Merge branch 'master' into feature/dlpack-kdlcpu-support

Author: oleksandr-pavlyk

.github/workflows/CODEOWNERS

@@ -0,0 +1 @@
+*       @oleksandr-pavlyk @ndgrigorian

.github/workflows/conda-package.yml

@@ -62,12 +62,12 @@ jobs:
             $CHANNELS \
             conda-recipe
       - name: Upload artifact
-        uses: actions/[email protected].4
+        uses: actions/[email protected].5
         with:
           name: ${{ env.PACKAGE_NAME }} ${{ runner.os }} Python ${{ matrix.python }}
           path: /usr/share/miniconda/conda-bld/linux-64/${{ env.PACKAGE_NAME }}-*.tar.bz2
       - name: Upload wheels artifact
-        uses: actions/[email protected].4
+        uses: actions/[email protected].5
         with:
           name: ${{ env.PACKAGE_NAME }} ${{ runner.os }} Wheels Python ${{ matrix.python }}
           path: ${{ env.WHEELS_OUTPUT_FOLDER }}${{ env.PACKAGE_NAME }}-*.whl
@@ -111,12 +111,12 @@ jobs:
           OVERRIDE_INTEL_IPO: 1   # IPO requires more resources that GH actions VM provides
         run: conda build --no-test --python ${{ matrix.python }} -c dppy/label/bootstrap -c ${{ env.INTEL_CHANNEL }} -c conda-forge --override-channels conda-recipe
       - name: Upload artifact
-        uses: actions/[email protected].4
+        uses: actions/[email protected].5
         with:
           name: ${{ env.PACKAGE_NAME }} ${{ runner.os }} Python ${{ matrix.python }}
           path: ${{ env.conda-bld }}${{ env.PACKAGE_NAME }}-*.tar.bz2
       - name: Upload wheels artifact
-        uses: actions/[email protected].4
+        uses: actions/[email protected].5
         with:
           name: ${{ env.PACKAGE_NAME }} ${{ runner.os }} Wheels Python ${{ matrix.python }}
           path: ${{ env.WHEELS_OUTPUT_FOLDER }}${{ env.PACKAGE_NAME }}-*.whl

.github/workflows/generate-docs.yml

@@ -104,7 +104,7 @@ jobs:
           git push tokened_docs gh-pages
       - name: Save built docs as an artifact
         if: ${{ github.event.pull_request && github.event.pull_request.head.repo.fork && github.event.action != 'closed'}}
-        uses: actions/[email protected].4
+        uses: actions/[email protected].5
         with:
           name: ${{ env.PACKAGE_NAME }} rendered documentation
           path: ~/docs

.github/workflows/openssf-scorecard.yml

@@ -60,7 +60,7 @@ jobs:
       # Upload the results as artifacts (optional). Commenting out will disable uploads of run results in SARIF
       # format to the repository Actions tab.
       - name: "Upload artifact"
-        uses: actions/upload-artifact@0b2256b8c012f0828dc542b3febcab082c67f72b # v4.3.4
+        uses: actions/upload-artifact@89ef406dd8d7e03cfd12d9e0a4a378f454709029 # v4.3.5
         with:
           name: SARIF file
           path: results.sarif

conda-recipe/build.sh

@@ -27,26 +27,6 @@ CMAKE_ARGS="${CMAKE_ARGS} -DDPCTL_LEVEL_ZERO_INCLUDE_DIR=${PREFIX}/include/level
 # -wnx flags mean: --wheel --no-isolation --skip-dependency-check
 ${PYTHON} -m build -w -n -x
 
-pushd dist
-${PYTHON} -m wheel unpack -d dpctl_wheel dpctl*.whl
-export lib_name=libDPCTLSyclInterface
-export so_full_path=$(find dpctl_wheel -regextype posix-extended -regex "^.*${lib_name}\.so")
-export sox_full_path=$(find dpctl_wheel -regextype posix-extended -regex "^.*${lib_name}\.so\.[0-9]*$")
-export soxxx_full_path=$(find dpctl_wheel -regextype posix-extended -regex "^.*${lib_name}\.so\.[0-9]*\.[0-9]*$")
-
-rm -rf ${so_full_path} ${soxxx_full_path}
-
-export so_name=$(basename ${so_full_path})
-export sox_name=$(basename ${sox_full_path})
-export soxxx_name=$(basename ${soxxx_full_path})
-export wheel_path=$(dirname $(dirname ${so_full_path}))
-
-# deal with hard copies
-${PYTHON} -m wheel pack ${wheel_path}
-
-rm -rf dpctl_wheel
-popd
-
 ${PYTHON} -m wheel tags --remove --build "$GIT_DESCRIBE_NUMBER" \
     --platform-tag "manylinux_${GLIBC_MAJOR}_${GLIBC_MINOR}_x86_64" \
     dist/dpctl*.whl
@@ -59,15 +39,6 @@ ${PYTHON} -m pip install dist/dpctl*.whl \
     --prefix "${PREFIX}" \
     -vv
 
