D, 0x0D,0x0E,0x0F,0x1F}) : ((v128u8){0x03,0x02,0x01,0x19, 0x07,0x06,0x05,0x1B, 0x0B,0x0A,0x09,0x1D, 0x0F,0x0E,0x0D,0x1F}));
if ( (BE_BYTESWAP == BESwapDst) || (BE_BYTESWAP == BESwapSrcDst) )
{
dstHi = vec_perm(dstHi, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x01,0x02,0x03,0x19, 0x05,0x06,0x07,0x1B, 0x09,0x0A,0x0B,0x1D, 0x0D,0x0E,0x0F,0x1F}) : ((v128u8){0x03,0x02,0x01,0x19, 0x07,0x06,0x05,0x1B, 0x0B,0x0A,0x09,0x1D, 0x0F,0x0E,0x0D,0x1F}));
}
else
{
dstHi = vec_perm(dstHi, srcAlphaBits, (SWAP_RB) ? ((v128u8){0x19,0x03,0x02,0x01, 0x1B,0x07,0x06,0x05, 0x1D,0x0B,0x0A,0x09, 0x1F,0x0F,0x0E,0x0D}) : ((v128u8){0x19,0x01,0x02,0x03, 0x1B,0x05,0x06,0x07, 0x1D,0x09,0x0A,0x0B, 0x1F,0x0D,0x0E,0x0F}));
}
}
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
FORCEINLINE void ColorspaceConvert555XTo666X_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi)
{
const v128u16 srcAlphaBits16 = {0, 0, 0, 0, 0, 0, 0, 0};
ColorspaceConvert555To6665_AltiVec<SWAP_RB>(srcColor, srcAlphaBits16, dstLo, dstHi);
ColorspaceConvert555To6665_AltiVec<SWAP_RB, BE_BYTESWAP>(srcColor, srcAlphaBits16, dstLo, dstHi);
}
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
FORCEINLINE void ColorspaceConvert555To8888Opaque_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi)
{
const v128u16 srcAlphaBits16 = {0xFF00, 0xFF00, 0xFF00, 0xFF00, 0xFF00, 0xFF00, 0xFF00, 0xFF00};
ColorspaceConvert555To8888_AltiVec<SWAP_RB>(srcColor, srcAlphaBits16, dstLo, dstHi);
const v128u16 srcAlphaBits16 = {0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF, 0xFFFF};
ColorspaceConvert555To8888_AltiVec<SWAP_RB, BE_BYTESWAP>(srcColor, srcAlphaBits16, dstLo, dstHi);
}
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
FORCEINLINE void ColorspaceConvert555To6665Opaque_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi)
{
const v128u16 srcAlphaBits16 = {0x1F00, 0x1F00, 0x1F00, 0x1F00, 0x1F00, 0x1F00, 0x1F00, 0x1F00};
ColorspaceConvert555To6665_AltiVec<SWAP_RB>(srcColor, srcAlphaBits16, dstLo, dstHi);
const v128u16 srcAlphaBits16 = {0x1F1F, 0x1F1F, 0x1F1F, 0x1F1F, 0x1F1F, 0x1F1F, 0x1F1F, 0x1F1F};
ColorspaceConvert555To6665_AltiVec<SWAP_RB, BE_BYTESWAP>(srcColor, srcAlphaBits16, dstLo, dstHi);
}
template <bool SWAP_RB>
@ -210,7 +262,7 @@ FORCEINLINE v128u32 ColorspaceCopy32_AltiVec(const v128u32 &src)
return src;
}
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
static size_t ColorspaceConvertBuffer555To8888Opaque_AltiVec(const u16 *__restrict src, u32 *__restrict dst, const size_t pixCountVec128)
{
size_t i = 0;
@ -219,7 +271,7 @@ static size_t ColorspaceConvertBuffer555To8888Opaque_AltiVec(const u16 *__restri
{
v128u32 dstConvertedLo, dstConvertedHi;
ColorspaceConvert555To8888Opaque_AltiVec<SWAP_RB>( vec_ld(0, src+i), dstConvertedLo, dstConvertedHi );
ColorspaceConvert555To8888Opaque_AltiVec<SWAP_RB, BE_BYTESWAP>( vec_ld(0, src+i), dstConvertedLo, dstConvertedHi );
vec_st(dstConvertedHi, 0, dst+i);
vec_st(dstConvertedLo, 16, dst+i);
}
@ -227,7 +279,7 @@ static size_t ColorspaceConvertBuffer555To8888Opaque_AltiVec(const u16 *__restri
return i;
}
template <bool SWAP_RB>
template <bool SWAP_RB, BESwapFlags BE_BYTESWAP>
size_t ColorspaceConvertBuffer555To6665Opaque_AltiVec(const u16 *__restrict src, u32 *__restrict dst, size_t pixCountVec128)
{
size_t i = 0;
@ -236,7 +288,7 @@ size_t ColorspaceConvertBuffer555To6665Opaque_AltiVec(const u16 *__restrict src,
{
v128u32 dstConvertedLo, dstConvertedHi;
ColorspaceConvert555To6665Opaque_AltiVec<SWAP_RB>( vec_ld(0, src+i), dstConvertedLo, dstConvertedHi );
ColorspaceConvert555To6665Opaque_AltiVec<SWAP_RB, BE_BYTESWAP>( vec_ld(0, src+i), dstConvertedLo, dstConvertedHi );
vec_st(dstConvertedHi, 0, dst+i);
vec_st(dstConvertedLo, 16, dst+i);
}
@ -424,24 +476,28 @@ size_t ColorspaceCopyBuffer32_AltiVec(const u32 *src, u32 *dst, size_t pixCountV
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AltiVec::ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AltiVec<false>(src, dst, pixCount);
return ColorspaceConvertBuffer555To8888Opaque_AltiVec<false, BE_BYTESWAP>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AltiVec::ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_AltiVec<true>(src, dst, pixCount);
return ColorspaceConvertBuffer555To8888Opaque_AltiVec<true, BE_BYTESWAP>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AltiVec::ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AltiVec<false>(src, dst, pixCount);
return ColorspaceConvertBuffer555To6665Opaque_AltiVec<false, BE_BYTESWAP>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_AltiVec::ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_AltiVec<true>(src, dst, pixCount);
return ColorspaceConvertBuffer555To6665Opaque_AltiVec<true, BE_BYTESWAP>(src, dst, pixCount);
