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

23
lib/codegen/transform/peephole.cc
View File

@@ -1,7 +1,9 @@
#include <algorithm>
#include <iostream>
#include "triton/ir/module.h"
#include "triton/ir/function.h"
#include "triton/codegen/transform/peephole.h"
#include "triton/codegen/analysis/layout.h"
namespace triton {
namespace codegen{
@@ -109,9 +111,18 @@ bool peephole::rewrite_load_to_shared(ir::instruction *value, ir::builder& build
ir::value *ptr = ld->get_pointer_operand();
ir::value *msk = ld->get_mask_operand();
ir::value *val = ld->get_false_value_operand();
ir::value* new_load = builder.create_masked_load_async(ptr, msk, val);
copy_to_shared->replace_all_uses_with(new_load);
return true;
analysis::scanline_layout* layout = layouts_->get(ptr)->to_scanline();
int nts = layout->nts(layout->get_order()[0]);
int dtsize = value->get_type()->get_scalar_ty()->get_primitive_size_in_bits() / 8;
if(nts*dtsize >= 4){
ir::value* new_load = builder.create_masked_load_async(ptr, msk, val);
copy_to_shared->replace_all_uses_with(new_load);
return true;
}
return false;
// analysis::scanline_layout* layout = layouts_->get(ptr)->to_scanline();
// std::cout << layout->nts(layout->get_order(0)) << std::endl;
// return true;
}
@@ -216,11 +227,11 @@ void peephole::run(ir::module &mod) {
bool was_modified = false;
was_modified = was_modified || rewrite_mult(i, builder);
// was_modified = was_modified || rewrite_cts_cfs(i, builder);
was_modified = was_modified || rewrite_trans_phi(i, builder);
// was_modified = was_modified || rewrite_trans_phi(i, builder);
was_modified = was_modified || rewrite_unit_red(i, builder);
was_modified = was_modified || rewrite_gep_ptr_min_off_plus_off(i, builder);
// if(tgt_->as_nvidia()->sm() >= 80)
// was_modified = was_modified || rewrite_load_to_shared(i, builder);
if(tgt_->as_nvidia()->sm() >= 80)
was_modified = was_modified || rewrite_load_to_shared(i, builder);
if(was_modified)
seen.insert(i);
}