Merge remote-tracking branch 'origin/sycl' into cts-spirv-be-2

Author: KornevNikita

.github/workflows/sycl-linux-run-tests.yml

@@ -210,27 +210,6 @@ jobs:
           devops
     - name: Register cleanup after job is finished
       uses: ./devops/actions/cleanup
-    - name: Checkout E2E tests
-      if: inputs.tests_selector == 'e2e'
-      uses: ./devops/actions/cached_checkout
-      with:
-        path: llvm
-        ref: ${{ inputs.ref || github.sha }}
-        merge_ref: ${{ inputs.merge_ref }}
-        cache_path: "/__w/repo_cache/"
-    - name: Checkout SYCL CTS tests
-      if: inputs.tests_selector == 'cts' && inputs.cts_testing_mode != 'run-only'
-      uses: ./devops/actions/cached_checkout
-      with:
-        path: khronos_sycl_cts
-        repository: 'KhronosGroup/SYCL-CTS'
-        ref: 'main'
-        default_branch: 'main'
-        cache_path: "/__w/repo_cache/"
-    - name: SYCL CTS GIT submodules init
-      if: inputs.tests_selector == 'cts' && inputs.cts_testing_mode != 'run-only'
-      run: |
-        git -C khronos_sycl_cts submodule update --init
     - name: Install drivers
       if: inputs.install_igc_driver == 'true' || inputs.install_dev_igc_driver == 'true'
       env:
@@ -312,178 +291,26 @@ jobs:
             cat /usr/local/lib/igc/IGCTAG.txt
           fi
 
