Merge branch 'const_buffer_rework'

This commit is contained in:
degasus
2013-10-13 14:14:32 +02:00
24 changed files with 392 additions and 893 deletions
@@ -14,15 +14,12 @@
#include "Globals.h"
#include "PixelShaderGen.h"
#include "PixelShaderCache.h"
#include "PixelShaderManager.h"
#include "ConfigManager.h"
extern int frameCount;
// See comment near the bottom of this file.
float psconstants[C_PENVCONST_END*4];
bool pscbufchanged = true;
namespace DX11
{
@@ -339,15 +336,15 @@ ID3D11PixelShader* PixelShaderCache::GetClearProgram()
ID3D11Buffer* &PixelShaderCache::GetConstantBuffer()
{
// TODO: divide the global variables of the generated shaders into about 5 constant buffers to speed this up
if (pscbufchanged)
if (PixelShaderManager::dirty)
{
D3D11_MAPPED_SUBRESOURCE map;
D3D::context->Map(pscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
memcpy(map.pData, psconstants, sizeof(psconstants));
memcpy(map.pData, &PixelShaderManager::constants, sizeof(PixelShaderConstants));
D3D::context->Unmap(pscbuf, 0);
pscbufchanged = false;
PixelShaderManager::dirty = false;
ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(psconstants));
ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(PixelShaderConstants));
}
return pscbuf;
}
@@ -364,7 +361,7 @@ public:
void PixelShaderCache::Init()
{
unsigned int cbsize = ((sizeof(psconstants))&(~0xf))+0x10; // must be a multiple of 16
unsigned int cbsize = ((sizeof(PixelShaderConstants))&(~0xf))+0x10; // must be a multiple of 16
D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
D3D::device->CreateBuffer(&cbdesc, NULL, &pscbuf);
CHECK(pscbuf!=NULL, "Create pixel shader constant buffer");
@@ -536,50 +533,4 @@ bool PixelShaderCache::InsertByteCode(const PixelShaderUid &uid, const void* byt
return true;
}
// These are "callbacks" from VideoCommon and thus must be outside namespace DX11.
// This will have to be changed when we merge.
// HACK to avoid some invasive VideoCommon changes
// these values are hardcoded, they depend on internal D3DCompile behavior; TODO: Solve this with D3DReflect or something
// offset given in floats, table index is float4
static const unsigned int ps_constant_offset_table[] = {
0, 4, 8, 12, // C_COLORS, 16
16, 20, 24, 28, // C_KCOLORS, 16
32, // C_ALPHA, 4
36, 40, 44, 48, 52, 56, 60, 64, // C_TEXDIMS, 32
68, 72, // C_ZBIAS, 8
76, 80, // C_INDTEXSCALE, 8
84, 88, 92, 96, 100, 104, // C_INDTEXMTX, 24
108, 112, 116, // C_FOG, 12
120, 124, 128, 132, 136, // C_PLIGHTS0, 20
140, 144, 148, 152, 156, // C_PLIGHTS1, 20
160, 164, 168, 172, 176, // C_PLIGHTS2, 20
180, 184, 188, 192, 196, // C_PLIGHTS3, 20
200, 204, 208, 212, 216, // C_PLIGHTS4, 20
220, 224, 228, 232, 236, // C_PLIGHTS5, 20
240, 244, 248, 252, 256, // C_PLIGHTS6, 20
260, 264, 268, 272, 276, // C_PLIGHTS7, 20
280, 284, 288, 292 // C_PMATERIALS, 16
};
void Renderer::SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
{
psconstants[ps_constant_offset_table[const_number] ] = f1;
psconstants[ps_constant_offset_table[const_number]+1] = f2;
psconstants[ps_constant_offset_table[const_number]+2] = f3;
psconstants[ps_constant_offset_table[const_number]+3] = f4;
pscbufchanged = true;
}
void Renderer::SetPSConstant4fv(unsigned int const_number, const float* f)
{
memcpy(&psconstants[ps_constant_offset_table[const_number]], f, sizeof(float)*4);
pscbufchanged = true;
}
void Renderer::SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float* f)
{
memcpy(&psconstants[ps_constant_offset_table[const_number]], f, sizeof(float)*4*count);
pscbufchanged = true;
}
} // DX11
