VertexDecoder *DrawEngineCommon::GetVertexDecoder(u32 vtype) {
	auto iter = decoderMap_.find(vtype);
	if (iter != decoderMap_.end())
		return iter->second;
	VertexDecoder *dec = new VertexDecoder();
	dec->SetVertexType(vtype, decOptions_, decJitCache_);
	decoderMap_[vtype] = dec;
	return dec;
}

// Just to get something on the screen, we'll just not subdivide correctly.
void TransformDrawEngine::DrawBezier(int ucount, int vcount) {
	u16 indices[3 * 3 * 6];

	static bool reported = false;
	if (!reported) {
		Reporting::ReportMessage("Unsupported bezier curve");
		reported = true;
	}

	// if (gstate.patchprimitive)
	// Generate indices for a rectangular mesh.
	int c = 0;
	for (int y = 0; y < 3; y++) {
		for (int x = 0; x < 3; x++) {
			indices[c++] = y * 4 + x;
			indices[c++] = y * 4 + x + 1;
			indices[c++] = (y + 1) * 4 + x + 1;
			indices[c++] = (y + 1) * 4 + x + 1;
			indices[c++] = (y + 1) * 4 + x;
			indices[c++] = y * 4 + x;
		}
	}

	// We are free to use the "decoded" buffer here.
	// Let's split it into two to get a second buffer, there's enough space.
	u8 *decoded2 = decoded + 65536 * 24;

	// Alright, now for the vertex data.
	// For now, we will simply inject UVs.

	float customUV[4 * 4 * 2];
	for (int y = 0; y < 4; y++) {
		for (int x = 0; x < 4; x++) {
			customUV[(y * 4 + x) * 2 + 0] = (float)x/3.0f;
			customUV[(y * 4 + x) * 2 + 1] = (float)y/3.0f;
		}
	}

	if (!(gstate.vertType & GE_VTYPE_TC_MASK)) {
		VertexDecoder *dec = GetVertexDecoder(gstate.vertType);
		dec->SetVertexType(gstate.vertType);
		u32 newVertType = dec->InjectUVs(decoded2, Memory::GetPointer(gstate_c.vertexAddr), customUV, 16);
		SubmitPrim(decoded2, &indices[0], GE_PRIM_TRIANGLES, c, newVertType, GE_VTYPE_IDX_16BIT, 0);
	} else {
		SubmitPrim(Memory::GetPointer(gstate_c.vertexAddr), &indices[0], GE_PRIM_TRIANGLES, c, gstate.vertType, GE_VTYPE_IDX_16BIT, 0);
	}
	Flush();  // as our vertex storage here is temporary, it will only survive one draw.
}

void TransformUnit::SubmitPrimitive(void* vertices, void* indices, u32 prim_type, int vertex_count, u32 vertex_type, int *bytesRead)
{
	// TODO: Cache VertexDecoder objects
	VertexDecoder vdecoder;
	VertexDecoderOptions options;
	memset(&options, 0, sizeof(options));
	options.expandAllUVtoFloat = false;
	vdecoder.SetVertexType(vertex_type, options);
	const DecVtxFormat& vtxfmt = vdecoder.GetDecVtxFmt();

	if (bytesRead)
		*bytesRead = vertex_count * vdecoder.VertexSize();

	// Frame skipping.
	if (gstate_c.skipDrawReason & SKIPDRAW_SKIPFRAME) {
		return;
	}

	u16 index_lower_bound = 0;
	u16 index_upper_bound = vertex_count - 1;
	bool indices_16bit = (vertex_type & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	bool indices_32bit = (vertex_type & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_32BIT;
	u8 *indices8 = (u8 *)indices;
	u16 *indices16 = (u16 *)indices;
	u32 *indices32 = (u32 *)indices;
	if (indices)
		GetIndexBounds(indices, vertex_count, vertex_type, &index_lower_bound, &index_upper_bound);
	vdecoder.DecodeVerts(buf, vertices, index_lower_bound, index_upper_bound);

	VertexReader vreader(buf, vtxfmt, vertex_type);

	const int max_vtcs_per_prim = 3;
	int vtcs_per_prim = 0;

	switch (prim_type) {
	case GE_PRIM_POINTS: vtcs_per_prim = 1; break;
	case GE_PRIM_LINES: vtcs_per_prim = 2; break;
	case GE_PRIM_TRIANGLES: vtcs_per_prim = 3; break;
	case GE_PRIM_RECTANGLES: vtcs_per_prim = 2; break;
	}

	VertexData data[max_vtcs_per_prim];

	// TODO: Do this in two passes - first process the vertices (before indexing/stripping),
	// then resolve the indices. This lets us avoid transforming shared vertices twice.

	switch (prim_type) {
	case GE_PRIM_POINTS:
	case GE_PRIM_LINES:
	case GE_PRIM_TRIANGLES:
	case GE_PRIM_RECTANGLES:
		{
			for (int vtx = 0; vtx < vertex_count; vtx += vtcs_per_prim) {
				for (int i = 0; i < vtcs_per_prim; ++i) {
					if (indices) {
						if (indices_32bit) {
							vreader.Goto(indices32[vtx + i]);
						} else if (indices_16bit) {
							vreader.Goto(indices16[vtx + i]);
						} else {
							vreader.Goto(indices8[vtx + i]);
						}
					} else {
						vreader.Goto(vtx+i);
					}

					data[i] = ReadVertex(vreader);
					if (outside_range_flag)
						break;
				}
				if (outside_range_flag) {
					outside_range_flag = false;
					continue;
				}

				switch (prim_type) {
				case GE_PRIM_TRIANGLES:
				{
					if (!gstate.isCullEnabled() || gstate.isModeClear()) {
						Clipper::ProcessTriangle(data[0], data[1], data[2]);
						Clipper::ProcessTriangle(data[2], data[1], data[0]);
					} else if (!gstate.getCullMode())
						Clipper::ProcessTriangle(data[2], data[1], data[0]);
					else
						Clipper::ProcessTriangle(data[0], data[1], data[2]);
					break;
				}

				case GE_PRIM_RECTANGLES:
					Clipper::ProcessRect(data[0], data[1]);
					break;

				case GE_PRIM_LINES:
					Clipper::ProcessLine(data[0], data[1]);
					break;

				case GE_PRIM_POINTS:
					Clipper::ProcessPoint(data[0]);
					break;
				}
//.........这里部分代码省略.........

