Skip to content

Do reductions properly in tensor contraction tasks #803

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
magnatelee merged 2 commits into from
Feb 21, 2023
Merged

Do reductions properly in tensor contraction tasks #803

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fc016d0
Do reductions properly in tensor contraction tasks
magnatelee Feb 16, 2023
057edd2
Rename the boolean flag to make the meaning clearer
magnatelee Feb 17, 2023
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
1 change: 0 additions & 1 deletion cunumeric_cpp.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,7 +198,6 @@ if(Legion_USE_OpenMP)
src/cunumeric/matrix/transpose_omp.cc
src/cunumeric/matrix/trilu_omp.cc
src/cunumeric/matrix/trsm_omp.cc
src/cunumeric/matrix/util_omp.cc
src/cunumeric/random/rand_omp.cc
src/cunumeric/search/argwhere_omp.cc
src/cunumeric/search/nonzero_omp.cc
Expand Down
18 changes: 12 additions & 6 deletions src/cunumeric/matrix/contract.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,8 @@ struct ContractImplBody<VariantKind::CPU, LegateTypeCode::FLOAT_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_s(&lhs, lhs_ndim, lhs_shape, lhs_data, lhs_strides);
Expand Down Expand Up @@ -80,7 +81,8 @@ struct ContractImplBody<VariantKind::CPU, LegateTypeCode::DOUBLE_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_d(&lhs, lhs_ndim, lhs_shape, lhs_data, lhs_strides);
Expand Down Expand Up @@ -111,7 +113,8 @@ struct ContractImplBody<VariantKind::CPU, LegateTypeCode::HALF_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
// TBLIS doesn't handle half-precision floating point directly, so we have to go through a
// conversion to single-precision.
Expand Down Expand Up @@ -145,7 +148,8 @@ struct ContractImplBody<VariantKind::CPU, LegateTypeCode::HALF_LT> {
rhs2_ndim,
rhs2_shape,
rhs2_copy_strides.data(),
rhs2_modes);
rhs2_modes,
lhs_overwritable);

float_tensor_to_half(lhs_data, lhs_copy_data, lhs_ndim, lhs_shape, lhs_strides);
}
Expand All @@ -167,7 +171,8 @@ struct ContractImplBody<VariantKind::CPU, LegateTypeCode::COMPLEX64_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_c(
Expand Down Expand Up @@ -209,7 +214,8 @@ struct ContractImplBody<VariantKind::CPU, LegateTypeCode::COMPLEX128_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_z(
Expand Down
36 changes: 24 additions & 12 deletions src/cunumeric/matrix/contract.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -80,7 +80,8 @@ __host__ void contract(T* lhs_data,
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
// Initialization
auto handle = get_cutensor();
Expand Down Expand Up @@ -135,7 +136,8 @@ __host__ void contract(T* lhs_data,
cutensorContractionPlan_t plan;
CHECK_CUTENSOR(cutensorInitContractionPlan(handle, &plan, &desc, &find, work_size));
const typename contract_helper<T>::scalar_t alpha = 1.0;
const typename contract_helper<T>::scalar_t beta = 0.0;
// lhs_overwritable being true means that the contraciton tasks can overwrite the lhs
const typename contract_helper<T>::scalar_t beta = lhs_overwritable ? 0.0 : 1.0;
CHECK_CUTENSOR(cutensorContraction(handle,
&plan,
&alpha,
Expand Down Expand Up @@ -167,7 +169,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::HALF_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
contract(lhs_data,
lhs_ndim,
Expand All @@ -183,7 +186,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::HALF_LT> {
rhs2_ndim,
rhs2_shape,
rhs2_strides,
rhs2_modes);
rhs2_modes,
lhs_overwritable);
}
};

Expand All @@ -203,7 +207,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::FLOAT_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
contract(lhs_data,
lhs_ndim,
Expand All @@ -219,7 +224,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::FLOAT_LT> {
rhs2_ndim,
rhs2_shape,
rhs2_strides,
rhs2_modes);
rhs2_modes,
lhs_overwritable);
}
};

Expand All @@ -239,7 +245,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::DOUBLE_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
contract(lhs_data,
lhs_ndim,
Expand All @@ -255,7 +262,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::DOUBLE_LT> {
rhs2_ndim,
rhs2_shape,
rhs2_strides,
rhs2_modes);
rhs2_modes,
lhs_overwritable);
}
};

Expand All @@ -275,7 +283,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::COMPLEX64_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
contract(lhs_data,
lhs_ndim,
Expand All @@ -291,7 +300,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::COMPLEX64_LT> {
rhs2_ndim,
rhs2_shape,
rhs2_strides,
rhs2_modes);
rhs2_modes,
lhs_overwritable);
}
};