@@ -51,15 +51,6 @@ public:
bool SaveScreenshot(const std::string &filename, const TargetRectangle &rc);
static bool CheckForResize();
void SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
void SetPSConstant4fv(unsigned int const_number, const float *f);
void SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
void SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
void SetVSConstant4fv(unsigned int const_number, const float *f);
void SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float *f);
void SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
};
}
@@ -12,14 +12,10 @@
#include "D3DShader.h"
#include "Globals.h"
#include "VertexShaderCache.h"
#include "VertexShaderManager.h"
#include "ConfigManager.h"
// See comment near the bottom of this file
static unsigned int vs_constant_offset_table[C_VENVCONST_END];
float vsconstants[C_VENVCONST_END*4];
bool vscbufchanged = true;
namespace DX11 {
VertexShaderCache::VSCache VertexShaderCache::vshaders;
@@ -44,15 +40,15 @@ ID3D11Buffer* vscbuf = NULL;
ID3D11Buffer* &VertexShaderCache::GetConstantBuffer()
{
// TODO: divide the global variables of the generated shaders into about 5 constant buffers to speed this up
if (vscbufchanged)
if (VertexShaderManager::dirty)
{
D3D11_MAPPED_SUBRESOURCE map;
D3D::context->Map(vscbuf, 0, D3D11_MAP_WRITE_DISCARD, 0, &map);
memcpy(map.pData, vsconstants, sizeof(vsconstants));
memcpy(map.pData, &VertexShaderManager::constants, sizeof(VertexShaderConstants));
D3D::context->Unmap(vscbuf, 0);
vscbufchanged = false;
VertexShaderManager::dirty = false;
ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(vsconstants));
ADDSTAT(stats.thisFrame.bytesUniformStreamed, sizeof(VertexShaderConstants));
}
return vscbuf;
}
@@ -116,7 +112,7 @@ void VertexShaderCache::Init()
{ "COLOR", 0, DXGI_FORMAT_R8G8B8A8_UNORM, 0, 12, D3D11_INPUT_PER_VERTEX_DATA, 0 },
};
unsigned int cbsize = ((sizeof(vsconstants))&(~0xf))+0x10; // must be a multiple of 16
unsigned int cbsize = ((sizeof(VertexShaderConstants))&(~0xf))+0x10; // must be a multiple of 16
D3D11_BUFFER_DESC cbdesc = CD3D11_BUFFER_DESC(cbsize, D3D11_BIND_CONSTANT_BUFFER, D3D11_USAGE_DYNAMIC, D3D11_CPU_ACCESS_WRITE);
HRESULT hr = D3D::device->CreateBuffer(&cbdesc, NULL, &vscbuf);
CHECK(hr==S_OK, "Create vertex shader constant buffer (size=%u)", cbsize);
@@ -141,19 +137,6 @@ void VertexShaderCache::Init()
Clear();
// these values are hardcoded, they depend on internal D3DCompile behavior
// TODO: Do this with D3DReflect or something instead
unsigned int k;
for (k = 0;k < 6;k++) vs_constant_offset_table[C_POSNORMALMATRIX+k] = 0+4*k;
for (k = 0;k < 4;k++) vs_constant_offset_table[C_PROJECTION+k] = 24+4*k;
for (k = 0;k < 4;k++) vs_constant_offset_table[C_MATERIALS+k] = 40+4*k;
for (k = 0;k < 40;k++) vs_constant_offset_table[C_LIGHTS+k] = 56+4*k;
for (k = 0;k < 24;k++) vs_constant_offset_table[C_TEXMATRICES+k] = 216+4*k;
for (k = 0;k < 64;k++) vs_constant_offset_table[C_TRANSFORMMATRICES+k] = 312+4*k;
for (k = 0;k < 32;k++) vs_constant_offset_table[C_NORMALMATRICES+k] = 568+4*k;
for (k = 0;k < 64;k++) vs_constant_offset_table[C_POSTTRANSFORMMATRICES+k] = 696+4*k;
vs_constant_offset_table[C_DEPTHPARAMS] = 952;
if (!File::Exists(File::GetUserPath(D_SHADERCACHE_IDX)))
File::CreateDir(File::GetUserPath(D_SHADERCACHE_IDX).c_str());
@@ -277,39 +260,4 @@ bool VertexShaderCache::InsertByteCode(const VertexShaderUid &uid, D3DBlob* bcod
return true;
}
// These are "callbacks" from VideoCommon and thus must be outside namespace DX11.
// This will have to be changed when we merge.
