Release GIL around blocking oneMKL calls in BLAS/LAPACK

CHANGELOG.md

@@ -82,6 +82,8 @@ This release is compatible with NumPy 2.5.
 * Fixed `dpnp.tensor.top_k` aborting for `k=0` by returning empty result arrays without launching a zero-sized kernel [#3022](https://github.com/IntelPython/dpnp/pull/3022)
 * Fixed comparison functions (`dpnp.equal`, `dpnp.not_equal`, `dpnp.less`, `dpnp.less_equal`, `dpnp.greater`, `dpnp.greater_equal`) and `dpnp.divide` raising `OverflowError` when comparing an integer array against a Python integer scalar outside the array dtype's range [#3017](https://github.com/IntelPython/dpnp/pull/3017)
 * Fixed a crash in boolean-mask advanced indexing (`dpnp.ndarray` get/set item) when the selection is empty (e.g. a scalar `False` index that injects a length-0 axis) [#3019](https://github.com/IntelPython/dpnp/pull/3019)
+* Released the GIL before the remaining blocking OneMKL BLAS and LAPACK calls to prevent host tasks contention, completing the work started in [#2850](https://github.com/IntelPython/dpnp/pull/2850) [#3027](https://github.com/IntelPython/dpnp/pull/3027)
+
 
 ### Security
 

dpnp/backend/extensions/blas/dot.hpp

@@ -30,12 +30,15 @@
 
 #include <stdexcept>
 
+#include <pybind11/pybind11.h>
+
 #include "dot_common.hpp"
 
 namespace dpnp::extensions::blas
 {
 namespace mkl_blas = oneapi::mkl::blas;
 namespace type_utils = dpnp::tensor::type_utils;
+namespace py = pybind11;
 
 template <typename T>
 static sycl::event dot_impl(sycl::queue &exec_q,
@@ -58,6 +61,10 @@ static sycl::event dot_impl(sycl::queue &exec_q,
 
     sycl::event dot_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         dot_event = mkl_blas::column_major::dot(exec_q,
                                                 n, // size of the input vectors
                                                 x, // Pointer to vector x.

dpnp/backend/extensions/blas/dotc.hpp

@@ -30,12 +30,15 @@
 
 #include <stdexcept>
 
+#include <pybind11/pybind11.h>
+
 #include "dot_common.hpp"
 
 namespace dpnp::extensions::blas
 {
 namespace mkl_blas = oneapi::mkl::blas;
 namespace type_utils = dpnp::tensor::type_utils;
+namespace py = pybind11;
 
 template <typename T>
 static sycl::event dotc_impl(sycl::queue &exec_q,
@@ -58,6 +61,10 @@ static sycl::event dotc_impl(sycl::queue &exec_q,
 
     sycl::event dotc_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         dotc_event =
             mkl_blas::column_major::dotc(exec_q,
                                          n,    // size of the input vectors

dpnp/backend/extensions/blas/dotu.hpp

@@ -30,12 +30,15 @@
 
 #include <stdexcept>
 
+#include <pybind11/pybind11.h>
+
 #include "dot_common.hpp"
 
 namespace dpnp::extensions::blas
 {
 namespace mkl_blas = oneapi::mkl::blas;
 namespace type_utils = dpnp::tensor::type_utils;
+namespace py = pybind11;
 
 template <typename T>
 static sycl::event dotu_impl(sycl::queue &exec_q,
@@ -58,6 +61,10 @@ static sycl::event dotu_impl(sycl::queue &exec_q,
 
     sycl::event dotu_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         dotu_event =
             mkl_blas::column_major::dotu(exec_q,
                                          n,    // size of the input vectors

dpnp/backend/extensions/blas/gemm.cpp

@@ -113,6 +113,10 @@ static sycl::event gemm_impl(sycl::queue &exec_q,
                                                     c, ldc, deps);
             }
         };
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         gemm_event = gemm_func(
             exec_q,
             transA, // Defines the transpose operation for matrix A:

dpnp/backend/extensions/blas/gemm_batch.cpp

@@ -125,6 +125,10 @@ static sycl::event gemm_batch_impl(sycl::queue &exec_q,
                     strideb, beta, c, ldc, stridec, batch_size, deps);
             }
         };
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         gemm_batch_event = gemm_batch_func(
             exec_q,
             transA,     // Defines the transpose operation for matrix A:

dpnp/backend/extensions/lapack/geqrf.cpp

@@ -86,6 +86,10 @@ static sycl::event geqrf_impl(sycl::queue &exec_q,
 
     sycl::event geqrf_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         geqrf_event = mkl_lapack::geqrf(

