Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[XLA:GPU] Allow Priority Fusion to fuse small constants into Triton fusions. #17252

Merged
merged 1 commit into from
Sep 17, 2024
Merged

[XLA:GPU] Allow Priority Fusion to fuse small constants into Triton fusions. #17252

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 1 addition & 2 deletions xla/service/gpu/fusions/triton/triton_fusion_emitter_parametrized_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2211,8 +2211,7 @@ ENTRY main {
; CHECK: ENTRY
; CHECK-DAG: %[[P0:.*]] = f32[125,127]{1,0} parameter(0)
; CHECK-DAG: %[[P1:.*]] = f32[10,125,127]{2,1,0} parameter(1)
; CHECK-DAG: %[[C0:.*]] = f32[] constant(0)
; CHECK: ROOT %[[FUSION:.*]] = f32[125,127]{1,0} fusion(%[[P0]], %[[P1]], %[[C0]])
; CHECK: ROOT %[[FUSION:.*]] = f32[125,127]{1,0} fusion(%[[P0]], %[[P1]])
; CHECK-SAME: kind=kCustom
; CHECK-SAME: __triton
)";
Expand Down
3 changes: 2 additions & 1 deletion xla/service/gpu/transforms/priority_fusion.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1023,7 +1023,8 @@ absl::StatusOr<bool> PriorityFusion::Run(
for (auto* constant : constants) {
auto users = constant->users();
for (auto* user : users) {
if (IsFusible(*user) && CanEmitInputFusedScatter(*constant, *user)) {
if ((IsFusible(*user) || IsGenericTritonFusion(*user)) &&
CanEmitInputFusedScatter(*constant, *user)) {
Fuse(constant, user);
changed = true;
}
Expand Down
29 changes: 29 additions & 0 deletions xla/service/gpu/transforms/priority_fusion_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -818,6 +818,35 @@ TEST_F(PriorityFusionTest, FuseOnlySmallConstant) {
m::Add(m::Parameter(), m::Broadcast(m::Constant())))));
}

TEST_F(PriorityFusionTest, FuseSmallConstantIntoTritonFusion) {
auto module = *ParseAndReturnVerifiedModule(R"(
HloModule module

add {
Arg_0 = f32[] parameter(0)
Arg_1 = f32[] parameter(1)
ROOT add = f32[] add(Arg_0, Arg_1)
}

triton_computation {
param_0 = f32[32,64] parameter(0)
param_1 = f32[] parameter(1)
ROOT reduce = f32[32] reduce(param_0, param_1), dimensions={1}, to_apply=add
}

ENTRY main {
param_0 = f32[32,64] parameter(0)
c_0 = f32[] constant(0)
ROOT triton_softmax = f32[32] fusion(param_0, c_0), kind=kCustom, calls=triton_computation, backend_config={"fusion_backend_config": {"kind":"__triton","block_level_fusion_config":{"output_tile_sizes":["1"],"num_warps":"1"}}}
})");
EXPECT_THAT(priority_fusion_.Run(module.get()), IsOkAndHolds(true));

HloInstruction* root = module->entry_computation()->root_instruction();
ASSERT_THAT(root, GmockMatch(m::Fusion(m::Parameter())));
EXPECT_THAT(root->fused_expression_root(),
GmockMatch(m::Reduce(m::Parameter(), m::Constant())));
}

TEST_F(PriorityFusionTest, DoNotFuseProducerConsumerMergedTooLarge) {
auto module = *ParseAndReturnVerifiedModule(R"(
HloModule module
Expand Down
Loading