Add Render Buffer API + Viewport/Scissor + Update Example

examples/gputest/source/main.c

@@ -21,8 +21,9 @@
 #define TOPSCR_COPYFLAG 0x01001000
 #endif
 
-u32* gpuDepthBuf;
-u32* gpuColorBuf;
+#define CLEAR_COLOR 0x80FF80FF
+
+C3D_RenderBuf rbTop, rbBot;
 
 static DVLB_s *pVsh, *pGsh;
 static shaderProgram_s shader;
@@ -49,23 +50,14 @@ static const vertex_t vertices[] =
 	{{-0.5f, +0.5f, -4.0f}, {0.0f, 1.0f}, {1.0f, 0.0f, 0.0f}},
 };
 
-//#define BG_COLOR 0x68B0D8FF
-#define BG_COLOR 0x80FF80FF
-
 #define FOVY (2.0f/15)
 
-static inline void clearGpuBuffers(u32 color)
-{
-	GX_SetMemoryFill(NULL, gpuColorBuf, color, gpuColorBuf+0x2EE00, 0x201, gpuDepthBuf, 0x00000000, gpuDepthBuf+0x2EE00, 0x201);
-	gspWaitForPSC0();
-}
-
 static C3D_VBO myVbo;
 static C3D_Tex myTex;
 
 static void drawScene(float trX, float trY)
 {
-	GPU_SetViewport((u32*)osConvertVirtToPhys((u32)gpuDepthBuf), (u32*)osConvertVirtToPhys((u32)gpuColorBuf), 0, 0, TOPSCR_WIDTH, 400);
+	C3D_RenderBufBind(&rbTop);
 
 	C3D_TexBind(0, &myTex);
 
@@ -88,7 +80,7 @@ static void drawScene(float trX, float trY)
 
 static void drawSceneBottom(float trX, float trY)
 {
-	GPU_SetViewport((u32*)osConvertVirtToPhys((u32)gpuDepthBuf), (u32*)osConvertVirtToPhys((u32)gpuColorBuf), 0, 0, 240/**2*/, 320);
+	C3D_RenderBufBind(&rbBot);
 
 	C3D_TexBind(0, NULL);
 
@@ -119,8 +111,10 @@ int main()
 	printf("Testing...\n");
 #endif
 
-	gpuDepthBuf = (u32*)vramAlloc(400*240*2*sizeof(float));
-	gpuColorBuf = (u32*)vramAlloc(400*240*2*sizeof(u32));
+	C3D_RenderBufInit(&rbTop, TOPSCR_WIDTH, 400, GPU_RB_RGBA8, GPU_RB_DEPTH24_STENCIL8);
+	C3D_RenderBufInit(&rbBot, 240, 320, GPU_RB_RGBA8, GPU_RB_DEPTH24_STENCIL8);
+	rbTop.clearColor = CLEAR_COLOR;
+	rbBot.clearColor = CLEAR_COLOR;
 
 	C3D_Init(C3D_DEFAULT_CMDBUF_SIZE);
 
@@ -154,7 +148,9 @@ int main()
 	C3D_VBOInit(&myVbo, sizeof(vertices));
 	C3D_VBOAddData(&myVbo, vertices, sizeof(vertices), sizeof(vertices)/sizeof(vertex_t));
 
-	clearGpuBuffers(BG_COLOR);
+	// Clear buffers
+	C3D_RenderBufClear(&rbTop);
+	C3D_RenderBufClear(&rbBot);
 
 	float trX = 0, trY = 0;
 	float zDist = 0.1f;
@@ -188,30 +184,22 @@ int main()
 		float czDist = zDist*slider/2;
 
 		drawScene(trX-czDist, trY);
-
-		GX_SetDisplayTransfer(NULL, gpuColorBuf, GX_BUFFER_DIM(TOPSCR_WIDTH,400), (u32*)gfxGetFramebuffer(GFX_TOP, GFX_LEFT, NULL, NULL), GX_BUFFER_DIM(TOPSCR_WIDTH,400), TOPSCR_COPYFLAG);
-		gspWaitForPPF();
+		C3D_RenderBufTransfer(&rbTop, (u32*)gfxGetFramebuffer(GFX_TOP, GFX_LEFT, NULL, NULL), TOPSCR_COPYFLAG);
 
 		if (slider > 0.0f)
 		{
-			clearGpuBuffers(BG_COLOR);
-
+			C3D_RenderBufClear(&rbTop);
 			drawScene(trX+czDist, trY);
-
-			GX_SetDisplayTransfer(NULL, gpuColorBuf, GX_BUFFER_DIM(TOPSCR_WIDTH,400), (u32*)gfxGetFramebuffer(GFX_TOP, GFX_RIGHT, NULL, NULL), GX_BUFFER_DIM(TOPSCR_WIDTH,400), TOPSCR_COPYFLAG);
-			gspWaitForPPF();
+			C3D_RenderBufTransfer(&rbTop, (u32*)gfxGetFramebuffer(GFX_TOP, GFX_RIGHT, NULL, NULL), TOPSCR_COPYFLAG);
 		}
 
