Vulkan, OpenGL, Metal: Let draws bind a second vertex stream

Vulkan pipelines can have a second vertex binding, and draws a second vertex buffer for it. OpenGL input
layout entries can come from a second buffer, with its own stride, given with the draw. Metal gets a second
vertex buffer index. Nothing uses them yet.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Common/GPU/Metal/MetalRenderManager.h

@@ -49,6 +49,8 @@
 // bottom. Vertex data goes above all of them.
 enum {
 	METAL_VERTEX_BUFFER_INDEX = 16,
+	// A second vertex stream, for the pipelines with one.
+	METAL_VERTEX_BUFFER_INDEX1 = 17,
 	// Push constants (only compute shaders have them) go in a buffer of their own, set with setBytes.
 	METAL_PUSH_CONSTANT_INDEX = 15,
 	METAL_MAX_TEXTURES = 8,
@@ -527,8 +529,8 @@ private:
 		MTLDepthClipMode depthClip;
 		MTLViewport viewport;
 		MTLScissorRect scissor;
-		__unsafe_unretained id<MTLBuffer> vertexBuffers[METAL_VERTEX_BUFFER_INDEX + 1];
-		uint32_t vertexBufferOffsets[METAL_VERTEX_BUFFER_INDEX + 1];
+		__unsafe_unretained id<MTLBuffer> vertexBuffers[METAL_VERTEX_BUFFER_INDEX1 + 1];
+		uint32_t vertexBufferOffsets[METAL_VERTEX_BUFFER_INDEX1 + 1];
 		__unsafe_unretained id<MTLBuffer> fragmentBuffers[METAL_MAX_BUFFERS];
 		uint32_t fragmentBufferOffsets[METAL_MAX_BUFFERS];
 		__unsafe_unretained id<MTLTexture> fragmentTextures[METAL_MAX_TEXTURES];

Common/GPU/OpenGL/GLQueueRunner.cpp

@@ -1341,19 +1341,35 @@ void GLQueueRunner::PerformRenderPass(const GLRStep &step, bool first, bool last
 		case GLRRenderCommand::DRAW:
 		{
 			GLRInputLayout *layout = c.draw.inputLayout;
+			if (attrMask != layout->semanticsMask_) {
+				EnableDisableVertexArrays(attrMask, layout->semanticsMask_);
+				attrMask = layout->semanticsMask_;
+			}
+			// Stream 1 first, so stream 0 is left bound.
+			if (layout->stride1) {
+				_dbg_assert_(c.draw.vertexBuffer1 && !c.draw.vertexBuffer1->Mapped());
+				GLuint buf1 = c.draw.vertexBuffer1->buffer_;
+				if (buf1 != curArrayBuffer) {
+					glBindBuffer(GL_ARRAY_BUFFER, buf1);
+					curArrayBuffer = buf1;
+				}
+				for (const auto &entry : layout->entries) {
+					if (entry.stream == 1) {
+						glVertexAttribPointer(entry.location, entry.count, entry.type, entry.normalized, layout->stride1, (const void *)(c.draw.vertexOffset1 + entry.offset));
+					}
+				}
+			}
 			GLuint buf = c.draw.vertexBuffer->buffer_;
 			_dbg_assert_(!c.draw.vertexBuffer->Mapped());
 			if (buf != curArrayBuffer) {
 				glBindBuffer(GL_ARRAY_BUFFER, buf);
 				curArrayBuffer = buf;
 			}
-			if (attrMask != layout->semanticsMask_) {
-				EnableDisableVertexArrays(attrMask, layout->semanticsMask_);
-				attrMask = layout->semanticsMask_;
-			}
 			for (size_t i = 0; i < layout->entries.size(); i++) {
 				auto &entry = layout->entries[i];
-				glVertexAttribPointer(entry.location, entry.count, entry.type, entry.normalized, layout->stride, (const void *)(c.draw.vertexOffset + entry.offset));
+				if (entry.stream == 0) {
+					glVertexAttribPointer(entry.location, entry.count, entry.type, entry.normalized, layout->stride, (const void *)(c.draw.vertexOffset + entry.offset));
+				}
 			}
 			if (c.draw.indexBuffer) {
 				GLuint buf = c.draw.indexBuffer->buffer_;

Common/GPU/OpenGL/GLQueueRunner.h

@@ -109,8 +109,11 @@ struct GLRRenderData {
 			GLRInputLayout *inputLayout;
 			GLRBuffer *vertexBuffer;
 			GLRBuffer *indexBuffer;
+			// For the input layout's stream 1.
+			GLRBuffer *vertexBuffer1;
 			uint32_t vertexOffset;
 			uint32_t indexOffset;
+			uint32_t vertexOffset1;
 			GLenum mode;  // primitive
 			GLint first;
 			GLint count;