void TransformUnit::SubmitSpline(void* control_points, void* indices, int count_u, int count_v, int type_u, int type_v, GEPatchPrimType prim_type, u32 vertex_type) {
	VertexDecoder vdecoder;
	VertexDecoderOptions options;
	memset(&options, 0, sizeof(options));
	options.expandAllUVtoFloat = false;
	vdecoder.SetVertexType(vertex_type, options);
	const DecVtxFormat& vtxfmt = vdecoder.GetDecVtxFmt();

	static u8 buf[65536 * 48]; // yolo
	u16 index_lower_bound = 0;
	u16 index_upper_bound = count_u * count_v - 1;
	bool indices_16bit = (vertex_type & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	bool indices_32bit = (vertex_type & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_32BIT;
	u8 *indices8 = (u8 *)indices;
	u16 *indices16 = (u16 *)indices;
	u32 *indices32 = (u32 *)indices;
	if (indices)
		GetIndexBounds(indices, count_u*count_v, vertex_type, &index_lower_bound, &index_upper_bound);
	vdecoder.DecodeVerts(buf, control_points, index_lower_bound, index_upper_bound);

	VertexReader vreader(buf, vtxfmt, vertex_type);

	int num_patches_u = count_u - 3;
	int num_patches_v = count_v - 3;

	if (patchBufferSize_ < num_patches_u * num_patches_v) {
		if (patchBuffer_) {
			FreeAlignedMemory(patchBuffer_);
		}
		patchBuffer_ = (SplinePatch *)AllocateAlignedMemory(num_patches_u * num_patches_v, 16);
		patchBufferSize_ = num_patches_u * num_patches_v;
	}
	SplinePatch *patches = patchBuffer_;

	for (int patch_u = 0; patch_u < num_patches_u; ++patch_u) {
		for (int patch_v = 0; patch_v < num_patches_v; ++patch_v) {
			SplinePatch& patch = patches[patch_u + patch_v * num_patches_u];

			for (int point = 0; point < 16; ++point) {
				int idx = (patch_u + point%4) + (patch_v + point/4) * count_u;
				if (indices) {
					if (indices_32bit) {
						vreader.Goto(indices32[idx]);
					} else if (indices_16bit) {
						vreader.Goto(indices16[idx]);
					} else {
						vreader.Goto(indices8[idx]);
					}
				} else {
					vreader.Goto(idx);
				}

				patch.points[point] = ReadVertex(vreader);
			}
			patch.type = (type_u | (type_v<<2));
			if (patch_u != 0) patch.type &= ~START_OPEN_U;
			if (patch_v != 0) patch.type &= ~START_OPEN_V;
			if (patch_u != num_patches_u-1) patch.type &= ~END_OPEN_U;
			if (patch_v != num_patches_v-1) patch.type &= ~END_OPEN_V;
		}
	}

	for (int patch_idx = 0; patch_idx < num_patches_u*num_patches_v; ++patch_idx) {
		SplinePatch& patch = patches[patch_idx];

		// TODO: Should do actual patch subdivision instead of just drawing the control points!
		const int tile_min_u = (patch.type & START_OPEN_U) ? 0 : 1;
		const int tile_min_v = (patch.type & START_OPEN_V) ? 0 : 1;
		const int tile_max_u = (patch.type & END_OPEN_U) ? 3 : 2;
		const int tile_max_v = (patch.type & END_OPEN_V) ? 3 : 2;
		for (int tile_u = tile_min_u; tile_u < tile_max_u; ++tile_u) {
			for (int tile_v = tile_min_v; tile_v < tile_max_v; ++tile_v) {
				int point_index = tile_u + tile_v*4;

				VertexData v0 = patch.points[point_index];
				VertexData v1 = patch.points[point_index+1];
				VertexData v2 = patch.points[point_index+4];
				VertexData v3 = patch.points[point_index+5];

				// TODO: Backface culling etc
				Clipper::ProcessTriangle(v0, v1, v2);
				Clipper::ProcessTriangle(v2, v1, v0);
				Clipper::ProcessTriangle(v2, v1, v3);
				Clipper::ProcessTriangle(v3, v1, v2);
			}
		}
	}

	host->GPUNotifyDraw();
}

void DrawEngineCommon::SubmitBezier(const void *control_points, const void *indices, int tess_u, int tess_v, int count_u, int count_v, GEPatchPrimType prim_type, bool computeNormals, bool patchFacing, u32 vertType) {
	PROFILE_THIS_SCOPE("bezier");

	DispatchFlush();

	// TODO: Verify correct functionality with < 4.
	if (count_u < 4 || count_v < 4)
		return;

	u16 index_lower_bound = 0;
	u16 index_upper_bound = count_u * count_v - 1;
	bool indices_16bit = (vertType & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	const u8* indices8 = (const u8*)indices;
	const u16* indices16 = (const u16*)indices;
	if (indices)
		GetIndexBounds(indices, count_u*count_v, vertType, &index_lower_bound, &index_upper_bound);

	// Simplify away bones and morph before proceeding
	// There are normally not a lot of control points so just splitting decoded should be reasonably safe, although not great.
	SimpleVertex *simplified_control_points = (SimpleVertex *)(decoded + 65536 * 12);
	u8 *temp_buffer = decoded + 65536 * 18;

	u32 origVertType = vertType;
	vertType = NormalizeVertices((u8 *)simplified_control_points, temp_buffer, (u8 *)control_points, index_lower_bound, index_upper_bound, vertType);

	VertexDecoder *vdecoder = GetVertexDecoder(vertType);

	int vertexSize = vdecoder->VertexSize();
	if (vertexSize != sizeof(SimpleVertex)) {
		ERROR_LOG(G3D, "Something went really wrong, vertex size: %i vs %i", vertexSize, (int)sizeof(SimpleVertex));
	}

	// Bezier patches share less control points than spline patches. Otherwise they are pretty much the same (except bezier don't support the open/close thing)
	int num_patches_u = (count_u - 1) / 3;
	int num_patches_v = (count_v - 1) / 3;
	BezierPatch* patches = new BezierPatch[num_patches_u * num_patches_v];
	for (int patch_u = 0; patch_u < num_patches_u; patch_u++) {
		for (int patch_v = 0; patch_v < num_patches_v; patch_v++) {
			BezierPatch& patch = patches[patch_u + patch_v * num_patches_u];
			for (int point = 0; point < 16; ++point) {
				int idx = (patch_u * 3 + point % 4) + (patch_v * 3 + point / 4) * count_u;
				if (indices)
					patch.points[point] = simplified_control_points + (indices_16bit ? indices16[idx] : indices8[idx]);
				else
					patch.points[point] = simplified_control_points + idx;
			}
			patch.u_index = patch_u * 3;
			patch.v_index = patch_v * 3;
			patch.index = patch_v * num_patches_u + patch_u;
			patch.primType = prim_type;
			patch.computeNormals = computeNormals;
			patch.patchFacing = patchFacing;
		}
	}

	int count = 0;
	u8 *dest = splineBuffer;

