zzogl-pg:

* properry separate both GLSL implementation
* glsl4: Use a define for logz instead of extra math computation. Much more easier to understand


git-svn-id: http://pcsx2.googlecode.com/svn/trunk@5364 96395faa-99c1-11dd-bbfe-3dabce05a288

plugins/zzogl-pg/opengl/CMakeLists.txt

@@ -47,7 +47,7 @@ endif(CMAKE_BUILD_TYPE STREQUAL Release)
 
 # Select the shader API 
 if(GLSL_API)
-    add_definitions(-DGLSL_API -DGLSL4_API -DOGL4_LOG)
+    add_definitions(-DGLSL4_API -DOGL4_LOG)
     #add_definitions(-DGLSL_API)
 else(GLSL_API)
 	add_definitions(-DNVIDIA_CG_API)

plugins/zzogl-pg/opengl/ZZDepthTargets.cpp

@@ -226,7 +226,7 @@ void CDepthTarget::Update(int context, CRenderTarget* prndr)
 		vdepth.z = vdepth.w = 0;
 	}
 
-#ifdef GLSL_API
+#if defined(GLSL_API) || defined(GLSL4_API)
 	assert(ppsBitBltDepth.sBitBltZ != -1);
 #else
 	assert(ppsBitBltDepth.sBitBltZ != 0);

plugins/zzogl-pg/opengl/ZZoglCreate.cpp

@@ -475,10 +475,10 @@ bool ZZCreate(int _width, int _height)
 	g_fiGPU_TEXWIDTH = 1.0f / GPU_TEXWIDTH;
 
 	// FIXME: not clean maybe re integrate the function in shader files --greg
-#ifndef GLSL_API
+#if defined(GLSL_API) || defined(GLSL4_API)
-	if (!CreateOpenShadersFile()) return false;
-#else
 	if (!ZZshCreateOpenShadersFile()) return false;
+#else
+	if (!CreateOpenShadersFile()) return false;
 #endif
 
 	GL_REPORT_ERROR();

plugins/zzogl-pg/opengl/ZZoglShaders.cpp

@@ -18,7 +18,7 @@
  */
 
 // By default enable nvidia cg api
-#if !defined(GLSL_API) && !defined(NVIDIA_CG_API)
+#if !defined(GLSL_API) && !defined(NVIDIA_CG_API) && !defined(GLSL4_API)
 #define NVIDIA_CG_API
 #endif
 

plugins/zzogl-pg/opengl/ZZoglShaders.h

@@ -35,7 +35,7 @@
 #include "GS.h"
 
 // By default enable nvidia cg api
-#if !defined(GLSL_API) && !defined(NVIDIA_CG_API)
+#if !defined(GLSL_API) && !defined(NVIDIA_CG_API) && !defined(GLSL4_API)
 #define NVIDIA_CG_API
 #endif
 // --------------------------- API abstraction level --------------------------------
@@ -55,7 +55,6 @@
 #define sZero			0			// Zero program
 
 #define SAFE_RELEASE_PROG(x) 	{ if( (x) != NULL ) { cgDestroyProgram(x); x = NULL; } }
-inline bool ZZshActiveParameter(ZZshParameter param) {return (param !=NULL); }
 
 #endif					// end NVIDIA cg-toolkit API
 
@@ -68,7 +67,10 @@ inline bool ZZshActiveParameter(ZZshParameter param) {return (param !=NULL); }
 // Set it to 0 to diable context usage, 1 -- to enable. FFX-1 have a strange issue with ClampExt.
 #define NOCONTEXT		0
 
-#ifdef GLSL_API
+#if defined(GLSL_API)
+
+#define MAX_ACTIVE_UNIFORMS 600
+#define MAX_ACTIVE_SHADERS 400
 
 enum ZZshPARAMTYPE {
 	ZZ_UNDEFINED,
@@ -89,6 +91,13 @@ typedef struct {
 	bool		Constant;	// Uniform could be constants, does not change at program flow
 	bool 		Settled;	// Check if Uniform value was set.
 } ZZshParamInfo;
+const ZZshParamInfo  qZero = {ShName:"", type:ZZ_UNDEFINED, fvalue:{0}, sampler: -1, texid: 0, Constant: false, Settled: false};
+
+#define SAFE_RELEASE_PROG(x) 	{ /*don't know what to do*/ }
+
+#endif
+
+#if defined(GLSL_API) || defined(GLSL4_API)
 
 typedef struct {
 	void*	 	link;
@@ -103,21 +112,13 @@ typedef struct {
 #define ZZshError		int
 #define ZZshIndex		GLuint
 
