Implement render queue system

include/c3d/renderqueue.h

@@ -0,0 +1,42 @@
+#pragma once
+#include "renderbuffer.h"
+
+typedef struct C3D_RenderTarget_tag C3D_RenderTarget;
+
+struct C3D_RenderTarget_tag
+{
+	C3D_RenderTarget *next, *prev, *link, *frame[2];
+	C3D_RenderBuf renderBuf;
+	u32 transferFlags;
+
+	u8 clearBits;
+	bool drawOk, transferOk;
+
+	bool linked;
+	gfxScreen_t screen;
+	gfx3dSide_t side;
+};
+
+// Flags for C3D_FrameBegin
+enum
+{
+	C3D_FRAME_SYNCDRAW = BIT(0), // Do not render the frame until the previous has finished rendering
+	C3D_FRAME_NONBLOCK = BIT(1), // Return false instead of waiting for the GPU to finish rendering
+};
+
+bool C3D_FrameBegin(u8 flags);
+bool C3D_FrameDrawOn(C3D_RenderTarget* target);
+void C3D_FrameEnd(u8 flags);
+
+// Flags for C3D_RenderTargetSetClear (only C3D_CLEAR_ALL implemented atm)
+enum
+{
+	C3D_CLEAR_COLOR = BIT(0),
+	C3D_CLEAR_DEPTH = BIT(1),
+	C3D_CLEAR_ALL   = C3D_CLEAR_COLOR | C3D_CLEAR_DEPTH,
+};
+
+C3D_RenderTarget* C3D_RenderTargetCreate(int width, int height, int colorFmt, int depthFmt);
+void C3D_RenderTargetDelete(C3D_RenderTarget* target);
+void C3D_RenderTargetSetClear(C3D_RenderTarget* target, u32 clearBits, u32 clearColor, u32 clearDepth);
+void C3D_RenderTargetSetOutput(C3D_RenderTarget* target, gfxScreen_t screen, gfx3dSide_t side, u32 transferFlags);

include/citro3d.h

@@ -23,6 +23,7 @@ extern "C" {
 #include "c3d/light.h"
 
 #include "c3d/renderbuffer.h"
+#include "c3d/renderqueue.h"
 
 #ifdef __cplusplus
 }

source/base.c

@@ -73,13 +73,14 @@ bool C3D_Init(size_t cmdBufSize)
 	if (ctx->flags & C3DiF_Active)
 		return false;
 
-	ctx->cmdBufSize = cmdBufSize;
-	ctx->cmdBuf = linearAlloc(cmdBufSize);
+	ctx->cmdBufSize = cmdBufSize/8; // Half of the size of the cmdbuf, in words
+	ctx->cmdBuf = (u32*)linearAlloc(cmdBufSize);
 	if (!ctx->cmdBuf) return false;
 
 	GPUCMD_SetBuffer(ctx->cmdBuf, ctx->cmdBufSize, 0);
 
 	ctx->flags = C3DiF_Active | C3DiF_TexEnvBuf | C3DiF_TexEnvAll | C3DiF_Effect | C3DiF_TexAll;
+	ctx->renderQueueExit = NULL;
 
 	// TODO: replace with direct struct access
 	C3D_DepthMap(-1.0f, 0.0f);
@@ -149,6 +150,7 @@ void C3Di_UpdateContext(void)
 		{
 			ctx->flags &= ~C3DiF_DrawUsed;
 			GPUCMD_AddWrite(GPUREG_FRAMEBUFFER_FLUSH, 1);
+			GPUCMD_AddWrite(GPUREG_EARLYDEPTH_CLEAR, 1);
 		}
 		C3Di_RenderBufBind(ctx->rb);
 	}
@@ -251,13 +253,10 @@ void C3Di_UpdateContext(void)
 	C3D_UpdateUniforms(GPU_GEOMETRY_SHADER);
 }
 
-void C3D_FlushAsync(void)
+void C3Di_FinalizeFrame(u32** pBuf, u32* pSize)
 {
 	C3D_Context* ctx = C3Di_GetContext();
 
-	if (!(ctx->flags & C3DiF_Active))
-		return;
-
 	if (ctx->flags & C3DiF_DrawUsed)
 	{
 		ctx->flags &= ~C3DiF_DrawUsed;
@@ -267,8 +266,30 @@ void C3D_FlushAsync(void)
 	}
 
 	GPUCMD_Finalize();
-	GPUCMD_FlushAndRun();
-	GPUCMD_SetBuffer(ctx->cmdBuf, ctx->cmdBufSize, 0);
+	GPUCMD_GetBuffer(pBuf, NULL, pSize);
+	*pSize *= 4;
+
+	ctx->flags ^= C3DiF_CmdBuffer;
+	u32* buf = ctx->cmdBuf;
+	if (ctx->flags & C3DiF_CmdBuffer)
+		buf += ctx->cmdBufSize;
+	GPUCMD_SetBuffer(buf, ctx->cmdBufSize, 0);
+}
+
+void C3D_FlushAsync(void)
+{
+	if (!(C3Di_GetContext()->flags & C3DiF_Active))
+		return;
+
+	u32* cmdBuf;
+	u32 cmdBufSize;
+	C3Di_FinalizeFrame(&cmdBuf, &cmdBufSize);
+
+	//take advantage of GX_FlushCacheRegions to flush gsp heap
+	extern u32 __ctru_linear_heap;
+	extern u32 __ctru_linear_heap_size;
+	GX_FlushCacheRegions(cmdBuf, cmdBufSize, (u32 *) __ctru_linear_heap, __ctru_linear_heap_size, NULL, 0);
+	GX_ProcessCommandList(cmdBuf, cmdBufSize, 0x0);
 }
 
 void C3D_Fini(void)
@@ -278,6 +299,9 @@ void C3D_Fini(void)
 	if (!(ctx->flags & C3DiF_Active))
 		return;
 
+	if (ctx->renderQueueExit)
+		ctx->renderQueueExit();
+
 	aptUnhook(&hookCookie);
 	linearFree(ctx->cmdBuf);
 	ctx->flags = 0;

source/context.h

@@ -26,7 +26,7 @@ typedef struct
 
 typedef struct
 {
-	void* cmdBuf;
+	u32* cmdBuf;
 	size_t cmdBufSize;
 
 	u32 flags;
@@ -49,6 +49,8 @@ typedef struct
 	u16 fixedAttribDirty, fixedAttribEverDirty;
 	C3D_FVec fixedAttribs[12];
 
+	void (* renderQueueExit)(void);
+
 } C3D_Context;
 
 enum
@@ -66,6 +68,7 @@ enum
 	C3DiF_LightEnv = BIT(10),
 	C3DiF_VshCode = BIT(11),
 	C3DiF_GshCode = BIT(12),
+	C3DiF_CmdBuffer = BIT(13),
 
 #define C3DiF_Tex(n) BIT(23+(n))
 	C3DiF_TexAll = 7 << 23,
@@ -91,3 +94,5 @@ 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);
+
+void C3Di_FinalizeFrame(u32** pBuf, u32* pSize);