	// Simple approximation of the real tesselation factor.
	// We shouldn't really split up into separate 4x4 patches, instead we should do something that works
	// like the splines, so we subdivide across the whole "mega-patch".
	if (num_patches_u == 0) num_patches_u = 1;
	if (num_patches_v == 0) num_patches_v = 1;
	if (tess_u < 4) tess_u = 4;
	if (tess_v < 4) tess_v = 4;

	u16 *inds = quadIndices_;
	int maxVertices = SPLINE_BUFFER_SIZE / vertexSize;
	for (int patch_idx = 0; patch_idx < num_patches_u*num_patches_v; ++patch_idx) {
		BezierPatch& patch = patches[patch_idx];
		TesselateBezierPatch(dest, inds, count, tess_u, tess_v, patch, origVertType, maxVertices);
	}
	delete[] patches;

	u32 vertTypeWithIndex16 = (vertType & ~GE_VTYPE_IDX_MASK) | GE_VTYPE_IDX_16BIT;

	UVScale prevUVScale;
	if (g_Config.bPrescaleUV) {
		// We scaled during Normalize already so let's turn it off when drawing.
		prevUVScale = gstate_c.uv;
		gstate_c.uv.uScale = 1.0f;
		gstate_c.uv.vScale = 1.0f;
		gstate_c.uv.uOff = 0;
		gstate_c.uv.vOff = 0;
	}

	int bytesRead;
	DispatchSubmitPrim(splineBuffer, quadIndices_, primType[prim_type], count, vertTypeWithIndex16, &bytesRead);

	DispatchFlush();

	if (g_Config.bPrescaleUV) {
		gstate_c.uv = prevUVScale;
	}
}

// TODO: This probably is not the best interface.
bool TransformUnit::GetCurrentSimpleVertices(int count, std::vector<GPUDebugVertex> &vertices, std::vector<u16> &indices) {
	// This is always for the current vertices.
	u16 indexLowerBound = 0;
	u16 indexUpperBound = count - 1;

	if ((gstate.vertType & GE_VTYPE_IDX_MASK) != GE_VTYPE_IDX_NONE) {
		const u8 *inds = Memory::GetPointer(gstate_c.indexAddr);
		const u16 *inds16 = (const u16 *)inds;

		if (inds) {
			GetIndexBounds(inds, count, gstate.vertType, &indexLowerBound, &indexUpperBound);
			indices.resize(count);
			switch (gstate.vertType & GE_VTYPE_IDX_MASK) {
			case GE_VTYPE_IDX_16BIT:
				for (int i = 0; i < count; ++i) {
					indices[i] = inds16[i];
				}
				break;
			case GE_VTYPE_IDX_8BIT:
				for (int i = 0; i < count; ++i) {
					indices[i] = inds[i];
				}
				break;
			default:
				return false;
			}
		} else {
			indices.clear();
		}
	} else {
		indices.clear();
	}

	static std::vector<u32> temp_buffer;
	static std::vector<SimpleVertex> simpleVertices;
	temp_buffer.resize(65536 * 24 / sizeof(u32));
	simpleVertices.resize(indexUpperBound + 1);

	VertexDecoder vdecoder;
	VertexDecoderOptions options;
	memset(&options, 0, sizeof(options));
	options.expandAllUVtoFloat = false;  // TODO: True should be fine here
	vdecoder.SetVertexType(gstate.vertType, options);
	DrawEngineCommon::NormalizeVertices((u8 *)(&simpleVertices[0]), (u8 *)(&temp_buffer[0]), Memory::GetPointer(gstate_c.vertexAddr), &vdecoder, indexLowerBound, indexUpperBound, gstate.vertType);

	float world[16];
	float view[16];
	float worldview[16];
	float worldviewproj[16];
	ConvertMatrix4x3To4x4(world, gstate.worldMatrix);
	ConvertMatrix4x3To4x4(view, gstate.viewMatrix);
	Matrix4ByMatrix4(worldview, world, view);
	Matrix4ByMatrix4(worldviewproj, worldview, gstate.projMatrix);

	vertices.resize(indexUpperBound + 1);
	for (int i = indexLowerBound; i <= indexUpperBound; ++i) {
		const SimpleVertex &vert = simpleVertices[i];

		if (gstate.isModeThrough()) {
			if (gstate.vertType & GE_VTYPE_TC_MASK) {
				vertices[i].u = vert.uv[0];
				vertices[i].v = vert.uv[1];
			} else {
				vertices[i].u = 0.0f;
				vertices[i].v = 0.0f;
			}
			vertices[i].x = vert.pos.x;
			vertices[i].y = vert.pos.y;
			vertices[i].z = vert.pos.z;
			if (gstate.vertType & GE_VTYPE_COL_MASK) {
				memcpy(vertices[i].c, vert.color, sizeof(vertices[i].c));
			} else {
				memset(vertices[i].c, 0, sizeof(vertices[i].c));
			}
		} else {
			float clipPos[4];
			Vec3ByMatrix44(clipPos, vert.pos.AsArray(), worldviewproj);
			ScreenCoords screenPos = ClipToScreen(clipPos);
			DrawingCoords drawPos = ScreenToDrawing(screenPos);

			if (gstate.vertType & GE_VTYPE_TC_MASK) {
				vertices[i].u = vert.uv[0];
				vertices[i].v = vert.uv[1];
			} else {
				vertices[i].u = 0.0f;
				vertices[i].v = 0.0f;
			}
			vertices[i].x = drawPos.x;
			vertices[i].y = drawPos.y;
			vertices[i].z = 1.0;
			if (gstate.vertType & GE_VTYPE_COL_MASK) {
				memcpy(vertices[i].c, vert.color, sizeof(vertices[i].c));
			} else {
				memset(vertices[i].c, 0, sizeof(vertices[i].c));
			}
		}
	}

	return true;
//.........这里部分代码省略.........

void TransformDrawEngine::SubmitSpline(void* control_points, void* indices, int count_u, int count_v, int type_u, int type_v, u32 prim_type, u32 vertex_type)
{
	Flush();

	if (prim_type != GE_PATCHPRIM_TRIANGLES) {
		// Only triangles supported!
		return;
	}

	// We're not actually going to decode, only reshuffle.
	VertexDecoder vdecoder;
	vdecoder.SetVertexType(vertex_type);

	int undecodedVertexSize = vdecoder.VertexSize();

	const DecVtxFormat& vtxfmt = vdecoder.GetDecVtxFmt();

	u16 index_lower_bound = 0;
	u16 index_upper_bound = count_u * count_v - 1;
	bool indices_16bit = (vertex_type & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	u8* indices8 = (u8*)indices;
	u16* indices16 = (u16*)indices;
	if (indices)
		GetIndexBounds(indices, count_u*count_v, vertex_type, &index_lower_bound, &index_upper_bound);

	int num_patches_u = count_u - 3;
	int num_patches_v = count_v - 3;