}
size_t ColorspaceHandler_AltiVec::ConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount) const
@ -524,23 +580,59 @@ size_t ColorspaceHandler_AltiVec::CopyBuffer32_SwapRB(const u32 *src, u32 *dst,
return ColorspaceCopyBuffer32_AltiVec<true>(src, dst, pixCount);
}
template void ColorspaceConvert555To8888_AltiVec<true>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<false>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<true, BESwapNone>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<false, BESwapNone>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<true, BESwapSrc>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<false, BESwapSrc>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<true, BESwapDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<false, BESwapDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<true, BESwapSrcDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888_AltiVec<false, BESwapSrcDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<true>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<false>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<true, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<false, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<true, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<false, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<true, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<false, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<true, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo888X_AltiVec<false, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<true>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<false>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<true, BESwapNone>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<false, BESwapNone>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<true, BESwapSrc>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<false, BESwapSrc>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<true, BESwapDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<false, BESwapDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<true, BESwapSrcDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665_AltiVec<false, BESwapSrcDst>(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<true>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<false>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<true, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<false, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<true, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<false, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<true, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<false, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<true, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555XTo666X_AltiVec<false, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<true>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<false>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<true, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<false, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<true, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<false, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<true, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<false, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<true, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To8888Opaque_AltiVec<false, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<true>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<false>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<true, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<false, BESwapNone>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<true, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<false, BESwapSrc>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<true, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<false, BESwapDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<true, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template void ColorspaceConvert555To6665Opaque_AltiVec<false, BESwapSrcDst>(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template v128u32 ColorspaceConvert8888To6665_AltiVec<true>(const v128u32 &src);
template v128u32 ColorspaceConvert8888To6665_AltiVec<false>(const v128u32 &src);

22
desmume/src/utils/colorspacehandler/colorspacehandler_AltiVec.h
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2017 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -24,12 +24,12 @@
#warning This header requires PowerPC AltiVec support.
