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

Author: cperkinsintel

.github/workflows/pr-code-format.yml

@@ -7,6 +7,9 @@ on:
   pull_request:
     branches:
       - main
+      - sycl
+      - sycl-devops-pr/**
+      - sycl-rel-**
       - 'users/**'
 
 jobs:

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

@@ -212,6 +212,12 @@ jobs:
       if: always() && !cancelled() && contains(inputs.changes, 'libdevice')
       run: |
         cmake --build $GITHUB_WORKSPACE/build --target check-libdevice
+    - name: Check E2E test requirements
+      if: always() && !cancelled() && !contains(inputs.changes, 'sycl')
+      run: |
+        # TODO consider moving this to Dockerfile.
+        export LD_LIBRARY_PATH=/usr/local/cuda/compat/:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
+        LIT_OPTS="--allow-empty-runs" LIT_FILTER="e2e_test_requirements" cmake --build $GITHUB_WORKSPACE/build --target check-sycl
     - name: Install
       if: ${{ always() && !cancelled() && steps.build.conclusion == 'success' }}
       # TODO replace utility installation with a single CMake target

buildbot/configure.py

@@ -64,7 +64,7 @@ def do_configure(args):
 
     sycl_enable_xpti_tracing = "ON"
     xpti_enable_werror = "OFF"
-    llvm_enable_zstd = "OFF"
+    llvm_enable_zstd = "ON"
 
     if sys.platform != "darwin":
         sycl_enabled_backends.append("level_zero")
@@ -134,8 +134,6 @@ def do_configure(args):
 
         # For clang-format, clang-tidy and code coverage
         llvm_enable_projects += ";clang-tools-extra;compiler-rt"
-        # Build with zstd enabled on CI.
-        llvm_enable_zstd = "ON"
         if sys.platform != "darwin":
             # libclc is required for CI validation
             libclc_enabled = True

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -398,6 +398,10 @@ def err_drv_sycl_missing_amdgpu_arch : Error<
   "missing AMDGPU architecture for SYCL offloading; specify it with '-Xsycl-target-backend%select{|=%1}0 --offload-arch=<arch-name>'">;
 def err_drv_sycl_thinlto_split_off: Error<
   "'%0' is not supported when '%1' is set with '-fsycl'">;
+def err_drv_sycl_offload_arch_new_driver: Error<
+  "'--offload-arch' is supported when '-fsycl' is set with '--offload-new-driver'">;
+def err_drv_sycl_offload_arch_missing_value : Error<
+  "must pass in an explicit cpu or gpu architecture to '--offload-arch'">;
 def warn_drv_sycl_offload_target_duplicate : Warning<
   "SYCL offloading target '%0' is similar to target '%1' already specified; "
   "will be ignored">, InGroup<SyclTarget>;

clang/include/clang/Driver/Action.h

@@ -698,6 +698,14 @@ class OffloadWrapperJobAction : public JobAction {
   // Get the compilation step setting.
   bool getCompileStep() const { return CompileStep; }
 
+  // Set the individual wrapping setting.  This is used to tell the wrapper job
+  // action that the wrapping (and subsequent compile step) should be done
+  // with for-each instead of using -batch.
+  void setWrapIndividualFiles() { WrapIndividualFiles = true; }
+
+  // Get the individual wrapping setting.
+  bool getWrapIndividualFiles() const { return WrapIndividualFiles; }
+
   // Set the offload kind for the current wrapping job action.  Default usage
   // is to use the kind of the current toolchain.
   void setOffloadKind(OffloadKind SetKind) { Kind = SetKind; }
@@ -707,6 +715,7 @@ class OffloadWrapperJobAction : public JobAction {
 
 private:
   bool CompileStep = true;
+  bool WrapIndividualFiles = false;
   OffloadKind Kind = OFK_None;
 };
 

clang/lib/Driver/Driver.cpp

@@ -1191,12 +1191,13 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
   llvm::StringMap<llvm::DenseSet<StringRef>> DerivedArchs;
   llvm::StringMap<StringRef> FoundNormalizedTriples;
   llvm::SmallVector<llvm::Triple, 4> UniqueSYCLTriplesVec;
+  // StringSet to contain SYCL target triples.
+  llvm::StringSet<> SYCLTriples;
   if (HasSYCLTargetsOption) {
     // At this point, we know we have a valid combination
     // of -fsycl*target options passed
     Arg *SYCLTargetsValues = SYCLTargets;
     if (SYCLTargetsValues) {
-      llvm::StringSet<> SYCLTriples;
       if (SYCLTargetsValues->getNumValues()) {
 