-const ZZshParamInfo  qZero = {ShName:"", type:ZZ_UNDEFINED, fvalue:{0}, sampler: -1, texid: 0, Constant: false, Settled: false};
-
 #define pZero			0
 
 const ZZshShaderLink sZero = {link: NULL, isFragment: false};
-
-inline bool ZZshActiveParameter(ZZshParameter param) {return (param > -1); }
-#ifndef GLSL4_API
-#define SAFE_RELEASE_PROG(x) 	{ /*don't know what to do*/ }
 #endif
 
 // ---------------------------
 
-#endif
-
 extern float4 g_vdepth;
 extern float4 vlogz;
 
@@ -160,15 +161,13 @@ struct ConstantUniform {
 			float g_fMult[4];
 			// VS
 			float g_fZ[4];
-			float g_fZMin[4];
-			float g_fZNorm[4];
 			// PS
 			float g_fExactColor[4];
 		};
-		float linear[8*4];
+		float linear[6*4];
 	};
 	void SettleFloat(uint indice, const float* v) {
-		assert(indice + 3 < 8*4);
+		assert(indice + 3 < 6*4);
 		linear[indice+0] = v[0];
 		linear[indice+1] = v[1];
 		linear[indice+2] = v[2];
@@ -223,8 +222,6 @@ struct VertexUniform {
 
 
 
-//const static char* g_pPsTexWrap[] = { "-DREPEAT", "-DCLAMP", "-DREGION_REPEAT", NULL };
-
 enum ZZshShaderType {ZZ_SH_ZERO, ZZ_SH_REGULAR, ZZ_SH_REGULAR_FOG, ZZ_SH_TEXTURE, ZZ_SH_TEXTURE_FOG, ZZ_SH_CRTC, ZZ_SH_NONE};
 // We have "compatible" shaders, as RegularFogVS and RegularFogPS. if don't need to wory about incompatible shaders
 // It used only in GLSL mode. 
@@ -235,9 +232,6 @@ extern int 		g_nPixelShaderVer;
 extern ZZshShaderLink 	pvs[16], g_vsprog, g_psprog;
 extern ZZshParameter 	g_vparamPosXY[2], g_fparamFogColor;
 
-#define MAX_ACTIVE_UNIFORMS 600
-#define MAX_ACTIVE_SHADERS 400
-
 #ifndef GLSL4_API
 struct FRAGMENTSHADER
 {
@@ -511,8 +505,6 @@ struct COMMONSHADER
 		g_fc0            = (ZZshParameter)offsetof(struct ConstantUniform, g_fc0) /4;
 		g_fMult          = (ZZshParameter)offsetof(struct ConstantUniform, g_fMult) /4;
 		g_fZ             = (ZZshParameter)offsetof(struct ConstantUniform, g_fZ) /4;
-		g_fZMin          = (ZZshParameter)offsetof(struct ConstantUniform, g_fZMin) /4;
-		g_fZNorm         = (ZZshParameter)offsetof(struct ConstantUniform, g_fZNorm) /4;
 		g_fExactColor    = (ZZshParameter)offsetof(struct ConstantUniform, g_fExactColor) /4;
 
