Vulkan Samples 阅读 -- Performance（一）: Multi Threaded & Instancing & Indirect Drawing

Multi Threaded

prepare

• vkCreateFence: Create a fence for synchronization
• loadAssets: 加载模型和星空纹理
• setupPipelineLayout: 这里用常量赋值矩阵
  • vkCreatePipelineLayout
• preparePipelines:
  • vkCreateGraphicsPipelines: 画模型
  • vkCreateGraphicsPipelines: 画星空
• prepareMultiThreadedRenderer: 多线程commandbuffer的创建
  • commandBufferAllocateInfo
  • vkAllocateCommandBuffers: 主线程buffer
  • vkAllocateCommandBuffers: 子线程画背景buffer
  • vkAllocateCommandBuffers: 子线程uibuffer
  • threadData.resize(numThreads);
  • loop numThreads
    • commandPoolCreateInfo
    • vkCreateCommandPool: 创建线程使用的池
    • thread->commandBuffer.resize(numObjectsPerThread)
    • commandBufferAllocateInfo: 每个渲染对象对应的子线程创建commandBuffer
    • vkAllocateCommandBuffers
    • thread->pushConstBlock.resize(numObjectsPerThread)
    • thread->objectData.resize(numObjectsPerThread)
    • loop numObjectsPerThread
      • thread->objectData: 写数据(随机参数)
• updateMatrices
  • MVP矩阵
  • frustum.update(matrices.projection * matrices.view)：视锥类要注意

render

• draw：多线程主要在draw
  • 等待信号

    VkResult fenceRes;
    do {
    	fenceRes = vkWaitForFences(device, 1, &renderFence, VK_TRUE, 100000000);
    } while (fenceRes == VK_TIMEOUT);
    VK_CHECK_RESULT(fenceRes);
    vkResetFences(device, 1, &renderFence);
    

• VulkanExampleBase::prepareFrame：正常
• updateCommandBuffers：这里是多线程额外增加的操作
  • 先利用线程池更新子线程的secondary command buffers
  • 再将这些buffer放入 primary command buffer
  • 最后提交queue去渲染
  • vkBeginCommandBuffer
  • vkCmdBeginRenderPass – primaryCommandBuffer：主线程不包含任何渲染命令，只负责从secondary command buffers读取结果
  • VkCommandBufferInheritanceInfo：从 secondary command buffers中继承数据
  • updateSecondaryCommandBuffers
    • vkBeginCommandBuffer：画星空的buffer
    • vkCmdSetViewport
    • vkCmdSetScissor
    • vkCmdBindPipeline
    • vkCmdPushConstants
    • draw
    • vkEndCommandBuffer
    • vkBeginCommandBuffer：画ui
    • vkCmdSetViewport
    • vkCmdSetScissor
    • vkCmdBindPipeline
    • drawUI
    • vkEndCommandBuffer
  • commandBuffers.push_back(secondaryCommandBuffers.background)：将画星空的buffer放入主线程使用的buffer
  • loop：threadPool.threads[t]->addJob([=] { threadRenderCode(t, i, inheritanceInfo); });
    • 循环利用线程渲染物体（很多ufo）
    • threadRenderCode
      • vkBeginCommandBuffer
      • vkCmdSetViewport
      • vkCmdSetScissor
      • vkCmdBindPipeline
      • vkCmdPushConstants
      • vkCmdBindVertexBuffers
      • vkCmdBindIndexBuffer
      • vkCmdDrawIndexed
      • vkEndCommandBuffer
  • threadPool.wait()
  • loop commandBuffers存入所有物体的command
  • commandBuffers.push_back(secondaryCommandBuffers.ui)：将画UI放入commandBuffers
  • vkCmdExecuteCommands–primaryCommandBuffer
  • vkCmdEndRenderPass – primaryCommandBuffer
• vkQueueSubmit：正常
  • submitInfo.pCommandBuffers = &primaryCommandBuffer：只需提交主线程buffer即可
• VulkanExampleBase::submitFrame：正常
• updateMatrices

shader

• 基础渲染方式

小结

   多线程渲染和正常渲染流程有些不同，在prepare阶段不用设置Descriptor相关内容。渲染时注意，主线程只负责从子线程中读取结果，而不负责任何渲染任务。

Instancing

prepare

• loadAssets: 加载模型和纹理
• prepareInstanceData
  • std::vector instanceData：这里初始化所有instance，Instanced data is static, copy to device local memory
  • createBuffer：创建stagingBuffer HOST
  • createBuffer：创建stagingBuffer DEVICE
  • createCommandBuffer：Copy to staging buffer
  • vkCmdCopyBuffer
  • flushCommandBuffer
  • vkDestroyBuffer
  • vkFreeMemory
• prepareUniformBuffers
  • updateUniformBuffer
• setupDescriptorSetLayout
  • vkCreateDescriptorSetLayout
  • vkCreatePipelineLayout
• preparePipelines:
  • vkCreateGraphicsPipelines: 画Instancing
    • vertexBindingDescriptionCount控制vs的执行次数

