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

Merge pull request #2720 from booto/efb-copy 

[WIP] Video: respect stride of efb copies to xfb
This commit is contained in:
Markus Wick 2015-07-25 12:03:03 +02:00
parent 9e4dae4a03 efd250494d
commit 96b77aaa81
15 changed files with 60 additions and 63 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
Source/Core/VideoBackends/D3D/FramebufferManager.cpp
View File

@ -159,10 +159,10 @@ FramebufferManager::~FramebufferManager()
SAFE_RELEASE(m_efb.resolved_depth_tex);
}
void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
{
u8* dst = Memory::GetPointer(xfbAddr);
s_xfbEncoder.Encode(dst, fbWidth, fbHeight, sourceRc, Gamma);
s_xfbEncoder.Encode(dst, fbStride, fbHeight, sourceRc, Gamma);
}
XFBSourceBase* FramebufferManager::CreateXFBSource(unsigned int target_width, unsigned int target_height, unsigned int layers)

2
Source/Core/VideoBackends/D3D/FramebufferManager.h
View File

@ -83,7 +83,7 @@ private:
XFBSourceBase* CreateXFBSource(unsigned int target_width, unsigned int target_height, unsigned int layers) override;
void GetTargetSize(unsigned int *width, unsigned int *height) override;
void CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
void CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
static struct Efb
{

8
Source/Core/VideoBackends/D3D/XFBEncoder.cpp
View File

@ -288,7 +288,7 @@ void XFBEncoder::Encode(u8* dst, u32 width, u32 height, const EFBRectangle& srcR
D3D::stateman->PushDepthState(m_xfbEncodeDepthState);
D3D::stateman->PushRasterizerState(m_xfbEncodeRastState);
D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, FLOAT(width/2), FLOAT(height));
D3D11_VIEWPORT vp = CD3D11_VIEWPORT(0.f, 0.f, FLOAT(width/4), FLOAT(height));
D3D::context->RSSetViewports(1, &vp);
D3D::stateman->SetInputLayout(m_quadLayout);
@ -300,7 +300,7 @@ void XFBEncoder::Encode(u8* dst, u32 width, u32 height, const EFBRectangle& srcR
TargetRectangle targetRect = g_renderer->ConvertEFBRectangle(srcRect);
XFBEncodeParams params = { 0 };
params.Width = FLOAT(width);
params.Width = FLOAT(width/2);
params.Height = FLOAT(height);
params.TexLeft = FLOAT(targetRect.left) / g_renderer->GetTargetWidth();
params.TexTop = FLOAT(targetRect.top) / g_renderer->GetTargetHeight();
@ -325,7 +325,7 @@ void XFBEncoder::Encode(u8* dst, u32 width, u32 height, const EFBRectangle& srcR
// Copy to staging buffer
D3D11_BOX srcBox = CD3D11_BOX(0, 0, 0, width/2, height, 1);
D3D11_BOX srcBox = CD3D11_BOX(0, 0, 0, width/4, height, 1);
D3D::context->CopySubresourceRegion(m_outStage, 0, 0, 0, 0, m_out, 0, &srcBox);
// Clean up state
@ -353,7 +353,7 @@ void XFBEncoder::Encode(u8* dst, u32 width, u32 height, const EFBRectangle& srcR
u8* src = (u8*)map.pData;
for (unsigned int y = 0; y < height; ++y)
{
memcpy(dst, src, 2*width);
memcpy(dst, src, width);
dst += bpmem.copyMipMapStrideChannels*32;
src += map.RowPitch;
}

4
Source/Core/VideoBackends/OGL/FramebufferManager.cpp
View File

@ -481,7 +481,7 @@ GLuint FramebufferManager::GetEFBDepthTexture(const EFBRectangle& sourceRc)
}
}
void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
{
u8* xfb_in_ram = Memory::GetPointer(xfbAddr);
if (!xfb_in_ram)
@ -491,7 +491,7 @@ void FramebufferManager::CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, c
}
TargetRectangle targetRc = g_renderer->ConvertEFBRectangle(sourceRc);
TextureConverter::EncodeToRamYUYV(ResolveAndGetRenderTarget(sourceRc), targetRc, xfb_in_ram, fbWidth, fbHeight);
TextureConverter::EncodeToRamYUYV(ResolveAndGetRenderTarget(sourceRc), targetRc, xfb_in_ram, sourceRc.GetWidth(), fbStride, fbHeight);
}
void FramebufferManager::SetFramebuffer(GLuint fb)

2
Source/Core/VideoBackends/OGL/FramebufferManager.h
View File

@ -97,7 +97,7 @@ private:
XFBSourceBase* CreateXFBSource(unsigned int target_width, unsigned int target_height, unsigned int layers) override;
void GetTargetSize(unsigned int *width, unsigned int *height) override;
void CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
void CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma) override;
static int m_targetWidth;
static int m_targetHeight;

