Skipped the ZCompLoc pass if the result can be determined at compile time. Brings back the speed lost by r146b02615c07.

This commit is contained in:
skidau
2012-08-06 09:29:01 +10:00
parent 86a1899b9e
commit 0efd4e5c29
6 changed files with 43 additions and 43 deletions

View File

@ -26,8 +26,7 @@
#include "BPMemory.h" #include "BPMemory.h"
#include "VideoConfig.h" #include "VideoConfig.h"
#include "NativeVertexFormat.h" #include "NativeVertexFormat.h"
#include "PixelShaderManager.h"
static ALPHA_PRETEST_RESULT AlphaPreTest();
static void StageHash(int stage, u32* out) static void StageHash(int stage, u32* out)
{ {
@ -117,7 +116,7 @@ void GetPixelShaderId(PIXELSHADERUID *uid, PSGRENDER_MODE PSGRenderMode, u32 com
uid->values[0] |= MustWriteToDepth << 11; // 1 uid->values[0] |= MustWriteToDepth << 11; // 1
uid->values[0] |= PixelLightingEnabled << 12; // 1 uid->values[0] |= PixelLightingEnabled << 12; // 1
if (!PixelLightingEnabled) uid->values[0] |= xfregs.numTexGen.numTexGens << 13; // 4 if (!PixelLightingEnabled) uid->values[0] |= xfregs.numTexGen.numTexGens << 13; // 4
ALPHA_PRETEST_RESULT alphaPreTest = AlphaPreTest(); ALPHA_PRETEST_RESULT alphaPreTest = PixelShaderManager::AlphaPreTest();
uid->values[0] |= ((u32)alphaPreTest) << 17; // 2 uid->values[0] |= ((u32)alphaPreTest) << 17; // 2
uid->values[0] |= ZCompLocEnabled << 18; // 2 uid->values[0] |= ZCompLocEnabled << 18; // 2
if (((alphaPreTest == ALPHAPT_ALWAYSFAIL) || (alphaPreTest == ALPHAPT_ALWAYSPASS && PSGRenderMode == PSGRENDER_DSTALPHA_ALPHA_PASS)) && !DepthTextureEnable) if (((alphaPreTest == ALPHAPT_ALWAYSFAIL) || (alphaPreTest == ALPHAPT_ALWAYSPASS && PSGRenderMode == PSGRENDER_DSTALPHA_ALPHA_PASS)) && !DepthTextureEnable)
@ -636,7 +635,7 @@ const char *GeneratePixelShaderCode(PSGRENDER_MODE PSGRenderMode, API_TYPE ApiTy
WRITE(p, "float zCoord = " I_ZBIAS "[1].x + (uv2.w / uv3.w) * " I_ZBIAS "[1].y;\n"); WRITE(p, "float zCoord = " I_ZBIAS "[1].x + (uv2.w / uv3.w) * " I_ZBIAS "[1].y;\n");
} }
char* pmainstart = p; char* pmainstart = p;
if(PSGRenderMode == PSGRENDER_ZCOMPLOCK && !DepthTextureEnable) if(PSGRenderMode == PSGRENDER_ZCOMPLOC && !DepthTextureEnable)
{ {
// Within ZCompLoc pass, make this an empty function // Within ZCompLoc pass, make this an empty function
WRITE(p, "ocol0 = 0;\n"); WRITE(p, "ocol0 = 0;\n");
@ -646,7 +645,7 @@ const char *GeneratePixelShaderCode(PSGRENDER_MODE PSGRenderMode, API_TYPE ApiTy
WRITE(p, "}\n"); WRITE(p, "}\n");
return text; return text;
} }
ALPHA_PRETEST_RESULT Pretest = AlphaPreTest(); ALPHA_PRETEST_RESULT Pretest = PixelShaderManager::AlphaPreTest();
// Test if we can predict the alpha test result or if we are in depth only mode // Test if we can predict the alpha test result or if we are in depth only mode
if((Pretest != ALPHAPT_UNDEFINED) && !DepthTextureEnable) if((Pretest != ALPHAPT_UNDEFINED) && !DepthTextureEnable)
{ {
@ -1169,39 +1168,6 @@ static const char *tevAlphaFunclogicTable[] =
" != ", // xor " != ", // xor
" == " // xnor " == " // xnor
}; };
static ALPHA_PRETEST_RESULT AlphaPreTest()
{
u32 op = bpmem.alphaFunc.logic;
u32 comp[2] = {bpmem.alphaFunc.comp0, bpmem.alphaFunc.comp1};
// First kill all the simple cases
switch(op)
{
case 0: // AND
if (comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_ALWAYS) return ALPHAPT_ALWAYSPASS;
if (comp[0] == ALPHACMP_NEVER || comp[1] == ALPHACMP_NEVER) return ALPHAPT_ALWAYSFAIL;
break;
case 1: // OR
if (comp[0] == ALPHACMP_ALWAYS || comp[1] == ALPHACMP_ALWAYS) return ALPHAPT_ALWAYSPASS;
if (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_NEVER)return ALPHAPT_ALWAYSFAIL;
break;
case 2: // XOR
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_NEVER) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_ALWAYS))
return ALPHAPT_ALWAYSPASS;
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_ALWAYS) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_NEVER))
return ALPHAPT_ALWAYSFAIL;
break;
case 3: // XNOR
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_NEVER) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_ALWAYS))
return ALPHAPT_ALWAYSFAIL;
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_ALWAYS) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_NEVER))
return ALPHAPT_ALWAYSPASS;
break;
default: PanicAlert("bad logic for alpha test? %08x", op);
}
return ALPHAPT_UNDEFINED;
}
static void WriteAlphaTest(char *&p, API_TYPE ApiType,PSGRENDER_MODE PSGRenderMode) static void WriteAlphaTest(char *&p, API_TYPE ApiType,PSGRENDER_MODE PSGRenderMode)
{ {

