Merge branch 'sycl' into cperkins-kernel_compiler-sycl-cache

Author: cperkinsintel

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

@@ -83,22 +83,22 @@ jobs:
           - name: Intel
             runner: '["Linux", "gen12"]'
             image: ghcr.io/intel/llvm/ubuntu2204_intel_drivers:latest
-            image_options: -u 1001 --device=/dev/dri --privileged --cap-add SYS_ADMIN
+            image_options: -u 1001 --device=/dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged --cap-add SYS_ADMIN
             target_devices: level_zero:gpu;opencl:gpu;opencl:cpu
             reset_intel_gpu: true
             extra_lit_opts: --param gpu-intel-gen12=True
           - name: E2E tests on Intel Arc A-Series Graphics
             runner: '["Linux", "arc"]'
             image: ghcr.io/intel/llvm/ubuntu2204_intel_drivers:latest
-            image_options: -u 1001 --device=/dev/dri --privileged --cap-add SYS_ADMIN
+            image_options: -u 1001 --device=/dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged --cap-add SYS_ADMIN
             target_devices: level_zero:gpu;opencl:gpu
             reset_intel_gpu: true
             extra_lit_opts: --param matrix-xmx8=True --param gpu-intel-dg2=True
             env: '{"LIT_FILTER":${{ needs.determine_arc_tests.outputs.arc_tests }} }'
           - name: E2E tests with dev igc on Intel Arc A-Series Graphics
             runner: '["Linux", "arc"]'
             image: ghcr.io/intel/llvm/ubuntu2204_intel_drivers:devigc
-            image_options: -u 1001 --device=/dev/dri --privileged --cap-add SYS_ADMIN
+            image_options: -u 1001 --device=/dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged --cap-add SYS_ADMIN
             target_devices: level_zero:gpu;opencl:gpu
             reset_intel_gpu: true
             extra_lit_opts: --param matrix-xmx8=True --param gpu-intel-dg2=True
@@ -117,11 +117,11 @@ jobs:
       env: ${{ matrix.env || '{}' }}
 
       # Do not install drivers on AMD and CUDA runners.
-      install_igc_driver: |
+      install_igc_driver: >-
         ${{ !contains(matrix.target_devices, 'ext_oneapi_cuda') &&
         !contains(matrix.target_devices, 'ext_oneapi_hip') &&
         contains(needs.detect_changes.outputs.filters, 'drivers') }}
-      install_dev_igc_driver: |
+      install_dev_igc_driver: >-
         ${{ !contains(matrix.target_devices, 'ext_oneapi_cuda') &&
         !contains(matrix.target_devices, 'ext_oneapi_hip') &&
         matrix.use_igc_dev && contains(needs.detect_changes.outputs.filters, 'devigccfg') ||

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

@@ -151,7 +151,7 @@ permissions:
 
 jobs:
   run:
-    if: inputs.skip_run == 'false'
+    if: github.event_name == 'workflow_dispatch' || inputs.skip_run == 'false'
     name: ${{ inputs.name }}
     runs-on: ${{ fromJSON(inputs.runner) }}
     container:

.github/workflows/sycl-nightly.yml

@@ -53,7 +53,7 @@ jobs:
           - name: Intel L0 GPU
             runner: '["Linux", "gen12"]'
             image: ghcr.io/intel/llvm/ubuntu2204_intel_drivers:latest
-            image_options: -u 1001 --device=/dev/dri --privileged --cap-add SYS_ADMIN
+            image_options: -u 1001 --device=/dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged --cap-add SYS_ADMIN
             target_devices: level_zero:gpu
             reset_intel_gpu: true
             tests_selector: e2e
@@ -62,7 +62,7 @@ jobs:
           - name: Intel OCL GPU
             runner: '["Linux", "gen12"]'
             image: ghcr.io/intel/llvm/ubuntu2204_intel_drivers:latest
-            image_options: -u 1001 --device=/dev/dri --privileged --cap-add SYS_ADMIN
+            image_options: -u 1001 --device=/dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged --cap-add SYS_ADMIN
             target_devices: opencl:gpu
             reset_intel_gpu: true
             tests_selector: e2e

.github/workflows/sycl-post-commit.yml

@@ -87,7 +87,7 @@ jobs:
       name: ${{ matrix.name }}
       runner: ${{ matrix. runner }}
       image: ${{ matrix.image || 'ghcr.io/intel/llvm/ubuntu2204_intel_drivers:latest' }}
-      image_options: ${{ matrix.image_options || '-u 1001 --device=/dev/dri --privileged --cap-add SYS_ADMIN' }}
+      image_options: ${{ matrix.image_options || '-u 1001 --device=/dev/dri -v /dev/dri/by-path:/dev/dri/by-path --privileged --cap-add SYS_ADMIN' }}
       target_devices: ${{ matrix.target_devices || 'level_zero:gpu' }}
       reset_intel_gpu: ${{ matrix.reset_intel_gpu }}
 

