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

[CODEGEN] Improvements and bugfixes (#463)

Browse Source
This commit is contained in:
Philippe Tillet
2022-02-24 14:56:24 -08:00
committed by GitHub
parent a9dfdcaaa9
commit 98ed7db8c1
14 changed files with 154 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
lib/codegen/selection/generator.cc
View File

@@ -119,7 +119,7 @@ Value* geper::operator()(Value *ptr, Value* off, const std::string& name){
#define icmp_ult(...) builder_->CreateICmpULT(__VA_ARGS__)
#define insert_elt(...) builder_->CreateInsertElement(__VA_ARGS__)
#define intrinsic(...) builder_->CreateIntrinsic(__VA_ARGS__)
#define load(...) builder_->CreateLoad(__VA_ARGS__)
#define load(ptr) builder_->CreateLoad(ptr->getType()->getPointerElementType(), ptr)
#define lshr(...) builder_->CreateLShr(__VA_ARGS__)
#define max_num(...) builder_->CreateMaxNum(__VA_ARGS__)
#define min_num(...) builder_->CreateMinNum(__VA_ARGS__)
@@ -576,18 +576,19 @@ void generator::visit_cast_inst(ir::cast_inst* x) {
// <> BF16
if(ret_sca_ty->is_bf16_ty() || op_sca_ty->is_bf16_ty()){
// FP32 -> BF16
if(op_sca_ty->is_fp32_ty())
// for(size_t i = 0; i < x_idxs.size(); i++)
// vals_[x][x_idxs[i + 0]] = fp32_to_bf16(vals_[op][op_idxs[i + 0]]);
if(op_sca_ty->is_fp32_ty()){
for (indices_t idx: idxs_.at(x)) {
Value *arg = vals_[x->get_operand(0)][idx];
vals_[x][idx] = fp32_to_bf16(arg); // cast(cvt(x->get_op()), arg, ty);
}
return;
}
// BF16 -> FP32
if(ret_sca_ty->is_fp32_ty())
if(ret_sca_ty->is_fp32_ty()){
for(size_t i = 0; i < x_idxs.size(); i++)
vals_[x][x_idxs[i + 0]] = bf16_to_fp32(vals_[op][op_idxs[i + 0]]);
return;
return;
}
}
@@ -697,12 +698,13 @@ void generator::visit_load_inst(ir::load_inst* x){
std::ostringstream asm_oss;
asm_oss << "@$" << n_words; // predicate
asm_oss << " ld";
// std::cout << x->get_is_volatile() << std::endl;
if(x->get_is_volatile())
asm_oss << ".volatile";
asm_oss << ".global";
if (x->get_cache_modifier() == ir::load_inst::CA) asm_oss << ".ca";
if (x->get_cache_modifier() == ir::load_inst::CG) asm_oss << ".cg";
if (x->get_eviction_policy() == ir::load_inst::EVICT_LAST) asm_oss << ".L1::evict_last";
if (x->get_eviction_policy() == ir::load_inst::EVICT_FIRST) asm_oss << ".L1::evict_first";
if(n_words > 1)
asm_oss << ".v" << n_words; // vector width
asm_oss << ".b" << width; // word size

3
lib/codegen/transform/peephole.cc
View File

@@ -123,7 +123,7 @@ bool peephole::rewrite_load_to_shared(ir::instruction *value, ir::builder& build
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, ld->get_cache_modifier());
ir::value* new_load = builder.create_masked_load_async(ptr, msk, val, ld->get_cache_modifier(), ld->get_eviction_policy());
copy_to_shared->replace_all_uses_with(new_load);
return true;
}
@@ -215,6 +215,7 @@ bool peephole::rewrite_select_masked_load(ir::instruction *value, ir::builder& b
if_value->get_mask_operand(),
select->get_else_value_op(),
if_value->get_cache_modifier(),
if_value->get_eviction_policy(),
if_value->get_is_volatile());
select->replace_all_uses_with(new_load);
return true;

10
lib/codegen/transform/pipeline.cc
View File

@@ -178,7 +178,7 @@ void pipeline::run(ir::module &mod) {
false_value = remat_false_value;
} else
false_value = builder.create_splat(ir::undef_value::get(ty->get_scalar_ty()), ty->get_block_shapes());
first_loads[0] = builder.create_masked_load(first_ptrs[0], first_masks[0], false_value, load->get_cache_modifier(), load->get_is_volatile());
first_loads[0] = builder.create_masked_load(first_ptrs[0], first_masks[0], false_value, load->get_cache_modifier(), load->get_eviction_policy(), load->get_is_volatile());
for (int stage = 1; stage < num_stages-1; ++stage) {
// mask is the loop condition of the previous iteration
@@ -193,7 +193,7 @@ void pipeline::run(ir::module &mod) {
first_masks[stage] = builder.create_and(first_masks[stage], remat_mask);
false_value = remat_false_value;
}
first_loads[stage] = builder.create_masked_load(first_ptrs[stage], first_masks[stage], false_value, load->get_cache_modifier(), load->get_is_volatile());
first_loads[stage] = builder.create_masked_load(first_ptrs[stage], first_masks[stage], false_value, load->get_cache_modifier(), load->get_eviction_policy(), load->get_is_volatile());
}
// create new phis for induction variables
@@ -222,7 +222,7 @@ void pipeline::run(ir::module &mod) {
next_mask = builder.create_and(next_mask, remat_mask);
false_value = remat_false_value;
}
ir::value* next_load = builder.create_masked_load(next_ptr, next_mask, false_value, load->get_cache_modifier(), load->get_is_volatile());
ir::value* next_load = builder.create_masked_load(next_ptr, next_mask, false_value, load->get_cache_modifier(), load->get_eviction_policy(), load->get_is_volatile());
// phi node
@@ -257,7 +257,7 @@ void pipeline::run(ir::module &mod) {
}
else
false_value = builder.create_splat(ir::undef_value::get(ty->get_scalar_ty()), ty->get_block_shapes());
ir::value* first_load = builder.create_masked_load(first_ptr, first_mask, false_value, load->get_cache_modifier(), load->get_is_volatile());
ir::value* first_load = builder.create_masked_load(first_ptr, first_mask, false_value, load->get_cache_modifier(), load->get_eviction_policy(), load->get_is_volatile());
// pre-fetch next iteration
builder.set_insert_point(block->get_inst_list().back());
ir::value* next_ptr = ptr->get_value_for_block(block);
@@ -268,7 +268,7 @@ void pipeline::run(ir::module &mod) {
next_mask = builder.create_and(next_mask, remat_mask);
false_value = remat_false_value;
}
ir::value* next_load = builder.create_masked_load(next_ptr, next_mask, false_value, load->get_cache_modifier(), load->get_is_volatile());
ir::value* next_load = builder.create_masked_load(next_ptr, next_mask, false_value, load->get_cache_modifier(), load->get_eviction_policy(), load->get_is_volatile());
// phi node
builder.set_insert_point(block->get_first_non_phi());
ir::phi_node* new_load = builder.create_phi(ty, 2);