	// TODO: Do something less idiotic to manage this buffer
	HWSplinePatch* patches = new HWSplinePatch[num_patches_u * num_patches_v];
	for (int patch_u = 0; patch_u < num_patches_u; ++patch_u) {
		for (int patch_v = 0; patch_v < num_patches_v; ++patch_v) {
			HWSplinePatch& patch = patches[patch_u + patch_v * num_patches_u];

			for (int point = 0; point < 16; ++point) {
				int idx = (patch_u + point%4) + (patch_v + point/4) * count_u;
				if (indices)
					patch.points[point] = (u8 *)control_points + undecodedVertexSize * (indices_16bit ? indices16[idx] : indices8[idx]);
				else
					patch.points[point] = (u8 *)control_points + undecodedVertexSize * idx;
			}
			patch.type = (type_u | (type_v<<2));
			if (patch_u != 0) patch.type &= ~START_OPEN_U;
			if (patch_v != 0) patch.type &= ~START_OPEN_V;
			if (patch_u != num_patches_u-1) patch.type &= ~END_OPEN_U;
			if (patch_v != num_patches_v-1) patch.type &= ~END_OPEN_V;
		}
	}

	u8 *decoded2 = decoded + 65536 * 24;

	int count = 0;
	u8 *dest = decoded2;

	for (int patch_idx = 0; patch_idx < num_patches_u*num_patches_v; ++patch_idx) {
		HWSplinePatch& patch = patches[patch_idx];

		// TODO: Should do actual patch subdivision instead of just drawing the control points!
		const int tile_min_u = (patch.type & START_OPEN_U) ? 0 : 1;
		const int tile_min_v = (patch.type & START_OPEN_V) ? 0 : 1;
		const int tile_max_u = (patch.type & END_OPEN_U) ? 3 : 2;
		const int tile_max_v = (patch.type & END_OPEN_V) ? 3 : 2;
		for (int tile_u = tile_min_u; tile_u < tile_max_u; ++tile_u) {
			for (int tile_v = tile_min_v; tile_v < tile_max_v; ++tile_v) {
				int point_index = tile_u + tile_v*4;

				u8 *v0 = patch.points[point_index];
				u8 *v1 = patch.points[point_index+1];
				u8 *v2 = patch.points[point_index+4];
				u8 *v3 = patch.points[point_index+5];

				// TODO: Insert UVs where applicable. Actually subdivide.
				CopyTriangle(dest, v0, v1, v2, undecodedVertexSize);
				CopyTriangle(dest, v2, v1, v0, undecodedVertexSize);
				CopyTriangle(dest, v2, v1, v3, undecodedVertexSize);
				CopyTriangle(dest, v3, v1, v2, undecodedVertexSize);
				count += 12;
			}
		}
	}
	delete[] patches;

	u32 vertTypeWithoutIndex = vertex_type & ~GE_VTYPE_IDX_MASK;

	SubmitPrim(decoded2, 0, GE_PRIM_TRIANGLES, count, vertTypeWithoutIndex, GE_VTYPE_IDX_NONE, 0);
	Flush();
}

// TODO: This probably is not the best interface.
// Also, we should try to merge this into the similar function in DrawEngineCommon.
bool TransformUnit::GetCurrentSimpleVertices(int count, std::vector<GPUDebugVertex> &vertices, std::vector<u16> &indices) {
	// This is always for the current vertices.
	u16 indexLowerBound = 0;
	u16 indexUpperBound = count - 1;

	if ((gstate.vertType & GE_VTYPE_IDX_MASK) != GE_VTYPE_IDX_NONE) {
		const u8 *inds = Memory::GetPointer(gstate_c.indexAddr);
		const u16 *inds16 = (const u16 *)inds;
		const u32 *inds32 = (const u32 *)inds;

		if (inds) {
			GetIndexBounds(inds, count, gstate.vertType, &indexLowerBound, &indexUpperBound);
			indices.resize(count);
			switch (gstate.vertType & GE_VTYPE_IDX_MASK) {
			case GE_VTYPE_IDX_8BIT:
				for (int i = 0; i < count; ++i) {
					indices[i] = inds[i];
				}
				break;
			case GE_VTYPE_IDX_16BIT:
				for (int i = 0; i < count; ++i) {
					indices[i] = inds16[i];
				}
				break;
			case GE_VTYPE_IDX_32BIT:
				WARN_LOG_REPORT_ONCE(simpleIndexes32, G3D, "SimpleVertices: Decoding 32-bit indexes");
				for (int i = 0; i < count; ++i) {
					// These aren't documented and should be rare.  Let's bounds check each one.
					if (inds32[i] != (u16)inds32[i]) {
						ERROR_LOG_REPORT_ONCE(simpleIndexes32Bounds, G3D, "SimpleVertices: Index outside 16-bit range");
					}
					indices[i] = (u16)inds32[i];
				}
				break;
			}
		} else {
			indices.clear();
		}
	} else {
		indices.clear();
	}

	static std::vector<u32> temp_buffer;
	static std::vector<SimpleVertex> simpleVertices;
	temp_buffer.resize(65536 * 24 / sizeof(u32));
	simpleVertices.resize(indexUpperBound + 1);

	VertexDecoder vdecoder;
	VertexDecoderOptions options{};
	vdecoder.SetVertexType(gstate.vertType, options);
	DrawEngineCommon::NormalizeVertices((u8 *)(&simpleVertices[0]), (u8 *)(&temp_buffer[0]), Memory::GetPointer(gstate_c.vertexAddr), &vdecoder, indexLowerBound, indexUpperBound, gstate.vertType);

	float world[16];
	float view[16];
	float worldview[16];
	float worldviewproj[16];
	ConvertMatrix4x3To4x4(world, gstate.worldMatrix);
	ConvertMatrix4x3To4x4(view, gstate.viewMatrix);
	Matrix4ByMatrix4(worldview, world, view);
	Matrix4ByMatrix4(worldviewproj, worldview, gstate.projMatrix);

	vertices.resize(indexUpperBound + 1);
	for (int i = indexLowerBound; i <= indexUpperBound; ++i) {
		const SimpleVertex &vert = simpleVertices[i];

		if (gstate.isModeThrough()) {
			if (gstate.vertType & GE_VTYPE_TC_MASK) {
				vertices[i].u = vert.uv[0];
				vertices[i].v = vert.uv[1];
			} else {
				vertices[i].u = 0.0f;
				vertices[i].v = 0.0f;
			}
			vertices[i].x = vert.pos.x;
			vertices[i].y = vert.pos.y;
			vertices[i].z = vert.pos.z;
			if (gstate.vertType & GE_VTYPE_COL_MASK) {
				memcpy(vertices[i].c, vert.color, sizeof(vertices[i].c));
			} else {
				memset(vertices[i].c, 0, sizeof(vertices[i].c));
			}
		} else {
			float clipPos[4];
			Vec3ByMatrix44(clipPos, vert.pos.AsArray(), worldviewproj);
			ScreenCoords screenPos = ClipToScreen(clipPos);
			DrawingCoords drawPos = ScreenToDrawing(screenPos);

			if (gstate.vertType & GE_VTYPE_TC_MASK) {
				vertices[i].u = vert.uv[0] * (float)gstate.getTextureWidth(0);
				vertices[i].v = vert.uv[1] * (float)gstate.getTextureHeight(0);
			} else {
				vertices[i].u = 0.0f;
				vertices[i].v = 0.0f;
			}
			vertices[i].x = drawPos.x;
			vertices[i].y = drawPos.y;
			vertices[i].z = drawPos.z;
			if (gstate.vertType & GE_VTYPE_COL_MASK) {
//.........这里部分代码省略.........

