Implement proper shader program DVLE constant management

source/base.c

@@ -49,9 +49,12 @@ static void C3Di_AptEventHook(APT_HookType hookType, void* param)
 		case APTHOOK_ONRESTORE:
 		{
 			ctx->flags |= C3DiF_AttrInfo | C3DiF_BufInfo | C3DiF_Effect | C3DiF_RenderBuf
-				| C3DiF_Viewport | C3DiF_Scissor | C3DiF_Program
+				| C3DiF_Viewport | C3DiF_Scissor | C3DiF_Program | C3DiF_VshCode | C3DiF_GshCode
 				| C3DiF_TexAll | C3DiF_TexEnvBuf | C3DiF_TexEnvAll | C3DiF_LightEnv;
 
+			C3Di_DirtyUniforms(GPU_VERTEX_SHADER);
+			C3Di_DirtyUniforms(GPU_GEOMETRY_SHADER);
+
 			C3D_LightEnv* env = ctx->lightEnv;
 			if (env)
 				env->Dirty(env);
@@ -129,8 +132,8 @@ void C3Di_UpdateContext(void)
 
 	if (ctx->flags & C3DiF_Program)
 	{
-		ctx->flags &= ~C3DiF_Program;
+		shaderProgramConfigure(ctx->program, (ctx->flags & C3DiF_VshCode) != 0, (ctx->flags & C3DiF_GshCode) != 0);
-		shaderProgramUse(ctx->program);
+		ctx->flags &= ~(C3DiF_Program | C3DiF_VshCode | C3DiF_GshCode);
 	}
 
 	if (ctx->flags & C3DiF_RenderBuf)
@@ -270,6 +273,27 @@ void C3D_BindProgram(shaderProgram_s* program)
 	if (!(ctx->flags & C3DiF_Active))
 		return;
 
+	shaderProgram_s* oldProg = ctx->program;
+	shaderInstance_s* newGsh = program->geometryShader;
+	if (oldProg != program)
+	{
 		ctx->program = program;
 		ctx->flags |= C3DiF_Program;
+
+		if (oldProg)
+		{
+			if (oldProg->vertexShader->dvle->dvlp != program->vertexShader->dvle->dvlp)
+				ctx->flags |= C3DiF_VshCode;
+			shaderInstance_s* oldGsh = oldProg->geometryShader;
+			if (newGsh && (!oldGsh || oldGsh->dvle->dvlp != newGsh->dvle->dvlp))
+				ctx->flags |= C3DiF_GshCode;
+		} else
+			ctx->flags |= C3DiF_VshCode | C3DiF_GshCode;
+	}
+
+	C3Di_LoadShaderUniforms(program->vertexShader);
+	if (newGsh)
+		C3Di_LoadShaderUniforms(newGsh);
+	else
+		C3Di_ClearShaderUniforms(GPU_GEOMETRY_SHADER);
 }

source/context.h

@@ -59,6 +59,8 @@ enum
 	C3DiF_Program = BIT(8),
 	C3DiF_TexEnvBuf = BIT(9),
 	C3DiF_LightEnv = BIT(10),
+	C3DiF_VshCode = BIT(11),
+	C3DiF_GshCode = BIT(12),
 
 #define C3DiF_Tex(n) BIT(23+(n))
 	C3DiF_TexAll = 7 << 23,
@@ -80,3 +82,7 @@ void C3Di_EffectBind(C3D_Effect* effect);
 void C3Di_RenderBufBind(C3D_RenderBuf* rb);
 
 void C3Di_LightMtlBlend(C3D_Light* light);
+
+void C3Di_DirtyUniforms(GPU_SHADER_TYPE type);
+void C3Di_LoadShaderUniforms(shaderInstance_s* si);
+void C3Di_ClearShaderUniforms(GPU_SHADER_TYPE type);

source/uniforms.c

@@ -9,12 +9,35 @@ bool C3D_FVUnifDirty[2][C3D_FVUNIF_COUNT];
 bool C3D_IVUnifDirty[2][C3D_IVUNIF_COUNT];
 bool C3D_BoolUnifsDirty[2];
 
+static struct
+{
+	bool dirty;
+	int count;
+	float24Uniform_s* data;
+} C3Di_ShaderFVecData[2];
+
+static bool C3Di_FVUnifEverDirty[2][C3D_FVUNIF_COUNT];
+static bool C3Di_IVUnifEverDirty[2][C3D_IVUNIF_COUNT];
+
 void C3D_UpdateUniforms(GPU_SHADER_TYPE type)
 {
 	int offset = type == GPU_GEOMETRY_SHADER ? (GPUREG_GSH_BOOLUNIFORM-GPUREG_VSH_BOOLUNIFORM) : 0;
+	int i = 0;
+
+	// Update FVec uniforms that come from shader constants
+	if (C3Di_ShaderFVecData[type].dirty)
+	{
+		while (i < C3Di_ShaderFVecData[type].count)
+		{
+			float24Uniform_s* u = &C3Di_ShaderFVecData[type].data[i++];
+			GPUCMD_AddIncrementalWrites(GPUREG_VSH_FLOATUNIFORM_CONFIG+offset, (u32*)u, 4);
+			C3D_FVUnifDirty[type][u->id] = false;
+		}
+		C3Di_ShaderFVecData[type].dirty = false;
+		i = 0;
+	}
 
