Skip to content

Remove MatmulRole::OUTPUT_AUX #2289

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

Merged
jacobhinkle merged 3 commits into from
May 23, 2024
Merged

Remove MatmulRole::OUTPUT_AUX #2289

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
317249f
Remove MatmulRole::OUTPUT_AUX
jacobhinkle May 22, 2024
1516c7a
Remove comment about AUX
jacobhinkle May 22, 2024
7d6ac69
Merge remote-tracking branch 'origin/main' into remove_output_aux
jacobhinkle May 23, 2024
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
6 changes: 2 additions & 4 deletions csrc/mma_type.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,14 +33,12 @@ enum class MatmulDomain { M = 0, N, K, Batch };
//! INPUT_B - a producer of MMA input B
//! INPUT_C - a producer of a tensor used in fusion epilogue,
//! for example tensor used in beta scaling fusion
//! OUTPUT_D - the main consumer of MMA op results
//! OUTPUT_AUX - fusion outputs that are consumers of OUTPUT_D
//! OUTPUT_D - fusion outputs that have the matmul as a dependency
//!
//! Naming convention is based on the following formula:
//! D = alpha * A x B + beta * C
//! AUX = relu(D)
//! Note: bias vector tensors will be assigned to INPUT_C role.
enum class MatmulRole { INPUT_A = 0, INPUT_B, OUTPUT_D, INPUT_C, OUTPUT_AUX };
enum class MatmulRole { INPUT_A = 0, INPUT_B, INPUT_C, OUTPUT_D };

//! The expected number of occurances of core TensorView roles in fusion
static constexpr size_t MATMUL_CORE_ROLES_EXPECTED_COUNT = 1;
Expand Down
6 changes: 0 additions & 6 deletions csrc/scheduler/matmul_utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -273,12 +273,6 @@ std::string isMatmulFusionDefinitionSupported(
tvs_with_roles.insert(entry->second.begin(), entry->second.end());
}

// Non-core output roles are optional, no requirements for definitions
entry = roles_map.find(MatmulRole::OUTPUT_AUX);
if (entry != roles_map.end()) {
tvs_with_roles.insert(entry->second.begin(), entry->second.end());
}

const auto in_out_tvs_count =
fusion_inputs_tvs.size() + fusion_outputs_tvs.size();
if (in_out_tvs_count != tvs_with_roles.size()) {
Expand Down
19 changes: 2 additions & 17 deletions csrc/scheduler/mma_utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1231,27 +1231,12 @@ RolesMapOpt getTensorRoles(
}
}

// NOTE: sort output roles in descending order by uses() size, and
// if equal then by name() to ensure the stable ordering of tensor
// views in collections assigned to the supported roles
std::sort(storage.begin(), storage.end(), [](TensorView* a, TensorView* b) {
return (a->uses().size() == b->uses().size())
? (a->name() < b->name())
: (a->uses().size() > b->uses().size());
});

Comment on lines -1234 to -1242

@jacobhinkle jacobhinkle May 23, 2024

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I realized this is redundant since we now do this for all roles, not just outputs.

if (!storage.empty()) {
// NOTE: currently, we pick as a reference tensor one with `m` and `n`
// IterDomains and the most uses
auto pos = storage.begin();
tensor_roles[MatmulRole::OUTPUT_D].push_back(*pos);
for (++pos; pos != storage.end(); ++pos) {
tensor_roles[MatmulRole::OUTPUT_AUX].push_back(*pos);
}
tensor_roles[MatmulRole::OUTPUT_D] = storage;
}

for (auto& [role, tvs] : tensor_roles) {
// NOTE: sort input roles in descending order by uses() size, and
// NOTE: sort role tvs in descending order by uses() size, and
// if equal then by name() to ensure the stable ordering of tensor
// views in collections assigned to the supported roles
std::sort(tvs.begin(), tvs.end(), [](TensorView* a, TensorView* b) {
Expand Down