Replace sycl::free with sycl_free_noexcept (#2058) (#2060)

* Replace sycl::free with sycl_free_noexcept

* Update changelog

Author: antonwolfy

CHANGELOG.md

@@ -103,6 +103,7 @@ In addition, this release completes implementation of `dpnp.fft` module and adds
 * Updated `Array Manipulation Routines` page in documentation to add missing functions and to remove duplicate entries [#2033](https://github.com/IntelPython/dpnp/pull/2033)
 * `dpnp` uses pybind11 2.13.6 [#2041](https://github.com/IntelPython/dpnp/pull/2041)
 * Updated `dpnp.fft` backend to depend on `INTEL_MKL_VERSION` flag to ensures that the appropriate code segment is executed based on the version of OneMKL [#2035](https://github.com/IntelPython/dpnp/pull/2035)
+* Use `dpctl::tensor::alloc_utils::sycl_free_noexcept` instead of `sycl::free` in `host_task` tasks associated with life-time management of temporary USM allocations [#2058](https://github.com/IntelPython/dpnp/pull/2058)
 
 ### Fixed
 

dpnp/backend/extensions/elementwise_functions/elementwise_functions.hpp

@@ -41,6 +41,7 @@
 #include "utils/memory_overlap.hpp"
 #include "utils/offset_utils.hpp"
 #include "utils/output_validation.hpp"
+#include "utils/sycl_alloc_utils.hpp"
 #include "utils/type_dispatch.hpp"
 
 namespace py = pybind11;
@@ -232,8 +233,9 @@ std::pair<sycl::event, sycl::event>
     auto ctx = q.get_context();
     sycl::event tmp_cleanup_ev = q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(strided_fn_ev);
+        using dpctl::tensor::alloc_utils::sycl_free_noexcept;
         cgh.host_task(
-            [ctx, shape_strides]() { sycl::free(shape_strides, ctx); });
+            [ctx, shape_strides]() { sycl_free_noexcept(shape_strides, ctx); });
     });
     host_tasks.push_back(tmp_cleanup_ev);
 
@@ -558,8 +560,9 @@ std::pair<sycl::event, sycl::event> py_binary_ufunc(
 
     sycl::event tmp_cleanup_ev = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(strided_fn_ev);
+        using dpctl::tensor::alloc_utils::sycl_free_noexcept;
         cgh.host_task(
-            [ctx, shape_strides]() { sycl::free(shape_strides, ctx); });
+            [ctx, shape_strides]() { sycl_free_noexcept(shape_strides, ctx); });
     });
 
     host_tasks.push_back(tmp_cleanup_ev);
@@ -810,8 +813,9 @@ std::pair<sycl::event, sycl::event>
 
     sycl::event tmp_cleanup_ev = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(strided_fn_ev);
+        using dpctl::tensor::alloc_utils::sycl_free_noexcept;
         cgh.host_task(
-            [ctx, shape_strides]() { sycl::free(shape_strides, ctx); });
+            [ctx, shape_strides]() { sycl_free_noexcept(shape_strides, ctx); });
     });
 
     host_tasks.push_back(tmp_cleanup_ev);

dpnp/backend/extensions/lapack/common_helpers.hpp

@@ -31,6 +31,9 @@
 #include <cstring>
 #include <stdexcept>
 
+// dpctl tensor headers
+#include "utils/sycl_alloc_utils.hpp"
+
 namespace dpnp::extensions::lapack::helper
 {
 namespace py = pybind11;
@@ -78,7 +81,7 @@ inline std::int64_t *alloc_ipiv(const std::int64_t n, sycl::queue &exec_q)
         }
     } catch (sycl::exception const &e) {
         if (ipiv != nullptr)
-            sycl::free(ipiv, exec_q);
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(ipiv, exec_q);
         throw std::runtime_error(
             std::string(
                 "Unexpected SYCL exception caught during ipiv allocation: ") +
@@ -122,7 +125,7 @@ inline T *alloc_scratchpad(std::int64_t scratchpad_size, sycl::queue &exec_q)
         }
     } catch (sycl::exception const &e) {
         if (scratchpad != nullptr) {
-            sycl::free(scratchpad, exec_q);
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, exec_q);
         }
         throw std::runtime_error(std::string("Unexpected SYCL exception caught "
                                              "during scratchpad allocation: ") +

dpnp/backend/extensions/lapack/geqrf.cpp

@@ -27,6 +27,7 @@
 
 // dpctl tensor headers
 #include "utils/memory_overlap.hpp"
+#include "utils/sycl_alloc_utils.hpp"
 #include "utils/type_utils.hpp"
 