Expand All @@ -311,7 +321,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::COMPLEX128_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
contract(lhs_data,
lhs_ndim,
Expand All @@ -327,7 +338,8 @@ struct ContractImplBody<VariantKind::GPU, LegateTypeCode::COMPLEX128_LT> {
rhs2_ndim,
rhs2_shape,
rhs2_strides,
rhs2_modes);
rhs2_modes,
lhs_overwritable);
}
};

Expand Down
27 changes: 16 additions & 11 deletions src/cunumeric/matrix/contract_omp.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,6 @@
#include "cunumeric/matrix/contract.h"
#include "cunumeric/matrix/contract_template.inl"
#include "cunumeric/matrix/util.h"
#include "cunumeric/matrix/util_omp.h"

#include <tblis/tblis.h>
#include <omp.h>
Expand All @@ -43,7 +42,8 @@ struct ContractImplBody<VariantKind::OMP, LegateTypeCode::FLOAT_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_s(&lhs, lhs_ndim, lhs_shape, lhs_data, lhs_strides);
Expand Down Expand Up @@ -74,7 +74,8 @@ struct ContractImplBody<VariantKind::OMP, LegateTypeCode::DOUBLE_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_d(&lhs, lhs_ndim, lhs_shape, lhs_data, lhs_strides);
Expand Down Expand Up @@ -105,25 +106,26 @@ struct ContractImplBody<VariantKind::OMP, LegateTypeCode::HALF_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
// TBLIS doesn't handle half-precision floating point directly, so we have to go through a
// conversion to single-precision.

std::vector<int64_t> lhs_copy_strides(lhs_ndim);
int64_t lhs_size = calculate_volume(lhs_ndim, lhs_shape, lhs_copy_strides.data());
float* lhs_copy_data = allocate_buffer(lhs_size);
half_tensor_to_float_omp(lhs_copy_data, lhs_data, lhs_ndim, lhs_shape, lhs_strides);
half_tensor_to_float(lhs_copy_data, lhs_data, lhs_ndim, lhs_shape, lhs_strides);

std::vector<int64_t> rhs1_copy_strides(rhs1_ndim);
int64_t rhs1_size = calculate_volume(rhs1_ndim, rhs1_shape, rhs1_copy_strides.data());
float* rhs1_copy_data = allocate_buffer(rhs1_size);
half_tensor_to_float_omp(rhs1_copy_data, rhs1_data, rhs1_ndim, rhs1_shape, rhs1_strides);
half_tensor_to_float(rhs1_copy_data, rhs1_data, rhs1_ndim, rhs1_shape, rhs1_strides);

std::vector<int64_t> rhs2_copy_strides(rhs2_ndim);
int64_t rhs2_size = calculate_volume(rhs2_ndim, rhs2_shape, rhs2_copy_strides.data());
float* rhs2_copy_data = allocate_buffer(rhs2_size);
half_tensor_to_float_omp(rhs2_copy_data, rhs2_data, rhs2_ndim, rhs2_shape, rhs2_strides);
half_tensor_to_float(rhs2_copy_data, rhs2_data, rhs2_ndim, rhs2_shape, rhs2_strides);

ContractImplBody<VariantKind::OMP, LegateTypeCode::FLOAT_LT>{}(lhs_copy_data,
lhs_ndim,
Expand All @@ -139,9 +141,10 @@ struct ContractImplBody<VariantKind::OMP, LegateTypeCode::HALF_LT> {
rhs2_ndim,
rhs2_shape,
rhs2_copy_strides.data(),
rhs2_modes);
rhs2_modes,
lhs_overwritable);

float_tensor_to_half_omp(lhs_data, lhs_copy_data, lhs_ndim, lhs_shape, lhs_strides);
float_tensor_to_half(lhs_data, lhs_copy_data, lhs_ndim, lhs_shape, lhs_strides);
}
};

Expand All @@ -161,7 +164,8 @@ struct ContractImplBody<VariantKind::OMP, LegateTypeCode::COMPLEX64_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_c(
Expand Down Expand Up @@ -203,7 +207,8 @@ struct ContractImplBody<VariantKind::OMP, LegateTypeCode::COMPLEX128_LT> {
size_t rhs2_ndim,
int64_t* rhs2_shape,
int64_t* rhs2_strides,
int32_t* rhs2_modes)
int32_t* rhs2_modes,
bool lhs_overwritable)
{
tblis_tensor lhs;
tblis_init_tensor_z(
Expand Down
3 changes: 2 additions & 1 deletion src/cunumeric/matrix/contract_template.inl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -183,7 +183,8 @@ struct ContractImpl {
rhs2_shape.size(),
rhs2_shape.data(),
rhs2_strides.data(),
rhs2_modes.data());
rhs2_modes.data(),
args.lhs.is_readable());

#if 0 // debugging output
std::cout << "end contract kernel:" << std::endl;
Expand Down
Loading