3a4c4c88fa
When USD sends work to the GPU (dispatch) incorrect datatype conversions can send incorrect vulkan commands. This happens on AMD GPUs. This PR will patch USD to limit the max allowed limit. A better solution would be to only use uint32_t in this function, but there is no `GfVec3u` data type and I didn't want to do to many changes.  **TODO** - [x] Validate with @ZedDB that this compiles and solves all the issues - [x] Add other platform maintainers as reviewers. This is a Windows + Linux issue. - [ ] Rebuild the USD library including this patch. Pull Request: https://projects.blender.org/blender/blender/pulls/140102
20 lines
1.0 KiB
Diff
20 lines
1.0 KiB
Diff
diff --git a/pxr/imaging/hgiVulkan/computeCmds.cpp b/pxr/imaging/hgiVulkan/computeCmds.cpp
|
|
index cbf9704fd..74a81c787 100644
|
|
--- a/pxr/imaging/hgiVulkan/computeCmds.cpp
|
|
+++ b/pxr/imaging/hgiVulkan/computeCmds.cpp
|
|
@@ -119,10 +119,11 @@ HgiVulkanComputeCmds::Dispatch(int dimX, int dimY)
|
|
// Determine device's num compute work group limits
|
|
const VkPhysicalDeviceLimits limits =
|
|
_hgi->GetCapabilities()->vkDeviceProperties.limits;
|
|
+ const uint32_t maxAllowedLimit = (1 << 31) - 1;
|
|
const GfVec3i maxNumWorkGroups = GfVec3i(
|
|
- limits.maxComputeWorkGroupCount[0],
|
|
- limits.maxComputeWorkGroupCount[1],
|
|
- limits.maxComputeWorkGroupCount[2]);
|
|
+ std::min<uint32_t>(maxAllowedLimit, limits.maxComputeWorkGroupCount[0]),
|
|
+ std::min<uint32_t>(maxAllowedLimit, limits.maxComputeWorkGroupCount[1]),
|
|
+ std::min<uint32_t>(maxAllowedLimit, limits.maxComputeWorkGroupCount[2]));
|
|
|
|
if (numWorkGroupsX > maxNumWorkGroups[0]) {
|
|
TF_WARN("Max number of work group available from device is %i, larger "
|