         // Multiple targets are currently not supported when using
@@ -1296,6 +1297,109 @@ void Driver::CreateOffloadingDeviceToolChains(Compilation &C,
         Diag(clang::diag::warn_drv_empty_joined_argument)
             << SYCLTargetsValues->getAsString(C.getInputArgs());
     }
+  }
+  // If the user specified --offload-arch, deduce the offloading
+  // target triple(s) from the set of architecture(s).
+  // Create a toolchain for each valid triple.
+  // We do not support SYCL offloading if any of the inputs is a
+  // .cu (for CUDA type) or .hip (for HIP type) file.
+  else if (HasValidSYCLRuntime &&
+           C.getInputArgs().hasArg(options::OPT_offload_arch_EQ) && !IsHIP &&
+           !IsCuda) {
+    // SYCL offloading to AOT Targets with '--offload-arch'
+    // is currently enabled only with '--offload-new-driver' option.
+    // Emit a diagnostic if '--offload-arch' is invoked without
+    // '--offload-new driver' option.
+    if (!C.getInputArgs().hasFlag(options::OPT_offload_new_driver,
+                                  options::OPT_no_offload_new_driver, false)) {
+      Diag(clang::diag::err_drv_sycl_offload_arch_new_driver);
+      return;
+    }
+    const ToolChain *HostTC = C.getSingleOffloadToolChain<Action::OFK_Host>();
+    auto AMDTriple = getHIPOffloadTargetTriple(*this, C.getInputArgs());
+    auto NVPTXTriple = getNVIDIAOffloadTargetTriple(*this, C.getInputArgs(),
+                                                    HostTC->getTriple());
+
+    // Attempt to deduce the offloading triple from the set of architectures.
+    // We need to temporarily create these toolchains so that we can access
+    // tools for inferring architectures.
+    llvm::DenseSet<StringRef> Archs;
+    if (NVPTXTriple) {
+      auto TempTC = std::make_unique<toolchains::CudaToolChain>(
+          *this, *NVPTXTriple, *HostTC, C.getInputArgs(), Action::OFK_None);
+      for (StringRef Arch :
+           getOffloadArchs(C, C.getArgs(), Action::OFK_SYCL, &*TempTC, true))
+        Archs.insert(Arch);
+    }
+    if (AMDTriple) {
+      auto TempTC = std::make_unique<toolchains::AMDGPUOpenMPToolChain>(
+          *this, *AMDTriple, *HostTC, C.getInputArgs());
+      for (StringRef Arch :
+           getOffloadArchs(C, C.getArgs(), Action::OFK_SYCL, &*TempTC, true))
+        Archs.insert(Arch);
+    }
+    if (!AMDTriple && !NVPTXTriple) {
+      for (StringRef Arch :
+           getOffloadArchs(C, C.getArgs(), Action::OFK_SYCL, nullptr, true))
+        Archs.insert(Arch);
+    }
+    for (StringRef Arch : Archs) {
+      if (NVPTXTriple && IsSYCLSupportedNVidiaGPUArch(StringToOffloadArch(
+                             getProcessorFromTargetID(*NVPTXTriple, Arch)))) {
+        DerivedArchs[NVPTXTriple->getTriple()].insert(Arch);
+      } else if (AMDTriple &&
+                 IsSYCLSupportedAMDGPUArch(StringToOffloadArch(
+                     getProcessorFromTargetID(*AMDTriple, Arch)))) {
+        DerivedArchs[AMDTriple->getTriple()].insert(Arch);
+      } else if (IsSYCLSupportedIntelCPUArch(StringToOffloadArchSYCL(Arch))) {
+        DerivedArchs[MakeSYCLDeviceTriple("spir64_x86_64").getTriple()].insert(
+            Arch);
+      } else if (IsSYCLSupportedIntelGPUArch(StringToOffloadArchSYCL(Arch))) {
+        StringRef IntelGPUArch;
+        // For Intel Graphics AOT target, valid values for '--offload-arch'
+        // are mapped to valid device names accepted by OCLOC (the Intel GPU AOT
+        // compiler) via the '-device' option. The mapIntelGPUArchName
+        // function maps the accepted values for '--offload-arch' to enable SYCL
+        // offloading to Intel GPUs and the corresponding '-device' value passed
+        // to OCLOC.
+        IntelGPUArch = mapIntelGPUArchName(Arch).data();
+        DerivedArchs[MakeSYCLDeviceTriple("spir64_gen").getTriple()].insert(
+            IntelGPUArch);
+      } else {
+        Diag(clang::diag::err_drv_invalid_sycl_target) << Arch;
+        return;
+      }
+    }
+    // Emit an error if architecture value is not provided
+    // to --offload-arch.
+    if (Archs.empty()) {
+      Diag(clang::diag::err_drv_sycl_offload_arch_missing_value);
+      return;
+    }
+
+    for (const auto &TripleAndArchs : DerivedArchs)
