[SYCL][Graph][Doc] Add SYCL-Graph usage guide and example doc

.github/CODEOWNERS

@@ -103,7 +103,7 @@ buildbot/ @intel/dpcpp-devops-reviewers
 devops/ @intel/dpcpp-devops-reviewers
 
 # Kernel fusion JIT compiler
-sycl-fusion/ @intel/dpcpp-kernel-fusion-reviewers
+sycl-jit/ @intel/dpcpp-kernel-fusion-reviewers
 sycl/doc/design/KernelFusionJIT.md @intel/dpcpp-kernel-fusion-reviewers
 sycl/doc/extensions/experimental/sycl_ext_codeplay_kernel_fusion.asciidoc @intel/dpcpp-kernel-fusion-reviewers
 sycl/include/sycl/ext/codeplay/experimental/fusion_properties.hpp  @intel/dpcpp-kernel-fusion-reviewers

.github/workflows/sycl-detect-changes.yml

@@ -30,9 +30,9 @@ jobs:
             clang: &clang
               - *llvm
               - 'clang/**'
-            sycl_fusion: &sycl-fusion
+            sycl_jit: &sycl-jit
               - *llvm
-              - 'sycl-fusion/**'
+              - 'sycl-jit/**'
             xptifw: &xptifw
               - 'xptifw/**'
             libclc: &libclc
@@ -41,7 +41,7 @@ jobs:
               - 'libclc/**'
             sycl: &sycl
               - *clang
-              - *sycl-fusion
+              - *sycl-jit
               - *llvm_spirv
               - *xptifw
               - *libclc
@@ -84,7 +84,7 @@ jobs:
               return '${{ steps.changes.outputs.changes }}';
             }
             // Treat everything as changed for huge PRs.
-            return ["llvm", "llvm_spirv", "clang", "sycl_fusion", "xptifw", "libclc", "sycl", "ci", "esimd"];
+            return ["llvm", "llvm_spirv", "clang", "sycl_jit", "xptifw", "libclc", "sycl", "ci", "esimd"];
 
       - run: echo '${{ steps.result.outputs.result }}'
 

buildbot/configure.py

@@ -31,10 +31,10 @@ def do_configure(args):
     libclc_amd_target_names = ";amdgcn--amdhsa"
     libclc_nvidia_target_names = ";nvptx64--nvidiacl"
 
-    sycl_enable_fusion = "OFF"
-    if not args.disable_fusion:
-        llvm_external_projects += ";sycl-fusion"
-        sycl_enable_fusion = "ON"
+    sycl_enable_jit = "OFF"
+    if not args.disable_jit:
+        llvm_external_projects += ";sycl-jit"
+        sycl_enable_jit = "ON"
 
     if args.llvm_external_projects:
         llvm_external_projects += ";" + args.llvm_external_projects.replace(",", ";")
@@ -45,7 +45,7 @@ def do_configure(args):
     xpti_dir = os.path.join(abs_src_dir, "xpti")
     xptifw_dir = os.path.join(abs_src_dir, "xptifw")
     libdevice_dir = os.path.join(abs_src_dir, "libdevice")
-    fusion_dir = os.path.join(abs_src_dir, "sycl-fusion")
+    jit_dir = os.path.join(abs_src_dir, "sycl-jit")
     llvm_targets_to_build = args.host_target
     llvm_enable_projects = "clang;" + llvm_external_projects
     libclc_build_native = "OFF"
@@ -174,7 +174,7 @@ def do_configure(args):
         "-DXPTI_SOURCE_DIR={}".format(xpti_dir),
         "-DLLVM_EXTERNAL_XPTIFW_SOURCE_DIR={}".format(xptifw_dir),
         "-DLLVM_EXTERNAL_LIBDEVICE_SOURCE_DIR={}".format(libdevice_dir),
-        "-DLLVM_EXTERNAL_SYCL_FUSION_SOURCE_DIR={}".format(fusion_dir),
+        "-DLLVM_EXTERNAL_SYCL_JIT_SOURCE_DIR={}".format(jit_dir),
         "-DLLVM_ENABLE_PROJECTS={}".format(llvm_enable_projects),
         "-DSYCL_BUILD_PI_HIP_PLATFORM={}".format(sycl_build_pi_hip_platform),
         "-DLLVM_BUILD_TOOLS=ON",
@@ -189,7 +189,7 @@ def do_configure(args):
         "-DXPTI_ENABLE_WERROR={}".format(xpti_enable_werror),
         "-DSYCL_CLANG_EXTRA_FLAGS={}".format(sycl_clang_extra_flags),
         "-DSYCL_ENABLE_PLUGINS={}".format(";".join(set(sycl_enabled_plugins))),
-        "-DSYCL_ENABLE_KERNEL_FUSION={}".format(sycl_enable_fusion),
+        "-DSYCL_ENABLE_EXTENSION_JIT={}".format(sycl_enable_jit),
         "-DSYCL_ENABLE_MAJOR_RELEASE_PREVIEW_LIB={}".format(sycl_preview_lib),
         "-DBUG_REPORT_URL=https://github.com/intel/llvm/issues",
     ]
@@ -379,9 +379,9 @@ def main():
         help="Disable building of the SYCL runtime major release preview library",
     )
     parser.add_argument(
-        "--disable-fusion",
+        "--disable-jit",
         action="store_true",
-        help="Disable the kernel fusion JIT compiler",
+        help="Disable the kernel JIT compiler for AMD and Nvidia",
     )
     parser.add_argument(
         "--add_security_flags",