 		// Setup the constant buffer
@@ -520,18 +512,11 @@ struct COMMONSHADER
 		// Set Z-test, log or no log;
 		if (conf.settings().no_logz) {
 			g_vdepth = float4( 255.0 /256.0f,  255.0/65536.0f, 255.0f/(65535.0f*256.0f), 1.0f/(65536.0f*65536.0f));
-			vlogz = float4( 1.0f, 0.0f, 0.0f, 0.0f);
 		}
 		else {
 			g_vdepth = float4( 256.0f*65536.0f, 65536.0f, 256.0f, 65536.0f*65536.0f);
-			vlogz = float4( 0.0f, 1.0f, 0.0f, 0.0f);
 		}
 		uniform_buffer_constant.SettleFloat(g_fZ, g_vdepth );
-		uniform_buffer_constant.SettleFloat(g_fZMin, vlogz );
-
-		const float g_filog32 = 0.999f / (32.0f * logf(2.0f));
-		float4 vnorm = float4(g_filog32, 0, 0,0);
-		uniform_buffer_constant.SettleFloat(g_fZNorm, vnorm);
 
 		uniform_buffer_constant.SettleFloat(g_fBilinear, float4(-0.2f, -0.65f, 0.9f, 1.0f / 32767.0f ) );
 		uniform_buffer_constant.SettleFloat(g_fZBias, float4(1.0f/256.0f, 1.0004f, 1, 0.5f) );
@@ -542,7 +527,7 @@ struct COMMONSHADER
 	}
 
 	ZZshParameter g_fparamFogColor, g_vparamPosXY;
-	ZZshParameter g_fBilinear, g_fZBias, g_fc0, g_fMult, g_fZ, g_fZMin, g_fZNorm, g_fExactColor;
+	ZZshParameter g_fBilinear, g_fZBias, g_fc0, g_fMult, g_fZ, g_fExactColor;
 	uint context;
 
 	GlobalUniform	uniform_buffer[ZZSH_CTX_ALL];

plugins/zzogl-pg/opengl/ZZoglShadersGLSL4.cpp

@@ -48,7 +48,6 @@
 #include "Util.h"
 #include "ZZoglShaders.h"
 #include "zpipe.h"
-#include <math.h>
 #include <map>
 #include <fcntl.h>			// this for open(). Maybe linux-specific
 #include <sys/mman.h>			// and this for mmap
@@ -75,11 +74,7 @@
 // #define ENABLE_MARKER // Fire some marker for opengl Debugger (apitrace, gdebugger)
 //------------------ Constants
 
-// Used in a logarithmic Z-test, as (1-o(1))/log(MAX_U32).
-const float g_filog32 = 0.999f / (32.0f * logf(2.0f));
-
 const static char* g_pTexTypes[] = { "32", "tex32", "clut32", "tex32to16", "tex16to8h" };
-const static char* g_pShaders[4] = { "full", "reduced", "accurate", "accurate-reduced" };
 const static char* g_pPsTexWrap[] = { "#define REPEAT 1\n", "#define CLAMP 1\n", "#define REGION_REPEAT 1\n", "\n" };
 const int GLSL_VERSION = 330;
 
@@ -146,16 +141,6 @@ bool ZZshCheckProfilesSupport() {
 	return true;
 }
 
-// Error handler. Setup in ZZogl_Create once.
-void HandleCgError(ZZshContext ctx, ZZshError err, void* appdata)
-{/*
-	ZZLog::Error_Log("%s->%s: %s", ShaderCallerName, ShaderHandleName, cgGetErrorString(err));
-	const char* listing = cgGetLastListing(g_cgcontext);
-	if (listing != NULL)
-		ZZLog::Debug_Log("	last listing: %s", listing);
-*/
-}
-
 bool ZZshStartUsingShaders() {
 
 	ZZLog::Error_Log("Creating effects.");
@@ -173,7 +158,6 @@ bool ZZshStartUsingShaders() {
 	memset(&temp, 0, sizeof(temp));
 	temp.wms = 3; temp.wmt = 3;
 
-	g_nPixelShaderVer = 0;//SHADER_ACCURATE;
 	// test
 	bool bFailed;
 	FRAGMENTSHADER* pfrag = ZZshLoadShadeEffect(0, 1, 1, 1, 1, temp, 0, &bFailed);
@@ -185,7 +169,6 @@ bool ZZshStartUsingShaders() {
 	ZZLog::Error_Log("Creating extra effects.");
 	B_G(ZZshLoadExtraEffects(), return false);
 
