[SYCL] [Driver] Add offload-arch support for SYCL offloading (#15624)

Implement `--offload-arch` option to enable SYCL offloading to `Intel
CPUs`, `Intel GPUs`, `NVidia` and `AMD GPUs`.

`--offload-arch` implementation is currently supported in the new driver
model ( i.e. with `--offload-new-driver` option)

**Example usage:**
```
clang++   --offload-new-driver  -fsycl  --offload-arch=bdw         // Offload SYCL code to Intel GPU
clang++   --offload-new-driver  -fsycl  --offload-arch=broadwell   // Offload SYCL code to Intel CPU
clang++   --offload-new-driver  -fsycl  --offload-arch=sm_80       // Offload SYCL code to NVidia GPU
clang++   --offload-new-driver  -fsycl  --offload-arch=gfx700      // Offload SYCL code to AMD GPU
```

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/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.

clang/lib/Driver/ToolChains/SYCL.cpp

@@ -27,6 +27,125 @@ using namespace clang::driver::tools;
 using namespace clang;
 using namespace llvm::opt;
 
+// Struct that relates an AOT target value with
+// Intel CPUs and Intel GPUs.
+struct StringToOffloadArchSYCLMap {
+  const char *ArchName;
+  SYCLSupportedIntelArchs IntelArch;
+};
+
+// Mapping of supported SYCL offloading architectures.
+static const StringToOffloadArchSYCLMap StringToArchNamesMap[] = {
+    // Intel CPU mapping.
+    {"skylake-avx512", SYCLSupportedIntelArchs::SKYLAKEAVX512},
+    {"core-avx2", SYCLSupportedIntelArchs::COREAVX2},
+    {"corei7-avx", SYCLSupportedIntelArchs::COREI7AVX},
+    {"corei7", SYCLSupportedIntelArchs::COREI7},
+    {"westmere", SYCLSupportedIntelArchs::WESTMERE},
+    {"sandybridge", SYCLSupportedIntelArchs::SANDYBRIDGE},
+    {"ivybridge", SYCLSupportedIntelArchs::IVYBRIDGE},
+    {"broadwell", SYCLSupportedIntelArchs::BROADWELL},
+    {"coffeelake", SYCLSupportedIntelArchs::COFFEELAKE},
+    {"alderlake", SYCLSupportedIntelArchs::ALDERLAKE},
+    {"skylake", SYCLSupportedIntelArchs::SKYLAKE},
+    {"skx", SYCLSupportedIntelArchs::SKX},
+    {"cascadelake", SYCLSupportedIntelArchs::CASCADELAKE},
+    {"icelake-client", SYCLSupportedIntelArchs::ICELAKECLIENT},
+    {"icelake-server", SYCLSupportedIntelArchs::ICELAKESERVER},
+    {"sapphirerapids", SYCLSupportedIntelArchs::SAPPHIRERAPIDS},
+    {"graniterapids", SYCLSupportedIntelArchs::GRANITERAPIDS},
+    // Intel GPU mapping.
+    {"bdw", SYCLSupportedIntelArchs::BDW},
+    {"skl", SYCLSupportedIntelArchs::SKL},
+    {"kbl", SYCLSupportedIntelArchs::KBL},
+    {"cfl", SYCLSupportedIntelArchs::CFL},
+    {"apl", SYCLSupportedIntelArchs::APL},
+    {"bxt", SYCLSupportedIntelArchs::BXT},
+    {"glk", SYCLSupportedIntelArchs::GLK},
+    {"whl", SYCLSupportedIntelArchs::WHL},
+    {"aml", SYCLSupportedIntelArchs::AML},
+    {"cml", SYCLSupportedIntelArchs::CML},
+    {"icllp", SYCLSupportedIntelArchs::ICLLP},
+    {"icl", SYCLSupportedIntelArchs::ICL},
+    {"ehl", SYCLSupportedIntelArchs::EHL},
+    {"jsl", SYCLSupportedIntelArchs::JSL},
+    {"tgllp", SYCLSupportedIntelArchs::TGLLP},
+    {"tgl", SYCLSupportedIntelArchs::TGL},
+    {"rkl", SYCLSupportedIntelArchs::RKL},
+    {"adl_s", SYCLSupportedIntelArchs::ADL_S},
+    {"rpl_s", SYCLSupportedIntelArchs::RPL_S},
+    {"adl_p", SYCLSupportedIntelArchs::ADL_P},
+    {"adl_n", SYCLSupportedIntelArchs::ADL_N},
