Decouple attribute and buffer management

examples/gputest/source/main.c

@@ -73,7 +73,7 @@ static void drawScene(float trX, float trY)
 	MtxStack_Update(&projMtx);
 	MtxStack_Update(&mdlvMtx);
 
-	C3D_DrawArray(&myVbo, GPU_TRIANGLES);
+	C3D_DrawArrays(GPU_TRIANGLES, 0, myVbo.vertexCount);
 
 	C3D_Flush();
 }
@@ -96,7 +96,7 @@ static void drawSceneBottom(float trX, float trY)
 	MtxStack_Update(&projMtx);
 	MtxStack_Update(&mdlvMtx);
 
-	C3D_DrawArray(&myVbo, GPU_TRIANGLES);
+	C3D_DrawArrays(GPU_TRIANGLES, 0, myVbo.vertexCount);
 
 	C3D_Flush();
 }
@@ -139,16 +139,20 @@ int main()
 	// Configure attributes
 	C3D_AttrInfo* attrInfo = C3D_GetAttrInfo();
 	AttrInfo_Init(attrInfo);
-	AttrInfo_AddParam(attrInfo, GPU_FLOAT, 3); // position
+	AttrInfo_AddLoader(attrInfo, 0, GPU_FLOAT, 3); // position
-	AttrInfo_AddParam(attrInfo, GPU_FLOAT, 2); // texcoord
+	AttrInfo_AddLoader(attrInfo, 1, GPU_FLOAT, 2); // texcoord
-	AttrInfo_AddParam(attrInfo, GPU_FLOAT, 3); // vertex color
+	AttrInfo_AddLoader(attrInfo, 2, GPU_FLOAT, 3); // vertex color
-	AttrInfo_AddBuffer(attrInfo, 0, sizeof(vertex_t), 3, 0x210);
 
 	// Configure VBO
 	C3D_VBOInit(&myVbo, sizeof(vertices));
 	C3D_VBOAddData(&myVbo, vertices, sizeof(vertices), sizeof(vertices)/sizeof(vertex_t));
 
-	// Clear buffers
+	// Configure buffers
+	C3D_BufInfo* bufInfo = C3D_GetBufInfo();
+	BufInfo_Init(bufInfo);
+	BufInfo_Add(bufInfo, myVbo.data, sizeof(vertex_t), 3, 0x210);
+
+	// Clear renderbuffers
 	C3D_RenderBufClear(&rbTop);
 	C3D_RenderBufClear(&rbBot);
 

include/c3d/attribs.h

@@ -3,29 +3,14 @@
 
 typedef struct
 {
-	u32 offset;
 	u32 flags[2];
-} C3D_AttrBufCfg;
-
-typedef struct
-{
-	// Base physical address intentionally left out
-	// Write to 0x0201 instead of 0x0200
-	u32 flags[2];
-	C3D_AttrBufCfg buffers[12];
-} C3D_AttrCfg;
-
-typedef struct
-{
-	C3D_AttrCfg cfg;
 	u64 permutation;
 	int attrCount;
-	int bufCount;
 } C3D_AttrInfo;
 
 void AttrInfo_Init(C3D_AttrInfo* info);
-bool AttrInfo_AddParam(C3D_AttrInfo* info, GPU_FORMATS format, int count);
+int  AttrInfo_AddLoader(C3D_AttrInfo* info, int regId, GPU_FORMATS format, int count);
-bool AttrInfo_AddBuffer(C3D_AttrInfo* info, ptrdiff_t offset, ptrdiff_t stride, int attribCount, u64 permutation);
+int  AttrInfo_AddFixed(C3D_AttrInfo* info, int regId);
 
 C3D_AttrInfo* C3D_GetAttrInfo(void);
 void C3D_SetAttrInfo(C3D_AttrInfo* info);

include/c3d/base.h

@@ -12,8 +12,8 @@ void C3D_BindProgram(shaderProgram_s* program);
 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_DrawArrays(GPU_Primitive_t primitive, int first, int size);
-void C3D_DrawElements(C3D_IBO* ibo, GPU_Primitive_t primitive);
+void C3D_DrawElements(GPU_Primitive_t primitive, int count, int type, const void* indices);
 
 static inline void C3D_FlushAwait(void)
 {

include/c3d/buffers.h

@@ -1,6 +1,25 @@
 #pragma once
 #include "types.h"
 