dpnp/backend/extensions/lapack/geqrf_batch.cpp

@@ -94,6 +94,10 @@ static sycl::event geqrf_batch_impl(sycl::queue &exec_q,
 
     sycl::event geqrf_batch_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         geqrf_batch_event = mkl_lapack::geqrf_batch(

dpnp/backend/extensions/lapack/gesv.cpp

@@ -112,6 +112,10 @@ static sycl::event gesv_impl(sycl::queue &exec_q,
 #if defined(USE_ONEMATH)
     sycl::event getrf_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         getrf_event = mkl_lapack::getrf(
             exec_q,
             n,          // The order of the square matrix A (0 ≤ n).
@@ -169,6 +173,10 @@ static sycl::event gesv_impl(sycl::queue &exec_q,
     }
 #else
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         comp_event = mkl_lapack::gesv(
             exec_q,
             n,    // The order of the square matrix A

dpnp/backend/extensions/lapack/gesvd.cpp

@@ -99,6 +99,10 @@ static sycl::event gesvd_impl(sycl::queue &exec_q,
 
     sycl::event gesvd_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         gesvd_event = mkl_lapack::gesvd(
             exec_q,
             jobu,  // Character specifying how to compute the matrix U:

dpnp/backend/extensions/lapack/getrf.cpp

@@ -87,6 +87,10 @@ static sycl::event getrf_impl(sycl::queue &exec_q,
 
     sycl::event getrf_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         getrf_event = mkl_lapack::getrf(

dpnp/backend/extensions/lapack/getrf_batch.cpp

@@ -96,6 +96,10 @@ static sycl::event getrf_batch_impl(sycl::queue &exec_q,
 
     sycl::event getrf_batch_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         getrf_batch_event = mkl_lapack::getrf_batch(

dpnp/backend/extensions/lapack/heevd.cpp

@@ -42,6 +42,7 @@
 
 namespace dpnp::extensions::lapack
 {
+namespace py = pybind11;
 namespace mkl_lapack = oneapi::mkl::lapack;
 namespace type_utils = dpnp::tensor::type_utils;
 
@@ -73,6 +74,10 @@ static sycl::event heevd_impl(sycl::queue &exec_q,
 
     sycl::event heevd_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         heevd_event = mkl_lapack::heevd(
             exec_q,
             jobz, // 'jobz == job::vec' means eigenvalues and eigenvectors are

dpnp/backend/extensions/lapack/orgqr_batch.cpp

@@ -96,6 +96,10 @@ static sycl::event orgqr_batch_impl(sycl::queue &exec_q,
 
     sycl::event orgqr_batch_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         orgqr_batch_event = mkl_lapack::orgqr_batch(

dpnp/backend/extensions/lapack/potrf.cpp

@@ -83,6 +83,10 @@ static sycl::event potrf_impl(sycl::queue &exec_q,
 
     sycl::event potrf_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         potrf_event = mkl_lapack::potrf(

dpnp/backend/extensions/lapack/potrf_batch.cpp

@@ -90,6 +90,10 @@ static sycl::event potrf_batch_impl(sycl::queue &exec_q,
 
     sycl::event potrf_batch_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         potrf_batch_event = mkl_lapack::potrf_batch(

dpnp/backend/extensions/lapack/syevd.cpp

@@ -42,6 +42,7 @@
 
 namespace dpnp::extensions::lapack
 {
+namespace py = pybind11;
 namespace mkl_lapack = oneapi::mkl::lapack;
 namespace type_utils = dpnp::tensor::type_utils;
 
@@ -73,6 +74,10 @@ static sycl::event syevd_impl(sycl::queue &exec_q,
 
     sycl::event syevd_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         syevd_event = mkl_lapack::syevd(
             exec_q,
             jobz, // 'jobz == job::vec' means eigenvalues and eigenvectors are

dpnp/backend/extensions/lapack/ungqr.cpp

@@ -87,6 +87,10 @@ static sycl::event ungqr_impl(sycl::queue &exec_q,
 
     sycl::event ungqr_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         ungqr_event = mkl_lapack::ungqr(

dpnp/backend/extensions/lapack/ungqr_batch.cpp

@@ -96,6 +96,10 @@ static sycl::event ungqr_batch_impl(sycl::queue &exec_q,
 
     sycl::event ungqr_batch_event;
     try {
+        // Release GIL to avoid serialization of host task submissions
+        // to the same queue in OneMKL
+        py::gil_scoped_release lock{};
+
         scratchpad = sycl::malloc_device<T>(scratchpad_size, exec_q);
 
         ungqr_batch_event = mkl_lapack::ungqr_batch(