Write usm_host_allocator that wraps call to free in try/catch

Wrote dpctl::tensor::offset_utils::usm_host_allocator<T> to allocate
USM-host memory as storage to std::vector.

Replaced uses of sycl::usm_memory<T, sycl::alloc::kind::host>. The
new class derives from this, but overrides deallocate method to
wrap call to base::deallocate in try/except. The exception, if
caught, is printed but otherwise ignored, consistent like this is
done on USMDeleter class used in dpctl.memory

This is to work around sporadic crashes due to unhandled exception
thrown by openCL::CPU driver, which appears to be benign.

The issue was reported to CPU driver team, with native reproducer
(compiler LLVM jira ticket 58387).

dpctl/tensor/libtensor/include/utils/offset_utils.hpp

@@ -27,6 +27,7 @@
 #pragma once
 
 #include <algorithm>
+#include <exception>
 #include <sycl/sycl.hpp>
 #include <tuple>
 #include <vector>
@@ -81,6 +82,30 @@ std::vector<T, A> concat(std::vector<T, A> lhs, Vs &&...vs)
 
 } // namespace detail
 
+template <typename T>
+class usm_host_allocator : public sycl::usm_allocator<T, sycl::usm::alloc::host>
+{
+public:
+    using baseT = sycl::usm_allocator<T, sycl::usm::alloc::host>;
+    using baseT::baseT;
+
+    template <typename U> struct rebind
+    {
+        typedef usm_host_allocator<U> other;
+    };
+
+    void deallocate(T *ptr, size_t n)
+    {
+        try {
+            baseT::deallocate(ptr, n);
+        } catch (const std::exception &e) {
+            std::cerr
+                << "Exception caught in `usm_host_allocator::deallocate`: "
+                << e.what() << std::endl;
+        }
+    }
+};
+
 template <typename indT, typename... Vs>
 std::tuple<indT *, size_t, sycl::event>
 device_allocate_and_pack(sycl::queue &q,
@@ -90,8 +115,7 @@ device_allocate_and_pack(sycl::queue &q,
 
     // memory transfer optimization, use USM-host for temporary speeds up
     // transfer to device, especially on dGPUs
-    using usm_host_allocatorT =
-        sycl::usm_allocator<indT, sycl::usm::alloc::host>;
+    using usm_host_allocatorT = usm_host_allocator<indT>;
     using shT = std::vector<indT, usm_host_allocatorT>;
 
     usm_host_allocatorT usm_host_allocator(q);

dpctl/tensor/libtensor/source/integer_advanced_indexing.cpp

@@ -35,6 +35,7 @@
 #include "dpctl4pybind11.hpp"
 #include "kernels/integer_advanced_indexing.hpp"
 #include "utils/memory_overlap.hpp"
+#include "utils/offset_utils.hpp"
 #include "utils/output_validation.hpp"
 #include "utils/type_dispatch.hpp"
 #include "utils/type_utils.hpp"
@@ -91,15 +92,15 @@ _populate_kernel_params(sycl::queue &exec_q,
 {
 
     using usm_host_allocator_T =
-        sycl::usm_allocator<char *, sycl::usm::alloc::host>;
+        dpctl::tensor::offset_utils::usm_host_allocator<char *>;
     using ptrT = std::vector<char *, usm_host_allocator_T>;
 
     usm_host_allocator_T ptr_allocator(exec_q);
     std::shared_ptr<ptrT> host_ind_ptrs_shp =
         std::make_shared<ptrT>(k, ptr_allocator);
 
     using usm_host_allocatorT =
-        sycl::usm_allocator<py::ssize_t, sycl::usm::alloc::host>;
+        dpctl::tensor::offset_utils::usm_host_allocator<py::ssize_t>;
     using shT = std::vector<py::ssize_t, usm_host_allocatorT>;
 
     usm_host_allocatorT sz_allocator(exec_q);

dpctl/tensor/libtensor/source/triul_ctor.cpp

@@ -32,6 +32,7 @@
 #include "kernels/constructors.hpp"
 #include "simplify_iteration_space.hpp"
 #include "utils/memory_overlap.hpp"
+#include "utils/offset_utils.hpp"
 #include "utils/output_validation.hpp"
 #include "utils/type_dispatch.hpp"
 
@@ -150,7 +151,7 @@ usm_ndarray_triul(sycl::queue &exec_q,
     nd += 2;
 
     using usm_host_allocatorT =
-        sycl::usm_allocator<py::ssize_t, sycl::usm::alloc::host>;
+        dpctl::tensor::offset_utils::usm_host_allocator<py::ssize_t>;
     using usmshT = std::vector<py::ssize_t, usm_host_allocatorT>;
 
     usm_host_allocatorT allocator(exec_q);