JittedVertexDecoder VertexDecoderJitCache::Compile(const VertexDecoder &dec) {
	dec_ = &dec;
	const u8 *start = this->GetCodePtr();

#ifdef _M_IX86
	// Store register values
	PUSH(ESI);
	PUSH(EDI);
	PUSH(EBX);
	PUSH(EBP);

	// Read parameters
	int offset = 4;
	MOV(32, R(srcReg), MDisp(ESP, 16 + offset + 0));
	MOV(32, R(dstReg), MDisp(ESP, 16 + offset + 4));
	MOV(32, R(counterReg), MDisp(ESP, 16 + offset + 8));
#endif

	// Save XMM4/XMM5 which apparently can be problematic?
	// Actually, if they are, it must be a compiler bug because they SHOULD be ok.
	// So I won't bother.
	SUB(PTRBITS, R(ESP), Imm8(64));
	MOVUPS(MDisp(ESP, 0), XMM4);
	MOVUPS(MDisp(ESP, 16), XMM5);
	MOVUPS(MDisp(ESP, 32), XMM6);
	MOVUPS(MDisp(ESP, 48), XMM7);

	bool prescaleStep = false;
	// Look for prescaled texcoord steps
	for (int i = 0; i < dec.numSteps_; i++) {
		if (dec.steps_[i] == &VertexDecoder::Step_TcU8Prescale ||
			dec.steps_[i] == &VertexDecoder::Step_TcU16Prescale ||
			dec.steps_[i] == &VertexDecoder::Step_TcFloatPrescale) {
				prescaleStep = true;
		}
	}

	// Add code to convert matrices to 4x4.
	// Later we might want to do this when the matrices are loaded instead.
	// This is mostly proof of concept.
	int boneCount = 0;
	if (dec.weighttype && g_Config.bSoftwareSkinning) {
		for (int i = 0; i < 8; i++) {
			MOVUPS(XMM0, M((gstate.boneMatrix + 12 * i)));
			MOVUPS(XMM1, M((gstate.boneMatrix + 12 * i + 3)));
			MOVUPS(XMM2, M((gstate.boneMatrix + 12 * i + 3 * 2)));
			MOVUPS(XMM3, M((gstate.boneMatrix + 12 * i + 3 * 3)));
			ANDPS(XMM0, M(&threeMasks));
			ANDPS(XMM1, M(&threeMasks));
			ANDPS(XMM2, M(&threeMasks));
			ANDPS(XMM3, M(&threeMasks));
			ORPS(XMM3, M(&aOne));
			MOVAPS(M((bones + 16 * i)), XMM0);
			MOVAPS(M((bones + 16 * i + 4)), XMM1);
			MOVAPS(M((bones + 16 * i + 8)), XMM2);
			MOVAPS(M((bones + 16 * i + 12)), XMM3);
		}
	}

	// Keep the scale/offset in a few fp registers if we need it.
	if (prescaleStep) {
#ifdef _M_X64
		MOV(64, R(tempReg1), Imm64((u64)(&gstate_c.uv)));
#else
		MOV(32, R(tempReg1), Imm32((u32)(&gstate_c.uv)));
#endif
		MOVSS(fpScaleOffsetReg, MDisp(tempReg1, 0));
		MOVSS(fpScratchReg, MDisp(tempReg1, 4));
		UNPCKLPS(fpScaleOffsetReg, R(fpScratchReg));
		if ((dec.VertexType() & GE_VTYPE_TC_MASK) == GE_VTYPE_TC_8BIT) {
			MULPS(fpScaleOffsetReg, M(&by128));
		} else if ((dec.VertexType() & GE_VTYPE_TC_MASK) == GE_VTYPE_TC_16BIT) {
			MULPS(fpScaleOffsetReg, M(&by32768));
		}
		MOVSS(fpScratchReg, MDisp(tempReg1, 8));
		MOVSS(fpScratchReg2, MDisp(tempReg1, 12));
		UNPCKLPS(fpScratchReg, R(fpScratchReg2));
		UNPCKLPD(fpScaleOffsetReg, R(fpScratchReg));
	}

	// Let's not bother with a proper stack frame. We just grab the arguments and go.
	JumpTarget loopStart = GetCodePtr();
	for (int i = 0; i < dec.numSteps_; i++) {
		if (!CompileStep(dec, i)) {
			// Reset the code ptr and return zero to indicate that we failed.
			SetCodePtr(const_cast<u8 *>(start));
			return 0;
		}
	}

	ADD(PTRBITS, R(srcReg), Imm32(dec.VertexSize()));
	ADD(PTRBITS, R(dstReg), Imm32(dec.decFmt.stride));
	SUB(32, R(counterReg), Imm8(1));
	J_CC(CC_NZ, loopStart, true);

	MOVUPS(XMM4, MDisp(ESP, 0));
	MOVUPS(XMM5, MDisp(ESP, 16));
	MOVUPS(XMM6, MDisp(ESP, 32));
	MOVUPS(XMM7, MDisp(ESP, 48));
	ADD(PTRBITS, R(ESP), Imm8(64));
//.........这里部分代码省略.........

void DrawEngineCommon::SubmitBezier(const void *control_points, const void *indices, int tess_u, int tess_v, int count_u, int count_v, GEPatchPrimType prim_type, bool computeNormals, bool patchFacing, u32 vertType, int *bytesRead) {
	PROFILE_THIS_SCOPE("bezier");

	DispatchFlush();

	u16 index_lower_bound = 0;
	u16 index_upper_bound = count_u * count_v - 1;
	IndexConverter idxConv(vertType, indices);
	if (indices)
		GetIndexBounds(indices, count_u*count_v, vertType, &index_lower_bound, &index_upper_bound);

