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

Guard against division by zero for no VGPR usage (e.g., in an empty kernel) (#1528)

Browse Source 
* guard against division by zero for no VGPR usage (e.g., in an empty kernel)

* fix bracket format

* clean up parenthesis


[ROCm/hip commit: 73ca2b0083]
This commit is contained in:
Nick Curtis
2019-10-16 00:19:56 -05:00
committed by Maneesh Gupta
parent 22978cb1a3
commit d2e9718d23
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/hip/src/hip_module.cpp
+4 -3
View File
@@ -1326,17 +1326,18 @@ hipError_t ihipOccupancyMaxActiveBlocksPerMultiprocessor(
size_t numWavefronts = (blockSize + wavefrontSize - 1) / wavefrontSize;
size_t availableVGPRs = (prop.regsPerBlock / wavefrontSize / simdPerCU);
size_t vgprs_alu_occupancy = simdPerCU * std::min(maxWavesPerSimd, availableVGPRs / usedVGPRS);
size_t vgprs_alu_occupancy = simdPerCU * (usedVGPRS == 0 ? maxWavesPerSimd
: std::min(maxWavesPerSimd, availableVGPRs / usedVGPRS));
// Calculate blocks occupancy per CU based on VGPR usage
*numBlocks = vgprs_alu_occupancy / numWavefronts;
const size_t availableSGPRs = (prop.gcnArch < 800) ? 512 : 800;
size_t sgprs_alu_occupancy = simdPerCU * ((usedSGPRS == 0) ? maxWavesPerSimd
size_t sgprs_alu_occupancy = simdPerCU * (usedSGPRS == 0 ? maxWavesPerSimd
: std::min(maxWavesPerSimd, availableSGPRs / usedSGPRS));
// Calculate blocks occupancy per CU based on SGPR usage
*numBlocks = std::min(*numBlocks, (uint32_t) (sgprs_alu_occupancy / numWavefronts));
*numBlocks = std::min(*numBlocks, (uint32_t) (sgprs_alu_occupancy / numWavefronts));
size_t total_used_lds = usedLDS + dynSharedMemPerBlk;
if (total_used_lds != 0) {