// maps the constant numbers to float indices in the constant buffer
void Renderer::SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
{
vsconstants[vs_constant_offset_table[const_number] ] = f1;
vsconstants[vs_constant_offset_table[const_number]+1] = f2;
vsconstants[vs_constant_offset_table[const_number]+2] = f3;
vsconstants[vs_constant_offset_table[const_number]+3] = f4;
vscbufchanged = true;
}
void Renderer::SetVSConstant4fv(unsigned int const_number, const float* f)
{
memcpy(&vsconstants[vs_constant_offset_table[const_number]], f, sizeof(float)*4);
vscbufchanged = true;
}
void Renderer::SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float* f)
{
for (unsigned int i = 0; i < count; i++)
{
memcpy(&vsconstants[vs_constant_offset_table[const_number+i]], f+3*i, sizeof(float)*3);
vsconstants[vs_constant_offset_table[const_number+i]+3] = 0.f;
}
vscbufchanged = true;
}
void Renderer::SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float* f)
{
memcpy(&vsconstants[vs_constant_offset_table[const_number]], f, sizeof(float)*4*count);
vscbufchanged = true;
}
} // namespace DX11
@@ -3,7 +3,6 @@ set(SRCS Src/FramebufferManager.cpp
Src/main.cpp
Src/NativeVertexFormat.cpp
Src/PerfQuery.cpp
Src/PixelShaderCache.cpp
Src/PostProcessing.cpp
Src/ProgramShaderCache.cpp
Src/RasterFont.cpp
@@ -12,7 +11,6 @@ set(SRCS Src/FramebufferManager.cpp
Src/StreamBuffer.cpp
Src/TextureCache.cpp
Src/TextureConverter.cpp
Src/VertexShaderCache.cpp
Src/VertexManager.cpp)
set(LIBS videocommon
+1 -3
View File
@@ -151,7 +151,6 @@
<ClCompile Include="Src\main.cpp" />
<ClCompile Include="Src\NativeVertexFormat.cpp" />
<ClCompile Include="Src\PerfQuery.cpp" />
<ClCompile Include="Src\PixelShaderCache.cpp" />
<ClCompile Include="Src\PostProcessing.cpp" />
<ClCompile Include="Src\ProgramShaderCache.cpp" />
<ClCompile Include="Src\RasterFont.cpp" />
@@ -169,7 +168,6 @@
<ClCompile Include="Src\TextureCache.cpp" />
<ClCompile Include="Src\TextureConverter.cpp" />
<ClCompile Include="Src\VertexManager.cpp" />
<ClCompile Include="Src\VertexShaderCache.cpp" />
</ItemGroup>
<ItemGroup>
<ClInclude Include="Src\FramebufferManager.h" />
@@ -203,4 +201,4 @@
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
</ImportGroup>
</Project>
</Project>
@@ -1,104 +0,0 @@
// Copyright 2013 Dolphin Emulator Project
// Licensed under GPLv2
// Refer to the license.txt file included.
#include "Globals.h"
#include "GLUtil.h"
#include <cmath>
#include "Statistics.h"
#include "VideoConfig.h"
#include "ImageWrite.h"
#include "Common.h"
#include "Render.h"
#include "VertexShaderGen.h"
#include "ProgramShaderCache.h"
#include "PixelShaderManager.h"
#include "OnScreenDisplay.h"
#include "StringUtil.h"
#include "FileUtil.h"
#include "Debugger.h"
namespace OGL
{
void SetPSConstant4fvByName(const char * name, unsigned int offset, const float *f, const unsigned int count = 1)
{
ProgramShaderCache::PCacheEntry tmp = ProgramShaderCache::GetShaderProgram();
for (int a = 0; a < NUM_UNIFORMS; ++a)
{
if (!strcmp(name, UniformNames[a]))
{
if (tmp.shader.UniformLocations[a] == -1)
return;
else if (tmp.shader.UniformSize[a] <= offset)
return;
else
{
unsigned int maxcount= tmp.shader.UniformSize[a]-offset;
glUniform4fv(tmp.shader.UniformLocations[a] + offset, std::min(count, maxcount), f);
return;
}
}
}
}
// Renderer functions
void Renderer::SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
{
float const f[4] = {f1, f2, f3, f4};
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
ProgramShaderCache::SetMultiPSConstant4fv(const_number, f, 1);
return;
}
for (unsigned int a = 0; a < 10; ++a)
{
if (const_number >= PSVar_Loc[a].reg && const_number < (PSVar_Loc[a].reg + PSVar_Loc[a].size))
{
unsigned int offset = const_number - PSVar_Loc[a].reg;
SetPSConstant4fvByName(PSVar_Loc[a].name, offset, f);
return;
}
}
}
void Renderer::SetPSConstant4fv(unsigned int const_number, const float *f)
{
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
ProgramShaderCache::SetMultiPSConstant4fv(const_number, f, 1);
return;
}
for (unsigned int a = 0; a < 10; ++a)
{
if (const_number >= PSVar_Loc[a].reg && const_number < (PSVar_Loc[a].reg + PSVar_Loc[a].size))
{
unsigned int offset = const_number - PSVar_Loc[a].reg;
SetPSConstant4fvByName(PSVar_Loc[a].name, offset, f);
return;
}
}
}
void Renderer::SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float *f)
{
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
ProgramShaderCache::SetMultiPSConstant4fv(const_number, f, count);
return;
}
for (unsigned int a = 0; a < 10; ++a)
{
if (const_number >= PSVar_Loc[a].reg && const_number < (PSVar_Loc[a].reg + PSVar_Loc[a].size))