	VertexDecoder *origVDecoder = GetVertexDecoder((vertType & 0xFFFFFF) | (gstate.getUVGenMode() << 24));
	*bytesRead = count_u * count_v * origVDecoder->VertexSize();

	// Real hardware seems to draw nothing when given < 4 either U or V.
	// This would result in num_patches_u / num_patches_v being 0.
	if (count_u < 4 || count_v < 4) {
		return;
	}

	// Simplify away bones and morph before proceeding
	// There are normally not a lot of control points so just splitting decoded should be reasonably safe, although not great.
	SimpleVertex *simplified_control_points = (SimpleVertex *)(decoded + 65536 * 12);
	u8 *temp_buffer = decoded + 65536 * 18;

	u32 origVertType = vertType;
	vertType = NormalizeVertices((u8 *)simplified_control_points, temp_buffer, (u8 *)control_points, index_lower_bound, index_upper_bound, vertType);

	VertexDecoder *vdecoder = GetVertexDecoder(vertType);

	int vertexSize = vdecoder->VertexSize();
	if (vertexSize != sizeof(SimpleVertex)) {
		ERROR_LOG(G3D, "Something went really wrong, vertex size: %i vs %i", vertexSize, (int)sizeof(SimpleVertex));
	}

	float *pos = (float*)(decoded + 65536 * 18); // Size 4 float
	float *tex = pos + count_u * count_v * 4; // Size 4 float
	float *col = tex + count_u * count_v * 4; // Size 4 float
	const bool hasColor = (origVertType & GE_VTYPE_COL_MASK) != 0;
	const bool hasTexCoords = (origVertType & GE_VTYPE_TC_MASK) != 0;

	// Bezier patches share less control points than spline patches. Otherwise they are pretty much the same (except bezier don't support the open/close thing)
	int num_patches_u = (count_u - 1) / 3;
	int num_patches_v = (count_v - 1) / 3;
	BezierPatch *patches = nullptr;
	if (g_Config.bHardwareTessellation && g_Config.bHardwareTransform && !g_Config.bSoftwareRendering) {
		int posStride, texStride, colStride;
		tessDataTransfer->PrepareBuffers(pos, tex, col, posStride, texStride, colStride, count_u * count_v, hasColor, hasTexCoords);
		float *p = pos;
		float *t = tex;
		float *c = col;
		for (int idx = 0; idx < count_u * count_v; idx++) {
			SimpleVertex *point = simplified_control_points + (indices ? idxConv.convert(idx) : idx);
			memcpy(p, point->pos.AsArray(), 3 * sizeof(float));
			p += posStride;
			if (hasTexCoords) {
				memcpy(t, point->uv, 2 * sizeof(float));
				t += texStride;
			}
			if (hasColor) {
				memcpy(c, Vec4f::FromRGBA(point->color_32).AsArray(), 4 * sizeof(float));
				c += colStride;
			}
		}
		if (!hasColor) {
			SimpleVertex *point = simplified_control_points + (indices ? idxConv.convert(0) : 0);
			memcpy(col, Vec4f::FromRGBA(point->color_32).AsArray(), 4 * sizeof(float));
		}
	} else {
		patches = new BezierPatch[num_patches_u * num_patches_v];
		for (int patch_u = 0; patch_u < num_patches_u; patch_u++) {
			for (int patch_v = 0; patch_v < num_patches_v; patch_v++) {
				BezierPatch& patch = patches[patch_u + patch_v * num_patches_u];
				for (int point = 0; point < 16; ++point) {
					int idx = (patch_u * 3 + point % 4) + (patch_v * 3 + point / 4) * count_u;
					patch.points[point] = simplified_control_points + (indices ? idxConv.convert(idx) : idx);
				}
				patch.u_index = patch_u * 3;
				patch.v_index = patch_v * 3;
				patch.index = patch_v * num_patches_u + patch_u;
				patch.primType = prim_type;
				patch.computeNormals = computeNormals;
				patch.patchFacing = patchFacing;
			}
		}
	}

	int count = 0;
	u8 *dest = splineBuffer;

	// We shouldn't really split up into separate 4x4 patches, instead we should do something that works
	// like the splines, so we subdivide across the whole "mega-patch".

	// If specified as 0, uses 1.
	if (tess_u < 1) {
		tess_u = 1;
	}
	if (tess_v < 1) {
		tess_v = 1;
	}
//.........这里部分代码省略.........

void TransformDrawEngine::SubmitBezier(void* control_points, void* indices, int count_u, int count_v, GEPatchPrimType prim_type, u32 vertType) {
	Flush();

	if (prim_type != GE_PATCHPRIM_TRIANGLES) {
		// Only triangles supported!
		return;
	}

	u16 index_lower_bound = 0;
	u16 index_upper_bound = count_u * count_v - 1;
	bool indices_16bit = (vertType & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	const u8* indices8 = (const u8*)indices;
	const u16* indices16 = (const u16*)indices;
	if (indices)
		GetIndexBounds(indices, count_u*count_v, vertType, &index_lower_bound, &index_upper_bound);

	// Simplify away bones and morph before proceeding
	SimpleVertex *simplified_control_points = (SimpleVertex *)(decoded + 65536 * 12);
	u8 *temp_buffer = decoded + 65536 * 24;

	u32 origVertType = vertType;
	vertType = NormalizeVertices((u8 *)simplified_control_points, temp_buffer, (u8 *)control_points, index_lower_bound, index_upper_bound, vertType);

	VertexDecoder *vdecoder = GetVertexDecoder(vertType);

	int vertexSize = vdecoder->VertexSize();
	if (vertexSize != sizeof(SimpleVertex)) {
		ERROR_LOG(G3D, "Something went really wrong, vertex size: %i vs %i", vertexSize, (int)sizeof(SimpleVertex));
	}
	const DecVtxFormat& vtxfmt = vdecoder->GetDecVtxFmt();

	// Bezier patches share less control points than spline patches. Otherwise they are pretty much the same (except bezier don't support the open/close thing)
	int num_patches_u = (count_u - 1) / 3;
	int num_patches_v = (count_v - 1) / 3;
	BezierPatch* patches = new BezierPatch[num_patches_u * num_patches_v];
	for (int patch_u = 0; patch_u < num_patches_u; patch_u++) {
		for (int patch_v = 0; patch_v < num_patches_v; patch_v++) {
			BezierPatch& patch = patches[patch_u + patch_v * num_patches_u];
			for (int point = 0; point < 16; ++point) {
				int idx = (patch_u * 3 + point%4) + (patch_v * 3 + point/4) * count_u;
				if (indices)
					patch.points[point] = simplified_control_points + (indices_16bit ? indices16[idx] : indices8[idx]);
				else
					patch.points[point] = simplified_control_points + idx;
			}
			patch.u_index = patch_u * 3;
			patch.v_index = patch_v * 3;
		}
	}

	u8 *decoded2 = decoded + 65536 * 36;

	int count = 0;
	u8 *dest = decoded2;

	for (int patch_idx = 0; patch_idx < num_patches_u*num_patches_v; ++patch_idx) {
		BezierPatch& patch = patches[patch_idx];
		TesselateBezierPatch(dest, count, patch, origVertType);
	}
	delete[] patches;

	u32 vertTypeWithIndex16 = (vertType & ~GE_VTYPE_IDX_MASK) | GE_VTYPE_IDX_16BIT;