16
Source/Core/VideoBackends/OGL/Render.cpp
View File

@ -1456,10 +1456,16 @@ void Renderer::SwapImpl(u32 xfbAddr, u32 fbWidth, u32 fbStride, u32 fbHeight, co
xfbSource = (const XFBSource*) xfbSourceList[i];
TargetRectangle drawRc;
TargetRectangle sourceRc;
sourceRc.left = xfbSource->sourceRc.left;
sourceRc.right = xfbSource->sourceRc.right;
sourceRc.top = xfbSource->sourceRc.top;
sourceRc.bottom = xfbSource->sourceRc.bottom;
if (g_ActiveConfig.bUseRealXFB)
{
drawRc = flipped_trc;
sourceRc.right -= fbStride - fbWidth;
}
else
{
@ -1481,18 +1487,12 @@ void Renderer::SwapImpl(u32 xfbAddr, u32 fbWidth, u32 fbStride, u32 fbHeight, co
//drawRc.bottom *= vScale;
//drawRc.left *= hScale;
//drawRc.right *= hScale;
sourceRc.right -= Renderer::EFBToScaledX(fbStride - fbWidth);
}
// Tell the OSD Menu about the current internal resolution
OSDInternalW = xfbSource->sourceRc.GetWidth(); OSDInternalH = xfbSource->sourceRc.GetHeight();
TargetRectangle sourceRc;
sourceRc.left = xfbSource->sourceRc.left;
sourceRc.right = xfbSource->sourceRc.right;
sourceRc.top = xfbSource->sourceRc.top;
sourceRc.bottom = xfbSource->sourceRc.bottom;
sourceRc.right -= Renderer::EFBToScaledX(fbStride - fbWidth);
BlitScreen(sourceRc, drawRc, xfbSource->texture, xfbSource->texWidth, xfbSource->texHeight);
}
}

3
Source/Core/VideoBackends/OGL/TextureCache.cpp
View File

@ -269,7 +269,8 @@ void TextureCache::TCacheEntry::FromRenderTarget(u32 dstAddr, unsigned int dstFo
isIntensity,
dstFormat,
scaleByHalf,
srcRect);
srcRect,
copyMipMapStrideChannels * 32);
u8* dst = Memory::GetPointer(dstAddr);
u64 const new_hash = GetHash64(dst,encoded_size,g_ActiveConfig.iSafeTextureCache_ColorSamples);

40
Source/Core/VideoBackends/OGL/TextureConverter.cpp
View File

@ -213,12 +213,12 @@ void Shutdown()
s_texConvFrameBuffer[1] = 0;
}
// dst_line_size, writeStride in bytes
static void EncodeToRamUsingShader(GLuint srcTexture,
u8* destAddr, int dstWidth, int dstHeight, int readStride,
bool linearFilter)
u8* destAddr, u32 dst_line_size, u32 dstHeight,
u32 writeStride, bool linearFilter)
{
// switch to texture converter frame buffer
// attach render buffer as color destination
FramebufferManager::SetFramebuffer(s_texConvFrameBuffer[0]);
@ -234,19 +234,13 @@ static void EncodeToRamUsingShader(GLuint srcTexture,
else
g_sampler_cache->BindNearestSampler(9);
glViewport(0, 0, (GLsizei)dstWidth, (GLsizei)dstHeight);
glViewport(0, 0, (GLsizei)(dst_line_size / 4), (GLsizei)dstHeight);
glDrawArrays(GL_TRIANGLE_STRIP, 0, 4);
// .. and then read back the results.
// TODO: make this less slow.
int dstSize = dst_line_size * dstHeight;
int writeStride = bpmem.copyMipMapStrideChannels * 32;
int dstSize = dstWidth*dstHeight*4;
int readHeight = readStride / dstWidth / 4; // 4 bytes per pixel
int readLoops = dstHeight / readHeight;
if (writeStride != readStride && readLoops > 1)
if ((writeStride != dst_line_size) && (dstHeight > 1))
{
// writing to a texture of a different size
// also copy more then one block line, so the different strides matters
@ -255,13 +249,13 @@ static void EncodeToRamUsingShader(GLuint srcTexture,
// CPU overhead because of the pbo
glBindBuffer(GL_PIXEL_PACK_BUFFER, s_PBO);
glBufferData(GL_PIXEL_PACK_BUFFER, dstSize, nullptr, GL_STREAM_READ);
glReadPixels(0, 0, (GLsizei)dstWidth, (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, nullptr);
glReadPixels(0, 0, (GLsizei)(dst_line_size / 4), (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, nullptr);
u8* pbo = (u8*)glMapBufferRange(GL_PIXEL_PACK_BUFFER, 0, dstSize, GL_MAP_READ_BIT);
for (int i = 0; i < readLoops; i++)
for (size_t i = 0; i < dstHeight; ++i)
{
memcpy(destAddr, pbo, readStride);
pbo += readStride;
memcpy(destAddr, pbo, dst_line_size);
pbo += dst_line_size;
destAddr += writeStride;
}
@ -270,11 +264,11 @@ static void EncodeToRamUsingShader(GLuint srcTexture,
}
else
{
glReadPixels(0, 0, (GLsizei)dstWidth, (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, destAddr);
glReadPixels(0, 0, (GLsizei)(dst_line_size / 4), (GLsizei)dstHeight, GL_BGRA, GL_UNSIGNED_BYTE, destAddr);
}
}
int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source)
int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source, u32 writeStride)
{
u32 format = copyfmt;
@ -323,13 +317,13 @@ int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer,
cacheLinesPerRow = numBlocksX;
EncodeToRamUsingShader(source_texture,
dest_ptr, cacheLinesPerRow * 8, numBlocksY, cacheLinesPerRow * 32,
bScaleByHalf > 0 && !bFromZBuffer);
dest_ptr, cacheLinesPerRow * 32, numBlocksY,
writeStride, bScaleByHalf > 0 && !bFromZBuffer);
return size_in_bytes; // TODO: D3D11 is calculating this value differently!
}
void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc, u8* destAddr, int dstWidth, int dstHeight)
void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc, u8* destAddr, u32 dstWidth, u32 dstStride, u32 dstHeight)
{
g_renderer->ResetAPIState();
@ -341,7 +335,7 @@ void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc, u8* des
// We enable linear filtering, because the GameCube does filtering in the vertical direction when
// yscale is enabled.
// Otherwise we get jaggies when a game uses yscaling (most PAL games)
EncodeToRamUsingShader(srcTexture, destAddr, dstWidth / 2, dstHeight, dstWidth*dstHeight*2, true);
EncodeToRamUsingShader(srcTexture, destAddr, dstWidth * 2, dstHeight, dstStride, true);
FramebufferManager::SetFramebuffer(0);
TextureCache::DisableStage(0);
g_renderer->RestoreAPIState();