-    - name: Download E2E Binaries
-      if: inputs.e2e_binaries_artifact != ''
-      uses: actions/download-artifact@v4
-      with:
-        name: ${{ inputs.e2e_binaries_artifact }}
-    - name: Extract E2E Binaries
-      if: inputs.e2e_binaries_artifact != ''
-      run: |
-        mkdir build-e2e
-        tar -I 'zstd' -xf e2e_binaries.tar.zst -C build-e2e
-
-    - name: Deduce E2E CMake options
-      if: inputs.tests_selector == 'e2e' && inputs.e2e_binaries_artifact == ''
-      id: cmake_opts
-      shell: bash
-      env:
-        CMAKE_EXTRA_ARGS: ${{ inputs.extra_cmake_args }}
-      run: |
-        if [ -n "$CMAKE_EXTRA_ARGS" ]; then
-          echo "opts=$CMAKE_EXTRA_ARGS" >> $GITHUB_OUTPUT
-        fi
-    - name: Configure E2E tests
-      if: inputs.tests_selector == 'e2e' && inputs.e2e_binaries_artifact == ''
-      run: |
-        cmake -GNinja -B./build-e2e -S./llvm/sycl/test-e2e -DCMAKE_CXX_COMPILER="$(which clang++)" -DLLVM_LIT="$PWD/llvm/llvm/utils/lit/lit.py" ${{ steps.cmake_opts.outputs.opts }}
-    - name: SYCL End-to-end tests
-      shell: bash {0}
+    - name: Run E2E Tests
       if: inputs.tests_selector == 'e2e'
-      env:
-        LIT_OPTS: -v --no-progress-bar --show-unsupported --show-pass --show-xfail --max-time 3600 --time-tests --param test-mode=${{ inputs.e2e_testing_mode }} --param sycl_devices=${{ inputs.target_devices }} ${{ inputs.extra_lit_opts }}
-      run: |
-        ninja -C build-e2e check-sycl-e2e > e2e.log 2>&1
-        exit_code=$?
-        cat e2e.log
-        if [ $exit_code -ne 0 ]; then
-          awk '/^Failed Tests|Unexpectedly Passed Tests|Unresolved tests|Testing Time/{flag=1}/FAILED: CMakeFiles/{flag=0}flag' e2e.log >> $GITHUB_STEP_SUMMARY
-        fi
-        exit $exit_code
-    - name: Build SYCL CTS tests
-      if: inputs.tests_selector == 'cts' && inputs.sycl_cts_artifact == ''
-      env:
-        CMAKE_EXTRA_ARGS: ${{ inputs.extra_cmake_args }}
-      run: |
-        cts_exclude_filter=""
-        # If CTS_TESTS_TO_BUILD is null - use filter
-        if [ -z "$CTS_TESTS_TO_BUILD" ]; then
-          if [ "${{ contains(inputs.cts_testing_mode, 'build-only')  }}" = "true" ]; then
-            cts_exclude_filter=$PWD/devops/cts_exclude_filter_compfails
-          elif [ "${{ contains(inputs.target_devices, 'opencl:cpu')  }}" = "true" ]; then
-            cts_exclude_filter=$PWD/devops/cts_exclude_filter_OCL_CPU
-          elif [ "${{ contains(inputs.target_devices, 'level_zero:gpu')  }}" = "true" ]; then
-            cts_exclude_filter=$PWD/devops/cts_exclude_filter_L0_GPU
-          fi
-
-          # List excluded SYCL CTS categories:
-          # SYCL_CTS_EXCLUDE_TEST_CATEGORIES - Optional file specifying a list
-          # of test categories to be excluded from the build.
-          echo "::group::Excluded test categories"
-          cat $cts_exclude_filter
-          echo "::endgroup::"
-        fi
-
-        cmake -GNinja -B./build-cts -S./khronos_sycl_cts -DCMAKE_CXX_COMPILER=$(which clang++) \
-        -DSYCL_IMPLEMENTATION=DPCPP \
-        -DSYCL_CTS_EXCLUDE_TEST_CATEGORIES="$cts_exclude_filter" \
-        -DSYCL_CTS_ENABLE_OPENCL_INTEROP_TESTS=OFF \
-        -DDPCPP_INSTALL_DIR="$(dirname $(which clang++))/.." \
-        $CMAKE_EXTRA_ARGS
-        # Ignore errors so that if one category build fails others still have a
-        # chance to finish and be executed at the run stage. Note that
-        # "test_conformance" target skips building "test_all" executable.
-        ninja -C build-cts -k0 $( [ -n "$CTS_TESTS_TO_BUILD" ] && echo "$CTS_TESTS_TO_BUILD" || echo "test_conformance")
-
-    - name: Pack SYCL-CTS binaries
-      if: always() && !cancelled() && inputs.cts_testing_mode == 'build-only'
-      run: tar -I 'zstd -9' -cf sycl_cts_bin.tar.zst -C ./build-cts/bin .
-
-    - name: Upload SYCL-CTS binaries
-      if: always() && !cancelled() && inputs.cts_testing_mode == 'build-only'
-      uses: actions/upload-artifact@v4
+      uses: ./devops/actions/run-tests/e2e
       with:
-        name: sycl_cts_bin
-        path: sycl_cts_bin.tar.zst
+        ref: ${{ inputs.ref || github.sha }}
+        merge_ref: ${{ inputs.merge_ref }}
+        e2e_binaries_artifact: ${{ inputs.e2e_binaries_artifact }}
+        extra_cmake_args: ${{ inputs.extra_cmake_args }}
+        e2e_testing_mode: ${{ inputs.e2e_testing_mode }}
+        target_devices: ${{ inputs.target_devices }}
+        extra_lit_opts: ${{ inputs.extra_lit_opts }}
+        artifact_suffix: ${{ inputs.artifact_suffix }}
         retention-days: ${{ inputs.retention-days }}
 
