Merge branch 'main' into revert-access-by

Author: npmiller

.github/scripts/compute_benchmarks.py

@@ -42,7 +42,7 @@ permissions:
 jobs:
   e2e-build-hw:
     # Run only on upstream; forks will not have the HW
-    if: github.repository == 'oneapi-src/unified-runtime'
+    # if: github.repository == 'oneapi-src/unified-runtime'
     name: Build SYCL, UR, run Compute Benchmarks
     strategy:
       matrix:
@@ -88,6 +88,9 @@ jobs:
       with:
         path: ur-repo
 
+    - name: Install pip packages
+      run: pip install -r ${{github.workspace}}/ur-repo/third_party/requirements.txt
+
     # We need to fetch special ref for proper PR's merge commit. Note, this ref may be absent if the PR is already merged.
     - name: Fetch PR's merge commit
       if: ${{ inputs.pr_no != 0 }}
@@ -131,44 +134,13 @@ jobs:
     - name: Build SYCL
       run: cmake --build ${{github.workspace}}/sycl_build -j
 
-    - name: Set additional env. vars
-      run: |
-        echo "${{github.workspace}}/sycl_build/bin" >> $GITHUB_PATH
-        echo "LD_LIBRARY_PATH=${{github.workspace}}/sycl_build/lib:$LD_LIBRARY_PATH" >> $GITHUB_ENV
-
-    # Running (newly built) sycl-ls sets up some extra variables
-    - name: Setup SYCL variables
-      run: |
-        which clang++ sycl-ls
-        SYCL_PI_TRACE=-1 sycl-ls
-
-    - name: Checkout Compute Benchmarks
-      uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # v4.1.1
-      with:
-        repository: intel/compute-benchmarks
-        path: compute-benchmarks-repo
-        submodules: recursive
-
-    - name: Configure Compute Benchmarks
-      run: >
-        cmake
-        -B ${{github.workspace}}/compute-benchmarks-build/
-        -S ${{github.workspace}}/compute-benchmarks-repo/
-        -DCMAKE_BUILD_TYPE=Release
-        -DBUILD_SYCL=ON
-        -DSYCL_COMPILER_ROOT=${{github.workspace}}/sycl_build
-        -DALLOW_WARNINGS=ON
-
-    - name: Build Compute Benchmarks
-      run: cmake --build ${{github.workspace}}/compute-benchmarks-build/ -j
-
     - name: Set oneAPI Device Selector
       run: |
         echo "ONEAPI_DEVICE_SELECTOR=${{ matrix.adapter.str_name }}:${{ matrix.adapter.unit }}" >> $GITHUB_ENV
 
-    - name: Run SYCL API Overhead benchmark
+    - name: Run benchmarks
       id: benchmarks
-      run: ${{ github.workspace }}/ur-repo/.github/scripts/compute_benchmarks.py ${{ github.workspace }}/compute-benchmarks-build/bin/ ${{ inputs.bench_script_params }}
+      run: numactl -N 0 ${{ github.workspace }}/ur-repo/scripts/benchmarks/main.py ~/bench_workdir ${{github.workspace}}/sycl_build ${{ inputs.bench_script_params }}
 
     - name: Add comment to PR
       uses: actions/github-script@60a0d83039c74a4aee543508d2ffcb1c3799cdea # v7.0.1
@@ -188,7 +160,7 @@ jobs:
           const test_status = '${{ steps.benchmarks.outcome }}';
           const job_status = '${{ job.status }}';
           const params = '${{ inputs.bench_script_params }}';
-          const body = `Compute Benchmarks ${adapter} run (with params: ${params}):\n${url}\nJob status: ${job_status}. Test status: ${test_status}.\n ${markdown}`;
+          const body = `Compute Benchmarks ${adapter} run (${params}):\n${url}\nJob status: ${job_status}. Test status: ${test_status}.\n ${markdown}`;
 
