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

[CODEGEN] Now padding shared memory for layout conversion (#468)

Browse Source
This commit is contained in:
Philippe Tillet
2022-03-03 22:19:05 -08:00
committed by GitHub
parent d9dd97492f
commit bb5765df5c
4 changed files with 62 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
lib/codegen/analysis/swizzle.cc
View File

@@ -14,41 +14,42 @@ void swizzle::run(ir::module &) {
max_phase_.clear();
for(auto &x: layouts_->get_all()){
shared_layout* layout = dynamic_cast<shared_layout*>(x.second);
if(!layout)
shared_layout* out_layout = dynamic_cast<shared_layout*>(x.second);
if(!out_layout)
continue;
ir::value* mma_dot_a = layout->hmma_dot_a();
ir::value* mma_dot_b = layout->hmma_dot_b();
if(!mma_dot_a && !mma_dot_b){
per_phase_[layout] = 1;
max_phase_[layout] = 1;
vec_[layout] = 1;
continue;
}
auto ord = layout->get_order();
scanline_layout* in_layout = dynamic_cast<scanline_layout*>(layout->get_arg_layout());
scanline_layout* in_layout = dynamic_cast<scanline_layout*>(out_layout->get_arg_layout());
if(!in_layout)
continue;
int dtsize = layout->get_type()->get_scalar_ty()->get_primitive_size_in_bits() / 8;
ir::value* mma_dot_a = out_layout->hmma_dot_a();
ir::value* mma_dot_b = out_layout->hmma_dot_b();
if(!mma_dot_a && !mma_dot_b){
per_phase_[out_layout] = 1;
max_phase_[out_layout] = 1;
vec_[out_layout] = 1;
continue;
}
auto ord = out_layout->get_order();
int dtsize = out_layout->get_type()->get_scalar_ty()->get_primitive_size_in_bits() / 8;
if(tgt_->as_nvidia() && tgt_->as_nvidia()->sm() < 80){
int inner = mma_dot_a ? 0 : 1;
per_phase_[layout] = std::max<int>(128 / (in_layout->mts(ord[0])*in_layout->nts(ord[0])*dtsize), 1);
max_phase_[layout] = (ord[inner] == 1 ? 8 : 4) / per_phase_[layout];
per_phase_[out_layout] = std::max<int>(128 / (in_layout->mts(ord[0])*in_layout->nts(ord[0])*dtsize), 1);
max_phase_[out_layout] = (ord[inner] == 1 ? 8 : 4) / per_phase_[out_layout];
if(mma_dot_a)
vec_[layout] = 2*layouts_->get(mma_dot_a)->to_mma()->rep(0);
vec_[out_layout] = 2*layouts_->get(mma_dot_a)->to_mma()->rep(0);
else
vec_[layout] = 2*layouts_->get(mma_dot_b)->to_mma()->rep(1);
vec_[out_layout] = 2*layouts_->get(mma_dot_b)->to_mma()->rep(1);
}
else {
if (!layout->allow_swizzle()) {
per_phase_[layout] = 1;
max_phase_[layout] = 1;
vec_[layout] = 1;
if (!out_layout->allow_swizzle()) {
per_phase_[out_layout] = 1;
max_phase_[out_layout] = 1;
vec_[out_layout] = 1;
} else {
per_phase_[layout] = std::max<int>(128 / (in_layout->mts(ord[0])*in_layout->nts(ord[0])*dtsize), 1);
max_phase_[layout] = layout->get_mma_strided() / per_phase_[layout];
vec_[layout] = layout->get_mma_vec();
per_phase_[out_layout] = std::max<int>(128 / (in_layout->mts(ord[0])*in_layout->nts(ord[0])*dtsize), 1);
max_phase_[out_layout] = out_layout->get_mma_strided() / per_phase_[out_layout];
vec_[out_layout] = out_layout->get_mma_vec();
}
}
}