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

[CODEGEN] Performance improvement on A100 (#125)

Browse Source 
Improved codegen for the Ampere GPUs.

    * Make the layout pass recognize the multistage pipelined pattern.
    * Now the pipeline pass can automate the multistage pipelining transformation.
    * Remove extra barriers (from the prefetch pass & WAR) on Ampere.
    * Update the code generator (generator.cc) to make Triton generate n-buffered shared memory loads/stores.
This commit is contained in:
daadaada
2021-06-21 14:25:13 +08:00
committed by Philippe Tillet
parent 5a51f3e529
commit d8d6b715c8
21 changed files with 855 additions and 174 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
lib/ir/print.cc
View File

@@ -77,6 +77,54 @@ void print(module &mod, std::ostream& os) {
}
}
void print(function &fn, std::ostream &os) {
//
}
void print(basic_block &bb, std::ostream &os) {
auto const &predecessors = bb.get_predecessors();
os << bb.get_name() << ":";
if(!predecessors.empty()){
os << " ";
os << "; preds = ";
auto const &predecessors = bb.get_predecessors();
for(ir::basic_block *pred: predecessors)
os << pred->get_name() << (pred!=predecessors.back()?", ":"");
}
os << std::endl;
for(ir::instruction *inst: bb.get_inst_list()){
print(*inst, os);
}
}
void print(instruction &instr, std::ostream &os) {
instruction *inst = &instr;
os << " ";
if(!inst->get_type()->is_void_ty()){
os << instr.get_name();
os << " = ";
}
ir::type* type = inst->get_type();
os << inst->repr() << " " << type->repr();
ir::instruction::ops_t ops = inst->ops();
size_t num_ops = inst->get_num_operands();
if(num_ops > 0)
os << " ";;
for(unsigned i = 0; i < num_ops; i++){
if(auto *x = dynamic_cast<ir::constant*>(ops[i]))
os << x->repr();
else
os << ops[i]->get_name();
os << (i < num_ops - 1?", ":"");
}
os << ";";
// os << " (";
// for(ir::user* usr: inst->get_users())
// os << get_name(usr, cnt++) << ", " ;
// os << " )";
os << std::endl;
}
}
}