4
Source/Core/VideoBackends/OGL/TextureConverter.h
View File

@ -19,12 +19,12 @@ void Init();
void Shutdown();
void EncodeToRamYUYV(GLuint srcTexture, const TargetRectangle& sourceRc,
u8* destAddr, int dstWidth, int dstHeight);
u8* destAddr, u32 dstWidth, u32 dstStride, u32 dstHeight);
void DecodeToTexture(u32 xfbAddr, int srcWidth, int srcHeight, GLuint destTexture);
// returns size of the encoded data (in bytes)
int EncodeToRamFromTexture(u32 address, GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source);
int EncodeToRamFromTexture(u32 address, GLuint source_texture, bool bFromZBuffer, bool bIsIntensityFmt, u32 copyfmt, int bScaleByHalf, const EFBRectangle& source, u32 writeStride);
}

7
Source/Core/VideoCommon/BPStructs.cpp
View File

@ -251,9 +251,10 @@ static void BPWritten(const BPCmd& bp)
height = MAX_XFB_HEIGHT;
}
u32 width = bpmem.copyMipMapStrideChannels << 4;
Renderer::RenderToXFB(destAddr, srcRect, width, height, s_gammaLUT[PE_copy.gamma]);
u32 stride = bpmem.copyMipMapStrideChannels << 5;
WARN_LOG(VIDEO, "RenderToXFB: destAddr: %08x | srcRect {%d %d %d %d} | fbWidth: %u | fbStride: %u | fbHeight: %u",
destAddr, srcRect.left, srcRect.top, srcRect.right, srcRect.bottom, bpmem.copyTexSrcWH.x + 1, stride, height);
Renderer::RenderToXFB(destAddr, srcRect, stride, height, s_gammaLUT[PE_copy.gamma]);
}
// Clear the rectangular region after copying it.

