zzogl: rework dump test, to avoid bad mix between u32/u8

glsl4: Replace some define with function (ogl4 support function pointer). Explain how depth is computed in vertex shader


git-svn-id: http://pcsx2.googlecode.com/svn/trunk@5233 96395faa-99c1-11dd-bbfe-3dabce05a288

cmake/BuildParameters.cmake

@@ -124,7 +124,7 @@ set(CMAKE_SHARED_LIBRARY_CXX_FLAGS "")
 #-------------------------------------------------------------------------------
 # Set some default compiler flags
 #-------------------------------------------------------------------------------
-set(DEFAULT_WARNINGS "-Wno-write-strings -Wno-format -Wno-unused-parameter -Wno-unused-value -Wstrict-aliasing")
+set(DEFAULT_WARNINGS "-Wno-write-strings -Wno-format -Wno-unused-parameter -Wno-unused-value -Wstrict-aliasing -Wno-unused-function")
 set(DEFAULT_GCC_FLAG "-m32 -msse -msse2 -march=i686 -pthread ${DEFAULT_WARNINGS}")
 set(DEFAULT_CPP_FLAG "${DEFAULT_GCC_FLAG} -Wno-invalid-offsetof")
 

plugins/GSdx/GSDeviceOGL.cpp

@@ -168,6 +168,7 @@ bool GSDeviceOGL::Create(GSWnd* wnd)
 		const GLubyte* s;
 		s = glGetString(GL_VERSION);
 		if (s == NULL) return false;
+		fprintf(stderr, "Supported Opengl version: %s\n", s);
 
 		GLuint dot = 0;
 		while (s[dot] != '\0' && s[dot] != '.') dot++;

plugins/zzogl-pg-cg/opengl/GSDump.cpp

@@ -58,13 +58,14 @@ void GSDump::Close()
 	}
 }
 
-void GSDump::Transfer(int index, const u8* mem, size_t size)
+void GSDump::Transfer(int index, const u32* mem, size_t size, u32 real_size)
 {
 	if(m_gs && size > 0)
 	{
 		fputc(0, m_gs);
 		fputc(index, m_gs);
 		fwrite(&size, 4, 1, m_gs);
+		fwrite(&real_size, 4, 1, m_gs);
 		fwrite(mem, size, 1, m_gs);
 	}
 }

plugins/zzogl-pg-cg/opengl/GSDump.h

@@ -56,7 +56,7 @@ public:
 	void Open(const string& fn, u32 crc, const freezeData& fd, u8* regs);
 	void Close();
 	void ReadFIFO(u32 size);
-	void Transfer(int index, const u8* mem, size_t size);
+	void Transfer(int index, const u32* mem, size_t size, u32 real_size);
 	void VSync(int field, bool last, u8* regs);
 	operator bool() {return m_gs != NULL;}
 };

plugins/zzogl-pg-cg/opengl/GSmain.cpp

@@ -399,7 +399,7 @@ static bool get_snapshot_filename(char *filename, char* path, const char* extens
 	{
 		snapshotnr++;
 
-		sprintf(filename, "%s/snap%03ld.%s", path, snapshotnr, extension);
+		sprintf(filename, "%s/snap%03d.%s", path, snapshotnr, extension);
 
 		bmpfile = fopen(filename, "rb");
 
@@ -627,6 +627,7 @@ struct Packet
 {
 	u8 type, param;
 	u32 size, addr;
+	u32 real_size;
 	vector<u32> buff;
 };
 
@@ -680,18 +681,20 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 				p->param = (u8)fgetc(fp);
 
 				fread(&p->size, 4, 1, fp);
+				fread(&p->real_size, 4, 1, fp);
 
 				switch(p->param)
 				{
 				case 0:
-					p->buff.resize(0x4000);
+					p->buff.resize(0x4000/4);
-					p->addr = 0x4000 - p->size;
+					//p->addr = 0x4000 - p->size;
-					fread(&p->buff[p->addr], p->size, 1, fp);
+					//fread(&p->buff[p->addr], p->size, 1, fp);
+					fread(&p->buff[0], p->size, 1, fp);
 					break;
 				case 1:
 				case 2:
 				case 3:
-					p->buff.resize(p->size);
+					p->buff.resize(p->size/4);
 					fread(&p->buff[0], p->size, 1, fp);
 					break;
 				}
@@ -712,7 +715,7 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 
 			case 3:
 