+typedef struct
+{
+	u32 offset;
+	u32 flags[2];
+} C3D_BufCfg;
+
+typedef struct
+{
+	u32 base_paddr;
+	int bufCount;
+	C3D_BufCfg buffers[12];
+} C3D_BufInfo;
+
+void BufInfo_Init(C3D_BufInfo* info);
+int  BufInfo_Add(C3D_BufInfo* info, const void* data, ptrdiff_t stride, int attribCount, u64 permutation);
+
+C3D_BufInfo* C3D_GetBufInfo(void);
+void C3D_SetBufInfo(C3D_BufInfo* info);
+
 typedef struct
 {
 	void* data;
@@ -12,7 +31,6 @@ typedef struct
 bool C3D_VBOInit(C3D_VBO* vbo, size_t capacity);
 bool C3D_VBOAddData(C3D_VBO* vbo, const void* data, size_t size, int vertexCount);
 void C3D_VBOFlush(C3D_VBO* vbo);
-void C3D_VBOBind(C3D_VBO* vbo);
 void C3D_VBODelete(C3D_VBO* vbo);
 
 typedef struct

source/attribs.c

@@ -5,32 +5,33 @@
 void AttrInfo_Init(C3D_AttrInfo* info)
 {
 	memset(info, 0, sizeof(*info));
-	info->cfg.flags[1] = 0xFFF << 16;
+	info->flags[1] = 0xFFF << 16;
 }
 
-bool AttrInfo_AddParam(C3D_AttrInfo* info, GPU_FORMATS format, int count)
+int AttrInfo_AddLoader(C3D_AttrInfo* info, int regId, GPU_FORMATS format, int count)
 {
-	if (info->attrCount == 12) return false;
+	if (info->attrCount == 12) return -1;
 	int id = info->attrCount++;
+	if (regId < 0) regId = id;
 	if (id < 8)
-		info->cfg.flags[0] |= GPU_ATTRIBFMT(id, count, format);
+		info->flags[0] |= GPU_ATTRIBFMT(id, count, format);
 	else
-		info->cfg.flags[1] |= GPU_ATTRIBFMT(id-8, count, format);
+		info->flags[1] |= GPU_ATTRIBFMT(id-8, count, format);
-	info->cfg.flags[1] &= ~BIT(id+16);
+
-	info->cfg.flags[1] = (info->cfg.flags[1] &~ 0xF0000000) | (id << 28);
+	info->flags[1] = (info->flags[1] &~ (0xF0000000 | BIT(id+16))) | (id << 28);
-	info->permutation |= id << (id*4);
+	info->permutation |= regId << (id*4);
-	return true;
+	return id;
 }
 
-bool AttrInfo_AddBuffer(C3D_AttrInfo* info, ptrdiff_t offset, ptrdiff_t stride, int attribCount, u64 permutation)
+int AttrInfo_AddFixed(C3D_AttrInfo* info, int regId)
 {
-	if (info->bufCount == 12) return false;
+	if (info->attrCount == 12) return -1;
-	int id = info->bufCount++;
+	int id = info->attrCount++;
-	C3D_AttrBufCfg* buf = &info->cfg.buffers[id];
+	if (regId < 0) regId = id;
-	buf->offset = offset;
+
-	buf->flags[0] = permutation & 0xFFFFFFFF;
+	info->flags[1] = (info->flags[1] &~ 0xF0000000) | (id << 28);
-	buf->flags[1] = (permutation >> 32) | (stride << 16) | (attribCount << 28);
+	info->permutation |= regId << (id*4);
-	return true;
+	return id;
 }
 
 C3D_AttrInfo* C3D_GetAttrInfo(void)
@@ -40,7 +41,7 @@ C3D_AttrInfo* C3D_GetAttrInfo(void)
 	if (!(ctx->flags & C3DiF_Active))
 		return NULL;
 