{
unsigned int offset = const_number - PSVar_Loc[a].reg;
SetPSConstant4fvByName(PSVar_Loc[a].name, offset, f, count);
return;
}
}
}
} // namespace OGL
@@ -10,18 +10,16 @@
#include "Statistics.h"
#include "ImageWrite.h"
#include "Render.h"
#include "PixelShaderManager.h"
#include "VertexShaderManager.h"
namespace OGL
{
static const u32 UBO_LENGTH = 32*1024*1024;
GLintptr ProgramShaderCache::s_vs_data_size;
GLintptr ProgramShaderCache::s_ps_data_size;
GLintptr ProgramShaderCache::s_vs_data_offset;
u8 *ProgramShaderCache::s_ubo_buffer;
u32 ProgramShaderCache::s_ubo_buffer_size;
bool ProgramShaderCache::s_ubo_dirty;
s32 ProgramShaderCache::s_ubo_align;
static StreamBuffer *s_buffer;
static int num_failures = 0;
@@ -36,6 +34,10 @@ UidChecker<VertexShaderUid,VertexShaderCode> ProgramShaderCache::vertex_uid_chec
static char s_glsl_header[1024] = "";
// Annoying sure, can be removed once we drop our UBO workaround
const char *UniformNames[NUM_UNIFORMS] =
{
// PIXEL SHADER UNIFORMS
@@ -61,6 +63,33 @@ const char *UniformNames[NUM_UNIFORMS] =
I_DEPTHPARAMS,
};
const static int PSVar_Loc[] = {
offsetof(PixelShaderConstants, colors)/16,
offsetof(PixelShaderConstants, kcolors)/16,
offsetof(PixelShaderConstants, alpha)/16,
offsetof(PixelShaderConstants, texdims)/16,
offsetof(PixelShaderConstants, zbias)/16,
offsetof(PixelShaderConstants, indtexscale)/16,
offsetof(PixelShaderConstants, indtexmtx)/16,
offsetof(PixelShaderConstants, fog)/16,
offsetof(PixelShaderConstants, plights)/16,
offsetof(PixelShaderConstants, pmaterials)/16,
};
const static int VSVar_Loc[] = {
offsetof(VertexShaderConstants, posnormalmatrix)/16,
offsetof(VertexShaderConstants, projection)/16,
offsetof(VertexShaderConstants, materials)/16,
offsetof(VertexShaderConstants, lights)/16,
offsetof(VertexShaderConstants, texmatrices)/16,
offsetof(VertexShaderConstants, transformmatrices)/16,
offsetof(VertexShaderConstants, normalmatrices)/16,
offsetof(VertexShaderConstants, posttransformmatrices)/16,
offsetof(VertexShaderConstants, depthparams)/16,
};
// End of UBO workaround
void SHADER::SetProgramVariables()
{
// glsl shader must be bind to set samplers
@@ -162,30 +191,43 @@ void SHADER::Bind()
}
}
void ProgramShaderCache::SetMultiPSConstant4fv(unsigned int offset, const float *f, unsigned int count)
{
s_ubo_dirty = true;
memcpy(s_ubo_buffer+(offset*4*sizeof(float)), f, count*4*sizeof(float));
}
void ProgramShaderCache::SetMultiVSConstant4fv(unsigned int offset, const float *f, unsigned int count)
{
s_ubo_dirty = true;
memcpy(s_ubo_buffer+(offset*4*sizeof(float))+s_vs_data_offset, f, count*4*sizeof(float));
}
void ProgramShaderCache::UploadConstants()
{
if(s_ubo_dirty) {
s_buffer->Alloc(s_ubo_buffer_size);
size_t offset = s_buffer->Upload(s_ubo_buffer, s_ubo_buffer_size);
glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_buffer->getBuffer(), offset, s_ps_data_size);
glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_buffer->getBuffer(), offset + s_vs_data_offset, s_vs_data_size);
s_ubo_dirty = false;
if(g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
if(PixelShaderManager::dirty || VertexShaderManager::dirty)
{
s_buffer->Alloc(s_ubo_buffer_size);
size_t offset = s_buffer->Upload((u8*)&PixelShaderManager::constants, ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align));
glBindBufferRange(GL_UNIFORM_BUFFER, 1, s_buffer->getBuffer(), offset, sizeof(PixelShaderConstants));
offset = s_buffer->Upload((u8*)&VertexShaderManager::constants, ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align));
glBindBufferRange(GL_UNIFORM_BUFFER, 2, s_buffer->getBuffer(), offset, sizeof(VertexShaderConstants));
PixelShaderManager::dirty = false;
VertexShaderManager::dirty = false;
ADDSTAT(stats.thisFrame.bytesUniformStreamed, s_ubo_buffer_size);
}
}
else
{
// UBO workaround
// this must be updated per shader switch, so also update it when it's not dirty
for (unsigned int a = 0; a < 10; ++a)
{
if(last_entry->shader.UniformSize[a] > 0)
glUniform4fv(last_entry->shader.UniformLocations[a], last_entry->shader.UniformSize[a], (float*) &PixelShaderManager::constants + 4*PSVar_Loc[a]);
}
for (unsigned int a = 0; a < 9; ++a)
{
if(last_entry->shader.UniformSize[a+10] > 0)
glUniform4fv(last_entry->shader.UniformLocations[a+10], last_entry->shader.UniformSize[a+10], (float*) &VertexShaderManager::constants + 4*VSVar_Loc[a]);
}
ADDSTAT(stats.thisFrame.bytesUniformStreamed, s_ubo_buffer_size);
}
}
GLuint ProgramShaderCache::GetCurrentProgram(void)
@@ -419,22 +461,14 @@ void ProgramShaderCache::Init(void)
// then the UBO will fail.
