[SYCL] [Driver] Add offload-arch support for SYCL offloading

clang/include/clang/Basic/DiagnosticDriverKinds.td

@@ -398,6 +398,8 @@ 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 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

@@ -1185,12 +1185,14 @@ 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
@@ -1290,6 +1292,101 @@ 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.
+  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(StringToOffloadArchSYCL(
+                             getProcessorFromTargetID(*NVPTXTriple, Arch)))) {
+        DerivedArchs[NVPTXTriple->getTriple()].insert(Arch);
+      } else if (AMDTriple &&
+                 IsSYCLSupportedAMDGPUArch(StringToOffloadArchSYCL(
+                     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;
+        IntelGPUArch = mapIntelGPUArchName(Arch).data();
+        DerivedArchs[MakeSYCLDeviceTriple("spir64_gen").getTriple()].insert(
+            IntelGPUArch);
+      } else {
+        Diag(clang::diag::err_drv_invalid_sycl_target) << Arch;
+        return;
+      }
+    }
+    // If the set is empty then we failed to find a native architecture.
+    if (Archs.empty()) {
+      Diag(clang::diag::err_drv_invalid_sycl_target) << "native";
+      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 +27,180 @@ using namespace clang::driver::tools;
 using namespace clang;
 using namespace llvm::opt;
 