-	ZZLog::Error_Log("Using %s shaders.", g_pShaders[g_nPixelShaderVer]);
 
 	return true;
 }
@@ -469,6 +452,9 @@ std::string BuildGlslMacro(bool writedepth, int texwrap = 3, bool testaem = fals
 	if (exactcolor) header += "#define EXACT_COLOR 1\n";
 	header += format("%s", g_pPsTexWrap[texwrap]);
 	//const char* AddAccurate  = (ps & SHADER_ACCURATE)?"#define ACCURATE_DECOMPRESSION 1\n":"";
+	if (conf.settings().no_logz) {
+		header += "#define NO_LOGZ 1\n";
+	}
 
 	return header;
 }
@@ -589,7 +575,7 @@ FRAGMENTSHADER* ZZshLoadShadeEffect(int type, int texfilter, int fog, int testae
 	else
 		texwrap = TEXWRAP_REPEAT_CLAMP;
 
-	int index = GET_SHADER_INDEX(type, texfilter, texwrap, fog, s_bWriteDepth, testaem, exactcolor, context, 0);
+	int index = GET_SHADER_INDEX(type, texfilter, texwrap, fog, s_bWriteDepth, testaem, exactcolor, 0, 0);
 
 	if( pbFailed != NULL ) *pbFailed = false;
 

plugins/zzogl-pg/opengl/ps2hw_gl4.glsl

@@ -31,6 +31,7 @@
 //#define WRITE_DEPTH // set if depth is also written in a MRT
 //#define ACCURATE_DECOMPRESSION // set for less capable hardware ATI Radeon 9000 series
 //#define EXACT_COLOR	// make sure the output color is clamped to 1/255 boundaries (for alpha testing)
+//#define NO_LOGZ       // disable logz
 
 #define PERSPECTIVE_CORRECT_TEX
 
@@ -58,7 +59,7 @@ struct vertex
 {
     vec4 color;
     TEX_DECL tex;
-    vec4 z;
+    vec4 Z;
     float fog;
 };
 
@@ -137,8 +138,6 @@ layout(std140, binding = 0) uniform constant_buffer
 	float4 g_fMult;
 	// Vertex
 	float4 g_fZ;				// transforms d3dcolor z into float z
-	float4 g_fZMin;
-	float4 g_fZNorm;
 	// Pixel
 	half4 g_fExactColor;
 };
@@ -567,7 +566,7 @@ half4 ps2FinalColor(half4 col)
 #ifdef WRITE_DEPTH
 void write_depth_target()
 {
-    FragData1 = PSin.z; 
+    FragData1 = PSin.Z; 
 }
 #else
 void write_depth_target() { }
@@ -665,7 +664,11 @@ void BitBltDepthPS() {
 	vec4 data;
 	data = texture(g_sMemory, ps2memcoord(PSin.tex.xy));
 	FragData0 = data + g_fZBias.y;
-	gl_FragDepth   = (log(g_fc0.y + dot(data, g_fBitBltZ)) * g_fOneColor.w) * g_fZMin.y + dot(data, g_fBitBltZ) * g_fZMin.x ;
+#ifdef NO_LOGZ
+	gl_FragDepth = dot(data, g_fBitBltZ);
+#else
+	gl_FragDepth = log(g_fc0.y + dot(data, g_fBitBltZ)) * g_fOneColor.w;
+#endif
 }
 
 void BitBltDepthMRTPS() {
@@ -673,7 +676,11 @@ void BitBltDepthMRTPS() {
 	data = texture(g_sMemory, ps2memcoord(PSin.tex.xy));
 	FragData0 = data + g_fZBias.y;
 	FragData1.x = g_fc0.x;
-	gl_FragDepth = (log(g_fc0.y + dot(data, g_fBitBltZ)) * g_fOneColor.w) * g_fZMin.y + dot(data, g_fBitBltZ) * g_fZMin.x ;
+#ifdef NO_LOGZ
+	gl_FragDepth = dot(data, g_fBitBltZ);
+#else
+	gl_FragDepth = log(g_fc0.y + dot(data, g_fBitBltZ)) * g_fOneColor.w;
+#endif
 }
 