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
GLint Align;
glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &Align);
glGetIntegerv(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT, &s_ubo_align);
s_ps_data_size = C_PENVCONST_END * sizeof(float) * 4;
s_vs_data_size = C_VENVCONST_END * sizeof(float) * 4;
s_vs_data_offset = ROUND_UP(s_ps_data_size, Align);
s_ubo_buffer_size = ROUND_UP(s_ps_data_size, Align) + ROUND_UP(s_vs_data_size, Align);
s_ubo_buffer_size = ROUND_UP(sizeof(PixelShaderConstants), s_ubo_align) + ROUND_UP(sizeof(VertexShaderConstants), s_ubo_align);
// We multiply by *4*4 because we need to get down to basic machine units.
// So multiply by four to get how many floats we have from vec4s
// Then once more to get bytes
s_buffer = new StreamBuffer(GL_UNIFORM_BUFFER, UBO_LENGTH);
s_ubo_buffer = new u8[s_ubo_buffer_size];
memset(s_ubo_buffer, 0, s_ubo_buffer_size);
s_ubo_dirty = true;
}
// Read our shader cache, only if supported
@@ -509,8 +543,6 @@ void ProgramShaderCache::Shutdown(void)
{
delete s_buffer;
s_buffer = 0;
delete [] s_ubo_buffer;
s_ubo_buffer = 0;
}
}
@@ -87,9 +87,6 @@ public:
static bool CompileShader(SHADER &shader, const char* vcode, const char* pcode);
static GLuint CompileSingleShader(GLuint type, const char *code);
static void SetMultiPSConstant4fv(unsigned int offset, const float *f, unsigned int count);
static void SetMultiVSConstant4fv(unsigned int offset, const float *f, unsigned int count);
static void UploadConstants();
static void Init(void);
@@ -110,12 +107,8 @@ private:
static UidChecker<PixelShaderUid,PixelShaderCode> pixel_uid_checker;
static UidChecker<VertexShaderUid,VertexShaderCode> vertex_uid_checker;
static GLintptr s_vs_data_size;
static GLintptr s_ps_data_size;
static GLintptr s_vs_data_offset;
static u8 *s_ubo_buffer;
static u32 s_ubo_buffer_size;
static bool s_ubo_dirty;
static s32 s_ubo_align;
};
} // namespace OGL
@@ -81,15 +81,6 @@ public:
bool SaveScreenshot(const std::string &filename, const TargetRectangle &rc);
void SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
void SetPSConstant4fv(unsigned int const_number, const float *f);
void SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
void SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4);
void SetVSConstant4fv(unsigned int const_number, const float *f);
void SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float *f);
void SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float *f);
private:
void UpdateEFBCache(EFBAccessType type, u32 cacheRectIdx, const EFBRectangle& efbPixelRc, const TargetRectangle& targetPixelRc, const u32* data);
};
@@ -324,7 +324,7 @@ int EncodeToRamFromTexture(u32 address,GLuint source_texture, bool bFromZBuffer,
};
texconv_shader.Bind();
glUniform4fv(texconv_shader.UniformLocations[C_COLORS], 2, params);
glUniform4fv(texconv_shader.UniformLocations[0], 2, params);
TargetRectangle scaledSource;
scaledSource.top = 0;
@@ -1,130 +0,0 @@
// Copyright 2013 Dolphin Emulator Project
// Licensed under GPLv2
// Refer to the license.txt file included.