View File

@ -111,7 +111,7 @@ enum PSGRENDER_MODE
PSGRENDER_NORMAL, // Render normally, without destination alpha PSGRENDER_NORMAL, // Render normally, without destination alpha
PSGRENDER_DSTALPHA_ALPHA_PASS, // Render normally first, then render again for alpha PSGRENDER_DSTALPHA_ALPHA_PASS, // Render normally first, then render again for alpha
PSGRENDER_DSTALPHA_DUAL_SOURCE_BLEND, // Use dual-source blending PSGRENDER_DSTALPHA_DUAL_SOURCE_BLEND, // Use dual-source blending
PSGRENDER_ZCOMPLOCK //Render to Depth Channel only with no depth dextures enabled PSGRENDER_ZCOMPLOC //Render to Depth Channel only with no depth dextures enabled
}; };
enum ALPHA_PRETEST_RESULT enum ALPHA_PRETEST_RESULT

View File

@ -466,6 +466,39 @@ void PixelShaderManager::SetMaterialColorChanged(int index)
nMaterialsChanged |= (1 << index); nMaterialsChanged |= (1 << index);
} }
ALPHA_PRETEST_RESULT PixelShaderManager::AlphaPreTest()
{
u32 op = bpmem.alphaFunc.logic;
u32 comp[2] = {bpmem.alphaFunc.comp0, bpmem.alphaFunc.comp1};
// First kill all the simple cases
switch(op)
{
case 0: // AND
if (comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_ALWAYS) return ALPHAPT_ALWAYSPASS;
if (comp[0] == ALPHACMP_NEVER || comp[1] == ALPHACMP_NEVER) return ALPHAPT_ALWAYSFAIL;
break;
case 1: // OR
if (comp[0] == ALPHACMP_ALWAYS || comp[1] == ALPHACMP_ALWAYS) return ALPHAPT_ALWAYSPASS;
if (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_NEVER)return ALPHAPT_ALWAYSFAIL;
break;
case 2: // XOR
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_NEVER) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_ALWAYS))
return ALPHAPT_ALWAYSPASS;
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_ALWAYS) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_NEVER))
return ALPHAPT_ALWAYSFAIL;
break;
case 3: // XNOR
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_NEVER) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_ALWAYS))
return ALPHAPT_ALWAYSFAIL;
if ((comp[0] == ALPHACMP_ALWAYS && comp[1] == ALPHACMP_ALWAYS) || (comp[0] == ALPHACMP_NEVER && comp[1] == ALPHACMP_NEVER))
return ALPHAPT_ALWAYSPASS;
break;
default: PanicAlert("bad logic for alpha test? %08x", op);
}
return ALPHAPT_UNDEFINED;
}
void PixelShaderManager::DoState(PointerWrap &p) void PixelShaderManager::DoState(PointerWrap &p)
{ {
p.Do(lastRGBAfull); p.Do(lastRGBAfull);

View File

@ -54,6 +54,7 @@ public:
static void SetColorMatrix(const float* pmatrix); static void SetColorMatrix(const float* pmatrix);
static void InvalidateXFRange(int start, int end); static void InvalidateXFRange(int start, int end);
static void SetMaterialColorChanged(int index); static void SetMaterialColorChanged(int index);
static ALPHA_PRETEST_RESULT AlphaPreTest();
}; };

View File

@ -187,9 +187,9 @@ void VertexManager::vFlush()
} }
bool useZcomploc = bpmem.zcontrol.zcomploc && bpmem.zmode.updateenable; bool useZcomploc = bpmem.zcontrol.zcomploc && bpmem.zmode.updateenable;
if(useZcomploc) if (useZcomploc && PixelShaderManager::AlphaPreTest() == ALPHAPT_UNDEFINED)
{ {
if (!PixelShaderCache::SetShader(PSGRENDER_ZCOMPLOCK, g_nativeVertexFmt->m_components)) if (!PixelShaderCache::SetShader(PSGRENDER_ZCOMPLOC, g_nativeVertexFmt->m_components))
{ {
GFX_DEBUGGER_PAUSE_LOG_AT(NEXT_ERROR,true,{printf("Fail to set pixel shader\n");}); GFX_DEBUGGER_PAUSE_LOG_AT(NEXT_ERROR,true,{printf("Fail to set pixel shader\n");});
goto shader_fail; goto shader_fail;

View File

@ -220,9 +220,9 @@ void VertexManager::vFlush()
g_renderer->RestoreState(); g_renderer->RestoreState();
} }
bool useZcomploc = bpmem.zcontrol.zcomploc && bpmem.zmode.updateenable; bool useZcomploc = bpmem.zcontrol.zcomploc && bpmem.zmode.updateenable;
if (useZcomploc) if (useZcomploc && PixelShaderManager::AlphaPreTest() == ALPHAPT_UNDEFINED)
{ {
ps = PixelShaderCache::SetShader(PSGRENDER_ZCOMPLOCK,g_nativeVertexFmt->m_components); ps = PixelShaderCache::SetShader(PSGRENDER_ZCOMPLOC,g_nativeVertexFmt->m_components);
if (ps) PixelShaderCache::SetCurrentShader(ps->glprogid); if (ps) PixelShaderCache::SetCurrentShader(ps->glprogid);
g_renderer->ApplyState(RSM_Zcomploc); g_renderer->ApplyState(RSM_Zcomploc);