lee/renderdoc
1
0
Fork 0
mirror of https://github.com/baldurk/renderdoc.git synced 2026-08-22 14:36:35 +00:00
Code Issues Packages Projects Releases Wiki Activity

Tweak selection of memory types for GPU readback on Vulkan

Browse Source
This commit is contained in:
baldurk
2023-03-23 16:16:02 +00:00
parent 0851a9f1ad
commit 129cb33266
1 changed files with 14 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files
renderdoc/driver/vulkan/wrappers/vk_resource_funcs.cpp
+14 -9
View File
@@ -686,20 +686,25 @@ VkResult WrappedVulkan::vkAllocateMemory(VkDevice device, const VkMemoryAllocate
{
record->memMapState->mapCoherent = (memProps & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) != 0;
// mark this memory as needing readback on the GPU if it's device-local on a discrete GPU,
// or host-uncached on any other type of GPU. On non-discrete GPUs it's unfortunately not
// always the case that the CPU can access memory at good speed, even if it's on the same
// chip. On discrete GPUs where the memory isn't device local it's CPU side so it should
// always be fast to access. In these cases we want to push the memory from the GPU to the
// CPU with a command buffer.
// some types of memory are faster to readback via a synchronous call to the GPU
if(Vulkan_GPUReadbackDeviceLocal())
{
// on discrete GPUs, all device local memory should be readback this way
if(m_PhysicalDeviceData.props.deviceType == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU)
record->memMapState->readbackOnGPU =
(memProps & VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) != 0;
else
record->memMapState->readbackOnGPU =
((memProps & VK_MEMORY_PROPERTY_HOST_CACHED_BIT) == 0);
// non-cached memory are generally always faster to readback on the GPU. Some GPUs allow
// faster readback via non-temporal loads but GPU copy speeds are comparable enough
record->memMapState->readbackOnGPU = ((memProps & VK_MEMORY_PROPERTY_HOST_CACHED_BIT) == 0);
#if ENABLED(RDOC_ANDROID)
// on Android all memory types are marked as device local. Types which are fast to read
// back directly from the CPU are CACHED but *not* COHERENT. Any types which are either
// only COHERENT, or are both CACHED and COHERENT, are still faster to readback via GPU
// copy.
record->memMapState->readbackOnGPU = (memProps & VK_MEMORY_PROPERTY_HOST_COHERENT_BIT) != 0;
#endif
// we need a wholeMemBuf to readback on the GPU
if(record->memMapState->readbackOnGPU && wholeMemBuf == VK_NULL_HANDLE)