Media/intel-graphics-compiler
1
0
Fork 0
mirror of https://github.com/intel/intel-graphics-compiler.git synced 2025-11-04 08:21:06 +08:00
Code Issues Packages Projects Releases Wiki Activity

Remove "cutlass" string parse

Browse Source 
This removes the workaround that ignores the value of VectorAliasBBThreshold if the kernel's name contains the strings "_Z" and "cutlass".
This commit is contained in:
Bzowski, Adam
2024-10-22 09:51:21 +00:00
committed by igcbot
parent 2eabaa10f9
commit c2b45279ea
1 changed files with 1 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
IGC/Compiler/CISACodeGen/VariableReuseAnalysis.cpp
View File

@ -1168,12 +1168,7 @@ void VariableReuseAnalysis::InsertElementAliasing(Function* F)
// =0x2: subvec aliasing for both isolated and non-isolated value)
const auto control = (m_pCtx->getVectorCoalescingControl() & 0x3);
// To avoid increasing GRF pressure, skip if F is too large or not an entry
// We remove the threshold when the code comes from CUTLASS,
// which often generates a large number of basic blocks.
const int32_t NumBBThreshold =
F->getName().str().substr(0, 2) == "_Z" && F->getName().str().find("cutlass") != std::string::npos
? std::numeric_limits<int32_t>::max()
: static_cast<int32_t>(IGC_GET_FLAG_VALUE(VectorAliasBBThreshold));
const int32_t NumBBThreshold = IGC_GET_FLAG_VALUE(VectorAliasBBThreshold);
MetaDataUtils* pMdUtils = getAnalysis<MetaDataUtilsWrapper>().getMetaDataUtils();
if (control == 0 || !isEntryFunc(pMdUtils, F) || getNumBBs(F) > NumBBThreshold) {
return;