commit

Author: maksimsab

.github/workflows/sycl-linux-precommit.yml

@@ -189,6 +189,11 @@ jobs:
             target_devices: level_zero:gpu;opencl:gpu;opencl:cpu
             extra_lit_opts: --param test-preview-mode=True
             binaries_artifact: e2e_bin_preview
+          - name: New Offload Model
+            runner: '[Linux, "gen12"]'
+            target_devices: opencl:cpu
+            extra_lit_opts: --param enable_new_offload_model=True
+            testing_mode: 'build-only'
 
           # We're in an ABI-breaking window, so these don't make sense for now.
           - name: ABI compatibility / sycl-rel-6_2
@@ -219,7 +224,7 @@ jobs:
       toolchain_artifact_filename: ${{ needs.build.outputs.toolchain_artifact_filename }}
       toolchain_decompress_command: ${{ needs.build.outputs.toolchain_decompress_command }}
       binaries_artifact: ${{ matrix.binaries_artifact || 'e2e_bin' }}
-      testing_mode: 'run-only'
+      testing_mode: ${{ matrix.testing_mode || 'run-only' }}
 
       # Do not install drivers on AMD and CUDA runners.
       install_igc_driver: >-

.github/workflows/sycl-nightly.yml

@@ -197,6 +197,28 @@ jobs:
       toolchain_artifact_filename: ${{ needs.ubuntu2404_oneapi_build.outputs.toolchain_artifact_filename }}
       toolchain_decompress_command: ${{ needs.ubuntu2404_oneapi_build.outputs.toolchain_decompress_command }}
 
+  e2e-with-new-offload-model:
+    needs: [ubuntu2204_build]
+    if: ${{ !cancelled() && needs.ubuntu2204_build.outputs.build_conclusion == 'success' }}
+    strategy:
+      fail-fast: false
+    uses: ./.github/workflows/sycl-linux-run-tests.yml
+    permissions:
+      contents: write
+      packages: read
+    with:
+      name: "E2E with enabled NewOffloadModel"
+      runner: ${{ inputs.runner }}
+      image_options: ${{ inputs.image_options }}
+      tests_selector: e2e
+      testing_mode: ${{ inputs.testing_mode }}
+      target_devices: ${{ inputs.target_devices }}
+      extra_lit_opts: "--param enable_new_offload_model=True"
+      tests_ref: ${{ github.sha }}
+      toolchain_artifact: ${{ needs.ubuntu2204_build.outputs.toolchain_artifact }}
+      toolchain_artifact_filename: ${{ needs.ubuntu2204_build.outputs.toolchain_artifact_filename }}
+      toolchain_decompress_command: ${{ needs.ubuntu2204_build.outputs.toolchain_decompress_command }}
+
   build-win:
     needs: get_date
     uses: ./.github/workflows/sycl-windows-build.yml
@@ -239,6 +261,22 @@ jobs:
       target_devices: ${{ matrix.target_devices }}
       toolchain_artifact_filename: ${{ needs.build-win.outputs.toolchain_artifact_filename }}
 
+  e2e-win-with-new-offload-model:
+    needs: [build-win]
+    if: ${{ !cancelled() && needs.build-win.outputs.build_conclusion == 'success' }}
+    strategy:
+      fail-fast: false
+    uses: ./.github/workflows/sycl-windows-run-tests.yml
+    permissions: read-all
+    with:
+      name: "E2E Windows with enabled NewOffloadModel"
+      runner: '["Windows", arc"]'
+      tests_selector: e2e
+      testing_mode: ${{ inputs.testing_mode }}
+      target_devices: level_zero:gpu
+      extra_lit_opts: "--param enable_new_offload_model=True"
+      toolchain_artifact_filename: ${{ needs.build-win.outputs.toolchain_artifact_filename }}
+
   cuda-aws-start:
     needs: [ubuntu2204_build]
     if: ${{ !cancelled() && needs.ubuntu2204_build.outputs.build_conclusion == 'success' }}

sycl/test-e2e/AOT/double.cpp

@@ -13,6 +13,9 @@
 // for pvc on windows.)
 // RUN: %if !windows %{ %clangxx -fsycl -fsycl-targets=intel_gpu_pvc -o %t.pvc.out %s %}
 
+// UNSUPPORTED: new-offload-model
+// UNSUPPORTED-TRACKER: CMPLRLLVM-71777
+
 // UNSUPPORTED: target-native_cpu
 // UNSUPPORTED-TRACKER: https://github.com/intel/llvm/issues/20142
 

sycl/test-e2e/AOT/reqd-sg-size.cpp

@@ -5,6 +5,9 @@
 // RUN: %clangxx -fsycl -fsycl-targets=intel_gpu_tgllp -o %t.tgllp.out %s
 // RUN: %clangxx -fsycl -fsycl-targets=spir64_x86_64 -o %t.x86.out %s
 
+// UNSUPPORTED: new-offload-model
+// UNSUPPORTED-TRACKER: CMPLRLLVM-71777
+
 // ocloc on windows does not have support for PVC, so this command will
 // result in an error when on windows. (In general, there is no support
 // for pvc on windows.)

sycl/test-e2e/Compression/compression_separate_compile.cpp

@@ -9,6 +9,9 @@
 // XFAIL: target-native_cpu
 // XFAIL-TRACKER: https://github.com/intel/llvm/issues/20397
 
+// UNSUPPORTED: new-offload-model
+// UNSUPPORTED-TRACKER: CMPLRLLVM-71777
+
 // CPU AOT targets host isa, so we compile everything on the run system instead.
 //////////////////////  Compile device images
 // RUN: %{run-aux} %clangxx -fsycl -fsycl-targets=spir64_x86_64 -fsycl-host-compiler=clang++ -fsycl-host-compiler-options='-std=c++17 -Wno-attributes -Wno-deprecated-declarations -fPIC -DENABLE_KERNEL1' -DENABLE_KERNEL1 -c %s -o %t_kernel1_aot.o

sycl/test-e2e/E2EExpr.py

@@ -48,6 +48,7 @@ class E2EExpr(BooleanExpression):
         "pdtracker",
         "ze_debug",
         "device-config-file",
+        "new-offload-model",
     }
 
     def __init__(self, string, variables, build_only_mode, final_unknown_value):

sycl/test-e2e/README.md

@@ -323,6 +323,7 @@ configure specific single test execution in the command line:
   by CMake variable HIP_LIBS_DIR (autodetected).
 * **run_launcher** - part of `%{run*}` expansion/substitution to alter execution
   of the test by, e.g., running it through Valgrind.
+* **enable_new_offload_model** - enables New Offload Model driver mode for all tests.
 
 Example:
 

sycl/test-e2e/SeparateCompile/test.cpp

@@ -1,3 +1,6 @@
+// UNSUPPORTED: new-offload-model
+// UNSUPPORTED-TRACKER: CMPLRLLVM-71777
+
 // REQUIRES: target-spir
 //
 // FIXME Disabled fallback assert as it'll require either online linking or

sycl/test-e2e/lit.cfg.py

@@ -984,6 +984,11 @@ def get_sycl_ls_verbose(sycl_device, env):
             "sycl-jit was not found. Tests requiring sycl-jit will be skipped."
         )
 
+# Check for enabled NewOffloadModel
+if lit_config.params.get("enable_new_offload_model", "False") != "False":
+    config.available_features.add("new-offload-model")
+    config.cxx_flags += "--offload-new-driver "
+
 # That has to be executed last so that all device-independent features have been
 # discovered already.
 config.sycl_dev_features = {}