-    - name: Download SYCL-CTS binaries
-      if: inputs.sycl_cts_artifact != ''
-      uses: actions/download-artifact@v4
-      with:
-        name: ${{ inputs.sycl_cts_artifact }}
-
-    - name: Extract SYCL-CTS binaries
-      if: inputs.sycl_cts_artifact != ''
-      run: |
-        mkdir -p build-cts/bin
-        tar -I 'zstd' -xf sycl_cts_bin.tar.zst -C build-cts/bin
-
-    - name: SYCL CTS List devices
-      # Proceed with execution even if the 'build' step did not succeed.
-      if: inputs.tests_selector == 'cts' && (always() && !cancelled()) && inputs.cts_testing_mode != 'build-only'
-      env:
-        ONEAPI_DEVICE_SELECTOR: ${{ inputs.target_devices }}
-      run: |
-        ./build-cts/bin/* --list-devices
-
-    # If the suite was built on another machine then the build contains the full
-    # set of tests. We have special files to filter out some test categories,
-    # see "devops/cts_exclude_filter_*". Each configuration has its own file, e.g.
-    # there is "cts_exclude_filter_OCL_CPU" for opencl:cpu device. Therefore,
-    # these files may differ from each other, so when there is a pre-built set of
-    # tests, we need to filter it according to the filter-file.
-    - name: Filter SYCL CTS test categories
-      if: inputs.sycl_cts_artifact != ''
-      shell: bash
-      run: |
-        cts_exclude_filter=""
-        if [ "${{ contains(inputs.target_devices, 'opencl:cpu')  }}" = "true" ]; then
-          cts_exclude_filter=$PWD/devops/cts_exclude_filter_OCL_CPU
-        elif [ "${{ contains(inputs.target_devices, 'level_zero:gpu')  }}" = "true" ]; then
-          cts_exclude_filter=$PWD/devops/cts_exclude_filter_L0_GPU
-        fi
-
-        while IFS= read -r line; do
-          if [[ $line != \#* ]]; then
-            rm "./build-cts/bin/test_$line"
-          fi
-        done < "$cts_exclude_filter"
-
-    - name: Run SYCL CTS tests
-      # Proceed with execution even if the previous two steps did not succeed.
-      if: inputs.tests_selector == 'cts' && (always() && !cancelled()) && inputs.cts_testing_mode != 'build-only'
-      env:
-        ONEAPI_DEVICE_SELECTOR: ${{ inputs.target_devices }}
-      # This job takes ~100min usually. But sometimes some test isn't
-      # responding, so the job reaches the 360min limit. Setting a lower one.
-      timeout-minutes: 150
-      # By-default GitHub actions execute the "run" shell script with -e option,
-      # so the execution terminates if any command returns a non-zero status.
-      # Since we're using a loop to run all test-binaries separately, some test
-      # may fail and terminate the execution. Setting "shell" value to override
-      # the default behavior.
-      # https://docs.github.com/en/actions/using-workflows/workflow-syntax-for-github-actions#custom-shell
-      shell: bash {0}
-      run: |
-        # Run each test category separately so that
-        #  - crash on one would not affect others
-        #  - multiple tests could be run in parallel
-        mkdir logs
-        find build-cts/bin/ -type f -print | \
-          xargs -t -I % -P 8 sh -c 'log=logs/$(basename %).log ; echo % >$log ; date >>$log ; timeout 60m % >>$log 2>&1 ; ret=$? ; echo "exit code: $ret" >>$log ; date >>$log  ; exit $ret'
-        ret=$?
-
-        for f in logs/* ; do
-          echo "::group::$f"
-          cat $f
-          echo "::endgroup::"
-        done
-
-        echo "::group::Fails:"
-        grep 'exit code: [^0]' -r logs
-        echo "::endgroup::"
-
-        grep 'exit code: [^0]' -r logs >> $GITHUB_STEP_SUMMARY
-
-        exit $ret
-    - name: Pack E2E binaries
-      if: ${{ always() && !cancelled() && inputs.e2e_testing_mode == 'build-only'}}
-      run: tar -I 'zstd -9' -cf e2e_binaries.tar.zst -C ./build-e2e .
-    - name: Upload E2E binaries
-      if: ${{ always() && !cancelled() && inputs.e2e_testing_mode == 'build-only'}}
-      uses: actions/upload-artifact@v4
+    - name: Run SYCL CTS Tests
+      if: inputs.tests_selector == 'cts'
+      uses: ./devops/actions/run-tests/cts
       with:
-        name: sycl_e2e_bin_${{ inputs.artifact_suffix }}
-        path: e2e_binaries.tar.zst
+        extra_cmake_args: ${{ inputs.extra_cmake_args }}
+        cts_testing_mode: ${{ inputs.cts_testing_mode }}
+        sycl_cts_artifact: ${{ inputs.sycl_cts_artifact }}
+        target_devices: ${{ inputs.target_devices }}
         retention-days: ${{ inputs.retention-days }}

clang/include/clang/Basic/Attr.td

@@ -2147,6 +2147,22 @@ def SYCLAddIRAnnotationsMember : InheritableAttr {
   let Documentation = [SYCLAddIRAnnotationsMemberDocs];
 }
 