+		C3D_RenderBufClear(&rbTop); // In theory this could be async but meh...
+		
 #ifndef DEBUG
-		clearGpuBuffers(BG_COLOR);
-
 		drawSceneBottom(trX, trY);
-
-		GX_SetDisplayTransfer(NULL, gpuColorBuf, GX_BUFFER_DIM(240,320), (u32*)gfxGetFramebuffer(GFX_BOTTOM, GFX_LEFT, NULL, NULL), GX_BUFFER_DIM(240,320), 0x1000);
-		gspWaitForPPF();
+		C3D_RenderBufTransfer(&rbBot, (u32*)gfxGetFramebuffer(GFX_BOTTOM, GFX_LEFT, NULL, NULL), 0x1000);
+		C3D_RenderBufClear(&rbBot); // Same here
 #endif
-
-		clearGpuBuffers(BG_COLOR);
 	}
 
 	C3D_Fini();

include/c3d/base.h

@@ -7,6 +7,9 @@ bool C3D_Init(size_t cmdBufSize);
 void C3D_FlushAsync(void);
 void C3D_Fini(void);
 
+void C3D_SetViewport(u32 x, u32 y, u32 w, u32 h);
+void C3D_SetScissor(GPU_SCISSORMODE mode, u32 x, u32 y, u32 w, u32 h);
+
 void C3D_DrawArray(C3D_VBO* vbo, GPU_Primitive_t primitive);
 void C3D_DrawElements(C3D_IBO* ibo, GPU_Primitive_t primitive);
 

include/c3d/renderbuffer.h

@@ -0,0 +1,44 @@
+#pragma once
+#include "types.h"
+
+typedef struct
+{
+	void *colorBuf, *depthBuf;
+	u16 colorFmt, depthFmt;
+	u16 width, height;
+	u32 clearColor, clearDepth;
+} C3D_RenderBuf;
+
+enum
+{
+	// Renderbuffer color formats
+	GPU_RB_RGBA8 = 0,
+	GPU_RB_RGB8,
+	GPU_RB_RGB5_A1,
+	GPU_RB_R5_G6_B5,
+	GPU_RB_RGBA4,
+
+	// Renderbuffer depth formats
+	GPU_RB_DEPTH16 = 0,
+	GPU_RB_DEPTHXX,
+	GPU_RB_DEPTH24,
+	GPU_RB_DEPTH24_STENCIL8,
+};
+
+bool C3D_RenderBufInit(C3D_RenderBuf* rb, int width, int height, int colorFmt, int depthFmt);
+void C3D_RenderBufClearAsync(C3D_RenderBuf* rb);
+void C3D_RenderBufTransferAsync(C3D_RenderBuf* rb, u32* frameBuf, u32 flags);
+void C3D_RenderBufBind(C3D_RenderBuf* rb);
+void C3D_RenderBufDelete(C3D_RenderBuf* rb);
+
+static inline void C3D_RenderBufClear(C3D_RenderBuf* rb)
+{
+	C3D_RenderBufClearAsync(rb);
+	gspWaitForPSC0();
+}
+
+static inline void C3D_RenderBufTransfer(C3D_RenderBuf* rb, u32* frameBuf, u32 flags)
+{
+	C3D_RenderBufTransferAsync(rb, frameBuf, flags);
+	gspWaitForPPF();
+}

include/citro3d.h

@@ -21,6 +21,8 @@ extern "C" {
 #include "c3d/effect.h"
 #include "c3d/texture.h"
 
+#include "c3d/renderbuffer.h"
+
 #ifdef __cplusplus
 }
 #endif

source/base.c

@@ -14,6 +14,13 @@ u32 C3Di_Float24(float f)
 	return (exp >= 0) ? (man | (exp << 16) | (s << 23)) : (s << 23);
 }
 