+      SYCLTriples.insert(TripleAndArchs.first());
+
+    for (const auto &Val : SYCLTriples) {
+      llvm::Triple SYCLTargetTriple(MakeSYCLDeviceTriple(Val.getKey()));
+      std::string NormalizedName = SYCLTargetTriple.normalize();
+
+      // Make sure we don't have a duplicate triple.
+      auto Duplicate = FoundNormalizedTriples.find(NormalizedName);
+      if (Duplicate != FoundNormalizedTriples.end()) {
+        Diag(clang::diag::warn_drv_sycl_offload_target_duplicate)
+            << Val.getKey() << Duplicate->second;
+        continue;
+      }
+
+      // Store the current triple so that we can check for duplicates in the
+      // following iterations.
+      FoundNormalizedTriples[NormalizedName] = Val.getKey();
+      UniqueSYCLTriplesVec.push_back(SYCLTargetTriple);
+    }
+
+    addSYCLDefaultTriple(C, UniqueSYCLTriplesVec);
+
   } else {
     // If -fsycl is supplied without -fsycl-targets we will assume SPIR-V.
     // For -fsycl-device-only, we also setup the implied triple as needed.
@@ -5432,9 +5536,58 @@ class OffloadingActionBuilder final {
                   BundlingActions, types::TY_Object);
               if (auto *OWA = dyn_cast<OffloadWrapperJobAction>(DeviceAction))
                 OWA->setOffloadKind(Action::OFK_Host);
+              // The Backend compilation step performed here is being done for
+              // creating FPGA archives.  The possible split binaries after
+              // sycl-post-link need to be individually wrapped as opposed to
+              // being passed into the clang-offload-wrapper via a table and
+              // using the -batch option - effectively creating a single
+              // binary.  The resulting archive created from -fsycl-link should
+              // not contain the singular binary, but should be individual
+              // binaries to be consumed later by either the -fsycl-link=image
+              // device compilation step or being linked into the final exe.
+              //
+              // Typical compile flow:
+              //               .bc
+              //                |
+              //     sycl-post-link -split=kernel
+              //                |
+              //       +--------+--------+
+              //       |        |        |
+              //     split1   split2   split3
+              //       |        |        |
+              // llvm-spirv llvm-spirv llvm-spirv
+              //       |        |        |
+              //     ocloc    ocloc    ocloc
+              //       |        |        |
+              //       +--------+--------+
+              //                |
+              //     clang-offload-wrapper -batch
+              //                |
+              //               .o
+              //
+              // Individual wrap compile flow:
+              //               .bc
+              //                |
+              //     sycl-post-link -split=kernel
+              //                |
+              //       +--------+--------+
+              //       |        |        |
+              //     split1   split2   split3
+              //       |        |        |
+              // llvm-spirv llvm-spirv llvm-spirv
+              //       |        |        |
+              //     ocloc    ocloc    ocloc
+              //       |        |        |
+              //     wrap     wrap     wrap
+              //       |        |        |
+              //      .o       .o       .o
+              //
               Action *CompiledDeviceAction =
-                  C.MakeAction<OffloadWrapperJobAction>(WrapperItems,
-                                                        types::TY_Object);
+                  C.MakeAction<OffloadWrapperJobAction>(FPGAAOTAction,
+                                                        types::TY_Tempfilelist);
+              if (auto *OWA =
+                      dyn_cast<OffloadWrapperJobAction>(CompiledDeviceAction))
+                OWA->setWrapIndividualFiles();
               addDeps(CompiledDeviceAction, TC, BoundArch);
             }
             addDeps(DeviceAction, TC, BoundArch);
@@ -5708,6 +5861,9 @@ class OffloadingActionBuilder final {
           };
 