	// Use all input bindings and attribute descriptions
	inputState.vertexBindingDescriptionCount = static_cast<uint32_t>(bindingDescriptions.size());
	inputState.vertexAttributeDescriptionCount = static_cast<uint32_t>(attributeDescriptions.size());

• vkCreateGraphicsPipelines: 画Planet

• vkCreateGraphicsPipelines：画Star

• setupDescriptorPool

• setupDescriptorSet

  • vkAllocateDescriptorSets: Instanced rocks
  • vkAllocateDescriptorSets: Planet
  • vkUpdateDescriptorSets

• buildCommandBuffers

  • loop drawCmdBuffers
    • vkBeginCommandBuffer
    • vkCmdBeginRenderPass
    • vkCmdSetViewport
    • vkCmdSetScissor
    • Star field
      • vkCmdBindDescriptorSets
      • vkCmdBindPipeline
      • vkCmdDraw
    • Planet
      • vkCmdBindDescriptorSets
      • vkCmdBindPipeline
      • draw
    • Instanced rocks
      • vkCmdBindDescriptorSets
      • vkCmdBindPipeline
      • vkCmdBindVertexBuffers: rock.vertices
      • vkCmdBindVertexBuffers: instanceBuffer.buffer
      • vkCmdBindIndexBuffer
      • vkCmdDrawIndexed
    • drawUI(drawCmdBuffers[i])
    • vkCmdEndRenderPass
    • vkEndCommandBuffer

render

• draw
• updateUniformBuffer

shader

• 基础渲染方式，但是在渲染instance rocks时将instance 的数据以VBO的形式传入管线

小结

   本节主要讲使用instance渲染更多物体。重点是自定义数据如何进入管线的（VBO）。

Indirect Drawing

prepare

• loadAssets: 加载模型和纹理
• prepareIndirectData
  • loop 模型初始化indirectCommand并计算indirectCommand个数
  • stagingBuffer: 创建stagingbuffer
  • copyBuffer:将stagingbuffer 复制到 INDIRECT_BUFFER
  • stagingBuffer.destroy
• prepareInstanceData: Indirect Drawing 在一次绘制调用中为不同的网格绘制多个实例
  • std::vector instanceData：这里初始化所有instance，Instanced data is static, copy to device local memory
  • createBuffer：创建stagingBuffer HOST
  • createBuffer：创建stagingBuffer DEVICE
  • copyBuffer
    • createCommandBuffer：Copy to staging buffer
    • vkCmdCopyBuffer
    • flushCommandBuffer
    • vkDestroyBuffer
    • vkFreeMemory
• prepareUniformBuffers
  • updateUniformBuffer
• setupDescriptorSetLayout
  • vkCreateDescriptorSetLayout
  • vkCreatePipelineLayout
• preparePipelines:
  • vkCreateGraphicsPipelines: 画Indirect (Instancing)
    • vertexBindingDescriptionCount控制vs的执行次数
  • vkCreateGraphicsPipelines: 画Ground
  • vkCreateGraphicsPipelines：画Skysphere
• setupDescriptorPool
• setupDescriptorSet
  • vkAllocateDescriptorSets
  • vkUpdateDescriptorSets
• buildCommandBuffers
  • loop drawCmdBuffers
    • vkBeginCommandBuffer
    • vkCmdBeginRenderPass
    • vkCmdSetViewport
    • vkCmdSetScissor
    • vkCmdBindDescriptorSets
    • Skysphere
      • vkCmdBindPipeline
      • draw
    • Ground
      • vkCmdBindPipeline
      • draw
    • Instanced
      • vkCmdBindPipeline
      • vkCmdBindVertexBuffers
      • vkCmdBindVertexBuffers
      • vkCmdBindIndexBuffer
      • if support
        • vkCmdDrawIndexedIndirect (间接渲染)
      • else
        • loop
          • vkCmdDrawIndexedIndirect
    • vkCmdEndRenderPass
    • vkEndCommandBuffer

render

• draw
• updateUniformBuffer

shader

• 基础渲染方式，但是在渲染instance rocks时将instance 的数据以VBO的形式传入管线

小结

   Indirect Drawing 是 instance 渲染一种扩展。能在一次渲染中为多个 mesh 绘制多个instance.