	SubmitPrim(decoded2, quadIndices_, GE_PRIM_TRIANGLES, count, vertTypeWithIndex16, -1, 0);
	Flush();
}

// This normalizes a set of vertices in any format to SimpleVertex format, by processing away morphing AND skinning.
// The rest of the transform pipeline like lighting will go as normal, either hardware or software.
// The implementation is initially a bit inefficient but shouldn't be a big deal.
// An intermediate buffer of not-easy-to-predict size is stored at bufPtr.
u32 TransformDrawEngine::NormalizeVertices(u8 *outPtr, u8 *bufPtr, const u8 *inPtr, int lowerBound, int upperBound, u32 vertType) {
	// First, decode the vertices into a GPU compatible format. This step can be eliminated but will need a separate
	// implementation of the vertex decoder.
	VertexDecoder *dec = GetVertexDecoder(vertType);
	dec->DecodeVerts(bufPtr, inPtr, lowerBound, upperBound);

	// OK, morphing eliminated but bones still remain to be taken care of.
	// Let's do a partial software transform where we only do skinning.

	VertexReader reader(bufPtr, dec->GetDecVtxFmt(), vertType);

	SimpleVertex *sverts = (SimpleVertex *)outPtr;	

	const u8 defaultColor[4] = {
		(u8)gstate.getMaterialAmbientR(),
		(u8)gstate.getMaterialAmbientG(),
		(u8)gstate.getMaterialAmbientB(),
		(u8)gstate.getMaterialAmbientA(),
	};

	// Let's have two separate loops, one for non skinning and one for skinning.
	if ((vertType & GE_VTYPE_WEIGHT_MASK) != GE_VTYPE_WEIGHT_NONE) {
		int numBoneWeights = vertTypeGetNumBoneWeights(vertType);
		for (int i = lowerBound; i <= upperBound; i++) {
			reader.Goto(i);
			SimpleVertex &sv = sverts[i];
			if (vertType & GE_VTYPE_TC_MASK) {
				reader.ReadUV(sv.uv);
			}

			if (vertType & GE_VTYPE_COL_MASK) {
				reader.ReadColor0_8888(sv.color);
			} else {
				memcpy(sv.color, defaultColor, 4);
			}

			float nrm[3], pos[3];
			float bnrm[3], bpos[3];

			if (vertType & GE_VTYPE_NRM_MASK) {
				// Normals are generated during tesselation anyway, not sure if any need to supply
				reader.ReadNrm(nrm);
			} else {
				nrm[0] = 0;
				nrm[1] = 0;
				nrm[2] = 1.0f;
			}
			reader.ReadPos(pos);

			// Apply skinning transform directly
			float weights[8];
			reader.ReadWeights(weights);
			// Skinning
			Vec3f psum(0,0,0);
			Vec3f nsum(0,0,0);
			for (int i = 0; i < numBoneWeights; i++) {
				if (weights[i] != 0.0f) {
					Vec3ByMatrix43(bpos, pos, gstate.boneMatrix+i*12);
					Vec3f tpos(bpos);
					psum += tpos * weights[i];

					Norm3ByMatrix43(bnrm, nrm, gstate.boneMatrix+i*12);
					Vec3f tnorm(bnrm);
					nsum += tnorm * weights[i];
				}
			}
			sv.pos = psum;
			sv.nrm = nsum;
		}
	} else {
		for (int i = lowerBound; i <= upperBound; i++) {
			reader.Goto(i);
			SimpleVertex &sv = sverts[i];
			if (vertType & GE_VTYPE_TC_MASK) {
				reader.ReadUV(sv.uv);
			} else {
				sv.uv[0] = 0;  // This will get filled in during tesselation
				sv.uv[1] = 0;
			}
			if (vertType & GE_VTYPE_COL_MASK) {
				reader.ReadColor0_8888(sv.color);
			} else {
				memcpy(sv.color, defaultColor, 4);
			}
			if (vertType & GE_VTYPE_NRM_MASK) {
				// Normals are generated during tesselation anyway, not sure if any need to supply
				reader.ReadNrm((float *)&sv.nrm);
			} else {
				sv.nrm.x = 0;
				sv.nrm.y = 0;
				sv.nrm.z = 1.0f;
			}
			reader.ReadPos((float *)&sv.pos);
		}
	}

//.........这里部分代码省略.........

void TransformUnit::SubmitSpline(void* control_points, void* indices, int count_u, int count_v, int type_u, int type_v, GEPatchPrimType prim_type, u32 vertex_type)
{
	VertexDecoder vdecoder;
	vdecoder.SetVertexType(vertex_type);
	const DecVtxFormat& vtxfmt = vdecoder.GetDecVtxFmt();

	static u8 buf[65536 * 48]; // yolo
	u16 index_lower_bound = 0;
	u16 index_upper_bound = count_u * count_v - 1;
	bool indices_16bit = (vertex_type & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	u8* indices8 = (u8*)indices;
	u16* indices16 = (u16*)indices;
	if (indices)
		GetIndexBounds(indices, count_u*count_v, vertex_type, &index_lower_bound, &index_upper_bound);
	vdecoder.DecodeVerts(buf, control_points, index_lower_bound, index_upper_bound);

	VertexReader vreader(buf, vtxfmt, vertex_type);

	int num_patches_u = count_u - 3;
	int num_patches_v = count_v - 3;

	// TODO: Do something less idiotic to manage this buffer
	SplinePatch* patches = new SplinePatch[num_patches_u * num_patches_v];

	for (int patch_u = 0; patch_u < num_patches_u; ++patch_u) {
		for (int patch_v = 0; patch_v < num_patches_v; ++patch_v) {
			SplinePatch& patch = patches[patch_u + patch_v * num_patches_u];

			for (int point = 0; point < 16; ++point) {
				int idx = (patch_u + point%4) + (patch_v + point/4) * count_u;
                if (indices)
                    vreader.Goto(indices_16bit ? indices16[idx] : indices8[idx]);
                else
                    vreader.Goto(idx);

				patch.points[point] = ReadVertex(vreader);
			}
			patch.type = (type_u | (type_v<<2));
			if (patch_u != 0) patch.type &= ~START_OPEN_U;
			if (patch_v != 0) patch.type &= ~START_OPEN_V;
			if (patch_u != num_patches_u-1) patch.type &= ~END_OPEN_U;
			if (patch_v != num_patches_v-1) patch.type &= ~END_OPEN_V;
		}
	}

	for (int patch_idx = 0; patch_idx < num_patches_u*num_patches_v; ++patch_idx) {
		SplinePatch& patch = patches[patch_idx];

