Skip to content

simplify blas backend when using with USE_ONEMATH_CUBLAS #2522

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
merged 3 commits into from
Jul 12, 2025
+3 −51
Merged

simplify blas backend when using with USE_ONEMATH_CUBLAS #2522

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
bb71239
simplify blas backend when using with USE_ONEMATH_CUBLAS
vtavana Jul 10, 2025
a240c4a
use constexpr for defining is_row_major
vtavana Jul 11, 2025
6de7918
Merge branch 'master' into blas-backend
vtavana Jul 12, 2025
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
18 changes: 1 addition & 17 deletions dpnp/backend/extensions/blas/gemm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,9 +55,7 @@ typedef sycl::event (*gemm_impl_fn_ptr_t)(sycl::queue &,
const std::int64_t,
char *,
const std::int64_t,
#if !defined(USE_ONEMATH_CUBLAS)
const bool,
#endif // !USE_ONEMATH_CUBLAS
const std::vector<sycl::event> &);

static gemm_impl_fn_ptr_t gemm_dispatch_table[dpctl_td_ns::num_types]
Expand All @@ -76,9 +74,7 @@ static sycl::event gemm_impl(sycl::queue &exec_q,
const std::int64_t ldb,
char *resultC,
const std::int64_t ldc,
#if !defined(USE_ONEMATH_CUBLAS)
const bool is_row_major,
#endif // !USE_ONEMATH_CUBLAS
const std::vector<sycl::event> &depends)
{
type_utils::validate_type_for_device<Tab>(exec_q);
Expand All @@ -100,11 +96,6 @@ static sycl::event gemm_impl(sycl::queue &exec_q,
const Tab *a, const std::int64_t lda, const Tab *b,
const std::int64_t ldb, Tab beta, Tc *c, const std::int64_t ldc,
const std::vector<sycl::event> &deps) -> sycl::event {
#if defined(USE_ONEMATH_CUBLAS)
return mkl_blas::column_major::gemm(q, transA, transB, m, n, k,
alpha, a, lda, b, ldb, beta, c,
ldc, deps);
#else
if (is_row_major) {
return mkl_blas::row_major::gemm(q, transA, transB, m, n, k,
alpha, a, lda, b, ldb, beta, c,
Expand All @@ -115,7 +106,6 @@ static sycl::event gemm_impl(sycl::queue &exec_q,
alpha, a, lda, b, ldb, beta,
c, ldc, deps);
}
#endif // USE_ONEMATH_CUBLAS
};
gemm_event = gemm_func(
exec_q,
Expand Down Expand Up @@ -242,7 +232,7 @@ std::tuple<sycl::event, sycl::event, bool>

// cuBLAS supports only column-major storage
#if defined(USE_ONEMATH_CUBLAS)
const bool is_row_major = false;
constexpr bool is_row_major = false;

transA = is_matrixA_c_contig ? oneapi::mkl::transpose::T
: oneapi::mkl::transpose::N;
Expand Down Expand Up @@ -320,15 +310,9 @@ std::tuple<sycl::event, sycl::event, bool>
const char *b_typeless_ptr = matrixB.get_data();
char *r_typeless_ptr = resultC.get_data();

#if defined(USE_ONEMATH_CUBLAS)
sycl::event gemm_ev =
gemm_fn(exec_q, transA, transB, m, n, k, a_typeless_ptr, lda,
b_typeless_ptr, ldb, r_typeless_ptr, ldc, depends);
#else
sycl::event gemm_ev = gemm_fn(exec_q, transA, transB, m, n, k,
a_typeless_ptr, lda, b_typeless_ptr, ldb,
r_typeless_ptr, ldc, is_row_major, depends);
#endif // USE_ONEMATH_CUBLAS

sycl::event args_ev = dpctl::utils::keep_args_alive(
exec_q, {matrixA, matrixB, resultC}, {gemm_ev});
Expand Down
19 changes: 1 addition & 18 deletions dpnp/backend/extensions/blas/gemm_batch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,9 +60,7 @@ typedef sycl::event (*gemm_batch_impl_fn_ptr_t)(
const char *,
const char *,
char *,
#if !defined(USE_ONEMATH_CUBLAS)
const bool,
#endif // !USE_ONEMATH_CUBLAS
const std::vector<sycl::event> &);

static gemm_batch_impl_fn_ptr_t
Expand All @@ -85,9 +83,7 @@ static sycl::event gemm_batch_impl(sycl::queue &exec_q,
const char *matrixA,
const char *matrixB,
char *resultC,
#if !defined(USE_ONEMATH_CUBLAS)
const bool is_row_major,
#endif // !USE_ONEMATH_CUBLAS
const std::vector<sycl::event> &depends)
{
type_utils::validate_type_for_device<Tab>(exec_q);
Expand All @@ -112,11 +108,6 @@ static sycl::event gemm_batch_impl(sycl::queue &exec_q,
Tc *c, const std::int64_t ldc, const std::int64_t stridec,
const std::int64_t batch_size,
const std::vector<sycl::event> &deps) -> sycl::event {
#if defined(USE_ONEMATH_CUBLAS)
return mkl_blas::column_major::gemm_batch(
q, transA, transB, m, n, k, alpha, a, lda, stridea, b, ldb,
strideb, beta, c, ldc, stridec, batch_size, deps);
#else
if (is_row_major) {
return mkl_blas::row_major::gemm_batch(
q, transA, transB, m, n, k, alpha, a, lda, stridea, b, ldb,
Expand All @@ -127,7 +118,6 @@ static sycl::event gemm_batch_impl(sycl::queue &exec_q,
q, transA, transB, m, n, k, alpha, a, lda, stridea, b, ldb,
strideb, beta, c, ldc, stridec, batch_size, deps);
}
#endif // USE_ONEMATH_CUBLAS
};
gemm_batch_event = gemm_batch_func(
exec_q,
Expand Down Expand Up @@ -317,7 +307,7 @@ std::tuple<sycl::event, sycl::event, bool>