#else
template<bool SWAP_RB> void ColorspaceConvert555To8888_AltiVec(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB> void ColorspaceConvert555XTo888X_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB> void ColorspaceConvert555To6665_AltiVec(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB> void ColorspaceConvert555XTo666X_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB> void ColorspaceConvert555To8888Opaque_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB> void ColorspaceConvert555To6665Opaque_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB, BESwapFlags BE_BYTESWAP> void ColorspaceConvert555To8888_AltiVec(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB, BESwapFlags BE_BYTESWAP> void ColorspaceConvert555XTo888X_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB, BESwapFlags BE_BYTESWAP> void ColorspaceConvert555To6665_AltiVec(const v128u16 &srcColor, const v128u16 &srcAlphaBits, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB, BESwapFlags BE_BYTESWAP> void ColorspaceConvert555XTo666X_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB, BESwapFlags BE_BYTESWAP> void ColorspaceConvert555To8888Opaque_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB, BESwapFlags BE_BYTESWAP> void ColorspaceConvert555To6665Opaque_AltiVec(const v128u16 &srcColor, v128u32 &dstLo, v128u32 &dstHi);
template<bool SWAP_RB> v128u32 ColorspaceConvert8888To6665_AltiVec(const v128u32 &src);
template<bool SWAP_RB> v128u32 ColorspaceConvert6665To8888_AltiVec(const v128u32 &src);
template<bool SWAP_RB> v128u16 ColorspaceConvert8888To5551_AltiVec(const v128u32 &srcLo, const v128u32 &srcHi);
@ -46,11 +46,11 @@ class ColorspaceHandler_AltiVec : public ColorspaceHandler
public:
ColorspaceHandler_AltiVec() {};
size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665_SwapRB(const u32 *src, u32 *dst, size_t pixCount) const;

10
desmume/src/utils/colorspacehandler/colorspacehandler_SSE2.cpp
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2019 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -936,41 +936,49 @@ size_t ColorspaceApplyIntensityToBuffer32_SSE2(u32 *dst, size_t pixCountVec128,
return i;
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_SSE2<false, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_SSE2<true, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_SSE2<false, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To8888Opaque_SSE2<true, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_SSE2<false, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_SSE2<true, false>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_SSE2<false, true>(src, dst, pixCount);
}
template <BESwapFlags BE_BYTESWAP>
size_t ColorspaceHandler_SSE2::ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const
{
return ColorspaceConvertBuffer555To6665Opaque_SSE2<true, true>(src, dst, pixCount);

18
desmume/src/utils/colorspacehandler/colorspacehandler_SSE2.h
View File

@ -1,5 +1,5 @@
/*
Copyright (C) 2016-2017 DeSmuME team
Copyright (C) 2016-2021 DeSmuME team
This file is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@ -47,15 +47,15 @@ class ColorspaceHandler_SSE2 : public ColorspaceHandler
public:
ColorspaceHandler_SSE2() {};
size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To8888Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
template<BESwapFlags BE_BYTESWAP> size_t ConvertBuffer555To6665Opaque_SwapRB_IsUnaligned(const u16 *__restrict src, u32 *__restrict dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665(const u32 *src, u32 *dst, size_t pixCount) const;
size_t ConvertBuffer8888To6665_SwapRB(const u32 *src, u32 *dst, size_t pixCount) const;