gaspersic/triton
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[CODEGEN] Major performance improvements on A100 (#70)

Browse Source 
Improved handling of asynchronous copy, scheduling and synchronization for A100. Now achieving CUTLASS-like performance on large square dense matrix multiplication tasks
This commit is contained in:
Philippe Tillet
2021-02-21 15:19:39 -08:00
committed by Philippe Tillet
parent 045ab5d62a
commit 5b83259592
31 changed files with 1331 additions and 1115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lib/codegen/analysis/align.cc
View File

@@ -312,7 +312,6 @@ std::vector<unsigned> align::populate_max_contiguous_gep(ir::getelementptr_inst*
if(rhs_cst_info[d].num_cst)
rvalue = lhs_max_contiguous[d];
result[d] = std::max(lvalue, rvalue);
// std::cout << "max contiguous: " << x->get_name() << " " << d << " " << result[d] << std::endl;
}
return add_to_cache(x, result, max_contiguous_);
}
@@ -527,8 +526,7 @@ void align::run(ir::module &mod) {
ir::for_each_value(mod, [this](ir::value* v) { populate(v); } );
// ir::for_each_value(mod, [this](ir::value* v) {
// if(dynamic_cast<ir::cast_inst*>(v) || dynamic_cast<ir::getelementptr_inst*>(v))
// std::cout << "ALIGN: " << v->get_name() << " " << starting_multiple_.at(v)[0] << " " << max_contiguous_.at(v)[0]
// << " " << starting_multiple_.at(v)[1] << " " << max_contiguous_.at(v)[1] << std::endl;
// std::cout << "ALIGN: " << v->get_name() << " " << max_contiguous_.at(v)[0] << " " << max_contiguous_.at(v)[1] << std::endl;
// });
}