Common/GPU/OpenGL/GLRenderManager.h

@@ -204,9 +204,12 @@ public:
 		GLenum type;
 		GLboolean normalized;
 		intptr_t offset;
+		// 1: from the draw's second vertex buffer, with stride1.
+		int stream = 0;
 	};
 	std::vector<Entry> entries;
 	int stride;
+	int stride1 = 0;
 	int semanticsMask_ = 0;
 };
 
@@ -355,13 +358,14 @@ public:
 		return step.create_program.program;
 	}
 
-	GLRInputLayout *CreateInputLayout(const std::vector<GLRInputLayout::Entry> &entries, int stride) {
+	GLRInputLayout *CreateInputLayout(const std::vector<GLRInputLayout::Entry> &entries, int stride, int stride1 = 0) {
 		std::lock_guard<std::mutex> lock(initStepsMutex_);
 		GLRInitStep &step = initSteps_.push_uninitialized();
 		step.stepType = GLRInitStepType::CREATE_INPUT_LAYOUT;
 		step.create_input_layout.inputLayout = new GLRInputLayout();
 		step.create_input_layout.inputLayout->entries = entries;
 		step.create_input_layout.inputLayout->stride = stride;
+		step.create_input_layout.inputLayout->stride1 = stride1;
 		for (auto &iter : step.create_input_layout.inputLayout->entries) {
 			step.create_input_layout.inputLayout->semanticsMask_ |= 1 << iter.location;
 		}
@@ -798,13 +802,16 @@ public:
 		data.clear.scissorH = scissorH;
 	}
 
-	void Draw(GLRInputLayout *inputLayout, GLRBuffer *vertexBuffer, uint32_t vertexOffset, GLenum mode, int first, int count) {
+	// vertexBuffer1 is for the input layout's stream 1 entries, if it has any.
+	void Draw(GLRInputLayout *inputLayout, GLRBuffer *vertexBuffer, uint32_t vertexOffset, GLenum mode, int first, int count, GLRBuffer *vertexBuffer1 = nullptr, uint32_t vertexOffset1 = 0) {
 		_dbg_assert_(vertexBuffer && curRenderStep_ && curRenderStep_->stepType == GLRStepType::RENDER);
 		GLRRenderData &data = curRenderStep_->commands.push_uninitialized();
 		data.cmd = GLRRenderCommand::DRAW;
 		data.draw.inputLayout = inputLayout;
 		data.draw.vertexOffset = vertexOffset;
 		data.draw.vertexBuffer = vertexBuffer;
+		data.draw.vertexBuffer1 = vertexBuffer1;
+		data.draw.vertexOffset1 = vertexOffset1;
 		data.draw.indexBuffer = nullptr;
 		data.draw.mode = mode;
 		data.draw.first = first;
@@ -815,13 +822,15 @@ public:
 	// Would really love to have a basevertex parameter, but impossible in unextended GLES, without glDrawElementsBaseVertex, unfortunately.
 	// If maxIndex is known (>= 0), it's passed on to the driver through glDrawRangeElements, which saves drivers that need the index
 	// range (like Panfrost) from scanning the index data on the CPU for every draw.
-	void DrawIndexed(GLRInputLayout *inputLayout, GLRBuffer *vertexBuffer, uint32_t vertexOffset, GLRBuffer *indexBuffer, uint32_t indexOffset, GLenum mode, int count, GLenum indexType, int instances = 1, int maxIndex = -1) {
+	void DrawIndexed(GLRInputLayout *inputLayout, GLRBuffer *vertexBuffer, uint32_t vertexOffset, GLRBuffer *indexBuffer, uint32_t indexOffset, GLenum mode, int count, GLenum indexType, int instances = 1, int maxIndex = -1, GLRBuffer *vertexBuffer1 = nullptr, uint32_t vertexOffset1 = 0) {
 		_dbg_assert_(vertexBuffer && indexBuffer && curRenderStep_ && curRenderStep_->stepType == GLRStepType::RENDER);
 		GLRRenderData &data = curRenderStep_->commands.push_uninitialized();
 		data.cmd = GLRRenderCommand::DRAW;
 		data.draw.inputLayout = inputLayout;
 		data.draw.vertexOffset = vertexOffset;
 		data.draw.vertexBuffer = vertexBuffer;
+		data.draw.vertexBuffer1 = vertexBuffer1;
+		data.draw.vertexOffset1 = vertexOffset1;
 		data.draw.indexBuffer = indexBuffer;
 		data.draw.indexOffset = indexOffset;
 		data.draw.mode = mode;