+def SYCLRegisteredKernels : InheritableAttr {
+  let Spellings = [CXX11<"__sycl_detail__", "__registered_kernels__">];
+  let Args = [VariadicExprArgument<"Args">];
+  let LangOpts = [SYCLIsDevice, SilentlyIgnoreSYCLIsHost];
+  let Subjects = SubjectList<[Empty], ErrorDiag, "Translation Unit Scope">;
+  let AdditionalMembers = SYCLAddIRAttrCommonMembers.MemberCode;
+  let Documentation = [SYCLAddIRAnnotationsMemberDocs];
+}
+
+def SYCLRegisteredKernelName : InheritableAttr {
+  let Spellings = [];
+  let Subjects = SubjectList<[Function]>;
+  let Args = [StringArgument<"RegName">];
+  let Documentation = [InternalOnly];
+}
+
 def C11NoReturn : InheritableAttr {
   let Spellings = [CustomKeyword<"_Noreturn">];
   let Subjects = SubjectList<[Function], ErrorDiag>;

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -12546,6 +12546,20 @@ def err_sycl_special_type_num_init_method : Error<
 def warn_launch_bounds_is_cuda_specific : Warning<
   "%0 attribute ignored, only applicable when targeting Nvidia devices">,
   InGroup<IgnoredAttributes>;
+def err_registered_kernels_num_of_args : Error<
+  "'__registered_kernels__' attribute must have at least one argument">;
+def err_registered_kernels_init_list : Error<
+  "argument to the '__registered_kernels__' attribute must be an "
+  "initializer list expression">;
+def err_registered_kernels_init_list_pair_values : Error<
+  "each initializer list argument to the '__registered_kernels__' attribute "
+  "must contain a pair of values">;
+def err_registered_kernels_resolve_function : Error<
+  "unable to resolve free function kernel '%0'">;
+def err_registered_kernels_name_already_registered : Error<
+  "free function kernel has already been registered with '%0'; cannot register with '%1'">;
+def err_not_sycl_free_function : Error<
+  "attempting to register a function that is not a SYCL free function as '%0'">;
 
 def warn_cuda_maxclusterrank_sm_90 : Warning<
   "'maxclusterrank' requires sm_90 or higher, CUDA arch provided: %0, ignoring "

clang/include/clang/Sema/SemaSYCL.h

@@ -252,8 +252,9 @@ class SemaSYCL : public SemaBase {
   // We need to store the list of the sycl_kernel functions and their associated
   // generated OpenCL Kernels so we can go back and re-name these after the
   // fact.
-  llvm::SmallVector<std::pair<const FunctionDecl *, FunctionDecl *>>
-      SyclKernelsToOpenCLKernels;
+  using KernelFDPairs =
+      llvm::SmallVector<std::pair<const FunctionDecl *, FunctionDecl *>>;
+  KernelFDPairs SyclKernelsToOpenCLKernels;
 
   // Used to suppress diagnostics during kernel construction, since these were
   // already emitted earlier. Diagnosing during Kernel emissions also skips the
@@ -296,11 +297,15 @@ class SemaSYCL : public SemaBase {
                               llvm::DenseSet<QualType> Visited,
                               ValueDecl *DeclToCheck);
 
+  const KernelFDPairs &getKernelFDPairs() { return SyclKernelsToOpenCLKernels; }
+
   void addSyclOpenCLKernel(const FunctionDecl *SyclKernel,
                            FunctionDecl *OpenCLKernel) {
     SyclKernelsToOpenCLKernels.emplace_back(SyclKernel, OpenCLKernel);
   }
 
+  void constructFreeFunctionKernel(FunctionDecl *FD, StringRef NameStr = "");
+
   void addSyclDeviceDecl(Decl *d) { SyclDeviceDecls.insert(d); }
   llvm::SetVector<Decl *> &syclDeviceDecls() { return SyclDeviceDecls; }
 
@@ -480,6 +485,7 @@ class SemaSYCL : public SemaBase {
   void handleSYCLIntelMaxWorkGroupsPerMultiprocessor(Decl *D,
                                                      const ParsedAttr &AL);
   void handleSYCLScopeAttr(Decl *D, const ParsedAttr &AL);
+  void handleSYCLRegisteredKernels(Decl *D, const ParsedAttr &AL);
 
   void checkSYCLAddIRAttributesFunctionAttrConflicts(Decl *D);
 