 #include "geqrf.hpp"
@@ -117,15 +118,17 @@ static sycl::event geqrf_impl(sycl::queue &exec_q,
     if (is_exception_caught) // an unexpected error occurs
     {
         if (scratchpad != nullptr) {
-            sycl::free(scratchpad, exec_q);
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, exec_q);
         }
         throw std::runtime_error(error_msg.str());
     }
 
     sycl::event clean_up_event = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(geqrf_event);
         auto ctx = exec_q.get_context();
-        cgh.host_task([ctx, scratchpad]() { sycl::free(scratchpad, ctx); });
+        cgh.host_task([ctx, scratchpad]() {
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, ctx);
+        });
     });
     host_task_events.push_back(clean_up_event);
 

dpnp/backend/extensions/lapack/geqrf_batch.cpp

@@ -27,6 +27,7 @@
 
 // dpctl tensor headers
 #include "utils/memory_overlap.hpp"
+#include "utils/sycl_alloc_utils.hpp"
 #include "utils/type_utils.hpp"
 
 #include "geqrf.hpp"
@@ -138,7 +139,7 @@ static sycl::event geqrf_batch_impl(sycl::queue &exec_q,
     if (is_exception_caught) // an unexpected error occurs
     {
         if (scratchpad != nullptr) {
-            sycl::free(scratchpad, exec_q);
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, exec_q);
         }
 
         throw std::runtime_error(error_msg.str());
@@ -147,7 +148,9 @@ static sycl::event geqrf_batch_impl(sycl::queue &exec_q,
     sycl::event clean_up_event = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(geqrf_batch_event);
         auto ctx = exec_q.get_context();
-        cgh.host_task([ctx, scratchpad]() { sycl::free(scratchpad, ctx); });
+        cgh.host_task([ctx, scratchpad]() {
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, ctx);
+        });
     });
     host_task_events.push_back(clean_up_event);
     return geqrf_batch_event;

dpnp/backend/extensions/lapack/gesv.cpp

@@ -39,6 +39,8 @@ namespace mkl_lapack = oneapi::mkl::lapack;
 namespace py = pybind11;
 namespace type_utils = dpctl::tensor::type_utils;
 
+using dpctl::tensor::alloc_utils::sycl_free_noexcept;
+
 typedef sycl::event (*gesv_impl_fn_ptr_t)(sycl::queue &,
                                           const std::int64_t,
                                           const std::int64_t,
@@ -93,7 +95,7 @@ static sycl::event gesv_impl(sycl::queue &exec_q,
         ipiv = helper::alloc_ipiv(n, exec_q);
     } catch (const std::exception &e) {
         if (scratchpad != nullptr)
-            sycl::free(scratchpad, exec_q);
+            sycl_free_noexcept(scratchpad, exec_q);
         throw;
     }
 
@@ -178,18 +180,18 @@ static sycl::event gesv_impl(sycl::queue &exec_q,
     if (is_exception_caught) // an unexpected error occurs
     {
         if (scratchpad != nullptr)
-            sycl::free(scratchpad, exec_q);
+            sycl_free_noexcept(scratchpad, exec_q);
         if (ipiv != nullptr)
-            sycl::free(ipiv, exec_q);
+            sycl_free_noexcept(ipiv, exec_q);
         throw std::runtime_error(error_msg.str());
     }
 
     sycl::event ht_ev = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(comp_event);
         auto ctx = exec_q.get_context();
         cgh.host_task([ctx, scratchpad, ipiv]() {
-            sycl::free(scratchpad, ctx);
-            sycl::free(ipiv, ctx);
+            sycl_free_noexcept(scratchpad, ctx);
+            sycl_free_noexcept(ipiv, ctx);
         });
     });
 

dpnp/backend/extensions/lapack/gesv_batch.cpp

@@ -39,6 +39,8 @@ namespace mkl_lapack = oneapi::mkl::lapack;
 namespace py = pybind11;
 namespace type_utils = dpctl::tensor::type_utils;
 
+using dpctl::tensor::alloc_utils::sycl_free_noexcept;
+
 typedef sycl::event (*gesv_batch_impl_fn_ptr_t)(
     sycl::queue &,
     const std::int64_t,
@@ -106,7 +108,7 @@ static sycl::event gesv_batch_impl(sycl::queue &exec_q,
         ipiv = helper::alloc_ipiv(batch_size * n, exec_q);
     } catch (const std::exception &e) {
         if (scratchpad != nullptr)
-            sycl::free(scratchpad, exec_q);
+            sycl_free_noexcept(scratchpad, exec_q);
         throw;
     }
 