clang/include/clang/Driver/Options.td

@@ -4189,6 +4189,10 @@ def fsycl_remove_unused_external_funcs : Flag<["-"], "fsycl-remove-unused-extern
   Group<sycl_Group>, HelpText<"Allow removal of unused `SYCL_EXTERNAL` functions (default)">;
 def fno_sycl_remove_unused_external_funcs : Flag<["-"], "fno-sycl-remove-unused-external-funcs">,
   Group<sycl_Group>, HelpText<"Prevent removal of unused `SYCL_EXTERNAL` functions">;
+def fsycl_allow_device_dependencies : Flag<["-"], "fsycl-allow-device-dependencies">,
+  Group<sycl_Group>, HelpText<"Allow dependencies between device code images">;
+def fno_sycl_allow_device_dependencies : Flag<["-"], "fno-sycl-allow-device-dependencies">,
+  Group<sycl_Group>, HelpText<"Do not allow dependencies between device code images (default)">;
 
 def fsave_optimization_record : Flag<["-"], "fsave-optimization-record">,
   Visibility<[ClangOption, FlangOption]>,

clang/lib/CodeGen/BackendUtil.cpp

@@ -56,6 +56,7 @@
 #include "llvm/SYCLLowerIR/RecordSYCLAspectNames.h"
 #include "llvm/SYCLLowerIR/SYCLAddOptLevelAttribute.h"
 #include "llvm/SYCLLowerIR/SYCLConditionalCallOnDevice.h"
+#include "llvm/SYCLLowerIR/SYCLCreateNVVMAnnotations.h"
 #include "llvm/SYCLLowerIR/SYCLPropagateAspectsUsage.h"
 #include "llvm/SYCLLowerIR/SYCLPropagateJointMatrixUsage.h"
 #include "llvm/SYCLLowerIR/SYCLVirtualFunctionsAnalysis.h"
@@ -1151,6 +1152,9 @@ void EmitAssemblyHelper::RunOptimizationPipeline(
       // and before cleaning up metadata)
       MPM.addPass(RecordSYCLAspectNamesPass());
 
+      if (TargetTriple.isNVPTX())
+        MPM.addPass(SYCLCreateNVVMAnnotationsPass());
+
       // Remove SYCL metadata added by the frontend, like sycl_aspects
       // Note, this pass should be at the end of the pipeline
       MPM.addPass(CleanupSYCLMetadataPass());