           Action *ExtractIRFilesAction = createExtractIRFilesAction();
+          // Device binaries that are individually wrapped when creating an
+          // FPGA Archive.
+          ActionList FPGAArchiveWrapperInputs;
 
           if (IsNVPTX || IsAMDGCN) {
             JobAction *FinAction =
@@ -5793,6 +5949,7 @@ class OffloadingActionBuilder final {
                 FileTableTformJobAction::COL_CODE,
                 FileTableTformJobAction::COL_CODE);
             WrapperInputs.push_back(ReplaceFilesAction);
+            FPGAArchiveWrapperInputs.push_back(BuildCodeAction);
           }
           if (SkipWrapper) {
             // Wrapper step not requested.
@@ -5827,8 +5984,11 @@ class OffloadingActionBuilder final {
                 if (auto *OWA = dyn_cast<OffloadWrapperJobAction>(DeviceAction))
                   OWA->setOffloadKind(Action::OFK_Host);
                 Action *CompiledDeviceAction =
-                    C.MakeAction<OffloadWrapperJobAction>(WrapperInputs,
-                                                          types::TY_Object);
+                    C.MakeAction<OffloadWrapperJobAction>(
+                        FPGAArchiveWrapperInputs, types::TY_Tempfilelist);
+                if (auto *OWA =
+                        dyn_cast<OffloadWrapperJobAction>(CompiledDeviceAction))
+                  OWA->setWrapIndividualFiles();
                 addDeps(CompiledDeviceAction, TC, nullptr);
               }
               addDeps(DeviceAction, TC, nullptr);
@@ -6357,7 +6517,7 @@ class OffloadingActionBuilder final {
           if (GpuInitHasErrors)
             return true;
 
-          int GenIndex = 0;
+          size_t GenIndex = 0;
           // Fill SYCLTargetInfoList
           for (auto &TT : SYCLTripleList) {
             auto TCIt = llvm::find_if(

clang/lib/Driver/ToolChains/Clang.cpp

@@ -10263,8 +10263,18 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
     const InputInfo &I = Inputs[0];
     assert(I.isFilename() && "Invalid input.");
 
-    if (I.getType() == types::TY_Tempfiletable ||
-        I.getType() == types::TY_Tempfilelist || IsEmbeddedIR)
+    // TODO: The embedded compilation step after the wrapping step restricts
+    // the ability to control the 'for each' methodology used when performing
+    // device code splitting.  We set the individual wrap behavior when we know
+    // the wrapping and compile step should be done individually.  Ideally this
+    // would be controlled at the JobAction creation, but we cannot do that
+    // until the compilation of the wrap is it's own JobAction.
+    bool IndividualWrapCompile = WrapperJob.getWrapIndividualFiles();
+    const InputInfo TempOutput(types::TY_LLVM_BC, WrapperFileName,
+                               WrapperFileName);
+    if (!IndividualWrapCompile &&
+        (I.getType() == types::TY_Tempfiletable ||
+         I.getType() == types::TY_Tempfilelist || IsEmbeddedIR))
       // Input files are passed via the batch job file table.
       WrapperArgs.push_back(C.getArgs().MakeArgString("-batch"));
     WrapperArgs.push_back(C.getArgs().MakeArgString(I.getFilename()));
@@ -10273,7 +10283,17 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
         JA, *this, ResponseFileSupport::None(),
         TCArgs.MakeArgString(getToolChain().GetProgramPath(getShortName())),
         WrapperArgs, std::nullopt);
-    C.addCommand(std::move(Cmd));
+
+    if (IndividualWrapCompile) {
+      // When wrapping FPGA device binaries for FPGA archives, create individual
+      // wrapped and compiled entries for the archive.
+      StringRef ParallelJobs =
+          C.getArgs().getLastArgValue(options::OPT_fsycl_max_parallel_jobs_EQ);
+      clang::driver::tools::SYCL::constructLLVMForeachCommand(
+          C, JA, std::move(Cmd), Inputs, TempOutput, this, "", "bc",
+          ParallelJobs);
+    } else
+      C.addCommand(std::move(Cmd));
 
     if (WrapperCompileEnabled) {
       // TODO Use TC.SelectTool().
@@ -10296,9 +10316,19 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
       SmallString<128> ClangPath(C.getDriver().Dir);
       llvm::sys::path::append(ClangPath, "clang");
       const char *Clang = C.getArgs().MakeArgString(ClangPath);
-      C.addCommand(std::make_unique<Command>(JA, *this,
-                                             ResponseFileSupport::None(), Clang,
-                                             ClangArgs, std::nullopt));
+      auto PostWrapCompileCmd =
+          std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
+                                    Clang, ClangArgs, std::nullopt);
+      if (IndividualWrapCompile) {
+        StringRef ParallelJobs = C.getArgs().getLastArgValue(
+            options::OPT_fsycl_max_parallel_jobs_EQ);
+        InputInfoList Inputs;
+        Inputs.push_back(TempOutput);
+        clang::driver::tools::SYCL::constructLLVMForeachCommand(
+            C, JA, std::move(PostWrapCompileCmd), Inputs, Output, this, "",
+            "bc", ParallelJobs);
+      } else
+        C.addCommand(std::move(PostWrapCompileCmd));
     }
     return;
   } // end of SYCL flavor of offload wrapper command creation