-				p->buff.resize(0x2000);
+				p->buff.resize(0x2000/4);
 
 				fread(&p->buff[0], 0x2000, 1, fp);
 
@@ -741,12 +744,14 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 				{
 				case 0:
 
+					//fprintf(stderr, "TRANSFER %d size %d\n", p->param, p->real_size);
 					switch(p->param)
 					{
-					case 0: GSgifTransfer1(&p->buff[0], p->addr); break;
+					//case 0: GSgifTransfer1((u32*)&p->buff[0], p->addr); break;
-					case 1: GSgifTransfer2(&p->buff[0], p->size / 16); break;
+					case 0: _GSgifTransfer<0>(&p->buff[0], p->real_size); break;
-					case 2: GSgifTransfer3(&p->buff[0], p->size / 16); break;
+					case 1: GSgifTransfer2((u32*)&p->buff[0], p->real_size); break;
-					case 3: GSgifTransfer(&p->buff[0], p->size / 16); break;
+					case 2: GSgifTransfer3((u32*)&p->buff[0], p->real_size); break;
+					case 3: GSgifTransfer((u32*)&p->buff[0], p->real_size); break;
 					}
 
 					break;
@@ -775,8 +780,8 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 				}
 			}
 			unsigned long end = timeGetTime();
-			fprintf(stderr, "The %d frames of the scene was render on %dms\n", frame_number, end - start);
+			fprintf(stderr, "The %ld frames of the scene was render on %ldms\n", frame_number, end - start);
-			fprintf(stderr, "A means of %fms by frame\n", (float)(end - start)/(float)frame_number);
+			fprintf(stderr, "A means of %fms by frame (limit 16ms/f)\n", (float)(end - start)/(float)frame_number);
 
 			sleep(1);
 			finished--;

plugins/zzogl-pg-cg/opengl/GifTransfer.cpp

@@ -76,6 +76,8 @@ template<int index> void _GSgifTransfer(const u32 *pMem, u32 size)
 
 #ifdef _DEBUG
 	gifTransferLog(index, pMem, size);
+	const u32* start = pMem;
+	u32 size_arg = size;
 
 	g_dump.Transfer(index, (const u8*)pMem, size);
 #endif
@@ -218,6 +220,10 @@ template<int index> void _GSgifTransfer(const u32 *pMem, u32 size)
 		}
 	}
 
+#ifdef _DEBUG
+	g_dump.Transfer(index, start, pMem - start, size_arg);
+#endif
+
 	// This is the case when not all data was readed from one try: VU1 has too much data.
 	// So we should redo reading from the start.
 	if (index == 0)

plugins/zzogl-pg/opengl/GSDump.cpp

@@ -58,13 +58,14 @@ void GSDump::Close()
 	}
 }
 
-void GSDump::Transfer(int index, const u8* mem, size_t size)
+void GSDump::Transfer(int index, const u32* mem, size_t size, u32 real_size)
 {
 	if(m_gs && size > 0)
 	{
 		fputc(0, m_gs);
 		fputc(index, m_gs);
 		fwrite(&size, 4, 1, m_gs);
+		fwrite(&real_size, 4, 1, m_gs);
 		fwrite(mem, size, 1, m_gs);
 	}
 }

plugins/zzogl-pg/opengl/GSDump.h

@@ -56,7 +56,7 @@ public:
 	void Open(const string& fn, u32 crc, const freezeData& fd, u8* regs);
 	void Close();
 	void ReadFIFO(u32 size);
-	void Transfer(int index, const u8* mem, size_t size);
+	void Transfer(int index, const u32* mem, size_t size, u32 real_size);
 	void VSync(int field, bool last, u8* regs);
 	operator bool() {return m_gs != NULL;}
 };

plugins/zzogl-pg/opengl/GSmain.cpp

@@ -403,7 +403,7 @@ static bool get_snapshot_filename(char *filename, char* path, const char* extens
 	{
 		snapshotnr++;
 