 	// Update FVec uniforms
-	int i = 0;
 	while (i < C3D_FVUNIF_COUNT)
 	{
 		if (!C3D_FVUnifDirty[type][i])
@@ -27,14 +50,6 @@ void C3D_UpdateUniforms(GPU_SHADER_TYPE type)
 		int j;
 		for (j = i; j < C3D_FVUNIF_COUNT && C3D_FVUnifDirty[type][j]; j ++);
 
-		/*
-		consoleClear();
-		printf("FVEC Uniform %02X--%02X\n", i, j-1);
-		int pp;
-		for (pp = i; pp < j; pp ++)
-			printf("%02X: (%f, %f, %f, %f)\n", pp, C3D_FVUnif[pp].x, C3D_FVUnif[pp].y, C3D_FVUnif[pp].z, C3D_FVUnif[pp].w);
-		*/
-
 		// Upload the uniforms
 		GPUCMD_AddWrite(GPUREG_VSH_FLOATUNIFORM_CONFIG+offset, 0x80000000|i);
 		GPUCMD_AddWrites(GPUREG_VSH_FLOATUNIFORM_DATA+offset, (u32*)&C3D_FVUnif[i], (j-i)*4);
@@ -42,7 +57,10 @@ void C3D_UpdateUniforms(GPU_SHADER_TYPE type)
 		// Clear the dirty flag
 		int k;
 		for (k = i; k < j; k ++)
+		{
 			C3D_FVUnifDirty[type][k] = false;
+			C3Di_FVUnifEverDirty[type][k] = true;
+		}
 
 		// Advance
 		i += j;
@@ -55,12 +73,58 @@ void C3D_UpdateUniforms(GPU_SHADER_TYPE type)
 
 		GPUCMD_AddWrite(GPUREG_VSH_INTUNIFORM_I0+offset+i, C3D_IVUnif[type][i]);
 		C3D_IVUnifDirty[type][i] = false;
+		C3Di_IVUnifEverDirty[type][i] = false;
 	}
 
 	// Update bool uniforms
 	if (C3D_BoolUnifsDirty[type])
 	{
-		GPUCMD_AddWrite(GPUREG_VSH_BOOLUNIFORM+offset, 0x7FFF0000 | (u32)C3D_BoolUnifs[type]);
+		GPUCMD_AddWrite(GPUREG_VSH_BOOLUNIFORM+offset, 0x7FFF0000 | (u32)~C3D_BoolUnifs[type]);
 		C3D_BoolUnifsDirty[type] = false;
 	}
 }
+
+void C3Di_DirtyUniforms(GPU_SHADER_TYPE type)
+{
+	int i;
+	C3D_BoolUnifsDirty[type] = true;
+	if (C3Di_ShaderFVecData[type].count)
+		C3Di_ShaderFVecData[type].dirty = true;
+	for (i = 0; i < C3D_FVUNIF_COUNT; i ++)
+		C3D_FVUnifDirty[type][i] = C3D_FVUnifDirty[type][i] || C3Di_FVUnifEverDirty[type][i];
+	for (i = 0; i < C3D_IVUNIF_COUNT; i ++)
+		C3D_IVUnifDirty[type][i] = C3D_IVUnifDirty[type][i] || C3Di_IVUnifEverDirty[type][i];
+}
+
+void C3Di_LoadShaderUniforms(shaderInstance_s* si)
+{
+	GPU_SHADER_TYPE type = si->dvle->type;
+	int i;
+	if (si->boolUniformMask)
+	{
+		C3D_BoolUnifs[type] &= ~si->boolUniformMask;
+		C3D_BoolUnifs[type] |= si->boolUniforms;
+		C3D_BoolUnifsDirty[type] = true;
+	}
+	if (si->intUniformMask)
+	{
+		for (i = 0; i < 4; i ++)
+		{
+			if (si->intUniformMask & BIT(i))
+			{
+				C3D_IVUnif[type][i] = si->intUniforms[i];
+				C3D_IVUnifDirty[type][i] = true;
+			}
+		}
+	}
+	C3Di_ShaderFVecData[type].dirty = true;
+	C3Di_ShaderFVecData[type].count = si->numFloat24Uniforms;
+	C3Di_ShaderFVecData[type].data = si->float24Uniforms;
+}
+
+void C3Di_ClearShaderUniforms(GPU_SHADER_TYPE type)
+{
+	C3Di_ShaderFVecData[type].dirty = false;
+	C3Di_ShaderFVecData[type].count = 0;
+	C3Di_ShaderFVecData[type].data = NULL;
+}