Optimize graphics performance using Frame Advisor render graphs
Introduction
What are render graphs?
Generating a render graph for your application
Understanding your render graph
Problem solving – unused resources
Problem solving – unwanted execution nodes
Problem solving – textures with excessive resolution
Problem solving – inefficient transfer workloads
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Focusing on transfer workloads

Your application may contain inefficient transfer workloads.

What are transfer workloads?

Transfer workloads use the GPU to move data between resources.

Transfer operations require power and bandwidth. For this reason, they should be made as small and efficient as possible – or preferably, avoided entirely.

Transfer workloads are initiated by specific API calls. These include:

  • vkCmdCopyBuffer(), which copies regions of a buffer
  • vkCmdClearColorImage(), which clears an image.

To find which API calls your application uses to start transfer workloads:

  • Open the render graph for your captured frames
  • Click a transfer node
  • Now move to the API Calls view (labeled “API Calls”)
  • Observe the API calls in use.

Problem: inefficient clear routines

vkCmdClearColorImage() is an inefficient way to clear an image.

A more efficient way to clear an image attachment is to clear the attachment at the start of a render pass. To do this:

  • Set the loadOp parameter in a VkAttachmentDescription structure to VK_ATTACHMENT_LOAD_OP_CLEAR
  • Attach the VkAttachmentDescription to a VkRenderPassCreateInfo
  • Supply this to API vkCreateRenderPass()

Problem: inefficient image resolutions

In the previous section, you saw an issue where unnecessarily large textures were inputs to a render pass. Similar problems can be seen in the context of transfer operations, which should operate over the smallest practicable area. To achieve this, change the values in the VkBufferCopy structure passed to vkCmdCopyBuffer().

Back
Next