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;
// });
}

17
lib/codegen/analysis/layout.cc
View File

@@ -118,15 +118,6 @@ data_layout::data_layout(id_t id,
// std::cout << max_contiguous[0] << " " << max_contiguous[1] << std::endl;
// std::cout << order_[0] << " " << order_[1] << std::endl;
}
if(is_recoalesce){
if(ptr.size() > 0){
// std::cout << "recoalesce: " << order_[0] << " " << order_[1] << " " << ptr.size() << std::endl;
// std::cout << max_contiguous[0] << " " << max_contiguous[1] << std::endl;
// if(order_[0] == 0)
// exit(1);
}
}
// std::cout << "---" << std::endl;
}
int data_layout::find_axis(int to_find) const {
@@ -213,14 +204,16 @@ scanline_layout::scanline_layout(size_t num_warps,
ir::value *ptr = nullptr;
for(ir::value *v: values)
for(ir::user *usr: v->get_users())
if(auto *st = dynamic_cast<ir::io_inst*>(usr))
ptr = st->get_pointer_operand();
if(auto *io = dynamic_cast<ir::io_inst*>(usr)){
if(!ptr || ptr->get_type()->get_tile_rank() < io->get_pointer_operand()->get_type()->get_tile_rank())
ptr = io->get_pointer_operand();
}
unsigned i = order_[0];
int contiguous = 1;
if(ptr){
int nbits = ptr->get_type()->get_pointer_element_ty()->get_scalar_ty()->get_primitive_size_in_bits();
contiguous = std::min<int>(align->contiguous(ptr)[i], 128 / nbits);
contiguous = std::min<int>(align->get(ptr, i), 128 / nbits);
}
nts_[i] = clamp(size / num_threads, 1, std::min<int>(contiguous, shape_[i]));