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

@@ -86,6 +86,17 @@ jobs:
         # Run E2E tests.
         export LIT_OPTS="-v --no-progress-bar --show-unsupported --show-pass --show-xfail --max-time 3600 --time-tests ${{ inputs.extra_lit_opts }}"
         cmake --build build-e2e --target check-sycl-e2e
+    - name: Detect hung tests
+      shell: powershell
+      run: |
+        $exitCode = 0
+        $hungTests = Get-Process | Where-Object { ($_.Path -match "llvm\\install") -or ($_.Path -match "llvm\\build-e2e") }
+        $hungTests | Foreach-Object {
+         $exitCode = 1
+         echo "Test $($_.Path) hung!"
+         Stop-Process -Force $_
+        }
+        exit $exitCode
     - name: Cleanup
       shell: cmd
       if: always()

clang/include/clang/Basic/LangOptions.def

@@ -307,7 +307,6 @@ LANGOPT(SYCLAllowFuncPtr  , 1, 0, "Allow function pointers in SYCL device code")
 LANGOPT(SYCLStdLayoutKernelParams, 1, 0, "Enable standard layout requirement for SYCL kernel parameters")
 LANGOPT(SYCLUnnamedLambda , 1, 0, "Allow unnamed lambda SYCL kernels")
 LANGOPT(SYCLForceInlineKernelLambda , 1, 0, "Force inline SYCL kernel lambdas in entry point")
-LANGOPT(SYCLAllowAllFeaturesInConstexpr, 1, 0, "Allow all C++ features in SYCL device code in manifestly constant-evaluated expressions")
 LANGOPT(SYCLESIMDForceStatelessMem, 1, 0, "Make accessors use USM memory in ESIMD kernels")
 LANGOPT(SYCLESIMDBuildHostCode, 1, 1, "Build the host implementation of ESIMD functions")
 ENUM_LANGOPT(SYCLVersion  , SYCLMajorVersion, 2, SYCL_None, "Version of the SYCL standard used")

clang/include/clang/Driver/Options.td

@@ -8628,11 +8628,6 @@ def fsycl_is_native_cpu : Flag<["-"], "fsycl-is-native-cpu">,
   HelpText<"Perform device compilation for Native CPU.">,
   Visibility<[CC1Option]>,
   MarshallingInfoFlag<LangOpts<"SYCLIsNativeCPU">>;
-defm sycl_allow_all_features_in_constexpr
-    : BoolFOption<
-          "sycl-allow-all-features-in-constexpr", LangOpts<"SYCLAllowAllFeaturesInConstexpr">,
-          DefaultFalse,
-          PosFlag<SetTrue, [], [CC1Option], "Allow all C++ features in SYCL device code in manifestly constant-evaluated expressions">, NegFlag<SetFalse>>;
 
 } // let Visibility = [CC1Option]
 

clang/lib/Driver/Driver.cpp

@@ -6350,7 +6350,7 @@ class OffloadingActionBuilder final {
           if (GpuInitHasErrors)
             return true;
 
-          int I = 0;
+          int GenIndex = 0;
           // Fill SYCLTargetInfoList
           for (auto &TT : SYCLTripleList) {
             auto TCIt = llvm::find_if(
@@ -6363,10 +6363,21 @@ class OffloadingActionBuilder final {
               // is the target device.
               if (TT.isSPIR() &&
                   TT.getSubArch() == llvm::Triple::SPIRSubArch_gen) {
-                StringRef Device(GpuArchList[I].second);
+                // Multiple spir64_gen targets are allowed to be used via the
+                // -fsycl-targets=spir64_gen and -fsycl-targets=intel_gpu_*
+                // specifiers. Using an index through the known GpuArchList
+                // values, increment through them accordingly to allow for
+                // the multiple settings as well as preventing re-use.
+                while (TT != GpuArchList[GenIndex].first &&
+                       GenIndex < GpuArchList.size())
+                  ++GenIndex;
+                if (GpuArchList[GenIndex].first != TT)
+                  // No match.
+                  continue;
+                StringRef Device(GpuArchList[GenIndex].second);
                 SYCLTargetInfoList.emplace_back(
                     *TCIt, Device.empty() ? nullptr : Device.data());
-                ++I;
+                ++GenIndex;
                 continue;
               }
               SYCLTargetInfoList.emplace_back(*TCIt, nullptr);
@@ -6380,7 +6391,6 @@ class OffloadingActionBuilder final {
               }
               assert(OffloadArch && "Failed to find matching arch.");
               SYCLTargetInfoList.emplace_back(*TCIt, OffloadArch);
-              ++I;
             }
           }
         }
@@ -6776,13 +6786,10 @@ class OffloadingActionBuilder final {
     // Do not use unbundler if the Host does not depend on device action.
     // Now that we have unbundled the object, when doing -fsycl-link we
     // want to continue the host link with the input object.
-    // For unbundling of an FPGA AOCX binary, we want to link with the original
-    // FPGA device archive.
     if ((OffloadKind == Action::OFK_None && CanUseBundler) ||
         (Args.hasArg(options::OPT_fsycl_link_EQ) && !HasFPGATarget) ||
         (HasFPGATarget && ((Args.hasArg(options::OPT_fsycl_link_EQ) &&
-                            HostAction->getType() == types::TY_Object) ||
-                           HostAction->getType() == types::TY_FPGA_AOCX)))
+                            HostAction->getType() == types::TY_Object))))
       if (auto *UA = dyn_cast<OffloadUnbundlingJobAction>(HostAction))
         HostAction = UA->getInputs().back();
 