+// Struct that relates an AOT target value with
+// Intel CPUs, Intel GPUs, AMD and NVidia GPUs.
+struct StringToOffloadArchSYCLMap {
+  const char *ArchName;
+  SYCLSupportedOffloadArchs IntelArch;
+};
+
+// Mapping of supported SYCL offloading architectures.
+static const StringToOffloadArchSYCLMap StringToArchNamesMap[] = {
+    // Intel CPU mapping.
+    {"skylake-avx512", SYCLSupportedOffloadArchs::SKYLAKEAVX512},
+    {"core-avx2", SYCLSupportedOffloadArchs::COREAVX2},
+    {"corei7-avx", SYCLSupportedOffloadArchs::COREI7AVX},
+    {"corei7", SYCLSupportedOffloadArchs::COREI7},
+    {"westmere", SYCLSupportedOffloadArchs::WESTMERE},
+    {"sandybridge", SYCLSupportedOffloadArchs::SANDYBRIDGE},
+    {"ivybridge", SYCLSupportedOffloadArchs::IVYBRIDGE},
+    {"broadwell", SYCLSupportedOffloadArchs::BROADWELL},
+    {"coffeelake", SYCLSupportedOffloadArchs::COFFEELAKE},
+    {"alderlake", SYCLSupportedOffloadArchs::ALDERLAKE},
+    {"skylake", SYCLSupportedOffloadArchs::SKYLAKE},
+    {"skx", SYCLSupportedOffloadArchs::SKX},
+    {"cascadelake", SYCLSupportedOffloadArchs::CASCADELAKE},
+    {"icelake-client", SYCLSupportedOffloadArchs::ICELAKECLIENT},
+    {"icelake-server", SYCLSupportedOffloadArchs::ICELAKESERVER},
+    {"sapphirerapids", SYCLSupportedOffloadArchs::SAPPHIRERAPIDS},
+    {"graniterapids", SYCLSupportedOffloadArchs::GRANITERAPIDS},
+    // Intel GPU mapping.
+    {"bdw", SYCLSupportedOffloadArchs::BDW},
+    {"skl", SYCLSupportedOffloadArchs::SKL},
+    {"kbl", SYCLSupportedOffloadArchs::KBL},
+    {"cfl", SYCLSupportedOffloadArchs::CFL},
+    {"apl", SYCLSupportedOffloadArchs::APL},
+    {"bxt", SYCLSupportedOffloadArchs::BXT},
+    {"glk", SYCLSupportedOffloadArchs::GLK},
+    {"whl", SYCLSupportedOffloadArchs::WHL},
+    {"aml", SYCLSupportedOffloadArchs::AML},
+    {"cml", SYCLSupportedOffloadArchs::CML},
+    {"icllp", SYCLSupportedOffloadArchs::ICLLP},
+    {"icl", SYCLSupportedOffloadArchs::ICL},
+    {"ehl", SYCLSupportedOffloadArchs::EHL},
+    {"jsl", SYCLSupportedOffloadArchs::JSL},
+    {"tgllp", SYCLSupportedOffloadArchs::TGLLP},
+    {"tgl", SYCLSupportedOffloadArchs::TGL},
+    {"rkl", SYCLSupportedOffloadArchs::RKL},
+    {"adl_s", SYCLSupportedOffloadArchs::ADL_S},
+    {"rpl_s", SYCLSupportedOffloadArchs::RPL_S},
+    {"adl_p", SYCLSupportedOffloadArchs::ADL_P},
+    {"adl_n", SYCLSupportedOffloadArchs::ADL_N},
+    {"dg1", SYCLSupportedOffloadArchs::DG1},
+    {"acm_g10", SYCLSupportedOffloadArchs::ACM_G10},
+    {"dg2_g10", SYCLSupportedOffloadArchs::DG2_G10},
+    {"acm_g11", SYCLSupportedOffloadArchs::ACM_G11},
+    {"dg2_g10", SYCLSupportedOffloadArchs::DG2_G10},
+    {"dg2_g11", SYCLSupportedOffloadArchs::DG2_G11},
+    {"acm_g12", SYCLSupportedOffloadArchs::ACM_G12},
+    {"dg2_g12", SYCLSupportedOffloadArchs::DG2_G12},
+    {"pvc", SYCLSupportedOffloadArchs::PVC},
+    {"pvc_vg", SYCLSupportedOffloadArchs::PVC_VG},
+    {"mtl_u", SYCLSupportedOffloadArchs::MTL_U},
+    {"mtl_s", SYCLSupportedOffloadArchs::MTL_S},
+    {"arl_u", SYCLSupportedOffloadArchs::ARL_U},
+    {"arl_s", SYCLSupportedOffloadArchs::ARL_S},
+    {"mtl_h", SYCLSupportedOffloadArchs::MTL_H},
+    {"arl_h", SYCLSupportedOffloadArchs::ARL_H},
+    {"bmg_g21", SYCLSupportedOffloadArchs::BMG_G21},
+    {"lnl_m", SYCLSupportedOffloadArchs::LNL_M},
+    // AMD GPU Mapping
+    {"gfx700", SYCLSupportedOffloadArchs::GFX700},
+    {"gfx701", SYCLSupportedOffloadArchs::GFX701},
+    {"gfx702", SYCLSupportedOffloadArchs::GFX702},
+    {"gfx801", SYCLSupportedOffloadArchs::GFX801},
+    {"gfx802", SYCLSupportedOffloadArchs::GFX802},
+    {"gfx803", SYCLSupportedOffloadArchs::GFX803},
+    {"gfx805", SYCLSupportedOffloadArchs::GFX805},
+    {"gfx810", SYCLSupportedOffloadArchs::GFX810},
+    {"gfx900", SYCLSupportedOffloadArchs::GFX900},
+    {"gfx902", SYCLSupportedOffloadArchs::GFX902},
+    {"gfx904", SYCLSupportedOffloadArchs::GFX904},
+    {"gfx906", SYCLSupportedOffloadArchs::GFX906},
+    {"gfx908", SYCLSupportedOffloadArchs::GFX908},
+    {"gfx909", SYCLSupportedOffloadArchs::GFX909},
+    {"gfx90a", SYCLSupportedOffloadArchs::GFX90A},
+    {"gfx90c", SYCLSupportedOffloadArchs::GFX90C},
+    {"gfx940", SYCLSupportedOffloadArchs::GFX940},
+    {"gfx941", SYCLSupportedOffloadArchs::GFX941},
+    {"gfx942", SYCLSupportedOffloadArchs::GFX942},
+    {"gfx1010", SYCLSupportedOffloadArchs::GFX1010},
+    {"gfx1011", SYCLSupportedOffloadArchs::GFX1011},
+    {"gfx1012", SYCLSupportedOffloadArchs::GFX1012},
+    {"gfx1013", SYCLSupportedOffloadArchs::GFX1013},
+    {"gfx1030", SYCLSupportedOffloadArchs::GFX1030},
+    {"gfx1031", SYCLSupportedOffloadArchs::GFX1031},
+    {"gfx1032", SYCLSupportedOffloadArchs::GFX1032},
+    {"gfx1033", SYCLSupportedOffloadArchs::GFX1033},
+    {"gfx1034", SYCLSupportedOffloadArchs::GFX1034},
+    {"gfx1035", SYCLSupportedOffloadArchs::GFX1035},
+    {"gfx1036", SYCLSupportedOffloadArchs::GFX1036},
+    {"gfx1100", SYCLSupportedOffloadArchs::GFX1100},
+    {"gfx1101", SYCLSupportedOffloadArchs::GFX1101},
+    {"gfx1102", SYCLSupportedOffloadArchs::GFX1102},
+    {"gfx1103", SYCLSupportedOffloadArchs::GFX1103},
+    {"gfx1150", SYCLSupportedOffloadArchs::GFX1150},
+    {"gfx1151", SYCLSupportedOffloadArchs::GFX1151},
+    {"gfx1200", SYCLSupportedOffloadArchs::GFX1200},
+    {"gfx1201", SYCLSupportedOffloadArchs::GFX1201},
+    // NVidia GPU Mapping.
+    {"sm_50", SYCLSupportedOffloadArchs::SM_50},
+    {"sm_52", SYCLSupportedOffloadArchs::SM_52},
+    {"sm_53", SYCLSupportedOffloadArchs::SM_53},
+    {"sm_60", SYCLSupportedOffloadArchs::SM_60},
+    {"sm_61", SYCLSupportedOffloadArchs::SM_61},
+    {"sm_62", SYCLSupportedOffloadArchs::SM_62},
+    {"sm_70", SYCLSupportedOffloadArchs::SM_70},
+    {"sm_72", SYCLSupportedOffloadArchs::SM_72},
+    {"sm_75", SYCLSupportedOffloadArchs::SM_75},
+    {"sm_80", SYCLSupportedOffloadArchs::SM_80},
+    {"sm_86", SYCLSupportedOffloadArchs::SM_86},
+    {"sm_87", SYCLSupportedOffloadArchs::SM_87},
+    {"sm_89", SYCLSupportedOffloadArchs::SM_89},
+    {"sm_90", SYCLSupportedOffloadArchs::SM_90},
+    {"sm_90a", SYCLSupportedOffloadArchs::SM_90A}};
+
+// Check if the user provided value for --offload-arch is a valid
+// SYCL supported AOT target.
+SYCLSupportedOffloadArchs
+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 SYCLSupportedOffloadArchs::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 + "/..");