-		sprintf(filename, "%s/snap%03ld.%s", path, snapshotnr, extension);
+		sprintf(filename, "%s/snap%03d.%s", path, snapshotnr, extension);
 
 		bmpfile = fopen(filename, "rb");
 
@@ -641,6 +641,7 @@ struct Packet
 {
 	u8 type, param;
 	u32 size, addr;
+	u32 real_size;
 	vector<u32> buff;
 };
 
@@ -694,18 +695,20 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 				p->param = (u8)fgetc(fp);
 
 				fread(&p->size, 4, 1, fp);
+				fread(&p->real_size, 4, 1, fp);
 
 				switch(p->param)
 				{
 				case 0:
-					p->buff.resize(0x4000);
+					p->buff.resize(0x4000/4);
-					p->addr = 0x4000 - p->size;
+					//p->addr = 0x4000 - p->size;
-					fread(&p->buff[p->addr], p->size, 1, fp);
+					//fread(&p->buff[p->addr], p->size, 1, fp);
+					fread(&p->buff[0], p->size, 1, fp);
 					break;
 				case 1:
 				case 2:
 				case 3:
-					p->buff.resize(p->size);
+					p->buff.resize(p->size/4);
 					fread(&p->buff[0], p->size, 1, fp);
 					break;
 				}
@@ -726,7 +729,7 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 
 			case 3:
 
-				p->buff.resize(0x2000);
+				p->buff.resize(0x2000/4);
 
 				fread(&p->buff[0], 0x2000, 1, fp);
 
@@ -755,12 +758,14 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 				{
 				case 0:
 
+					//fprintf(stderr, "TRANSFER %d size %d\n", p->param, p->real_size);
 					switch(p->param)
 					{
-					case 0: GSgifTransfer1(&p->buff[0], p->addr); break;
+					//case 0: GSgifTransfer1((u32*)&p->buff[0], p->addr); break;
-					case 1: GSgifTransfer2(&p->buff[0], p->size / 16); break;
+					case 0: _GSgifTransfer<0>(&p->buff[0], p->real_size); break;
-					case 2: GSgifTransfer3(&p->buff[0], p->size / 16); break;
+					case 1: GSgifTransfer2((u32*)&p->buff[0], p->real_size); break;
-					case 3: GSgifTransfer(&p->buff[0], p->size / 16); break;
+					case 2: GSgifTransfer3((u32*)&p->buff[0], p->real_size); break;
+					case 3: GSgifTransfer((u32*)&p->buff[0], p->real_size); break;
 					}
 
 					break;
@@ -789,8 +794,8 @@ EXPORT_C_(void) GSReplay(char* lpszCmdLine)
 				}
 			}
 			unsigned long end = timeGetTime();
-			fprintf(stderr, "The %d frames of the scene was render on %dms\n", frame_number, end - start);
+			fprintf(stderr, "The %ld frames of the scene was render on %ldms\n", frame_number, end - start);
-			fprintf(stderr, "A means of %fms by frame\n", (float)(end - start)/(float)frame_number);
+			fprintf(stderr, "A means of %fms by frame (limit 16ms/f)\n", (float)(end - start)/(float)frame_number);
 
 			sleep(1);
 			finished--;

plugins/zzogl-pg/opengl/GifTransfer.cpp

@@ -76,6 +76,8 @@ template<int index> void _GSgifTransfer(const u32 *pMem, u32 size)
 
 #ifdef _DEBUG
 	gifTransferLog(index, pMem, size);
+	const u32* start = pMem;
+	u32 size_arg = size;
 
 	g_dump.Transfer(index, (const u8*)pMem, size);
 #endif
@@ -222,6 +224,10 @@ template<int index> void _GSgifTransfer(const u32 *pMem, u32 size)
 		}
 	}
 