@@ -655,6 +661,10 @@ class SemaSYCL : public SemaBase {
   void addIntelReqdSubGroupSizeAttr(Decl *D, const AttributeCommonInfo &CI,
                                     Expr *E);
   void handleKernelEntryPointAttr(Decl *D, const ParsedAttr &AL);
+
+  // Used to check whether the function represented by FD is a SYCL
+  // free function kernel or not.
+  bool isFreeFunction(const FunctionDecl *FD);
 };
 
 } // namespace clang

clang/lib/CodeGen/CodeGenFunction.cpp

@@ -641,6 +641,12 @@ void CodeGenFunction::EmitKernelMetadata(const FunctionDecl *FD,
 
   llvm::LLVMContext &Context = getLLVMContext();
 
+  if (getLangOpts().SYCLIsDevice)
+    if (FD->hasAttr<SYCLRegisteredKernelNameAttr>())
+      CGM.SYCLAddRegKernelNamePairs(
+          FD->getAttr<SYCLRegisteredKernelNameAttr>()->getRegName(),
+          FD->getNameAsString());
+
   if (FD->hasAttr<OpenCLKernelAttr>() || FD->hasAttr<CUDAGlobalAttr>())
     CGM.GenKernelArgMetadata(Fn, FD, this);
 

clang/lib/CodeGen/CodeGenModule.cpp

@@ -1431,6 +1431,19 @@ void CodeGenModule::Release() {
         AspectEnumValsMD->addOperand(
             getAspectEnumValueMD(Context, TheModule.getContext(), ECD));
     }
+
+    if (!SYCLRegKernelNames.empty()) {
+      std::vector<llvm::Metadata *> Nodes;
+      llvm::LLVMContext &Ctx = TheModule.getContext();
+      for (auto MDKernelNames : SYCLRegKernelNames) {
+        llvm::Metadata *Vals[] = {MDKernelNames.first, MDKernelNames.second};
+        Nodes.push_back(llvm::MDTuple::get(Ctx, Vals));
+      }
+
+      llvm::NamedMDNode *SYCLRegKernelsMD =
+          TheModule.getOrInsertNamedMetadata("sycl_registered_kernels");
+      SYCLRegKernelsMD->addOperand(llvm::MDNode::get(Ctx, Nodes));
+    }
   }
 
   // HLSL related end of code gen work items.

clang/lib/CodeGen/CodeGenModule.h

@@ -456,6 +456,9 @@ class CodeGenModule : public CodeGenTypeCache {
   /// handled differently than regular annotations so they cannot share map.
   llvm::DenseMap<unsigned, llvm::Constant *> SYCLAnnotationArgs;
 
+  typedef std::pair<llvm::Metadata *, llvm::Metadata *> MetadataPair;
+  SmallVector<MetadataPair, 4> SYCLRegKernelNames;
+
   llvm::StringMap<llvm::GlobalVariable *> CFConstantStringMap;
 
   llvm::DenseMap<llvm::Constant *, llvm::GlobalVariable *> ConstantStringMap;
@@ -1483,6 +1486,12 @@ class CodeGenModule : public CodeGenTypeCache {
   llvm::Constant *EmitSYCLAnnotationArgs(
       SmallVectorImpl<std::pair<std::string, std::string>> &Pairs);
 
+  void SYCLAddRegKernelNamePairs(StringRef First, StringRef Second) {
+    SYCLRegKernelNames.push_back(
+        std::make_pair(llvm::MDString::get(getLLVMContext(), First),
+                       llvm::MDString::get(getLLVMContext(), Second)));
+  }
+
   /// Add attributes from add_ir_attributes_global_variable on TND to GV.
   void AddGlobalSYCLIRAttributes(llvm::GlobalVariable *GV,
                                  const RecordDecl *RD);

clang/lib/Sema/SemaDeclAttr.cpp

@@ -7479,6 +7479,9 @@ ProcessDeclAttribute(Sema &S, Scope *scope, Decl *D, const ParsedAttr &AL,
   case ParsedAttr::AT_SYCLAddIRAnnotationsMember:
     S.SYCL().handleSYCLAddIRAnnotationsMemberAttr(D, AL);
     break;
+  case ParsedAttr::AT_SYCLRegisteredKernels:
+    S.SYCL().handleSYCLRegisteredKernels(D, AL);
+    break;
 
   // Swift attributes.
   case ParsedAttr::AT_SwiftAsyncName: