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

[auto-tuning] now not compiling kernels that use too much shared memory

Browse Source
This commit is contained in:
Philippe Tillet
2019-09-05 21:03:09 -04:00
parent 1f8fd525b5
commit b79bcbaee8
1 changed files with 6 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
lib/runtime/function.cc
View File

@@ -169,6 +169,9 @@ function::caller function::autotune(driver::stream* stream, const grid_fn_ty& gr
}catch(const std::runtime_error& e) { }catch(const std::runtime_error& e) {
return; return;
} }
// kernel uses too much resources
if(!bin)
return;
// benchmark // benchmark
ir::function *tmp = ir->get_function_list()[0]; ir::function *tmp = ir->get_function_list()[0];
caller call(tmp, std::move(bin), opt); caller call(tmp, std::move(bin), opt);
@@ -201,7 +204,6 @@ std::unique_ptr<driver::module> function::make_bin(ir::module &module, driver::c
codegen::transform::peephole peephole; codegen::transform::peephole peephole;
codegen::transform::reassociate reassociate(&alignment_info, &grids); codegen::transform::reassociate reassociate(&alignment_info, &grids);
codegen::selection selection(&shmem_allocation, &grids, &shmem_info, &alignment_info, target.get()); codegen::selection selection(&shmem_allocation, &grids, &shmem_info, &alignment_info, target.get());
// run passes // run passes
peephole.run(module); peephole.run(module);
dce.run(module); dce.run(module);
@@ -210,11 +212,12 @@ std::unique_ptr<driver::module> function::make_bin(ir::module &module, driver::c
reassociate.run(module); reassociate.run(module);
dce.run(module); dce.run(module);
peephole.run(module); peephole.run(module);
if(target->is_gpu()){ if(target->is_gpu()){
shmem_info.run(module); shmem_info.run(module);
shmem_liveness.run(module); shmem_liveness.run(module);
shmem_allocation.run(); shmem_allocation.run();
if(shmem_allocation.get_allocated_size() > context->device()->max_shared_memory())
return std::unique_ptr<driver::module>();
shmem_barriers.run(module); shmem_barriers.run(module);
} }
dce.run(module); dce.run(module);
@@ -226,6 +229,7 @@ std::unique_ptr<driver::module> function::make_bin(ir::module &module, driver::c
selection.run(module, *llvm); selection.run(module, *llvm);
// return binary // return binary
std::unique_ptr<driver::module> res(driver::module::create(context, std::move(llvm))); std::unique_ptr<driver::module> res(driver::module::create(context, std::move(llvm)));
// done
return res; return res;
} }