-	ctx->flags |= C3DiF_AttrBuf;
+	ctx->flags |= C3DiF_AttrInfo;
 	return &ctx->attrInfo;
 }
 
@@ -51,13 +52,14 @@ void C3D_SetAttrInfo(C3D_AttrInfo* info)
 	if (!(ctx->flags & C3DiF_Active))
 		return;
 
+	if (info != &ctx->attrInfo)
 		memcpy(&ctx->attrInfo, info, sizeof(*info));
-	ctx->flags |= C3DiF_AttrBuf;
+	ctx->flags |= C3DiF_AttrInfo;
 }
 
 void C3Di_AttrInfoBind(C3D_AttrInfo* info)
 {
-	GPUCMD_AddIncrementalWrites(GPUREG_ATTRIBBUFFERS_FORMAT_LOW, (u32*)&info->cfg, sizeof(C3D_AttrCfg)/sizeof(u32));
+	GPUCMD_AddIncrementalWrites(GPUREG_ATTRIBBUFFERS_FORMAT_LOW, (u32*)info->flags, sizeof(info->flags)/sizeof(u32));
 	GPUCMD_AddMaskedWrite(GPUREG_VSH_INPUTBUFFER_CONFIG, 0xB, 0xA0000000 | (info->attrCount - 1));
 	GPUCMD_AddWrite(GPUREG_0242, info->attrCount - 1);
 	GPUCMD_AddIncrementalWrites(GPUREG_VSH_ATTRIBUTES_PERMUTATION_LOW, (u32*)&info->permutation, 2);

source/base.c

@@ -47,7 +47,7 @@ static void C3Di_AptEventHook(int hookType, void* param)
 		}
 		case APTHOOK_ONRESTORE:
 		{
-			ctx->flags |= C3DiF_AttrBuf | C3DiF_Effect | C3DiF_RenderBuf
+			ctx->flags |= C3DiF_AttrInfo | C3DiF_BufInfo | C3DiF_Effect | C3DiF_RenderBuf
 				| C3DiF_Viewport | C3DiF_Scissor | C3DiF_Program
 				| C3DiF_TexAll | C3DiF_TexEnvAll;
 			break;
@@ -150,12 +150,18 @@ void C3Di_UpdateContext(void)
 		GPUCMD_AddIncrementalWrites(GPUREG_SCISSORTEST_MODE, ctx->scissor, 3);
 	}
 
-	if (ctx->flags & C3DiF_AttrBuf)
+	if (ctx->flags & C3DiF_AttrInfo)
 	{
-		ctx->flags &= ~C3DiF_AttrBuf;
+		ctx->flags &= ~C3DiF_AttrInfo;
 		C3Di_AttrInfoBind(&ctx->attrInfo);
 	}
 
+	if (ctx->flags & C3DiF_BufInfo)
+	{
+		ctx->flags &= ~C3DiF_BufInfo;
+		C3Di_BufInfoBind(&ctx->bufInfo);
+	}
+
 	if (ctx->flags & C3DiF_Effect)
 	{
 		ctx->flags &= ~C3DiF_Effect;

