AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

SWDEV-270773 - Fix workgroup calculation logic for GWS initialization

Browse Source 
The existing workgroup calculation logic for GWS initialization is
incorrect. It tries to add together workgroups across dimensions,
leading to major under-count in 2D and 3D kernels. An (x,y,z) kernel
uses x * y * z blocks, not x + y + z.

In addition, the previous logic was incorrect for the case of launching
a single-threaded kernel. It calculated 0 workgroups, leading to
initializing GWS to -1.

Change-Id: I1bb20a0d5b6e0cc10ac55901c28d8f93aac61c09
This commit is contained in:
Joseph Greathouse
2021-01-31 01:16:48 -06:00
parent b7f0fd0354
commit 54d1d69c0a
2 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/device/rocm/rocvirtual.cpp
+3 -3
View File
@@ -2595,10 +2595,10 @@ void VirtualGPU::submitKernel(amd::NDRangeKernelCommand& vcmd) {
if (vcmd.cooperativeGroups()) {
// Initialize GWS if it's cooperative groups launch
uint32_t workgroups = 0;
uint32_t workgroups = 1;
for (uint i = 0; i < vcmd.sizes().dimensions(); i++) {
if ((vcmd.sizes().local()[i] != 0) && (vcmd.sizes().global()[i] != 1)) {
workgroups += (vcmd.sizes().global()[i] / vcmd.sizes().local()[i]);
if (vcmd.sizes().local()[i] != 0) {
workgroups *= (vcmd.sizes().global()[i] / vcmd.sizes().local()[i]);
}
}