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

SWDEV-305894 - Cooperative groups grid and multi grid sync support for gfx940+

Browse Source 
Change-Id: I35d72f1cb50c3a96eee56a612b72d641852b145f
This commit is contained in:
Sourabh Betigeri
2022-11-30 19:59:34 +00:00
committed by Sourabh Betigeri
parent 3296dd8636
commit 5d7f3f9f3c
6 changed files with 51 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rocclr/platform/command.hpp
+15 -1
View File
@@ -1050,7 +1050,8 @@ class NDRangeKernelCommand : public Command {
uint32_t numGrids_; //!< Total number of grids in multi GPU launch
uint64_t prevGridSum_; //!< A sum of previous grids to the current launch
uint64_t allGridSum_; //!< A sum of all grids in multi GPU launch
uint32_t firstDevice_; //!< Device index of the first device in the grid
uint32_t firstDevice_; //!< Device index of the first device in the gridc
uint32_t numWorkgroups_; //!< Total number of workgroups in the current launch
public:
enum {
@@ -1118,9 +1119,22 @@ class NDRangeKernelCommand : public Command {
//! Return the index of the first device in multi GPU launch
uint64_t firstDevice() const { return firstDevice_; }
uint32_t numWorkgroups() const { return numWorkgroups_; }
//! Set the local work size.
void setLocalWorkSize(const NDRange& local) { sizes_.local() = local; }
//! Set the number of workgroups
void setNumWorkgroups() {
uint32_t numWorkgroups = 1;
for (uint i = 0; i < sizes().dimensions(); i++) {
if (sizes().local()[i] != 0) {
numWorkgroups *= (sizes().global()[i] / sizes().local()[i]);
}
}
numWorkgroups_ = numWorkgroups;
}
int32_t captureAndValidate();
};