+u32 C3Di_FloatInv24(u32 val)
+{
+	// Too lazy to copy & paste & cleanup the libctru function
+	extern u32 computeInvValue(u32 val);
+	return computeInvValue(val);
+}
+
 static void C3Di_SetTex(GPU_TEXUNIT unit, C3D_Tex* tex)
 {
 	u32 reg[4];
@@ -81,6 +88,28 @@ bool C3D_Init(size_t cmdBufSize)
 	return true;
 }
 
+void C3D_SetViewport(u32 x, u32 y, u32 w, u32 h)
+{
+	C3D_Context* ctx = C3Di_GetContext();
+	ctx->flags |= C3DiF_Viewport | C3DiF_Scissor;
+	ctx->viewport[0] = C3Di_Float24((float)w/2);
+	ctx->viewport[1] = C3Di_FloatInv24(w);
+	ctx->viewport[2] = C3Di_Float24((float)h/2);
+	ctx->viewport[3] = C3Di_FloatInv24(h);
+	ctx->viewport[4] = (y << 16) | (x & 0xFFFF);
+	ctx->scissor[0] = GPU_SCISSOR_DISABLE;
+}
+
+void C3D_SetScissor(GPU_SCISSORMODE mode, u32 x, u32 y, u32 w, u32 h)
+{
+	C3D_Context* ctx = C3Di_GetContext();
+	ctx->flags |= C3DiF_Scissor;
+	ctx->scissor[0] = mode;
+	if (mode == GPU_SCISSOR_DISABLE) return;
+	ctx->scissor[1] = (y << 16) | (x & 0xFFFF);
+	ctx->scissor[2] = ((h-1) << 16) | ((w-1) & 0xFFFF);
+}
+
 void C3Di_UpdateContext(void)
 {
 	int i;
@@ -92,6 +121,25 @@ void C3Di_UpdateContext(void)
 		//GPU_FinishDrawing();
 	}
 
+	if (ctx->flags & C3DiF_RenderBuf)
+	{
+		ctx->flags &= ~C3DiF_RenderBuf;
+		C3Di_RenderBufBind(ctx->rb);
+	}
+
+	if (ctx->flags & C3DiF_Viewport)
+	{
+		ctx->flags &= ~C3DiF_Viewport;
+		GPUCMD_AddIncrementalWrites(GPUREG_0041, ctx->viewport, 4);
+		GPUCMD_AddWrite(GPUREG_0068, ctx->viewport[4]);
+	}
+
+	if (ctx->flags & C3DiF_Scissor)
+	{
+		ctx->flags &= ~C3DiF_Scissor;
+		GPUCMD_AddIncrementalWrites(GPUREG_SCISSORTEST_MODE, ctx->scissor, 3);
+	}
+
 	if (ctx->flags & C3DiF_AttrBuf)
 	{
 		ctx->flags &= ~C3DiF_AttrBuf;

source/context.h

@@ -6,6 +6,7 @@
 #include <c3d/texenv.h>
 #include <c3d/effect.h>
 #include <c3d/texture.h>
+#include <c3d/renderbuffer.h>
 
 typedef struct
 {
@@ -34,6 +35,10 @@ typedef struct
 	C3D_Tex* tex[3];
 	C3D_TexEnv texEnv[6];
 
+	C3D_RenderBuf* rb;
+	u32 viewport[5];
+	u32 scissor[3];
+
 } C3D_Context;
 
 enum
@@ -42,6 +47,9 @@ enum
 	C3DiF_NeedFinishDrawing = BIT(1),
 	C3DiF_AttrBuf = BIT(2),
 	C3DiF_Effect = BIT(3),
+	C3DiF_RenderBuf = BIT(4),
+	C3DiF_Viewport = BIT(5),
+	C3DiF_Scissor = BIT(6),
 
 #define C3DiF_Tex(n) BIT(23+(n))
 	C3DiF_TexAll = 7 << 23,
@@ -56,8 +64,10 @@ static inline C3D_Context* C3Di_GetContext(void)
 }
 
 u32 C3Di_Float24(float f);
+u32 C3Di_FloatInv24(u32 val);
 void C3Di_UpdateContext(void);
 void C3Di_IBOBind(C3D_IBO* ibo);
 void C3Di_AttrInfoBind(C3D_AttrInfo* info);
 void C3Di_TexEnvBind(int id, C3D_TexEnv* env);
 void C3Di_EffectBind(C3D_Effect* effect);
+void C3Di_RenderBufBind(C3D_RenderBuf* rb);