Common/GPU/Vulkan/VulkanQueueRunner.cpp

@@ -1374,8 +1374,9 @@ void VulkanQueueRunner::PerformRenderPass(const VKRStep &step, VkCommandBuffer c
 				_dbg_assert_(set != VK_NULL_HANDLE);
 				vkCmdBindDescriptorSets(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &set, c.drawIndexed.numUboOffsets, c.drawIndexed.uboOffsets);
 				vkCmdBindIndexBuffer(cmd, c.drawIndexed.ibuffer, c.drawIndexed.ioffset, VK_INDEX_TYPE_UINT16);
-				VkDeviceSize voffset = c.drawIndexed.voffset;
-				vkCmdBindVertexBuffers(cmd, 0, 1, &c.drawIndexed.vbuffer, &voffset);
+				const VkBuffer vbuffers[2] = { c.drawIndexed.vbuffer, c.drawIndexed.vbuffer1 };
+				const VkDeviceSize voffsets[2] = { c.drawIndexed.voffset, c.drawIndexed.voffset1 };
+				vkCmdBindVertexBuffers(cmd, 0, c.drawIndexed.vbuffer1 ? 2 : 1, vbuffers, voffsets);
 				vkCmdDrawIndexed(cmd, c.drawIndexed.count, c.drawIndexed.instances, 0, 0, 0);
 			}
 			break;
@@ -1387,7 +1388,9 @@ void VulkanQueueRunner::PerformRenderPass(const VKRStep &step, VkCommandBuffer c
 				_dbg_assert_(set != VK_NULL_HANDLE);
 				vkCmdBindDescriptorSets(cmd, VK_PIPELINE_BIND_POINT_GRAPHICS, pipelineLayout, 0, 1, &set, c.draw.numUboOffsets, c.draw.uboOffsets);
 				if (c.draw.vbuffer) {
-					vkCmdBindVertexBuffers(cmd, 0, 1, &c.draw.vbuffer, &c.draw.voffset);
+					const VkBuffer vbuffers[2] = { c.draw.vbuffer, c.draw.vbuffer1 };
+					const VkDeviceSize voffsets[2] = { c.draw.voffset, c.draw.voffset1 };
+					vkCmdBindVertexBuffers(cmd, 0, c.draw.vbuffer1 ? 2 : 1, vbuffers, voffsets);
 				}
 				vkCmdDraw(cmd, c.draw.count, 1, c.draw.offset, 0);
 			}

Common/GPU/Vulkan/VulkanQueueRunner.h

@@ -69,6 +69,9 @@ struct VkRenderData {
 			VkDeviceSize voffset;
 			uint32_t count;
 			uint32_t offset;
+			// For the second vertex stream, if any.
+			VkBuffer vbuffer1;
+			uint32_t voffset1;
 		} draw;
 		struct {
 			uint32_t descSetIndex;
@@ -80,6 +83,9 @@ struct VkRenderData {
 			uint32_t voffset;
 			uint32_t ioffset;
 			uint32_t count;
+			// For the second vertex stream, if any.
+			VkBuffer vbuffer1;
+			uint32_t voffset1;
 		} drawIndexed;
 		struct {
 			uint32_t clearColor;

Common/GPU/Vulkan/VulkanRenderManager.h

@@ -101,7 +101,8 @@ public:
 
 	VkPrimitiveTopology topology = VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST;
 	VkVertexInputAttributeDescription attrs[8]{};
-	VkVertexInputBindingDescription ibd{};
+	// The second is only used by pipelines with two vertex streams.
+	VkVertexInputBindingDescription ibd[2]{};
 	VkPipelineVertexInputStateCreateInfo vis{ VK_STRUCTURE_TYPE_PIPELINE_VERTEX_INPUT_STATE_CREATE_INFO };
 	VkPipelineViewportStateCreateInfo views{ VK_STRUCTURE_TYPE_PIPELINE_VIEWPORT_STATE_CREATE_INFO };
 