source/buffers.c

@@ -0,0 +1,55 @@
+#include <c3d/buffers.h>
+#include <string.h>
+#include "context.h"
+
+#define BUFFER_BASE_PADDR 0x18000000
+
+void BufInfo_Init(C3D_BufInfo* info)
+{
+	memset(info, 0, sizeof(*info));
+	info->base_paddr = BUFFER_BASE_PADDR;
+}
+
+int BufInfo_Add(C3D_BufInfo* info, const void* data, ptrdiff_t stride, int attribCount, u64 permutation)
+{
+	if (info->bufCount == 12) return -1;
+	int id = info->bufCount++;
+
+	u32 pa = osConvertVirtToPhys((u32)data);
+	if (pa < info->base_paddr) return -2;
+
+	C3D_BufCfg* buf = &info->buffers[id];
+	buf->offset = pa - info->base_paddr;
+	buf->flags[0] = permutation & 0xFFFFFFFF;
+	buf->flags[1] = (permutation >> 32) | (stride << 16) | (attribCount << 28);
+	return id;
+}
+
+C3D_BufInfo* C3D_GetBufInfo(void)
+{
+	C3D_Context* ctx = C3Di_GetContext();
+
+	if (!(ctx->flags & C3DiF_Active))
+		return NULL;
+
+	ctx->flags |= C3DiF_BufInfo;
+	return &ctx->bufInfo;
+}
+
+void C3D_SetBufInfo(C3D_BufInfo* info)
+{
+	C3D_Context* ctx = C3Di_GetContext();
+
+	if (!(ctx->flags & C3DiF_Active))
+		return;
+
+	if (info != &ctx->bufInfo)
+		memcpy(&ctx->bufInfo, info, sizeof(*info));
+	ctx->flags |= C3DiF_BufInfo;
+}
+
+void C3Di_BufInfoBind(C3D_BufInfo* info)
+{
+	GPUCMD_AddWrite(GPUREG_ATTRIBBUFFERS_LOC, info->base_paddr >> 3);
+	GPUCMD_AddIncrementalWrites(GPUREG_ATTRIBBUFFER0_CONFIG0, (u32*)info->buffers, sizeof(info->buffers)/sizeof(u32));
+}

source/context.h

@@ -1,8 +1,8 @@
 #pragma once
 #include <c3d/base.h>
 #include <c3d/uniforms.h>
-#include <c3d/buffers.h>
 #include <c3d/attribs.h>
+#include <c3d/buffers.h>
 #include <c3d/texenv.h>
 #include <c3d/effect.h>
 #include <c3d/texture.h>
@@ -29,8 +29,8 @@ typedef struct
 	u32 flags;
 	shaderProgram_s* program;
 
-	u32 vboPos;
 	C3D_AttrInfo attrInfo;
+	C3D_BufInfo bufInfo;
 	C3D_Effect effect;
 
 	C3D_Tex* tex[3];
@@ -46,12 +46,13 @@ enum
 {
 	C3DiF_Active = BIT(0),
 	C3DiF_NeedFinishDrawing = BIT(1),
-	C3DiF_AttrBuf = BIT(2),
+	C3DiF_AttrInfo = BIT(2),
-	C3DiF_Effect = BIT(3),
+	C3DiF_BufInfo = BIT(3),
-	C3DiF_RenderBuf = BIT(4),
+	C3DiF_Effect = BIT(4),
-	C3DiF_Viewport = BIT(5),
+	C3DiF_RenderBuf = BIT(5),
-	C3DiF_Scissor = BIT(6),
+	C3DiF_Viewport = BIT(6),
-	C3DiF_Program = BIT(7),
+	C3DiF_Scissor = BIT(7),
+	C3DiF_Program = BIT(8),
 
 #define C3DiF_Tex(n) BIT(23+(n))
 	C3DiF_TexAll = 7 << 23,
@@ -66,8 +67,8 @@ static inline C3D_Context* C3Di_GetContext(void)
 }
 
 void C3Di_UpdateContext(void);
-void C3Di_IBOBind(C3D_IBO* ibo);
 void C3Di_AttrInfoBind(C3D_AttrInfo* info);
+void C3Di_BufInfoBind(C3D_BufInfo* info);
 void C3Di_TexEnvBind(int id, C3D_TexEnv* env);
 void C3Di_EffectBind(C3D_Effect* effect);
 void C3Di_RenderBufBind(C3D_RenderBuf* rb);

source/drawArrays.c

@@ -1,8 +1,7 @@
 #include "context.h"
 
-void C3D_DrawArray(C3D_VBO* vbo, GPU_Primitive_t primitive)
+void C3D_DrawArrays(GPU_Primitive_t primitive, int first, int size)
 {
-	C3D_VBOBind(vbo);
 	C3Di_UpdateContext();
 
 	// Set primitive type
@@ -11,7 +10,9 @@ void C3D_DrawArray(C3D_VBO* vbo, GPU_Primitive_t primitive)
 	// The index buffer is not used, but 0x000F0227 is still required
 	GPUCMD_AddWrite(GPUREG_INDEXBUFFER_CONFIG, 0x80000000);
 	// Number of vertices
-	GPUCMD_AddWrite(GPUREG_NUMVERTICES, vbo->vertexCount);
+	GPUCMD_AddWrite(GPUREG_NUMVERTICES, size);
+	// First vertex
+	GPUCMD_AddWrite(GPUREG_DRAW_VERTEX_OFFSET, first);
 