// cuBLAS supports only column-major storage
#if defined(USE_ONEMATH_CUBLAS)
const bool is_row_major = false;
constexpr bool is_row_major = false;

transA = A_base_is_c_contig ? oneapi::mkl::transpose::T
: oneapi::mkl::transpose::N;
Expand Down Expand Up @@ -396,17 +386,10 @@ std::tuple<sycl::event, sycl::event, bool>
const char *b_typeless_ptr = matrixB.get_data();
char *r_typeless_ptr = resultC.get_data();

#if defined(USE_ONEMATH_CUBLAS)
sycl::event gemm_batch_ev =
gemm_batch_fn(exec_q, m, n, k, batch_size, lda, ldb, ldc, stridea,
strideb, stridec, transA, transB, a_typeless_ptr,
b_typeless_ptr, r_typeless_ptr, depends);
#else
sycl::event gemm_batch_ev =
gemm_batch_fn(exec_q, m, n, k, batch_size, lda, ldb, ldc, stridea,
strideb, stridec, transA, transB, a_typeless_ptr,
b_typeless_ptr, r_typeless_ptr, is_row_major, depends);
#endif // USE_ONEMATH_CUBLAS

sycl::event args_ev = dpctl::utils::keep_args_alive(
exec_q, {matrixA, matrixB, resultC}, {gemm_batch_ev});
Expand Down
17 changes: 1 addition & 16 deletions dpnp/backend/extensions/blas/gemv.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,9 +53,7 @@ typedef sycl::event (*gemv_impl_fn_ptr_t)(sycl::queue &,
const std::int64_t,
char *,
const std::int64_t,
#if !defined(USE_ONEMATH_CUBLAS)
const bool,
#endif // !USE_ONEMATH_CUBLAS
const std::vector<sycl::event> &);

static gemv_impl_fn_ptr_t gemv_dispatch_vector[dpctl_td_ns::num_types];
Expand All @@ -71,9 +69,7 @@ static sycl::event gemv_impl(sycl::queue &exec_q,
const std::int64_t incx,
char *vectorY,
const std::int64_t incy,
#if !defined(USE_ONEMATH_CUBLAS)
const bool is_row_major,
#endif // !USE_ONEMATH_CUBLAS
const std::vector<sycl::event> &depends)
{
type_utils::validate_type_for_device<T>(exec_q);
Expand All @@ -93,10 +89,6 @@ static sycl::event gemv_impl(sycl::queue &exec_q,
const std::int64_t lda, const T *x, const std::int64_t incx,
T beta, T *y, const std::int64_t incy,
const std::vector<sycl::event> &deps) -> sycl::event {
#if defined(USE_ONEMATH_CUBLAS)
return mkl_blas::column_major::gemv(q, transA, m, n, alpha, a, lda,
x, incx, beta, y, incy, deps);
#else
if (is_row_major) {
return mkl_blas::row_major::gemv(q, transA, m, n, alpha, a, lda,
x, incx, beta, y, incy, deps);
Expand All @@ -106,7 +98,6 @@ static sycl::event gemv_impl(sycl::queue &exec_q,
lda, x, incx, beta, y, incy,
deps);
}
#endif // USE_ONEMATH_CUBLAS
};
gemv_event = gemv_func(
exec_q,
Expand Down Expand Up @@ -196,7 +187,7 @@ std::pair<sycl::event, sycl::event>

// cuBLAS supports only column-major storage
#if defined(USE_ONEMATH_CUBLAS)
const bool is_row_major = false;
constexpr bool is_row_major = false;
std::int64_t m;
std::int64_t n;

Expand Down Expand Up @@ -304,15 +295,9 @@ std::pair<sycl::event, sycl::event>
y_typeless_ptr -= (y_shape[0] - 1) * std::abs(incy) * y_elemsize;
}

#if defined(USE_ONEMATH_CUBLAS)
sycl::event gemv_ev =
gemv_fn(exec_q, transA, m, n, a_typeless_ptr, lda, x_typeless_ptr, incx,
y_typeless_ptr, incy, depends);
#else
sycl::event gemv_ev =
gemv_fn(exec_q, transA, m, n, a_typeless_ptr, lda, x_typeless_ptr, incx,
y_typeless_ptr, incy, is_row_major, depends);
#endif // USE_ONEMATH_CUBLAS

sycl::event args_ev = dpctl::utils::keep_args_alive(
exec_q, {matrixA, vectorX, vectorY}, {gemv_ev});
Expand Down
Loading