 // static const float BlurKernel[9] = {
@@ -694,7 +701,7 @@ half4 BilinearFloat16(float2 tex0)
 }
 
 void CRTCTargInterPS() {
-	float finter = texture(g_sInterlace, PSin.z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
+	float finter = texture(g_sInterlace, PSin.Z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
 	float4 c = BilinearFloat16(PSin.tex.xy);
 	c.w = ( g_fc0.w*c.w * g_fOneColor.x + g_fOneColor.y ) * finter;
 	FragData0 = c;
@@ -709,7 +716,7 @@ void CRTCTargPS() {
 }
 
 void CRTCInterPS() {
-	float finter = texture(g_sInterlace, PSin.z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
+	float finter = texture(g_sInterlace, PSin.Z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
 	float2 filtcoord = trunc(PSin.tex.xy) * g_fInvTexDims.xy + g_fInvTexDims.zw;
 	half4 c = BilinearBitBlt(filtcoord);
 	c.w = (c.w * g_fOneColor.x + g_fOneColor.y)*finter;
@@ -718,7 +725,7 @@ void CRTCInterPS() {
 
 // simpler
 void CRTCInterPS_Nearest() {
-	float finter = texture(g_sInterlace, PSin.z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
+	float finter = texture(g_sInterlace, PSin.Z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
 	half4 c = texture(g_sMemory, ps2memcoord(PSin.tex.xy));
 	c.w = (c.w * g_fOneColor.x + g_fOneColor.y)*finter;
 	FragData0 = c;
@@ -739,7 +746,7 @@ void CRTCPS_Nearest() {
 }
 
 void CRTC24InterPS() {
-	float finter = texture(g_sInterlace, PSin.z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
+	float finter = texture(g_sInterlace, PSin.Z.yy).x * g_fOneColor.z + g_fOneColor.w + g_fc0.w;
 	float2 filtcoord = trunc(PSin.tex.xy) * g_fInvTexDims.xy + g_fInvTexDims.zw;
 
 	half4 c = texture(g_sMemory, ps2memcoord(filtcoord));
@@ -832,8 +839,8 @@ void SetTex() {
 
 void SetZ() {
 #ifdef WRITE_DEPTH
-	VSout.z = SecondaryColor * g_fZBias.x + g_fZBias.y;
+	VSout.Z = SecondaryColor * g_fZBias.x + g_fZBias.y;
-    VSout.z.w = g_fc0.y;
+    VSout.Z.w = 1.0f;
 #endif
 }
 
@@ -849,8 +856,14 @@ void SetPosition() {
     // position.z = log(g_fc0.y + dot(g_fZ, SecondaryColor.zyxw)) * g_fZNorm.x
     // position.z = log(1 + Z_INT) * 0.999f / (32 * log(2.0)) = log2(1 + Z_INT) * 0.999f / 32
     // log2(...) will range from 0 to 32
-	position.z = (log(g_fc0.y + dot(g_fZ, SecondaryColor.zyxw)) * g_fZNorm.x + g_fZNorm.y) * g_fZMin.y + dot(g_fZ, SecondaryColor.zyxw) * g_fZMin.x ;
+
-	position.w = g_fc0.y;
+	// position.z = (log(g_fc0.y + dot(g_fZ, SecondaryColor.zyxw)) * g_fZNorm.x + g_fZNorm.y) * g_fZMin.y + dot(g_fZ, SecondaryColor.zyxw) * g_fZMin.x ;
+#ifdef NO_LOGZ
+	position.z = dot(g_fZ, SecondaryColor.zyxw);
+#else
+	position.z = log2(1.0f + dot(g_fZ, SecondaryColor.zyxw)) * 0.999f/32.0f;
+#endif
+	position.w = 1.0f;
 
     gl_Position = position;
 }
@@ -896,7 +909,7 @@ void BitBltVS() {
 	gl_Position = position;
 
 	VSout.tex.xy = TexCoord.xy * g_fBitBltTex.xy + g_fBitBltTex.zw;
-	VSout.z.xy = position.xy * g_fBitBltTrans.xy + g_fBitBltTrans.zw;
+	VSout.Z.xy = position.xy * g_fBitBltTrans.xy + g_fBitBltTrans.zw;
 }
 
 #endif