source/renderbuffer.c

@@ -0,0 +1,99 @@
+#include "context.h"
+
+static const u8 colorFmtSizes[] = {2,1,0,0,0};
+static const u8 depthFmtSizes[] = {0,0,1,2};
+
+static inline u32 calcColorBufSize(u32 width, u32 height, u32 fmt)
+{
+	return width*height*(colorFmtSizes[fmt]+2);
+}
+
+static inline u32 calcDepthBufSize(u32 width, u32 height, u32 fmt)
+{
+	return width*height*(depthFmtSizes[fmt]+2);
+}
+
+bool C3D_RenderBufInit(C3D_RenderBuf* rb, int width, int height, int colorFmt, int depthFmt)
+{
+	if (rb->colorBuf || rb->depthBuf) return false;
+
+	rb->colorFmt = colorFmt;
+	rb->depthFmt = depthFmt;
+	rb->width = width;
+	rb->height = height;
+	rb->clearColor = rb->clearDepth = 0;
+
+	rb->colorBuf = vramAlloc(calcColorBufSize(width, height, colorFmt));
+	if (!rb->colorBuf) return false;
+
+	rb->depthBuf = vramAlloc(calcDepthBufSize(width, height, depthFmt));
+	if (!rb->depthBuf)
+	{
+		vramFree(rb->colorBuf);
+		rb->colorBuf = NULL;
+		return false;
+	}
+
+	return true;
+}
+
+void C3D_RenderBufClearAsync(C3D_RenderBuf* rb)
+{
+	u32 colorBufSize = calcColorBufSize(rb->width, rb->height, rb->colorFmt);
+	u32 depthBufSize = calcDepthBufSize(rb->width, rb->height, rb->depthFmt);
+	GX_SetMemoryFill(NULL,
+		(u32*)rb->colorBuf, rb->clearColor, (u32*)((u8*)rb->colorBuf+colorBufSize), BIT(0) | ((u32)colorFmtSizes[rb->colorFmt] << 8),
+		(u32*)rb->depthBuf, rb->clearDepth, (u32*)((u8*)rb->depthBuf+depthBufSize), BIT(0) | ((u32)colorFmtSizes[rb->depthFmt] << 8));
+}
+
+void C3D_RenderBufTransferAsync(C3D_RenderBuf* rb, u32* frameBuf, u32 flags)
+{
+	u32 dim = GX_BUFFER_DIM((u32)rb->width, (u32)rb->height);
+	GX_SetDisplayTransfer(NULL, (u32*)rb->colorBuf, dim, frameBuf, dim, flags);
+}
+
+void C3D_RenderBufBind(C3D_RenderBuf* rb)
+{
+	C3D_Context* ctx = C3Di_GetContext();
+	ctx->flags |= C3DiF_RenderBuf;
+	ctx->rb = rb;
+	C3D_SetViewport(0, 0, rb->width, rb->height);
+}
+
+void C3Di_RenderBufBind(C3D_RenderBuf* rb)
+{
+	u32 param[4];
+
+	param[0] = param[1] = 1;
+	GPUCMD_AddIncrementalWrites(GPUREG_0110, param, 2);
+
+	param[0] = osConvertVirtToPhys((u32)rb->depthBuf) >> 3;
+	param[1] = osConvertVirtToPhys((u32)rb->colorBuf) >> 3;
+	param[2] = 0x01000000 | (((u32)(rb->height-1) & 0xFFF) << 12) | (rb->width & 0xFFF);
+	GPUCMD_AddIncrementalWrites(GPUREG_DEPTHBUFFER_LOC, param, 3);
+
+	GPUCMD_AddWrite(GPUREG_006E, param[2]); //?
+	GPUCMD_AddWrite(GPUREG_DEPTHBUFFER_FORMAT, rb->depthFmt);
+	GPUCMD_AddWrite(GPUREG_COLORBUFFER_FORMAT, ((u32)rb->colorFmt << 16) | colorFmtSizes[rb->colorFmt]);
+	GPUCMD_AddWrite(GPUREG_011B, 0x00000000); //?
+
+	// "Enable depth buffer" (?)
+	param[0] = param[1] = 0xF;
+	param[2] = param[3] = 0x2;
+	GPUCMD_AddIncrementalWrites(GPUREG_0112, param, 4);
+}
+
+void C3D_RenderBufDelete(C3D_RenderBuf* rb)
+{
+	if (rb->colorBuf)
+	{
+		vramFree(rb->colorBuf);
+		rb->colorBuf = NULL;
+	}
+
+	if (rb->depthBuf)
+	{
+		vramFree(rb->depthBuf);
+		rb->depthBuf = NULL;
+	}
+}