+    {"dg1", SYCLSupportedIntelArchs::DG1},
+    {"acm_g10", SYCLSupportedIntelArchs::ACM_G10},
+    {"dg2_g10", SYCLSupportedIntelArchs::DG2_G10},
+    {"acm_g11", SYCLSupportedIntelArchs::ACM_G11},
+    {"dg2_g10", SYCLSupportedIntelArchs::DG2_G10},
+    {"dg2_g11", SYCLSupportedIntelArchs::DG2_G11},
+    {"acm_g12", SYCLSupportedIntelArchs::ACM_G12},
+    {"dg2_g12", SYCLSupportedIntelArchs::DG2_G12},
+    {"pvc", SYCLSupportedIntelArchs::PVC},
+    {"pvc_vg", SYCLSupportedIntelArchs::PVC_VG},
+    {"mtl_u", SYCLSupportedIntelArchs::MTL_U},
+    {"mtl_s", SYCLSupportedIntelArchs::MTL_S},
+    {"arl_u", SYCLSupportedIntelArchs::ARL_U},
+    {"arl_s", SYCLSupportedIntelArchs::ARL_S},
+    {"mtl_h", SYCLSupportedIntelArchs::MTL_H},
+    {"arl_h", SYCLSupportedIntelArchs::ARL_H},
+    {"bmg_g21", SYCLSupportedIntelArchs::BMG_G21},
+    {"lnl_m", SYCLSupportedIntelArchs::LNL_M}};
+
+// Check if the user provided value for --offload-arch is a valid
+// SYCL supported Intel AOT target.
+SYCLSupportedIntelArchs
+clang::driver::StringToOffloadArchSYCL(llvm::StringRef ArchNameAsString) {
+  auto result = std::find_if(
+      std::begin(StringToArchNamesMap), std::end(StringToArchNamesMap),
+      [ArchNameAsString](const StringToOffloadArchSYCLMap &map) {
+        return ArchNameAsString == map.ArchName;
+      });
+  if (result == std::end(StringToArchNamesMap))
+    return SYCLSupportedIntelArchs::UNKNOWN;
+  return result->IntelArch;
+}
+
+// This is a mapping between the user provided --offload-arch value for Intel
+// GPU targets and the spir64_gen device name accepted by OCLOC (the Intel GPU
+// AOT compiler).
+StringRef clang::driver::mapIntelGPUArchName(StringRef ArchName) {
+  StringRef Arch;
+  Arch = llvm::StringSwitch<StringRef>(ArchName)
+             .Case("bdw", "bdw")
+             .Case("skl", "skl")
+             .Case("kbl", "kbl")
+             .Case("cfl", "cfl")
+             .Cases("apl", "bxt", "apl")
+             .Case("glk", "glk")
+             .Case("whl", "whl")
+             .Case("aml", "aml")
+             .Case("cml", "cml")
+             .Cases("icllp", "icl", "icllp")
+             .Cases("ehl", "jsl", "ehl")
+             .Cases("tgllp", "tgl", "tgllp")
+             .Case("rkl", "rkl")
+             .Cases("adl_s", "rpl_s", "adl_s")
+             .Case("adl_p", "adl_p")
+             .Case("adl_n", "adl_n")
+             .Case("dg1", "dg1")
+             .Cases("acm_g10", "dg2_g10", "acm_g10")
+             .Cases("acm_g11", "dg2_g11", "acm_g11")
+             .Cases("acm_g12", "dg2_g12", "acm_g12")
+             .Case("pvc", "pvc")
+             .Case("pvc_vg", "pvc_vg")
+             .Cases("mtl_u", "mtl_s", "arl_u", "arl_s", "mtl_u")
+             .Case("mtl_h", "mtl_h")
+             .Case("arl_h", "arl_h")
+             .Case("bmg_g21", "bmg_g21")
+             .Case("lnl_m", "lnl_m")
+             .Default("");
+  return Arch;
+}
+
 SYCLInstallationDetector::SYCLInstallationDetector(const Driver &D)
     : D(D), InstallationCandidates() {
   InstallationCandidates.emplace_back(D.Dir + "/..");

clang/lib/Driver/ToolChains/SYCL.h

@@ -9,13 +9,122 @@
 #ifndef LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_SYCL_H
 #define LLVM_CLANG_LIB_DRIVER_TOOLCHAINS_SYCL_H
 