           github.rest.issues.createComment({
             issue_number: pr_no,

.github/workflows/benchmarks_compute.yml

@@ -21,7 +21,7 @@ jobs:
       config: ""
       unit: "gpu"
       # Failing tests
-      xfail: "ESIMD/preemption.cpp;Matrix/SG32/element_wise_all_ops.cpp;Matrix/SG32/get_coord_int8_matB.cpp;Matrix/element_wise_all_ops.cpp;Matrix/element_wise_all_ops_1d.cpp;Matrix/element_wise_all_ops_1d_cont.cpp;Matrix/element_wise_all_ops_scalar.cpp;Matrix/element_wise_ops.cpp;Matrix/get_coord_int8_matB.cpp;Matrix/joint_matrix_apply_bf16.cpp;Matrix/joint_matrix_apply_two_matrices.cpp;Matrix/joint_matrix_bfloat16.cpp;Matrix/joint_matrix_bfloat16_array.cpp;Matrix/joint_matrix_rowmajorA_rowmajorB.cpp;ProgramManager/uneven_kernel_split.cpp"
+      xfail: "ESIMD/regression/minmax.cpp;ESIMD/preemption.cpp;Matrix/SG32/element_wise_all_ops.cpp;Matrix/SG32/get_coord_int8_matB.cpp;Matrix/element_wise_all_ops.cpp;Matrix/element_wise_all_ops_1d.cpp;Matrix/element_wise_all_ops_1d_cont.cpp;Matrix/element_wise_all_ops_scalar.cpp;Matrix/element_wise_ops.cpp;Matrix/get_coord_int8_matB.cpp;Matrix/joint_matrix_apply_bf16.cpp;Matrix/joint_matrix_apply_two_matrices.cpp;Matrix/joint_matrix_bfloat16.cpp;Matrix/joint_matrix_bfloat16_array.cpp;Matrix/joint_matrix_rowmajorA_rowmajorB.cpp;ProgramManager/uneven_kernel_split.cpp"
       # Flaky tests
       filter_out: "UserDefinedReductions/user_defined_reductions.cpp"
       # These runners by default spawn upwards of 260 workers.

.github/workflows/e2e_level_zero.yml

@@ -283,6 +283,7 @@ typedef enum ur_structure_type_t {
     UR_STRUCTURE_TYPE_EXP_WIN32_HANDLE = 0x2004,                             ///< ::ur_exp_win32_handle_t
     UR_STRUCTURE_TYPE_EXP_SAMPLER_ADDR_MODES = 0x2005,                       ///< ::ur_exp_sampler_addr_modes_t
     UR_STRUCTURE_TYPE_EXP_SAMPLER_CUBEMAP_PROPERTIES = 0x2006,               ///< ::ur_exp_sampler_cubemap_properties_t
+    UR_STRUCTURE_TYPE_EXP_IMAGE_COPY_REGION = 0x2007,                        ///< ::ur_exp_image_copy_region_t
     UR_STRUCTURE_TYPE_EXP_ENQUEUE_NATIVE_COMMAND_PROPERTIES = 0x3000,        ///< ::ur_exp_enqueue_native_command_properties_t
     /// @cond
     UR_STRUCTURE_TYPE_FORCE_UINT32 = 0x7fffffff
@@ -7501,6 +7502,18 @@ typedef struct ur_exp_interop_semaphore_desc_t {
 
 } ur_exp_interop_semaphore_desc_t;
 
+///////////////////////////////////////////////////////////////////////////////
+/// @brief Describes the (sub-)regions and the extent to be copied
+typedef struct ur_exp_image_copy_region_t {
+    ur_structure_type_t stype;   ///< [in] type of this structure, must be
+                                 ///< ::UR_STRUCTURE_TYPE_EXP_IMAGE_COPY_REGION
+    const void *pNext;           ///< [in][optional] pointer to extension-specific structure
+    ur_rect_offset_t srcOffset;  ///< [in] the offset into the source image
+    ur_rect_offset_t dstOffset;  ///< [in] the offset into the destination image
+    ur_rect_region_t copyExtent; ///< [in] the extent (region) of the image to copy
+
+} ur_exp_image_copy_region_t;
+
 ///////////////////////////////////////////////////////////////////////////////
 /// @brief USM allocate pitched memory
 ///
@@ -7740,7 +7753,7 @@ urBindlessImagesSampledImageCreateExp(
 );
 
