only events from host task has to be passed in keep_args_alive()

Author: antonwolfy

.github/workflows/conda-package.yml

@@ -194,8 +194,10 @@ jobs:
       # TODO: run the whole scope once the issues on CPU are resolved
       - name: Run tests
         run: |
-          python -m pytest -q -ra --disable-warnings -vv ${{ env.TEST_SCOPE }}
+          python -m pytest -q -ra --disable-warnings -vv -s ${{ env.TEST_SCOPE }}
         working-directory: ${{ env.tests-path }}
+        env:
+          SYCL_QUEUE_THREAD_POOL_SIZE: 16
 
   test_windows:
     name: Test ['windows-latest', python='${{ matrix.python }}']
@@ -331,8 +333,10 @@ jobs:
       # TODO: run the whole scope once the issues on CPU are resolved
       - name: Run tests
         run: |
-          python -m pytest -q -ra --disable-warnings -vv ${{ env.TEST_SCOPE }}
+          python -m pytest -q -ra --disable-warnings -vv -s ${{ env.TEST_SCOPE }}
         working-directory: ${{ env.tests-path }}
+        env:
+          SYCL_QUEUE_THREAD_POOL_SIZE: 16
 
   upload:
     name: Upload ['${{ matrix.os }}', python='${{ matrix.python }}']

.github/workflows/generate_coverage.yaml

@@ -66,6 +66,7 @@ jobs:
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           COVERALLS_PARALLEL: true
+          SYCL_QUEUE_THREAD_POOL_SIZE: 6
 
   coveralls:
     name: Indicate completion to coveralls.io

dpnp/backend/extensions/vm/CMakeLists.txt

@@ -55,7 +55,9 @@ else()
   target_compile_options(${python_module_name} PRIVATE
     -fno-approx-func
     -fno-finite-math-only
+    -no-ipo
     )
+    target_link_options(${python_module_name} PRIVATE -no-ipo)
 endif()
 
 target_link_options(${python_module_name} PUBLIC -fsycl-device-code-split=per_kernel)
@@ -70,6 +72,17 @@ endif()
 
 target_link_libraries(${python_module_name} PUBLIC MKL::MKL_DPCPP)
 
+target_link_libraries(${python_module_name} PUBLIC oneDPL)
+
+if (UNIX)
+  # needed for STL headers with GCC < 11
+  target_compile_definitions(${python_module_name} PUBLIC _GLIBCXX_USE_TBB_PAR_BACKEND=0)
+endif()
+
+target_compile_definitions(${python_module_name} PUBLIC PSTL_USE_PARALLEL_POLICIES=0)
+# work-around for Windows at exit crash with predefined policies
+target_compile_definitions(${python_module_name} PUBLIC ONEDPL_USE_PREDEFINED_POLICIES=0)
+
 install(TARGETS ${python_module_name}
   DESTINATION "dpnp/backend/extensions/vm"
 )

dpnp/backend/extensions/vm/div.cpp

@@ -64,16 +64,46 @@ static sycl::event div_impl(sycl::queue exec_q,
 {
     type_utils::validate_type_for_device<T>(exec_q);
 
-    const T* a = reinterpret_cast<const T*>(in_a);
-    const T* b = reinterpret_cast<const T*>(in_b);
-    T* y = reinterpret_cast<T*>(out_y);
+    std::cerr << "enter div_impl" << std::endl;
 
-    return mkl_vm::div(exec_q,
+    const T* _a = reinterpret_cast<const T*>(in_a);
+    const T* _b = reinterpret_cast<const T*>(in_b);
+    T* _y = reinterpret_cast<T*>(out_y);
+
+    std::cerr << "casting is done" << std::endl;
+
+    T* a = sycl::malloc_device<T>(n, exec_q);
+    T* b = sycl::malloc_device<T>(n, exec_q);
+    T* y = sycl::malloc_device<T>(n, exec_q);
+
+    std::cerr << "malloc is done" << std::endl;
+
+    exec_q.copy(_a, a, n).wait();
+    exec_q.copy(_b, b, n).wait();
+    exec_q.copy(_y, y, n).wait();
+
+    std::cerr << "copy is done" << std::endl;
+
+    sycl::event ev = mkl_vm::div(exec_q,
                        n, // number of elements to be calculated
                        a, // pointer `a` containing 1st input vector of size n
                        b, // pointer `b` containing 2nd input vector of size n
                        y, // pointer `y` to the output vector of size n
                        depends);
+    ev.wait();
+
+    std::cerr << "div is done" << std::endl;
+
+    exec_q.copy(y, _y, n).wait();
+
+    std::cerr << "copy is done" << std::endl;
+
+    sycl::free(a, exec_q);
+    sycl::free(b, exec_q);
+    sycl::free(y, exec_q);
+
+    std::cerr << "leaving div_impl" << std::endl;
+    return sycl::event();
 }
 
 std::pair<sycl::event, sycl::event> div(sycl::queue exec_q,
@@ -175,9 +205,21 @@ std::pair<sycl::event, sycl::event> div(sycl::queue exec_q,
         throw py::value_error("No div implementation defined");
     }
     sycl::event sum_ev = div_fn(exec_q, src_nelems, src1_data, src2_data, dst_data, depends);
-
-    sycl::event ht_ev = dpctl::utils::keep_args_alive(exec_q, {src1, src2, dst}, {sum_ev});
-    return std::make_pair(ht_ev, sum_ev);
+    // sum_ev.wait();
+
+    // int* dummy = sycl::malloc_device<int>(1, exec_q);
+    // sycl::event cleanup_ev = exec_q.submit([&](sycl::handler& cgh) {
+    //     // cgh.depends_on(sum_ev);
+    //     auto ctx = exec_q.get_context();
+    //     cgh.host_task([dummy, ctx]() {
+    //         // dummy host task to pass into keep_args_alive
+    //         sycl::free(dummy, ctx);
+    //     });
+    // });
+
+    // sycl::event ht_ev = dpctl::utils::keep_args_alive(exec_q, {src1, src2, dst}, {sum_ev});
+    // return std::make_pair(ht_ev, sum_ev);
+    return std::make_pair(sycl::event(), sycl::event());
 }
 
 bool can_call_div(sycl::queue exec_q,