@@ -466,7 +467,8 @@ public:
 		return retval;
 	}
 
-	void Draw(int descSetIndex, int numUboOffsets, const uint32_t *uboOffsets, VkBuffer vbuffer, int voffset, int count, int offset = 0) {
+	// vbuffer1 is for the pipeline's second vertex stream, if it has one.
+	void Draw(int descSetIndex, int numUboOffsets, const uint32_t *uboOffsets, VkBuffer vbuffer, int voffset, int count, int offset = 0, VkBuffer vbuffer1 = VK_NULL_HANDLE, uint32_t voffset1 = 0) {
 		_dbg_assert_(curRenderStep_ && curRenderStep_->stepType == VKRStepType::RENDER && curStepHasViewport_ && curStepHasScissor_);
 		VkRenderData &data = curRenderStep_->commands.push_uninitialized();
 		data.cmd = VKRRenderCommand::DRAW;
@@ -475,6 +477,8 @@ public:
 		data.draw.descSetIndex = descSetIndex;
 		data.draw.vbuffer = vbuffer;
 		data.draw.voffset = voffset;
+		data.draw.vbuffer1 = vbuffer1;
+		data.draw.voffset1 = voffset1;
 		data.draw.numUboOffsets = numUboOffsets;
 		_dbg_assert_(numUboOffsets <= ARRAY_SIZE(data.draw.uboOffsets));
 		for (int i = 0; i < numUboOffsets; i++)
@@ -482,7 +486,7 @@ public:
 		curRenderStep_->render.numDraws++;
 	}
 
-	void DrawIndexed(int descSetIndex, int numUboOffsets, const uint32_t *uboOffsets, VkBuffer vbuffer, int voffset, VkBuffer ibuffer, int ioffset, int count, int numInstances) {
+	void DrawIndexed(int descSetIndex, int numUboOffsets, const uint32_t *uboOffsets, VkBuffer vbuffer, int voffset, VkBuffer ibuffer, int ioffset, int count, int numInstances, VkBuffer vbuffer1 = VK_NULL_HANDLE, uint32_t voffset1 = 0) {
 		_dbg_assert_(curRenderStep_ && curRenderStep_->stepType == VKRStepType::RENDER && curStepHasViewport_ && curStepHasScissor_);
 		VkRenderData &data = curRenderStep_->commands.push_uninitialized();
 		data.cmd = VKRRenderCommand::DRAW_INDEXED;
@@ -491,6 +495,8 @@ public:
 		data.drawIndexed.descSetIndex = descSetIndex;
 		data.drawIndexed.vbuffer = vbuffer;
 		data.drawIndexed.voffset = voffset;
+		data.drawIndexed.vbuffer1 = vbuffer1;
+		data.drawIndexed.voffset1 = voffset1;
 		data.drawIndexed.ibuffer = ibuffer;
 		data.drawIndexed.ioffset = ioffset;
 		data.drawIndexed.numUboOffsets = numUboOffsets;

Common/GPU/Vulkan/thin3d_vulkan.cpp

@@ -1320,13 +1320,13 @@ Pipeline *VKContext::CreateGraphicsPipeline(const PipelineDesc &desc, const char
 	_dbg_assert_((int)input->attributes.size() == (int)input->visc.vertexAttributeDescriptionCount);
 
 	pipeline->stride = input->binding.stride;
-	gDesc.ibd = input->binding;
+	gDesc.ibd[0] = input->binding;
 	for (size_t i = 0; i < input->attributes.size(); i++) {
 		gDesc.attrs[i] = input->attributes[i];
 	}
 	gDesc.vis.vertexAttributeDescriptionCount = input->visc.vertexAttributeDescriptionCount;
 	gDesc.vis.vertexBindingDescriptionCount = input->visc.vertexBindingDescriptionCount;
-	gDesc.vis.pVertexBindingDescriptions = &gDesc.ibd;
+	gDesc.vis.pVertexBindingDescriptions = gDesc.ibd;
 	gDesc.vis.pVertexAttributeDescriptions = gDesc.attrs;
 
 	gDesc.blend0 = blend->attachments[0];

GPU/Vulkan/PipelineManagerVulkan.cpp

@@ -312,7 +312,7 @@ static VulkanPipeline *CreateVulkanPipeline(VulkanRenderManager *renderManager,
 		vertexStride = (int)sizeof(TransformedVertex);
 	}
 
-	VkVertexInputBindingDescription &ibd = desc->ibd;
+	VkVertexInputBindingDescription &ibd = desc->ibd[0];
 	ibd.binding = 0;
 	ibd.inputRate = VK_VERTEX_INPUT_RATE_VERTEX;
 	ibd.stride = vertexStride;
@@ -320,7 +320,7 @@ static VulkanPipeline *CreateVulkanPipeline(VulkanRenderManager *renderManager,
 	VkPipelineVertexInputStateCreateInfo &vis = desc->vis;
 	vis.flags = 0;
 	vis.vertexBindingDescriptionCount = 1;
-	vis.pVertexBindingDescriptions = &desc->ibd;
+	vis.pVertexBindingDescriptions = desc->ibd;
 	vis.vertexAttributeDescriptionCount = attributeCount;
 	vis.pVertexAttributeDescriptions = attrs;