clang/lib/CodeGen/Targets/NVPTX.cpp

@@ -291,80 +291,6 @@ void NVPTXTargetCodeGenInfo::setTargetAttributes(
           addNVVMMetadata(F, "grid_constant", GridConstantParamIdxs);
       }
     }
-    bool HasMaxWorkGroupSize = false;
-    bool HasMinWorkGroupPerCU = false;
-    if (const auto *MWGS = FD->getAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
-      HasMaxWorkGroupSize = true;
-      // We must index-flip between SYCL's notation, X,Y,Z (aka dim0,dim1,dim2)
-      // with the fastest-moving dimension rightmost, to CUDA's, where X is the
-      // fastest-moving dimension.
-      addNVVMMetadata(F, "maxntidx", MWGS->getZDimVal());
-      addNVVMMetadata(F, "maxntidy", MWGS->getYDimVal());
-      addNVVMMetadata(F, "maxntidz", MWGS->getXDimVal());
-    }
-
-    if (const auto *RWGS = FD->getAttr<SYCLReqdWorkGroupSizeAttr>()) {
-      llvm::SmallVector<std::optional<int64_t>, 3> Ops;
-      // Index-flip and pad out any missing elements. Note the misleading
-      // nomenclature of the methods: getXDimVal doesn't return the X dimension;
-      // it returns the left-most dimension (dim0). This could correspond to
-      // CUDA's X, Y, or Z, depending on the number of operands provided.
-      if (auto Dim0 = RWGS->getXDimVal())
-        Ops.push_back(Dim0->getExtValue());
-      if (auto Dim1 = RWGS->getYDimVal())
-        Ops.push_back(Dim1->getExtValue());
-      if (auto Dim2 = RWGS->getZDimVal())
-        Ops.push_back(Dim2->getExtValue());
-      std::reverse(Ops.begin(), Ops.end());
-      Ops.append(3 - Ops.size(), std::nullopt);
-
-      // Work-group sizes (in NVVM annotations) must be positive and less than
-      // INT32_MAX, whereas SYCL can allow for larger work-group sizes (see
-      // -fno-sycl-id-queries-fit-in-int). If any dimension is too large for
-      // NVPTX, don't emit any annotation at all.
-      if (llvm::all_of(Ops, [](std::optional<int64_t> V) {
-            return !V || llvm::isUInt<31>(*V);
-          })) {
-        if (auto X = Ops[0])
-          addNVVMMetadata(F, "reqntidx", *X);
-        if (auto Y = Ops[1])
-          addNVVMMetadata(F, "reqntidy", *Y);
-        if (auto Z = Ops[2])
-          addNVVMMetadata(F, "reqntidz", *Z);
-      }
-    }
-
-    auto attrValue = [&](Expr *E) {
-      const auto *CE = cast<ConstantExpr>(E);
-      std::optional<llvm::APInt> Val = CE->getResultAsAPSInt();
-      return Val->getZExtValue();
-    };
-
-    if (const auto *MWGPCU =
-            FD->getAttr<SYCLIntelMinWorkGroupsPerComputeUnitAttr>()) {
-      if (!HasMaxWorkGroupSize && FD->hasAttr<OpenCLKernelAttr>()) {
-        M.getDiags().Report(D->getLocation(),
-                            diag::warn_launch_bounds_missing_attr)
-            << MWGPCU << 0;
-      } else {
-        // The value is guaranteed to be > 0, pass it to the metadata.
-        addNVVMMetadata(F, "minctasm", attrValue(MWGPCU->getValue()));
-        HasMinWorkGroupPerCU = true;
-      }
-    }
-
-    if (const auto *MWGPMP =
-            FD->getAttr<SYCLIntelMaxWorkGroupsPerMultiprocessorAttr>()) {
-      if ((!HasMaxWorkGroupSize || !HasMinWorkGroupPerCU) &&
-          FD->hasAttr<OpenCLKernelAttr>()) {
-        M.getDiags().Report(D->getLocation(),
-                            diag::warn_launch_bounds_missing_attr)
-            << MWGPMP << 1;
-      } else {
-        // The value is guaranteed to be > 0, pass it to the metadata.
-        addNVVMMetadata(F, "maxclusterrank", attrValue(MWGPMP->getValue()));
-      }
-    }
   }
 
   // Perform special handling in CUDA mode.

clang/lib/Driver/Driver.cpp

@@ -101,8 +101,6 @@
 #include <map>
 #include <memory>
 #include <optional>
-#include <regex>
-#include <sstream>
 #include <set>
 #include <utility>
 #if LLVM_ON_UNIX

clang/lib/Driver/ToolChains/Clang.cpp

@@ -10703,6 +10703,14 @@ static void addArgs(ArgStringList &DstArgs, const llvm::opt::ArgList &Alloc,
   }
 }
 