 ///////////////////////////////////////////////////////////////////////////////
-/// @brief Copy image data Host to Device or Device to Host
+/// @brief Copy image data Host to Device, Device to Host, or Device to Device
 ///
 /// @remarks
 ///   _Analogues_
@@ -7757,34 +7770,34 @@ urBindlessImagesSampledImageCreateExp(
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
 ///         + `NULL == hQueue`
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
-///         + `NULL == pDst`
 ///         + `NULL == pSrc`
-///         + `NULL == pImageFormat`
-///         + `NULL == pImageDesc`
+///         + `NULL == pDst`
+///         + `NULL == pSrcImageDesc`
+///         + `NULL == pDstImageDesc`
+///         + `NULL == pSrcImageFormat`
+///         + `NULL == pDstImageFormat`
+///         + `NULL == pCopyRegion`
 ///     - ::UR_RESULT_ERROR_INVALID_ENUMERATION
 ///         + `::UR_EXP_IMAGE_COPY_FLAGS_MASK & imageCopyFlags`
 ///     - ::UR_RESULT_ERROR_INVALID_QUEUE
 ///     - ::UR_RESULT_ERROR_INVALID_VALUE
 ///     - ::UR_RESULT_ERROR_INVALID_IMAGE_FORMAT_DESCRIPTOR
-///         + `pImageDesc && UR_MEM_TYPE_IMAGE_CUBEMAP_EXP < pImageDesc->type`
+///         + `pSrcImageDesc && UR_MEM_TYPE_IMAGE_CUBEMAP_EXP < pSrcImageDesc->type`
+///         + `pDstImageDesc && UR_MEM_TYPE_IMAGE_CUBEMAP_EXP < pDstImageDesc->type`
 ///     - ::UR_RESULT_ERROR_INVALID_IMAGE_SIZE
 ///     - ::UR_RESULT_ERROR_INVALID_OPERATION
 UR_APIEXPORT ur_result_t UR_APICALL
 urBindlessImagesImageCopyExp(
     ur_queue_handle_t hQueue,                 ///< [in] handle of the queue object
-    void *pDst,                               ///< [in] location the data will be copied to
     const void *pSrc,                         ///< [in] location the data will be copied from
-    const ur_image_format_t *pImageFormat,    ///< [in] pointer to image format specification
-    const ur_image_desc_t *pImageDesc,        ///< [in] pointer to image description
+    void *pDst,                               ///< [in] location the data will be copied to
+    const ur_image_desc_t *pSrcImageDesc,     ///< [in] pointer to image description
+    const ur_image_desc_t *pDstImageDesc,     ///< [in] pointer to image description
+    const ur_image_format_t *pSrcImageFormat, ///< [in] pointer to image format specification
+    const ur_image_format_t *pDstImageFormat, ///< [in] pointer to image format specification
+    ur_exp_image_copy_region_t *pCopyRegion,  ///< [in] Pointer to structure describing the (sub-)regions of source and
+                                              ///< destination images
     ur_exp_image_copy_flags_t imageCopyFlags, ///< [in] flags describing copy direction e.g. H2D or D2H
-    ur_rect_offset_t srcOffset,               ///< [in] defines the (x,y,z) source offset in pixels in the 1D, 2D, or 3D
-                                              ///< image
-    ur_rect_offset_t dstOffset,               ///< [in] defines the (x,y,z) destination offset in pixels in the 1D, 2D,
-                                              ///< or 3D image
-    ur_rect_region_t copyExtent,              ///< [in] defines the (width, height, depth) in pixels of the 1D, 2D, or 3D
-                                              ///< region to copy
-    ur_rect_region_t hostExtent,              ///< [in] defines the (width, height, depth) in pixels of the 1D, 2D, or 3D
-                                              ///< region on the host
     uint32_t numEventsInWaitList,             ///< [in] size of the event wait list
     const ur_event_handle_t *phEventWaitList, ///< [in][optional][range(0, numEventsInWaitList)] pointer to a list of
                                               ///< events that must be complete before this command can be executed.
@@ -11138,15 +11151,14 @@ typedef struct ur_bindless_images_sampled_image_create_exp_params_t {
 ///     allowing the callback the ability to modify the parameter's value
 typedef struct ur_bindless_images_image_copy_exp_params_t {
     ur_queue_handle_t *phQueue;
-    void **ppDst;
     const void **ppSrc;
-    const ur_image_format_t **ppImageFormat;
-    const ur_image_desc_t **ppImageDesc;
+    void **ppDst;
+    const ur_image_desc_t **ppSrcImageDesc;
+    const ur_image_desc_t **ppDstImageDesc;
+    const ur_image_format_t **ppSrcImageFormat;
+    const ur_image_format_t **ppDstImageFormat;
+    ur_exp_image_copy_region_t **ppCopyRegion;
     ur_exp_image_copy_flags_t *pimageCopyFlags;
-    ur_rect_offset_t *psrcOffset;
-    ur_rect_offset_t *pdstOffset;
-    ur_rect_region_t *pcopyExtent;
-    ur_rect_region_t *phostExtent;
     uint32_t *pnumEventsInWaitList;
     const ur_event_handle_t **pphEventWaitList;
     ur_event_handle_t **pphEvent;