 	// Unknown commands
 	GPUCMD_AddMaskedWrite(GPUREG_0253, 1, 0x00000001);

source/drawElements.c

@@ -1,16 +1,21 @@
 #include "context.h"
 
-void C3D_DrawElements(C3D_IBO* ibo, GPU_Primitive_t primitive)
+void C3D_DrawElements(GPU_Primitive_t primitive, int count, int type, const void* indices)
 {
+	C3D_Context* ctx = C3Di_GetContext();
+	u32 pa = osConvertVirtToPhys((u32)indices);
+	u32 base = ctx->bufInfo.base_paddr;
+	if (pa < base) return;
+
 	C3Di_UpdateContext();
 
 	// Set primitive type
 	GPUCMD_AddMaskedWrite(GPUREG_PRIMITIVE_CONFIG, 2, primitive);
 	GPUCMD_AddMaskedWrite(GPUREG_025F, 2, 0x00000001);
-	// Bind the IBO
+	// Configure the index buffer
-	C3Di_IBOBind(ibo);
+	GPUCMD_AddWrite(GPUREG_INDEXBUFFER_CONFIG, (pa - base) | (type << 31));
 	// Number of vertices
-	GPUCMD_AddWrite(GPUREG_NUMVERTICES, ibo->indexCount);
+	GPUCMD_AddWrite(GPUREG_NUMVERTICES, count);
 
 	// Unknown commands
 	GPUCMD_AddMaskedWrite(GPUREG_GEOSTAGE_CONFIG, 2, 0x00000100);

source/ibo.c

@@ -30,15 +30,6 @@ void C3D_IBOFlush(C3D_IBO* ibo)
 	GSPGPU_FlushDataCache(ibo->data, ibo->indexCount*stride);
 }
 
-void C3Di_IBOBind(C3D_IBO* ibo)
-{
-	C3D_Context* ctx = C3Di_GetContext();
-
-	u32 pa = osConvertVirtToPhys((u32)ibo->data);
-	u32 offset = (pa - ctx->vboPos) &~ BIT(31);
-	GPUCMD_AddWrite(GPUREG_INDEXBUFFER_CONFIG, offset | (ibo->format << 31));
-}
-
 void C3D_IBODelete(C3D_IBO* ibo)
 {
 	if (!ibo->data) return;

source/renderbuffer.c

@@ -41,7 +41,7 @@ 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(
+	GX_MemoryFill(
 		(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)depthFmtSizes[rb->depthFmt] << 8));
 }
@@ -49,7 +49,7 @@ void C3D_RenderBufClearAsync(C3D_RenderBuf* rb)
 void C3D_RenderBufTransferAsync(C3D_RenderBuf* rb, u32* frameBuf, u32 flags)
 {
 	u32 dim = GX_BUFFER_DIM((u32)rb->width, (u32)rb->height);
-	GX_SetDisplayTransfer((u32*)rb->colorBuf, dim, frameBuf, dim, flags);
+	GX_DisplayTransfer((u32*)rb->colorBuf, dim, frameBuf, dim, flags);
 }
 
 void C3D_RenderBufBind(C3D_RenderBuf* rb)

source/vbo.c

@@ -30,17 +30,6 @@ void C3D_VBOFlush(C3D_VBO* vbo)
 		GSPGPU_FlushDataCache(vbo->data, vbo->size);
 }
 
-void C3D_VBOBind(C3D_VBO* vbo)
-{
-	C3D_Context* ctx = C3Di_GetContext();
-
-	if (!(ctx->flags & C3DiF_Active))
-		return;
-
-	ctx->vboPos = osConvertVirtToPhys((u32)vbo->data);
-	GPUCMD_AddWrite(GPUREG_ATTRIBBUFFERS_LOC, ctx->vboPos >> 3);
-}
-
 void C3D_VBODelete(C3D_VBO* vbo)
 {
 	if (!vbo->data) return;