+static bool supportDynamicLinking(const llvm::opt::ArgList &TCArgs) {
+  if (TCArgs.hasFlag(options::OPT_fsycl_allow_device_dependencies,
+                     options::OPT_fno_sycl_allow_device_dependencies,
+                     false))
+    return true;
+  return false;
+}
+
 static void getNonTripleBasedSYCLPostLinkOpts(const ToolChain &TC,
                                               const JobAction &JA,
                                               const llvm::opt::ArgList &TCArgs,
@@ -10729,6 +10737,9 @@ static void getNonTripleBasedSYCLPostLinkOpts(const ToolChain &TC,
   if (TCArgs.hasFlag(options::OPT_fno_sycl_esimd_force_stateless_mem,
                      options::OPT_fsycl_esimd_force_stateless_mem, false))
     addArgs(PostLinkArgs, TCArgs, {"-lower-esimd-force-stateless-mem=false"});
+
+  if (supportDynamicLinking(TCArgs))
+    addArgs(PostLinkArgs, TCArgs, {"-support-dynamic-linking"});
 }
 
 // Add any sycl-post-link options that rely on a specific Triple in addition
@@ -10776,6 +10787,8 @@ static void getTripleBasedSYCLPostLinkOpts(const ToolChain &TC,
                        options::OPT_fsycl_remove_unused_external_funcs,
                        false) &&
        !isSYCLNativeCPU(TC)) &&
+      // When supporting dynamic linking, non-kernels in a device image can be called
+      !supportDynamicLinking(TCArgs) &&
       !Triple.isNVPTX() && !Triple.isAMDGPU())
     addArgs(PostLinkArgs, TCArgs, {"-emit-only-kernels-as-entry-points"});
 
@@ -11199,6 +11212,14 @@ void LinkerWrapper::ConstructJob(Compilation &C, const JobAction &JA,
     CmdArgs.push_back(Args.MakeArgString(
         Twine("-sycl-device-library-location=") + DeviceLibDir));
 
+    if (C.getDriver().isDumpDeviceCodeEnabled()) {
+      SmallString<128> DumpDir;
+      Arg *A = C.getArgs().getLastArg(options::OPT_fsycl_dump_device_code_EQ);
+      DumpDir = A ? A->getValue() : "";
+      CmdArgs.push_back(
+          Args.MakeArgString(Twine("-sycl-dump-device-code=") + DumpDir));
+    }
+
     auto appendOption = [](SmallString<128> &OptString, StringRef AddOpt) {
       if (!OptString.empty())
         OptString += " ";

clang/lib/Sema/SemaDeclAttr.cpp

@@ -4089,11 +4089,17 @@ bool static check32BitInt(const Expr *E, Sema &S, llvm::APSInt &I,
 
 void Sema::AddSYCLIntelMinWorkGroupsPerComputeUnitAttr(
     Decl *D, const AttributeCommonInfo &CI, Expr *E) {
-  if (Context.getLangOpts().SYCLIsDevice &&
-      !Context.getTargetInfo().getTriple().isNVPTX()) {
-    Diag(E->getBeginLoc(), diag::warn_launch_bounds_is_cuda_specific)
-        << CI << E->getSourceRange();
-    return;
+  if (Context.getLangOpts().SYCLIsDevice) {
+    if (!Context.getTargetInfo().getTriple().isNVPTX()) {
+      Diag(E->getBeginLoc(), diag::warn_launch_bounds_is_cuda_specific)
+          << CI << E->getSourceRange();
+      return;
+    }
+
+    if (!D->hasAttr<SYCLIntelMaxWorkGroupSizeAttr>()) {
+      Diag(CI.getLoc(), diag::warn_launch_bounds_missing_attr) << CI << 0;
+      return;
+    }
   }
   if (!E->isValueDependent()) {
     // Validate that we have an integer constant expression and then store the
@@ -4154,6 +4160,12 @@ void Sema::AddSYCLIntelMaxWorkGroupsPerMultiprocessorAttr(
           << CudaArchToString(SM) << CI << E->getSourceRange();
       return;
     }
+
+    if (!D->hasAttr<SYCLIntelMaxWorkGroupSizeAttr>() ||
+        !D->hasAttr<SYCLIntelMinWorkGroupsPerComputeUnitAttr>()) {
+      Diag(CI.getLoc(), diag::warn_launch_bounds_missing_attr) << CI << 1;
+      return;
+    }
   }
   if (!E->isValueDependent()) {
     // Validate that we have an integer constant expression and then store the