#include <math.h>
#include "Globals.h"
#include "VideoConfig.h"
#include "Statistics.h"
#include "GLUtil.h"
#include "Render.h"
#include "VertexShaderGen.h"
#include "VertexShaderManager.h"
#include "ProgramShaderCache.h"
#include "VertexManager.h"
#include "VertexLoader.h"
#include "XFMemory.h"
#include "ImageWrite.h"
#include "FileUtil.h"
#include "Debugger.h"
namespace OGL
{
void SetVSConstant4fvByName(const char * name, unsigned int offset, const float *f, const unsigned int count = 1)
{
ProgramShaderCache::PCacheEntry tmp = ProgramShaderCache::GetShaderProgram();
for (int a = 0; a < NUM_UNIFORMS; ++a)
{
if (!strcmp(name, UniformNames[a]))
{
if (tmp.shader.UniformLocations[a] == -1)
return;
else if (tmp.shader.UniformSize[a] <= offset)
return;
else
{
unsigned int maxcount= tmp.shader.UniformSize[a]-offset;
glUniform4fv(tmp.shader.UniformLocations[a] + offset, std::min(count, maxcount), f);
return;
}
}
}
}
void Renderer::SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4)
{
float const buf[4] = {f1, f2, f3, f4};
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
ProgramShaderCache::SetMultiVSConstant4fv(const_number, buf, 1);
return;
}
for (unsigned int a = 0; a < 9; ++a)
{
if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
{
unsigned int offset = const_number - VSVar_Loc[a].reg;
SetVSConstant4fvByName(VSVar_Loc[a].name, offset, buf);
return;
}
}
}
void Renderer::SetVSConstant4fv(unsigned int const_number, const float *f)
{
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
ProgramShaderCache::SetMultiVSConstant4fv(const_number, f, 1);
return;
}
for (unsigned int a = 0; a < 9; ++a)
{
if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
{
unsigned int offset = const_number - VSVar_Loc[a].reg;
SetVSConstant4fvByName(VSVar_Loc[a].name, offset, f);
return;
}
}
}
void Renderer::SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float *f)
{
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
ProgramShaderCache::SetMultiVSConstant4fv(const_number, f, count);
return;
}
for (unsigned int a = 0; a < 9; ++a)
{
if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
{
unsigned int offset = const_number - VSVar_Loc[a].reg;
SetVSConstant4fvByName(VSVar_Loc[a].name, offset, f, count);
return;
}
}
}
void Renderer::SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float *f)
{
float buf[4 * C_VENVCONST_END];
for (unsigned int i = 0; i < count; i++)
{
buf[4*i ] = *f++;
buf[4*i+1] = *f++;
buf[4*i+2] = *f++;
buf[4*i+3] = 0.f;
}
if (g_ActiveConfig.backend_info.bSupportsGLSLUBO)
{
ProgramShaderCache::SetMultiVSConstant4fv(const_number, buf, count);
return;
}
for (unsigned int a = 0; a < 9; ++a)
{
if (const_number >= VSVar_Loc[a].reg && const_number < ( VSVar_Loc[a].reg + VSVar_Loc[a].size))
{
unsigned int offset = const_number - VSVar_Loc[a].reg;
SetVSConstant4fvByName(VSVar_Loc[a].name, offset, buf, count);
return;
}
}
}
} // namespace OGL
+22 -13
View File
@@ -168,12 +168,16 @@ void BPWritten(const BPCmd& bp)
case BPMEM_IND_MTXA+6:
case BPMEM_IND_MTXB+6:
case BPMEM_IND_MTXC+6:
PixelShaderManager::SetIndMatrixChanged((bp.address - BPMEM_IND_MTXA) / 3);
if(bp.changes)
PixelShaderManager::SetIndMatrixChanged((bp.address - BPMEM_IND_MTXA) / 3);
break;
case BPMEM_RAS1_SS0: // Index Texture Coordinate Scale 0
PixelShaderManager::SetIndTexScaleChanged(0x03);
if(bp.changes)
PixelShaderManager::SetIndTexScaleChanged(false);
break;
case BPMEM_RAS1_SS1: // Index Texture Coordinate Scale 1
PixelShaderManager::SetIndTexScaleChanged(0x0c);
if(bp.changes)
PixelShaderManager::SetIndTexScaleChanged(true);
break;
// ----------------
// Scissor Control
@@ -220,7 +224,8 @@ void BPWritten(const BPCmd& bp)
case BPMEM_CONSTANTALPHA: // Set Destination Alpha
{
PRIM_LOG("constalpha: alp=%d, en=%d", bpmem.dstalpha.alpha, bpmem.dstalpha.enable);
PixelShaderManager::SetDestAlpha(bpmem.dstalpha);
if(bp.changes & 0xFF)
PixelShaderManager::SetDestAlpha();
if(bp.changes & 0x100)
SetBlendMode();
break;
@@ -341,29 +346,32 @@ void BPWritten(const BPCmd& bp)
