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

SWDEV-287351 - Fix hipStreamWithCUMask hang on Navi21 (#2939)

Browse Source 
- For gfx >= 10, one work group processor encompasses 2 CUs & hence the CUs need to be enabled in pair

Change-Id: I359df1b221b4400b260b02201b7b0385054784f3
This commit is contained in:
ROCm CI Service Account
2022-09-23 12:37:42 +05:30
committed by GitHub
parent 8c8499be77
commit 2b09d94ea2
4 changed files with 49 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/src/runtimeApi/stream/hipStreamGetCUMask.cpp
+7 -1
View File
@@ -124,7 +124,13 @@ int main(int argc, char* argv[]) {
cout << "info: CU mask for the default stream is: 0x" << ss.str().c_str() << endl;
vector<uint32_t> cuMask1(defaultCUMask);
cuMask1[0] = 0xe;
if (props.major >= 10) {
// For gfx >= 10, one work group processor encompasses 2 CUs &
// hence the CUs need to be enabled in pair
cuMask1[0] = 0xc;
} else {
cuMask1[0] = 0xe;
}
HIPCHECK(hipExtStreamCreateWithCUMask(&stream, cuMask1.size(), cuMask1.data()));
ss.str("");
tests/src/runtimeApi/stream/hipStreamWithCUMask.cpp
+15 -8
View File
@@ -72,6 +72,13 @@ int main(int argc, char* argv[]) {
hA[i] = 1.618f + i;
}
int cuCountPerGroup = 1;
if (props.major >= 10) {
cuCountPerGroup = 2; // For gfx >= 10, one work group processor encompasses 2 CUs
}
unsigned long mask = pow(2, cuCountPerGroup) - 1;
for (int np = 0; np < numPartition; np++) {
HIPCHECK(hipMalloc(&dA[np], Nbytes));
@@ -79,16 +86,16 @@ int main(int argc, char* argv[]) {
// make unique CU masks in the multiple of dwords for each stream
uint32_t temp = 0;
uint32_t bit_index = np;
for (int i = np; i < props.multiProcessorCount; i = i + 4) {
temp |= 1UL << bit_index;
uint32_t bit_index = cuCountPerGroup * np;
for (int i = np; i < props.multiProcessorCount; i = i + cuCountPerGroup * 4) {
temp |= mask << bit_index;
if (bit_index >= 32) {
cuMasks[np].push_back(temp);
temp = 0;
bit_index = np;
temp |= 1UL << bit_index;
bit_index = cuCountPerGroup * np;
temp |= mask << bit_index;
}
bit_index += 4;
bit_index += cuCountPerGroup * 4;
}
if (bit_index != 0) {
cuMasks[np].push_back(temp);
@@ -98,9 +105,9 @@ int main(int argc, char* argv[]) {
HIPCHECK(hipMemcpy(dA[np], hA, Nbytes, hipMemcpyHostToDevice));
ss[np] << std::hex;
ss[np] << std::hex << std::setfill('0');
for (int i = cuMasks[np].size() - 1; i >= 0; i--) {
ss[np] << cuMasks[np][i];
ss[np] << std::setw(8) << cuMasks[np][i];
}
}