clang/test/CodeGenSYCL/launch_bounds_nvptx.cpp

@@ -48,37 +48,6 @@ int main() {
 // CHECK: define dso_local void @{{.*}}kernel_name2() #0 {{.*}} !min_work_groups_per_cu ![[MWGPC:[0-9]+]] !max_work_groups_per_mp ![[MWGPM:[0-9]+]] !max_work_group_size ![[MWGS:[0-9]+]]
 // CHECK: define dso_local void @{{.*}}kernel_name3() #0 {{.*}} !min_work_groups_per_cu ![[MWGPC_MWGPM:[0-9]+]] !max_work_groups_per_mp ![[MWGPC_MWGPM]] !max_work_group_size ![[MWGS_2:[0-9]+]]
 
-// CHECK: {{.*}}@{{.*}}kernel_name1, !"maxntidx", i32 8}
-// CHECK: {{.*}}@{{.*}}kernel_name1, !"maxntidy", i32 4}
-// CHECK: {{.*}}@{{.*}}kernel_name1, !"maxntidz", i32 2}
-// CHECK: {{.*}}@{{.*}}kernel_name1, !"minctasm", i32 2}
-// CHECK: {{.*}}@{{.*}}kernel_name1, !"maxclusterrank", i32 4}
-// CHECK: {{.*}}@{{.*}}Foo{{.*}}, !"maxntidx", i32 8}
-// CHECK: {{.*}}@{{.*}}Foo{{.*}}, !"maxntidy", i32 4}
-// CHECK: {{.*}}@{{.*}}Foo{{.*}}, !"maxntidz", i32 2}
-// CHECK: {{.*}}@{{.*}}Foo{{.*}}, !"minctasm", i32 2}
-// CHECK: {{.*}}@{{.*}}Foo{{.*}}, !"maxclusterrank", i32 4}
-// CHECK: {{.*}}@{{.*}}kernel_name2, !"maxntidx", i32 8}
-// CHECK: {{.*}}@{{.*}}kernel_name2, !"maxntidy", i32 4}
-// CHECK: {{.*}}@{{.*}}kernel_name2, !"maxntidz", i32 2}
-// CHECK: {{.*}}@{{.*}}kernel_name2, !"minctasm", i32 2}
-// CHECK: {{.*}}@{{.*}}kernel_name2, !"maxclusterrank", i32 4}
-// CHECK: {{.*}}@{{.*}}main{{.*}}, !"maxntidx", i32 8}
-// CHECK: {{.*}}@{{.*}}main{{.*}}, !"maxntidy", i32 4}
-// CHECK: {{.*}}@{{.*}}main{{.*}}, !"maxntidz", i32 2}
-// CHECK: {{.*}}@{{.*}}main{{.*}}, !"minctasm", i32 2}
-// CHECK: {{.*}}@{{.*}}main{{.*}}, !"maxclusterrank", i32 4}
-// CHECK: {{.*}}@{{.*}}kernel_name3, !"maxntidx", i32 8}
-// CHECK: {{.*}}@{{.*}}kernel_name3, !"maxntidy", i32 4}
-// CHECK: {{.*}}@{{.*}}kernel_name3, !"maxntidz", i32 6}
-// CHECK: {{.*}}@{{.*}}kernel_name3, !"minctasm", i32 6}
-// CHECK: {{.*}}@{{.*}}kernel_name3, !"maxclusterrank", i32 6}
-// CHECK: {{.*}}@{{.*}}Functor{{.*}}, !"maxntidx", i32 8}
-// CHECK: {{.*}}@{{.*}}Functor{{.*}}, !"maxntidy", i32 4}
-// CHECK: {{.*}}@{{.*}}Functor{{.*}}, !"maxntidz", i32 6}
-// CHECK: {{.*}}@{{.*}}Functor{{.*}}, !"minctasm", i32 6}
-// CHECK: {{.*}}@{{.*}}Functor{{.*}}, !"maxclusterrank", i32 6}
-
 // CHECK: ![[MWGPC]] = !{i32 2}
 // CHECK: ![[MWGPM]] = !{i32 4}
 // CHECK: ![[MWGS]] = !{i32 8, i32 4, i32 2}