gl3: orphan im2D/im3D buffers at actual size on Switch

src/gl/gl3immed.cpp

@@ -133,8 +133,19 @@ im2DRenderPrimitive(PrimitiveType primType, void *vertices, int32 numVertices)
 #endif
 
 	glBindBuffer(GL_ARRAY_BUFFER, im2DVbo);
+#ifdef __SWITCH__
+	// Orphan with exactly the size we need, in one call that also uploads.
+	// The stock path orphans the whole STARTVERTICES buffer (~240KB) on every
+	// single 2D draw - even a 4-vertex sprite - and hundreds of those per frame
+	// exhaust nouveau's buffer allocator. It then hands back a buffer with no
+	// backing store and glBufferSubData memcpy's into null (Data Abort at 0).
+	// Orphaning itself must stay: without it the GPU may still be reading the
+	// buffer we overwrite, which mixes geometry between draws.
+	glBufferData(GL_ARRAY_BUFFER, numVertices*sizeof(Im2DVertex), vertices, GL_STREAM_DRAW);
+#else
 	glBufferData(GL_ARRAY_BUFFER, STARTVERTICES*sizeof(Im2DVertex), nil, GL_STREAM_DRAW);
 	glBufferSubData(GL_ARRAY_BUFFER, 0, numVertices*sizeof(Im2DVertex), vertices);
+#endif
 
 	if(im2dOverrideShader)
 		im2dOverrideShader->use();
@@ -163,12 +174,20 @@ im2DRenderIndexedPrimitive(PrimitiveType primType,
 #endif
 
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, im2DIbo);
+#ifdef __SWITCH__
+	glBufferData(GL_ELEMENT_ARRAY_BUFFER, numIndices*2, indices, GL_STREAM_DRAW);
+#else
 	glBufferData(GL_ELEMENT_ARRAY_BUFFER, STARTINDICES*2, nil, GL_STREAM_DRAW);
 	glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, numIndices*2, indices);
+#endif
 
 	glBindBuffer(GL_ARRAY_BUFFER, im2DVbo);
+#ifdef __SWITCH__
+	glBufferData(GL_ARRAY_BUFFER, numVertices*sizeof(Im2DVertex), vertices, GL_STREAM_DRAW);
+#else
 	glBufferData(GL_ARRAY_BUFFER, STARTVERTICES*sizeof(Im2DVertex), nil, GL_STREAM_DRAW);
 	glBufferSubData(GL_ARRAY_BUFFER, 0, numVertices*sizeof(Im2DVertex), vertices);
+#endif
 
 	if(im2dOverrideShader)
 		im2dOverrideShader->use();
@@ -263,8 +282,12 @@ im3DTransform(void *vertices, int32 numVertices, Matrix *world, uint32 flags)
 #endif
 
 	glBindBuffer(GL_ARRAY_BUFFER, im3DVbo);
+#ifdef __SWITCH__
+	glBufferData(GL_ARRAY_BUFFER, numVertices*sizeof(Im3DVertex), vertices, GL_STREAM_DRAW);
+#else
 	glBufferData(GL_ARRAY_BUFFER, STARTVERTICES*sizeof(Im3DVertex), nil, GL_STREAM_DRAW);
 	glBufferSubData(GL_ARRAY_BUFFER, 0, numVertices*sizeof(Im3DVertex), vertices);
+#endif
 #ifndef RW_GL_USE_VAOS
 	setAttribPointers(im3dattribDesc, 3);
 #endif
@@ -284,8 +307,12 @@ void
 im3DRenderIndexedPrimitive(PrimitiveType primType, void *indices, int32 numIndices)
 {
 	glBindBuffer(GL_ELEMENT_ARRAY_BUFFER, im3DIbo);
+#ifdef __SWITCH__
+	glBufferData(GL_ELEMENT_ARRAY_BUFFER, numIndices*2, indices, GL_STREAM_DRAW);
+#else
 	glBufferData(GL_ELEMENT_ARRAY_BUFFER, STARTINDICES*2, nil, GL_STREAM_DRAW);
 	glBufferSubData(GL_ELEMENT_ARRAY_BUFFER, 0, numIndices*2, indices);
+#endif
 
 	flushCache();
 	glDrawElements(primTypeMap[primType], numIndices,

src/gl/gl3raster.cpp

@@ -319,9 +319,15 @@ allocateDXT(Raster *raster, int32 dxt, int32 numLevels, bool32 hasAlpha)
 
 	glGenTextures(1, &natras->texid);
 	uint32 prev = bindTexture(natras->texid);
+#ifndef __SWITCH__
 	glTexImage2D(GL_TEXTURE_2D, 0, natras->internalFormat,
 	             raster->width, raster->height,
 	             0, natras->format, natras->type, nil);
+#else
+	// __SWITCH__: skip this level-0 allocation. rasterUnlock's
+	// glCompressedTexImage2D allocates+uploads level 0 anyway, so pre-allocating
+	// here just wastes a driver call per texture (~0.5ms each on nouveau).
+#endif
 	// TODO: allocate other levels...probably
 	glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, natras->numLevels-1);
 	natras->filterMode = 0;
@@ -946,14 +952,41 @@ readNativeTexture(Stream *stream)
 	natras = GETGL3RASTEREXT(raster);
 	tex->raster = raster;
 
+	int32 uploadLevels = numLevels;
+#ifdef __SWITCH__
+	// switch-mesa/nouveau charges a fixed ~0.7ms per texture-upload driver call
+	// regardless of data size. A txd's long tail of tiny mip levels (each an
+	// 8-16 byte block) then dominates load time and causes streaming hitches.
+	// Upload only the largest RW_SWITCH_TEX_MAXLEVELS level(s); bump this if
+	// distant-texture aliasing becomes noticeable.
+	#ifndef RW_SWITCH_TEX_MAXLEVELS
+	#define RW_SWITCH_TEX_MAXLEVELS 1
+	#endif
+	if(uploadLevels > RW_SWITCH_TEX_MAXLEVELS)
+		uploadLevels = RW_SWITCH_TEX_MAXLEVELS;
+#endif
+
 	uint32 size;
 	uint8 *data;
 	for(int32 i = 0; i < numLevels; i++){
 		size = stream->readU32();
+		if(i >= uploadLevels){
+			// skipped level: consume its bytes, don't upload
+			stream->seek(size);
+			continue;
+		}
 		data = raster->lock(i, Raster::LOCKWRITE|Raster::LOCKNOFETCH);
 		stream->read8(data, size);
 		raster->unlock(i);
 	}
+#ifdef __SWITCH__
+	if(uploadLevels < numLevels){
+		// cap the sampled mip level so undefined higher levels aren't read
+		uint32 prev = bindTexture(natras->texid);
+		glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAX_LEVEL, uploadLevels-1);
+		bindTexture(prev);
+	}
+#endif
 	return tex;
 }