case BPMEM_FOGRANGE+3:
case BPMEM_FOGRANGE+4:
case BPMEM_FOGRANGE+5:
if (!GetConfig(CONFIG_DISABLEFOG))
if (!GetConfig(CONFIG_DISABLEFOG) && bp.changes)
PixelShaderManager::SetFogRangeAdjustChanged();
break;
case BPMEM_FOGPARAM0:
case BPMEM_FOGBMAGNITUDE:
case BPMEM_FOGBEXPONENT:
case BPMEM_FOGPARAM3:
if (!GetConfig(CONFIG_DISABLEFOG))
if (!GetConfig(CONFIG_DISABLEFOG) && bp.changes)
PixelShaderManager::SetFogParamChanged();
break;
case BPMEM_FOGCOLOR: // Fog Color
if (!GetConfig(CONFIG_DISABLEFOG))
if (!GetConfig(CONFIG_DISABLEFOG) && bp.changes)
PixelShaderManager::SetFogColorChanged();
break;
case BPMEM_ALPHACOMPARE: // Compare Alpha Values
PRIM_LOG("alphacmp: ref0=%d, ref1=%d, comp0=%d, comp1=%d, logic=%d", bpmem.alpha_test.ref0,
bpmem.alpha_test.ref1, bpmem.alpha_test.comp0, bpmem.alpha_test.comp1, bpmem.alpha_test.logic);
PixelShaderManager::SetAlpha(bpmem.alpha_test);
g_renderer->SetColorMask();
if(bp.changes & 0xFFFF)
PixelShaderManager::SetAlpha();
if(bp.changes)
g_renderer->SetColorMask();
break;
case BPMEM_BIAS: // BIAS
PRIM_LOG("ztex bias=0x%x", bpmem.ztex1.bias);
PixelShaderManager::SetZTextureBias(bpmem.ztex1.bias);
if(bp.changes)
PixelShaderManager::SetZTextureBias();
break;
case BPMEM_ZTEX2: // Z Texture type
{
@@ -573,7 +581,8 @@ void BPWritten(const BPCmd& bp)
case BPMEM_SU_TSIZE+12:
case BPMEM_SU_SSIZE+14:
case BPMEM_SU_TSIZE+14:
PixelShaderManager::SetTexCoordChanged((bp.address - BPMEM_SU_SSIZE) >> 1);
if(bp.changes)
PixelShaderManager::SetTexCoordChanged((bp.address - BPMEM_SU_SSIZE) >> 1);
break;
// ------------------------
// BPMEM_TX_SETMODE0 - (Texture lookup and filtering mode) LOD/BIAS Clamp, MaxAnsio, LODBIAS, DiagLoad, Min Filter, Mag Filter, Wrap T, S
@@ -627,9 +636,9 @@ void BPWritten(const BPCmd& bp)
// don't compare with changes!
int num = (bp.address >> 1) & 0x3;
if ((bp.address & 1) == 0)
PixelShaderManager::SetColorChanged(bpmem.tevregs[num].low.type, num, false);
PixelShaderManager::SetColorChanged(bpmem.tevregs[num].low.type, num);
else
PixelShaderManager::SetColorChanged(bpmem.tevregs[num].high.type, num, true);
PixelShaderManager::SetColorChanged(bpmem.tevregs[num].high.type, num);
}
break;
@@ -0,0 +1,42 @@
// Copyright 2013 Dolphin Emulator Project
// Licensed under GPLv2
// Refer to the license.txt file included.
#ifndef _CONSTANTMANAGER_H
#define _CONSTANTMANAGER_H
// all constant buffer attributes must be 16 bytes aligned, so this are the only allowed components:
typedef float float4[4];
typedef u32 uint4[4];
typedef s32 int4[4];
struct PixelShaderConstants
{
float4 colors[4];
float4 kcolors[4];
float4 alpha;
float4 texdims[8];
float4 zbias[2];
float4 indtexscale[2];
float4 indtexmtx[6];
float4 fog[3];
// For pixel lighting
float4 plights[40];
float4 pmaterials[4];
};
struct VertexShaderConstants
{
float4 posnormalmatrix[6];
float4 projection[4];
float4 materials[4];
float4 lights[40];
float4 texmatrices[24];
float4 transformmatrices[64];
float4 normalmatrices[32];
float4 posttransformmatrices[64];
float4 depthparams;
};
#endif
+1 -13
View File
@@ -21,6 +21,7 @@
#define I_PLIGHTS "cPLights"
#define I_PMATERIALS "cPmtrl"
// TODO: get rid of them as they aren't used
#define C_COLORMATRIX 0 // 0
#define C_COLORS 0 // 0
#define C_KCOLORS (C_COLORS + 4) // 4
@@ -43,19 +44,6 @@ enum DSTALPHA_MODE
DSTALPHA_DUAL_SOURCE_BLEND // Use dual-source blending
};
// Annoying sure, can be removed once we get up to GLSL ~1.3
const s_svar PSVar_Loc[] = { {I_COLORS, C_COLORS, 4 },
{I_KCOLORS, C_KCOLORS, 4 },
{I_ALPHA, C_ALPHA, 1 },
{I_TEXDIMS, C_TEXDIMS, 8 },
{I_ZBIAS , C_ZBIAS, 2 },
{I_INDTEXSCALE , C_INDTEXSCALE, 2 },
{I_INDTEXMTX, C_INDTEXMTX, 6 },
{I_FOG, C_FOG, 3 },
{I_PLIGHTS, C_PLIGHTS, 40 },
{I_PMATERIALS, C_PMATERIALS, 4 },
};
#pragma pack(1)
struct pixel_shader_uid_data
{
File diff suppressed because it is too large Load Diff
@@ -8,13 +8,15 @@
#include "BPMemory.h"
#include "XFMemory.h"
#include "PixelShaderGen.h"
#include "ConstantManager.h"
class PointerWrap;
// The non-API dependent parts.