		// TODO: Should do actual patch subdivision instead of just drawing the control points!
		const int tile_min_u = (patch.type & START_OPEN_U) ? 0 : 1;
		const int tile_min_v = (patch.type & START_OPEN_V) ? 0 : 1;
		const int tile_max_u = (patch.type & END_OPEN_U) ? 3 : 2;
		const int tile_max_v = (patch.type & END_OPEN_V) ? 3 : 2;
		for (int tile_u = tile_min_u; tile_u < tile_max_u; ++tile_u) {
			for (int tile_v = tile_min_v; tile_v < tile_max_v; ++tile_v) {
				int point_index = tile_u + tile_v*4;

				VertexData v0 = patch.points[point_index];
				VertexData v1 = patch.points[point_index+1];
				VertexData v2 = patch.points[point_index+4];
				VertexData v3 = patch.points[point_index+5];

				// TODO: Backface culling etc
				Clipper::ProcessTriangle(v0, v1, v2);
				Clipper::ProcessTriangle(v2, v1, v0);
				Clipper::ProcessTriangle(v2, v1, v3);
				Clipper::ProcessTriangle(v3, v1, v2);
			}
		}
	}
	delete[] patches;
}

void TransformUnit::SubmitPrimitive(void* vertices, void* indices, u32 prim_type, int vertex_count, u32 vertex_type)
{
	// TODO: Cache VertexDecoder objects
	VertexDecoder vdecoder;
	vdecoder.SetVertexType(vertex_type);
	const DecVtxFormat& vtxfmt = vdecoder.GetDecVtxFmt();

	static u8 buf[65536 * 48]; // yolo
	u16 index_lower_bound = 0;
	u16 index_upper_bound = vertex_count - 1;
	bool indices_16bit = (vertex_type & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	u8* indices8 = (u8*)indices;
	u16* indices16 = (u16*)indices;
	if (indices)
		GetIndexBounds(indices, vertex_count, vertex_type, &index_lower_bound, &index_upper_bound);
	vdecoder.DecodeVerts(buf, vertices, index_lower_bound, index_upper_bound);

	VertexReader vreader(buf, vtxfmt, vertex_type);

	const int max_vtcs_per_prim = 3;
	int vtcs_per_prim = 0;
	if (prim_type == GE_PRIM_POINTS) vtcs_per_prim = 1;
	else if (prim_type == GE_PRIM_LINES) vtcs_per_prim = 2;
	else if (prim_type == GE_PRIM_TRIANGLES) vtcs_per_prim = 3;
	else if (prim_type == GE_PRIM_RECTANGLES) vtcs_per_prim = 2;
	else {
		// TODO: Unsupported
	}

	if (prim_type == GE_PRIM_POINTS || prim_type == GE_PRIM_LINES || prim_type == GE_PRIM_TRIANGLES || prim_type == GE_PRIM_RECTANGLES) {
		for (int vtx = 0; vtx < vertex_count; vtx += vtcs_per_prim) {
			VertexData data[max_vtcs_per_prim];

			for (int i = 0; i < vtcs_per_prim; ++i) {
				if (indices)
					vreader.Goto(indices_16bit ? indices16[vtx+i] : indices8[vtx+i]);
				else
					vreader.Goto(vtx+i);

				data[i] = ReadVertex(vreader);
				if (outside_range_flag)
					break;
			}
			if (outside_range_flag) {
				outside_range_flag = false;
				continue;
			}


			switch (prim_type) {
			case GE_PRIM_TRIANGLES:
			{
				if (!gstate.isCullEnabled() || gstate.isModeClear()) {
					Clipper::ProcessTriangle(data[0], data[1], data[2]);
					Clipper::ProcessTriangle(data[2], data[1], data[0]);
				} else if (!gstate.getCullMode())
					Clipper::ProcessTriangle(data[2], data[1], data[0]);
				else
					Clipper::ProcessTriangle(data[0], data[1], data[2]);
				break;
			}

			case GE_PRIM_RECTANGLES:
				Clipper::ProcessQuad(data[0], data[1]);
				break;
			}
		}
	} else if (prim_type == GE_PRIM_TRIANGLE_STRIP) {
		VertexData data[3];
		unsigned int skip_count = 2; // Don't draw a triangle when loading the first two vertices

		for (int vtx = 0; vtx < vertex_count; ++vtx) {
			if (indices)
				vreader.Goto(indices_16bit ? indices16[vtx] : indices8[vtx]);
			else
				vreader.Goto(vtx);

			data[vtx % 3] = ReadVertex(vreader);
			if (outside_range_flag) {
				// Drop all primitives containing the current vertex
				skip_count = 2;
				outside_range_flag = false;
				continue;
			}

			if (skip_count) {
				--skip_count;
				continue;
			}

			if (!gstate.isCullEnabled() || gstate.isModeClear()) {
				Clipper::ProcessTriangle(data[0], data[1], data[2]);
				Clipper::ProcessTriangle(data[2], data[1], data[0]);
			} else if ((!gstate.getCullMode()) ^ (vtx % 2)) {
				// We need to reverse the vertex order for each second primitive,
				// but we additionally need to do that for every primitive if CCW cullmode is used.
				Clipper::ProcessTriangle(data[2], data[1], data[0]);
			} else {
				Clipper::ProcessTriangle(data[0], data[1], data[2]);
			}
//.........这里部分代码省略.........

void TransformDrawEngine::SubmitSpline(void* control_points, void* indices, int count_u, int count_v, int type_u, int type_v, GEPatchPrimType prim_type, u32 vertType) {
	Flush();

	if (prim_type != GE_PATCHPRIM_TRIANGLES) {
		// Only triangles supported!
		return;
	}

	u16 index_lower_bound = 0;
	u16 index_upper_bound = count_u * count_v - 1;
	bool indices_16bit = (vertType & GE_VTYPE_IDX_MASK) == GE_VTYPE_IDX_16BIT;
	const u8* indices8 = (const u8*)indices;
	const u16* indices16 = (const u16*)indices;
	if (indices)
		GetIndexBounds(indices, count_u*count_v, vertType, &index_lower_bound, &index_upper_bound);

	// Simplify away bones and morph before proceeding
	SimpleVertex *simplified_control_points = (SimpleVertex *)(decoded + 65536 * 12);
	u8 *temp_buffer = decoded + 65536 * 24;
	
	u32 origVertType = vertType;
	vertType = NormalizeVertices((u8 *)simplif