@@ -7433,14 +7440,12 @@ void Driver::BuildActions(Compilation &C, DerivedArgList &Args,
 
   // For an FPGA archive, we add the unbundling step above to take care of
   // the device side, but also unbundle here to extract the host side
-  bool EarlyLink = false;
-  if (const Arg *A = Args.getLastArg(options::OPT_fsycl_link_EQ))
-    EarlyLink = A->getValue() == StringRef("early");
   for (auto &LI : LinkerInputs) {
     Action *UnbundlerInput = nullptr;
     auto wrapObject = [&] {
-      if (EarlyLink && Args.hasArg(options::OPT_fintelfpga)) {
-        // Only wrap the object with -fsycl-link=early
+      if (Args.hasArg(options::OPT_fsycl_link_EQ) &&
+          Args.hasArg(options::OPT_fintelfpga)) {
+        // Wrap the object when creating an FPGA AOCX or AOCR binary.
         auto *BC = C.MakeAction<OffloadWrapperJobAction>(LI, types::TY_LLVM_BC);
         auto *ASM = C.MakeAction<BackendJobAction>(BC, types::TY_PP_Asm);
         auto *OBJ = C.MakeAction<AssembleJobAction>(ASM, types::TY_Object);

clang/lib/Driver/ToolChains/Cuda.cpp

@@ -946,6 +946,15 @@ void CudaToolChain::addClangTargetOptions(
 
     if (DriverArgs.hasArg(options::OPT_fsycl_fp32_prec_sqrt))
       CC1Args.push_back("-fcuda-prec-sqrt");
+
+    bool FastRelaxedMath = DriverArgs.hasFlag(
+        options::OPT_ffast_math, options::OPT_fno_fast_math, false);
+    bool UnsafeMathOpt =
+        DriverArgs.hasFlag(options::OPT_funsafe_math_optimizations,
+                           options::OPT_fno_unsafe_math_optimizations, false);
+    if (FastRelaxedMath || UnsafeMathOpt)
+      CC1Args.append({"-mllvm", "--nvptx-prec-divf32=0", "-mllvm",
+                      "--nvptx-prec-sqrtf32=0"});
   } else {
     CC1Args.append(
         {"-fcuda-is-device", "-mllvm", "-enable-memcpyopt-without-libcalls"});

clang/lib/Driver/ToolChains/SYCL.cpp

@@ -579,6 +579,7 @@ const char *SYCL::Linker::constructLLVMLinkCommand(
     auto isSYCLDeviceLib = [&](const InputInfo &II) {
       const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>();
       const bool IsNVPTX = this->getToolChain().getTriple().isNVPTX();
+      const bool IsAMDGCN = this->getToolChain().getTriple().isAMDGCN();
       const bool IsFPGA = this->getToolChain().getTriple().isSPIR() &&
                           this->getToolChain().getTriple().getSubArch() ==
                               llvm::Triple::SPIRSubArch_fpga;
@@ -598,6 +599,9 @@ const char *SYCL::Linker::constructLLVMLinkCommand(
       // NativeCPU links against libclc (libspirv)
       if (IsSYCLNativeCPU && InputFilename.contains("libspirv"))
         return true;
+      // AMDGCN links against our libdevice (devicelib)
+      if (IsAMDGCN && InputFilename.starts_with("devicelib-"))
+        return true;
       // NVPTX links against our libclc (libspirv), our libdevice (devicelib),
       // and the CUDA libdevice
       if (IsNVPTX && (InputFilename.starts_with("devicelib-") ||