@@ -172,9 +174,9 @@ static sycl::event gesv_batch_impl(sycl::queue &exec_q,
         error_msg << ".";
 
         if (scratchpad != nullptr)
-            sycl::free(scratchpad, exec_q);
+            sycl_free_noexcept(scratchpad, exec_q);
         if (ipiv != nullptr)
-            sycl::free(ipiv, exec_q);
+            sycl_free_noexcept(ipiv, exec_q);
 
         throw LinAlgError(error_msg.str().c_str());
     } catch (mkl_lapack::exception const &e) {
@@ -218,7 +220,7 @@ static sycl::event gesv_batch_impl(sycl::queue &exec_q,
         ipiv = helper::alloc_ipiv_batch<T>(n, n_linear_streams, exec_q);
     } catch (const std::exception &e) {
         if (scratchpad != nullptr)
-            sycl::free(scratchpad, exec_q);
+            sycl_free_noexcept(scratchpad, exec_q);
         throw;
     }
 
@@ -281,9 +283,9 @@ static sycl::event gesv_batch_impl(sycl::queue &exec_q,
     if (is_exception_caught) // an unexpected error occurs
     {
         if (scratchpad != nullptr)
-            sycl::free(scratchpad, exec_q);
+            sycl_free_noexcept(scratchpad, exec_q);
         if (ipiv != nullptr)
-            sycl::free(ipiv, exec_q);
+            sycl_free_noexcept(ipiv, exec_q);
         throw std::runtime_error(error_msg.str());
     }
 
@@ -297,8 +299,8 @@ static sycl::event gesv_batch_impl(sycl::queue &exec_q,
 #endif // USE_ONEMKL_INTERFACES
         auto ctx = exec_q.get_context();
         cgh.host_task([ctx, scratchpad, ipiv]() {
-            sycl::free(scratchpad, ctx);
-            sycl::free(ipiv, ctx);
+            sycl_free_noexcept(scratchpad, ctx);
+            sycl_free_noexcept(ipiv, ctx);
         });
     });
 

dpnp/backend/extensions/lapack/gesv_common_utils.hpp

@@ -30,6 +30,7 @@
 // dpctl tensor headers
 #include "utils/memory_overlap.hpp"
 #include "utils/output_validation.hpp"
+#include "utils/sycl_alloc_utils.hpp"
 #include "utils/type_dispatch.hpp"
 
 #include "common_helpers.hpp"
@@ -159,10 +160,12 @@ inline void handle_lapack_exc(sycl::queue &exec_q,
         const auto threshold =
             std::numeric_limits<ThresholdType>::epsilon() * 100;
         if (std::abs(host_U) < threshold) {
+            using dpctl::tensor::alloc_utils::sycl_free_noexcept;
+
             if (scratchpad != nullptr)
-                sycl::free(scratchpad, exec_q);
+                sycl_free_noexcept(scratchpad, exec_q);
             if (ipiv != nullptr)
-                sycl::free(ipiv, exec_q);
+                sycl_free_noexcept(ipiv, exec_q);
             throw LinAlgError("The input coefficient matrix is singular.");
         }
         else {

dpnp/backend/extensions/lapack/gesvd.cpp

@@ -125,15 +125,17 @@ static sycl::event gesvd_impl(sycl::queue &exec_q,
     if (is_exception_caught) // an unexpected error occurs
     {
         if (scratchpad != nullptr) {
-            sycl::free(scratchpad, exec_q);
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, exec_q);
         }
         throw std::runtime_error(error_msg.str());
     }
 
     sycl::event ht_ev = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(gesvd_event);
         auto ctx = exec_q.get_context();
-        cgh.host_task([ctx, scratchpad]() { sycl::free(scratchpad, ctx); });
+        cgh.host_task([ctx, scratchpad]() {
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, ctx);
+        });
     });
 
     return ht_ev;

dpnp/backend/extensions/lapack/gesvd_batch.cpp

@@ -189,7 +189,7 @@ static sycl::event gesvd_batch_impl(sycl::queue &exec_q,
     if (is_exception_caught) // an unexpected error occurs
     {
         if (scratchpad != nullptr) {
-            sycl::free(scratchpad, exec_q);
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, exec_q);
         }
         throw std::runtime_error(error_msg.str());
     }
@@ -199,7 +199,9 @@ static sycl::event gesvd_batch_impl(sycl::queue &exec_q,
             cgh.depends_on(ev);
         }
         auto ctx = exec_q.get_context();
-        cgh.host_task([ctx, scratchpad]() { sycl::free(scratchpad, ctx); });
+        cgh.host_task([ctx, scratchpad]() {
+            dpctl::tensor::alloc_utils::sycl_free_noexcept(scratchpad, ctx);
+        });
     });
 
     return ht_ev;