-export libdir=$(find $PREFIX -name 'libDPCTLSyclInterface*' -exec dirname \{\} \;)
-
-# Recover symbolic links
-# libDPCTLSyclInterface.so.0 -> libDPCTLSyclInterface.so.0.17
-# libDPCTLSyclInterface.so -> libDPCTLSyclInterface.so.0
-mv ${libdir}/${sox_name} ${libdir}/${soxxx_name}
-ln -s ${libdir}/${soxxx_name} ${libdir}/${sox_name}
-ln -s ${libdir}/${sox_name} ${libdir}/${so_name}
-
 # Copy wheel package
 if [[ -d "${WHEELS_OUTPUT_FOLDER}" ]]; then
     cp dist/dpctl*.whl "${WHEELS_OUTPUT_FOLDER[@]}"

docs/doc_sources/api_reference/dpctl/tensor.indexing_functions.rst

@@ -15,3 +15,4 @@ by either integral arrays of indices or boolean mask arrays.
     place
     put
     take
+    take_along_axis

dpctl/apis/include/dpctl4pybind11.hpp

@@ -27,6 +27,7 @@
 
 #include "dpctl_capi.h"
 #include <complex>
+#include <exception>
 #include <memory>
 #include <pybind11/pybind11.h>
 #include <sycl/sycl.hpp>
@@ -748,6 +749,53 @@ class usm_memory : public py::object
             throw py::error_already_set();
     }
 
+    /*! @brief Create usm_memory object from shared pointer that manages
+     *  lifetime of the USM allocation.
+     */
+    usm_memory(void *usm_ptr,
+               size_t nbytes,
+               const sycl::queue &q,
+               std::shared_ptr<void> shptr)
+    {
+        auto const &api = ::dpctl::detail::dpctl_capi::get();
+        DPCTLSyclUSMRef usm_ref = reinterpret_cast<DPCTLSyclUSMRef>(usm_ptr);
+        auto q_uptr = std::make_unique<sycl::queue>(q);
+        DPCTLSyclQueueRef QRef =
+            reinterpret_cast<DPCTLSyclQueueRef>(q_uptr.get());
+
+        auto vacuous_destructor = []() {};
+        py::capsule mock_owner(vacuous_destructor);
+
+        // create memory object owned by mock_owner, it is a new reference
+        PyObject *_memory =
+            api.Memory_Make_(usm_ref, nbytes, QRef, mock_owner.ptr());
+        auto ref_count_decrementer = [](PyObject *o) noexcept { Py_DECREF(o); };
+
+        using py_uptrT =
+            std::unique_ptr<PyObject, decltype(ref_count_decrementer)>;
+
+        if (!_memory) {
+            throw py::error_already_set();
+        }
+
+        auto memory_uptr = py_uptrT(_memory, ref_count_decrementer);
+        std::shared_ptr<void> *opaque_ptr = new std::shared_ptr<void>(shptr);
+
+        Py_MemoryObject *memobj = reinterpret_cast<Py_MemoryObject *>(_memory);
+        // replace mock_owner capsule as the owner
+        memobj->refobj = Py_None;
+        // set opaque ptr field, usm_memory now knowns that USM is managed
+        // by smart pointer
+        memobj->_opaque_ptr = reinterpret_cast<void *>(opaque_ptr);
+
+        // _memory will delete created copies of sycl::queue, and
+        // std::shared_ptr and the deleter of the shared_ptr<void> is
+        // supposed to free the USM allocation
+        m_ptr = _memory;
+        q_uptr.release();
+        memory_uptr.release();
+    }
+
     sycl::queue get_queue() const
     {
         Py_MemoryObject *mem_obj = reinterpret_cast<Py_MemoryObject *>(m_ptr);

dpctl/memory/_opaque_smart_ptr.hpp

@@ -34,6 +34,7 @@
 #include "syclinterface/dpctl_sycl_types.h"
 #include <memory>
 #include <sycl/sycl.hpp>
+#include <utility>
 
 #include <exception>
 #include <iostream>

dpctl/tensor/__init__.py

@@ -60,7 +60,14 @@
 )
 from dpctl.tensor._device import Device
 from dpctl.tensor._dlpack import from_dlpack
-from dpctl.tensor._indexing_functions import extract, nonzero, place, put, take
+from dpctl.tensor._indexing_functions import (
+    extract,
+    nonzero,
+    place,
+    put,
+    take,
+    take_along_axis,
+)
 from dpctl.tensor._linear_algebra_functions import (
     matmul,
     matrix_transpose,
@@ -377,4 +384,5 @@
     "diff",
     "count_nonzero",
     "DLDeviceType",
+    "take_along_axis",
 ]

dpctl/tensor/_copy_utils.py

@@ -795,13 +795,18 @@ def _nonzero_impl(ary):
     return res
 
 
-def _take_multi_index(ary, inds, p):
+def _take_multi_index(ary, inds, p, mode=0):
     if not isinstance(ary, dpt.usm_ndarray):
         raise TypeError(
             f"Expecting type dpctl.tensor.usm_ndarray, got {type(ary)}"
         )
     ary_nd = ary.ndim
     p = normalize_axis_index(operator.index(p), ary_nd)
+    mode = operator.index(mode)
+    if mode not in [0, 1]:
+        raise ValueError(
+            "Invalid value for mode keyword, only 0 or 1 is supported"
+        )
     queues_ = [
         ary.sycl_queue,
     ]
@@ -860,7 +865,7 @@ def _take_multi_index(ary, inds, p):
         ind=inds,
         dst=res,
         axis_start=p,
-        mode=0,
+        mode=mode,
         sycl_queue=exec_q,
         depends=dep_ev,
     )