+#ifdef _DEBUG
+	g_dump.Transfer(index, start, pMem - start, size_arg);
+#endif
+
 	// This is the case when not all data was readed from one try: VU1 has too much data.
 	// So we should redo reading from the start.
 	if (index == 0)

plugins/zzogl-pg/opengl/ps2hw_gl4.glsl

@@ -57,12 +57,6 @@
 #define TEX_DECL vec4
 #endif
 
-#ifdef WRITE_DEPTH
-#define DOZWRITE(x) x
-#else
-#define DOZWRITE(x)
-#endif
-
 // NVidia CG-data types
 #define half2 vec2
 #define half3 vec3
@@ -602,38 +596,34 @@ half4 ps2FinalColor(half4 col)
 half4 ps2FinalColor(half4 col)
 {
 	return col * g_fOneColor.xxxy + g_fOneColor.zzzw;
-	//return vec4(1.0f,0.0f,1.0f, 0.8f);
 }
 #endif
 
+
 #ifdef FRAGMENT_SHADER 			// This is code only for FRAGMENTS (pixel shader)
 
+#ifdef WRITE_DEPTH
+void write_depth_target(vec4 z)
+{
+    FragData1 = z; 
+}
+#else
+void write_depth_target(vec4 z) { }
+#endif
+
 void RegularPS() {
 	// whenever outputting depth, make sure to mult by 255/256 and 1
 	FragData0 = ps2FinalColor(PSin.color);
-	DOZWRITE(FragData1 = PSin.z;)
+    write_depth_target(PSin.z);
 }
 
-#ifdef WRITE_DEPTH
-
 #define DECL_TEXPS(num, bit) \
 void Texture##num##bit##PS() \
 { \
 	FragData0 = ps2FinalColor(ps2CalcShade(ps2shade##num##bit(PSin.tex), PSin.color)); \
-	FragData1 = PSin.z; \
+	write_depth_target(PSin.z); \
 }
 
-#else
-
-#define DECL_TEXPS(num, bit) \
-void Texture##num##bit##PS() \
-{ \
-	FragData0 = ps2FinalColor(ps2CalcShade(ps2shade##num##bit(PSin.tex), PSin.color)); \
-}
-
-#endif
-
-
 #define DECL_TEXPS_(num) \
 DECL_TEXPS(num, _32) \
 DECL_TEXPS(num, _tex32) \
@@ -653,32 +643,18 @@ void RegularFogPS() {
 	c.xyz = mix(g_fFogColor.xyz, PSin.color.xyz, vec3(PSin.fog));
 	c.w = PSin.color.w;
 	FragData0 = ps2FinalColor(c);
-   	DOZWRITE(FragData1 = PSin.z;)
+    write_depth_target(PSin.z);
 }
 
-#ifdef WRITE_DEPTH
-
 #define DECL_TEXFOGPS(num, bit) \
 void TextureFog##num##bit##PS() \
 { \
 	half4 c = ps2CalcShade(ps2shade##num##bit(PSin.tex), PSin.color); \
 	c.xyz = mix(g_fFogColor.xyz, c.xyz, vec3(PSin.fog)); \
 	FragData0 = ps2FinalColor(c); \
-   	FragData1 = PSin.z; \
+	write_depth_target(PSin.z); \
 }
 
-#else
-
-#define DECL_TEXFOGPS(num, bit) \
-void TextureFog##num##bit##PS() \
-{ \
-	half4 c = ps2CalcShade(ps2shade##num##bit(PSin.tex), PSin.color); \
-	c.xyz = mix(g_fFogColor.xyz, c.xyz, vec3(PSin.fog)); \
-	FragData0 = ps2FinalColor(c); \
-}
-
-#endif
-
 #define DECL_TEXFOGPS_(num) \
 DECL_TEXFOGPS(num, _32) \
 DECL_TEXFOGPS(num, _tex32) \
@@ -876,70 +852,78 @@ void Convert32to16PS() {
 