class PixelShaderManager
{
static void SetPSTextureDims(int texid);
public:
static void Init();
static void Dirty();
@@ -24,23 +26,25 @@ public:
static void SetConstants(u32 components); // sets pixel shader constants
// constant management, should be called after memory is committed
static void SetColorChanged(int type, int index, bool high);
static void SetAlpha(const AlphaTest& alpha);
static void SetDestAlpha(const ConstantAlpha& alpha);
static void SetColorChanged(int type, int index);
static void SetAlpha();
static void SetDestAlpha();
static void SetTexDims(int texmapid, u32 width, u32 height, u32 wraps, u32 wrapt);
static void SetZTextureBias(u32 bias);
static void SetZTextureBias();
static void SetViewportChanged();
static void SetIndMatrixChanged(int matrixidx);
static void SetTevKSelChanged(int id);
static void SetZTextureTypeChanged();
static void SetIndTexScaleChanged(u8 stagemask);
static void SetIndTexScaleChanged(bool high);
static void SetTexCoordChanged(u8 texmapid);
static void SetFogColorChanged();
static void SetFogParamChanged();
static void SetFogRangeAdjustChanged();
static void SetColorMatrix(const float* pmatrix);
static void InvalidateXFRange(int start, int end);
static void SetMaterialColorChanged(int index);
static void SetMaterialColorChanged(int index, u32 color);
static PixelShaderConstants constants;
static bool dirty;
};
-11
View File
@@ -115,17 +115,6 @@ public:
static unsigned int GetPrevPixelFormat() { return prev_efb_format; }
static void StorePixelFormat(unsigned int new_format) { prev_efb_format = new_format; }
// TODO: doesn't belong here
virtual void SetPSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4) = 0;
virtual void SetPSConstant4fv(unsigned int const_number, const float *f) = 0;
virtual void SetMultiPSConstant4fv(unsigned int const_number, unsigned int count, const float *f) = 0;
// TODO: doesn't belong here
virtual void SetVSConstant4f(unsigned int const_number, float f1, float f2, float f3, float f4) = 0;
virtual void SetVSConstant4fv(unsigned int const_number, const float *f) = 0;
virtual void SetMultiVSConstant3fv(unsigned int const_number, unsigned int count, const float *f) = 0;
virtual void SetMultiVSConstant4fv(unsigned int const_number, unsigned int count, const float *f) = 0;
protected:
static void CalculateTargetScale(int x, int y, int &scaledX, int &scaledY);
@@ -867,10 +867,4 @@ const char *GenerateEncodingShader(u32 format,API_TYPE ApiType)
return text;
}
void SetShaderParameters(float width, float height, float offsetX, float offsetY, float widthStride, float heightStride,float buffW,float buffH)
{
g_renderer->SetPSConstant4f(C_COLORMATRIX, widthStride, heightStride, buffW, buffH);
g_renderer->SetPSConstant4f(C_COLORMATRIX + 1, width, (height - 1), offsetX, offsetY);
}
} // namespace
@@ -15,8 +15,6 @@ u16 GetEncodedSampleCount(u32 format);
const char *GenerateEncodingShader(u32 format, API_TYPE ApiType = API_OPENGL);
void SetShaderParameters(float width, float height, float offsetX, float offsetY, float widthStride, float heightStride,float buffW = 0.0f,float buffH = 0.0f);
}
#endif // _TEXTURECONVERSIONSHADER_H_
+1 -11
View File
@@ -42,6 +42,7 @@
#define I_POSTTRANSFORMMATRICES "cpostmtx"
#define I_DEPTHPARAMS "cDepth" // farZ, zRange, scaled viewport width, scaled viewport height
//TODO: get rid of them, they aren't used at all
#define C_POSNORMALMATRIX 0
#define C_PROJECTION (C_POSNORMALMATRIX + 6)
#define C_MATERIALS (C_PROJECTION + 4)
@@ -53,17 +54,6 @@
#define C_DEPTHPARAMS (C_POSTTRANSFORMMATRICES + 64)
#define C_VENVCONST_END (C_DEPTHPARAMS + 1)
const s_svar VSVar_Loc[] = { {I_POSNORMALMATRIX, C_POSNORMALMATRIX, 6 },
{I_PROJECTION , C_PROJECTION, 4 },
{I_MATERIALS, C_MATERIALS, 4 },
{I_LIGHTS, C_LIGHTS, 40 },
{I_TEXMATRICES, C_TEXMATRICES, 24 },
{I_TRANSFORMMATRICES , C_TRANSFORMMATRICES, 64 },
{I_NORMALMATRICES , C_NORMALMATRICES, 32 },
{I_POSTTRANSFORMMATRICES, C_POSTTRANSFORMMATRICES, 64 },
{I_DEPTHPARAMS, C_DEPTHPARAMS, 1 },
};
#pragma pack(1)
struct vertex_shader_uid_data

Some files were not shown because too many files have changed in this diff Show More