+#include "clang/Basic/Cuda.h"
 #include "clang/Driver/Options.h"
 #include "clang/Driver/Tool.h"
 #include "clang/Driver/ToolChain.h"
 
 namespace clang {
 namespace driver {
 
+// List of architectures (Intel CPUs and Intel GPUs)
+// that support SYCL offloading.
+enum class SYCLSupportedIntelArchs {
+  // Intel CPUs
+  UNKNOWN,
+  SKYLAKEAVX512,
+  COREAVX2,
+  COREI7AVX,
+  COREI7,
+  WESTMERE,
+  SANDYBRIDGE,
+  IVYBRIDGE,
+  BROADWELL,
+  COFFEELAKE,
+  ALDERLAKE,
+  SKYLAKE,
+  SKX,
+  CASCADELAKE,
+  ICELAKECLIENT,
+  ICELAKESERVER,
+  SAPPHIRERAPIDS,
+  GRANITERAPIDS,
+  // Intel GPUs
+  BDW,
+  SKL,
+  KBL,
+  CFL,
+  APL,
+  BXT,
+  GLK,
+  WHL,
+  AML,
+  CML,
+  ICLLP,
+  ICL,
+  EHL,
+  JSL,
+  TGLLP,
+  TGL,
+  RKL,
+  ADL_S,
+  RPL_S,
+  ADL_P,
+  ADL_N,
+  DG1,
+  ACM_G10,
+  DG2_G10,
+  ACM_G11,
+  DG2_G11,
+  ACM_G12,
+  DG2_G12,
+  PVC,
+  PVC_VG,
+  MTL_U,
+  MTL_S,
+  ARL_U,
+  ARL_S,
+  MTL_H,
+  ARL_H,
+  BMG_G21,
+  LNL_M,
+};
+
+// Check if the given Arch value is a Generic AMD GPU.
+// Currently GFX*_GENERIC AMD GPUs do not support SYCL offloading.
+// This list is used to filter out GFX*_GENERIC AMD GPUs in
+// `IsSYCLSupportedAMDGPUArch`.
+static inline bool IsAMDGenericGPUArch(OffloadArch Arch) {
+  return Arch == OffloadArch::GFX9_GENERIC ||
+         Arch == OffloadArch::GFX10_1_GENERIC ||
+         Arch == OffloadArch::GFX10_3_GENERIC ||
+         Arch == OffloadArch::GFX11_GENERIC ||
+         Arch == OffloadArch::GFX12_GENERIC;
+}
+
+// Check if the given Arch value is a valid SYCL supported AMD GPU.
+static inline bool IsSYCLSupportedAMDGPUArch(OffloadArch Arch) {
+  return Arch >= OffloadArch::GFX700 && Arch < OffloadArch::AMDGCNSPIRV &&
+         !IsAMDGenericGPUArch(Arch);
+}
+
+// Check if the given Arch value is a valid SYCL supported NVidia GPU.
+static inline bool IsSYCLSupportedNVidiaGPUArch(OffloadArch Arch) {
+  return Arch >= OffloadArch::SM_50 && Arch <= OffloadArch::SM_90a;
+}
+
+// Check if the given Arch value is a valid SYCL supported Intel CPU.
+static inline bool IsSYCLSupportedIntelCPUArch(SYCLSupportedIntelArchs Arch) {
+  return Arch >= SYCLSupportedIntelArchs::SKYLAKEAVX512 &&
+         Arch <= SYCLSupportedIntelArchs::GRANITERAPIDS;
+}
+
+// Check if the given Arch value is a valid SYCL supported Intel GPU.
+static inline bool IsSYCLSupportedIntelGPUArch(SYCLSupportedIntelArchs Arch) {
+  return Arch >= SYCLSupportedIntelArchs::BDW &&
+         Arch <= SYCLSupportedIntelArchs::LNL_M;
+}
+
+// Check if the user provided value for --offload-arch is a valid
+// SYCL supported Intel AOT target.
+SYCLSupportedIntelArchs
+StringToOffloadArchSYCL(llvm::StringRef ArchNameAsString);
+
+// This is a mapping between the user provided --offload-arch value for Intel
+// GPU targets and the spir64_gen device name accepted by OCLOC (the Intel GPU
+// AOT compiler).
+StringRef mapIntelGPUArchName(StringRef ArchName);
+
 class SYCLInstallationDetector {
 public:
   SYCLInstallationDetector(const Driver &D);