 #ifdef VERTEX_SHADER
 
-float4 OutPosition() {
+void SetColor() {
+	VSout.color = Color;
+}
+
+void SetTex() {
+#ifdef PERSPECTIVE_CORRECT_TEX
+	VSout.tex.xyz = TexCoord.xyz;
+#else
+	VSout.tex.xy = TexCoord.xy/TexCoord.z;
+#endif
+}
+
+void SetZ() {
+#ifdef WRITE_DEPTH
+	VSout.z = SecondaryColor * g_fZBias.x + g_fZBias.y;
+    VSout.z.w = g_fc0.y;
+#endif
+}
+
+float SetPosition() {
 	float4 position;
 	position.xy = vec2(Vert.xy) * g_fPosXY.xy + g_fPosXY.zw;
+    // FIXME: the factor in normal mode seem bogus. They don't have same order than in log mode. Or I failed to understand the logic
+    //// normal mode.
+    // -> dot(g_fZ, SecondaryColor.zyxw)) => reconstruct a float from normalized char. The float range from 0 to 1
+    // position.z = dot(g_fZ, SecondaryColor.zyxw);
+    //// logz mode
+    // -> dot(g_fZ, SecondaryColor.zyxw)) => reconstruct a float from normalized char. The float range from 0 to 2**32
+    // 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;
-	return position;
+
+    gl_Position = position;
+
+    // For fog
+    return position.z;
+}
+
+void SetFog(float z) {
+    VSout.fog = z * g_fBilinear.w;
 }
 
 // just smooth shadering
 void RegularVS() {
-	float4 position = OutPosition();
+    float z = SetPosition();
-	gl_Position = position;
+    SetColor();
-
+    SetZ();
-	VSout.color = Color;
-
-	DOZWRITE(VSout.z = SecondaryColor * g_fZBias.x + g_fZBias.y;)
-    DOZWRITE(VSout.z.w = g_fc0.y;)
 }
 
 // diffuse texture mapping
 void TextureVS() {
-	float4 position = OutPosition();
+    float z = SetPosition();
-	gl_Position = position;
+    SetColor();
-
+    SetTex();
-	VSout.color = Color;
+    SetZ();
-
-#ifdef PERSPECTIVE_CORRECT_TEX
-	VSout.tex.xyz = TexCoord.xyz;
-#else
-	VSout.tex.xy = TexCoord.xy/TexCoord.z;
-#endif
-
- 	DOZWRITE(VSout.z = SecondaryColor * g_fZBias.x + g_fZBias.y;)
-    DOZWRITE(VSout.z.w = g_fc0.y;)
 }
 
 void RegularFogVS() {
-	float4 position = OutPosition();
+    float z = SetPosition();
-	gl_Position = position;
+    SetColor();
-
+    SetZ();
-	VSout.color = Color;
+    SetFog(z);
-
-    VSout.fog = position.z * g_fBilinear.w;
-
-	DOZWRITE(VSout.z = SecondaryColor * g_fZBias.x + g_fZBias.y;)
-    DOZWRITE(VSout.z.w = g_fc0.y;)
 }
 
 void TextureFogVS() {
-	float4 position = OutPosition();
+    float z = SetPosition();
-	gl_Position = position;
+    SetColor();
-
+    SetTex();
-	VSout.color = Color;
+    SetZ();
-
+    SetFog(z);
-#ifdef PERSPECTIVE_CORRECT_TEX
-	VSout.tex.xyz = TexCoord.xyz;
-#else
-	VSout.tex.xy = TexCoord.xy/TexCoord.z;
-#endif
-
-    VSout.fog = position.z * g_fBilinear.w;
-
-	DOZWRITE(VSout.z = SecondaryColor * g_fZBias.x + g_fZBias.y;)
-    DOZWRITE(VSout.z.w = g_fc0.y;)
 }
 
 void BitBltVS() {