derek/gem5
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

gpu-compute: Update Requests for invalidations

Browse Source 
The SQC and TCC invalidations share a Request pointer which they both
modify. This can cause some problems, so use a different request pointer
for each invalidate. The setContext call is also removed as the value
being assigned to it is uninitialized.

Change-Id: I82ea7aa44a4f4515c1560993caa26cc6a89355af
This commit is contained in:
Matthew Poremba
2024-08-07 12:52:48 -07:00
parent db0d5f19cf
commit 84fedecafe
2 changed files with 9 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/gpu-compute/compute_unit.cc
View File

@@ -409,8 +409,6 @@ ComputeUnit::doInvalidate(RequestPtr req, int kernId){
// kern_id will be used in inv responses
gpuDynInst->kern_id = kernId;
// update contextId field
req->setContext(gpuDynInst->wfDynId);
injectGlobalMemFence(gpuDynInst, true, req);
}
@@ -438,8 +436,6 @@ ComputeUnit::doSQCInvalidate(RequestPtr req, int kernId){
// kern_id will be used in inv responses
gpuDynInst->kern_id = kernId;
// update contextId field
req->setContext(gpuDynInst->wfDynId);
gpuDynInst->staticInstruction()->setFlag(GPUStaticInst::Scalar);
scalarMemoryPipe.injectScalarMemFence(gpuDynInst, true, req);

14
src/gpu-compute/shader.cc
View File

@@ -214,19 +214,23 @@ Shader::prepareInvalidate(HSAQueueEntry *task) {
for (int i_cu = 0; i_cu < n_cu; ++i_cu) {
// create a request to hold INV info; the request's fields will
// be updated in cu before use
auto req = std::make_shared<Request>(0, 0, 0,
cuList[i_cu]->requestorId(),
0, -1);
auto tcc_req = std::make_shared<Request>(0, 0, 0,
cuList[i_cu]->requestorId(),
0, -1);
_dispatcher.updateInvCounter(kernId, +1);
// all necessary INV flags are all set now, call cu to execute
cuList[i_cu]->doInvalidate(req, task->dispatchId());
cuList[i_cu]->doInvalidate(tcc_req, task->dispatchId());
// A set of CUs share a single SQC cache. Send a single invalidate
// request to each SQC
auto sqc_req = std::make_shared<Request>(0, 0, 0,
cuList[i_cu]->requestorId(),
0, -1);
if ((i_cu % n_cu_per_sqc) == 0) {
cuList[i_cu]->doSQCInvalidate(req, task->dispatchId());
cuList[i_cu]->doSQCInvalidate(sqc_req, task->dispatchId());
}
// I don't like this. This is intrusive coding.