12
Source/Core/VideoCommon/FramebufferManagerBase.cpp
View File

@ -115,17 +115,17 @@ const XFBSourceBase* const* FramebufferManagerBase::GetVirtualXFBSource(u32 xfbA
return &m_overlappingXFBArray[0];
}
void FramebufferManagerBase::CopyToXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
void FramebufferManagerBase::CopyToXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc, float Gamma)
{
if (g_ActiveConfig.bUseRealXFB)
g_framebuffer_manager->CopyToRealXFB(xfbAddr, fbWidth, fbHeight, sourceRc,Gamma);
g_framebuffer_manager->CopyToRealXFB(xfbAddr, fbStride, fbHeight, sourceRc, Gamma);
else
CopyToVirtualXFB(xfbAddr, fbWidth, fbHeight, sourceRc,Gamma);
CopyToVirtualXFB(xfbAddr, fbStride, fbHeight, sourceRc, Gamma);
}
void FramebufferManagerBase::CopyToVirtualXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma)
void FramebufferManagerBase::CopyToVirtualXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc, float Gamma)
{
VirtualXFBListType::iterator vxfb = FindVirtualXFB(xfbAddr, fbWidth, fbHeight);
VirtualXFBListType::iterator vxfb = FindVirtualXFB(xfbAddr, sourceRc.GetWidth(), fbHeight);
if (m_virtualXFBList.end() == vxfb)
{
@ -165,7 +165,7 @@ void FramebufferManagerBase::CopyToVirtualXFB(u32 xfbAddr, u32 fbWidth, u32 fbHe
}
vxfb->xfbSource->srcAddr = vxfb->xfbAddr = xfbAddr;
vxfb->xfbSource->srcWidth = vxfb->xfbWidth = fbWidth;
vxfb->xfbSource->srcWidth = vxfb->xfbWidth = sourceRc.GetWidth();
vxfb->xfbSource->srcHeight = vxfb->xfbHeight = fbHeight;
vxfb->xfbSource->sourceRc = g_renderer->ConvertEFBRectangle(sourceRc);

4
Source/Core/VideoCommon/FramebufferManagerBase.h
View File

@ -45,7 +45,7 @@ public:
FramebufferManagerBase();
virtual ~FramebufferManagerBase();
static void CopyToXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma);
static void CopyToXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma);
static const XFBSourceBase* const* GetXFBSource(u32 xfbAddr, u32 fbWidth, u32 fbHeight, u32* xfbCount);
static void SetLastXfbWidth(unsigned int width) { s_last_xfb_width = width; }
@ -87,7 +87,7 @@ private:
static void ReplaceVirtualXFB();
// TODO: merge these virtual funcs, they are nearly all the same
virtual void CopyToRealXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma = 1.0f) = 0;
virtual void CopyToRealXFB(u32 xfbAddr, u32 fbStride, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma = 1.0f) = 0;
static void CopyToVirtualXFB(u32 xfbAddr, u32 fbWidth, u32 fbHeight, const EFBRectangle& sourceRc,float Gamma = 1.0f);
static const XFBSourceBase* const* GetRealXFBSource(u32 xfbAddr, u32 fbWidth, u32 fbHeight, u32* xfbCount);

8
Source/Core/VideoCommon/RenderBase.cpp
View File

@ -112,22 +112,22 @@ Renderer::~Renderer()
#endif
}
void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbWidth, u32 fbHeight, float Gamma)
void Renderer::RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbStride, u32 fbHeight, float Gamma)
{
CheckFifoRecording();
if (!fbWidth || !fbHeight)
if (!fbStride || !fbHeight)
return;
XFBWrited = true;
if (g_ActiveConfig.bUseXFB)
{
FramebufferManagerBase::CopyToXFB(xfbAddr, fbWidth, fbHeight, sourceRc, Gamma);
FramebufferManagerBase::CopyToXFB(xfbAddr, fbStride, fbHeight, sourceRc, Gamma);
}
else
{
Swap(xfbAddr, fbWidth, fbWidth, fbHeight, sourceRc, Gamma);
Swap(xfbAddr, sourceRc.GetWidth(), fbStride, fbHeight, sourceRc, Gamma);
}
}

2
Source/Core/VideoCommon/RenderBase.h
View File

@ -109,7 +109,7 @@ public:
virtual void ClearScreen(const EFBRectangle& rc, bool colorEnable, bool alphaEnable, bool zEnable, u32 color, u32 z) = 0;
virtual void ReinterpretPixelData(unsigned int convtype) = 0;
static void RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbWidth, u32 fbHeight, float Gamma = 1.0f);
static void RenderToXFB(u32 xfbAddr, const EFBRectangle& sourceRc, u32 fbStride, u32 fbHeight, float Gamma = 1.0f);
virtual u32 AccessEFB(EFBAccessType type, u32 x, u32 y, u32 poke_data) = 0;
virtual void PokeEFB(EFBAccessType type, const std::vector<EfbPokeData>& data);

7
Source/Core/VideoCommon/VideoCommon.h
View File

@ -19,9 +19,10 @@ enum
EFB_HEIGHT = 528,
};
// XFB width is decided by EFB copy operation. The VI can do horizontal
// scaling (TODO: emulate).
const u32 MAX_XFB_WIDTH = EFB_WIDTH;
// Max XFB width is 720. You can only copy out 640 wide areas of efb to XFB
// so you need multiple copies to do the full width.
// The VI can do horizontal scaling (TODO: emulate).
const u32 MAX_XFB_WIDTH = 720;
// Although EFB height is 528, 574-line XFB's can be created either with
// vertical scaling by the EFB copy operation or copying to multiple XFB's