[flang][cuda] Extends matching distance computation #91810

clementval · 2024-05-10T21:07:30Z

Extends the computation of the matching distance in the generic resolution to support options described in the table: https://docs.nvidia.com/hpc-sdk/archive/24.3/compilers/cuda-fortran-prog-guide/index.html#cfref-var-attr-unified-data

Options are added as language features in the SemanticsContext and a flag is added in bbc for testing purpose.

llvmbot · 2024-05-10T21:08:01Z

@llvm/pr-subscribers-flang-semantics

Author: Valentin Clement (バレンタインクレメン) (clementval)

Changes

Extends the computation of the matching distance in the generic resolution to support options described in the table: https://docs.nvidia.com/hpc-sdk/archive/24.3/compilers/cuda-fortran-prog-guide/index.html#cfref-var-attr-unified-data

Options are added as language features in the SemanticsContext and a flag is added in bbc for testing purpose.

Full diff: https://github.com/llvm/llvm-project/pull/91810.diff

8 Files Affected:

(modified) flang/include/flang/Common/Fortran-features.h (+3-1)
(modified) flang/include/flang/Common/Fortran.h (+3-1)
(modified) flang/lib/Common/Fortran.cpp (+26-11)
(modified) flang/lib/Semantics/check-call.cpp (+1-1)
(modified) flang/lib/Semantics/expression.cpp (+39-7)
(added) flang/test/Semantics/cuf14.cuf (+55)
(added) flang/test/Semantics/cuf15.cuf (+55)
(modified) flang/tools/bbc/bbc.cpp (+10)

diff --git a/flang/include/flang/Common/Fortran-features.h b/flang/include/flang/Common/Fortran-features.h
index 07ed7f43c1e73..f930490716fdc 100644
--- a/flang/include/flang/Common/Fortran-features.h
+++ b/flang/include/flang/Common/Fortran-features.h
@@ -49,7 +49,7 @@ ENUM_CLASS(LanguageFeature, BackslashEscapes, OldDebugLines,
     IndistinguishableSpecifics, SubroutineAndFunctionSpecifics,
     EmptySequenceType, NonSequenceCrayPointee, BranchIntoConstruct,
     BadBranchTarget, ConvertedArgument, HollerithPolymorphic, ListDirectedSize,
-    NonBindCInteroperability)
+    NonBindCInteroperability, GpuManaged, GpuUnified)
 
 // Portability and suspicious usage warnings
 ENUM_CLASS(UsageWarning, Portability, PointerToUndefinable,
@@ -81,6 +81,8 @@ class LanguageFeatureControl {
     disable_.set(LanguageFeature::OpenACC);
     disable_.set(LanguageFeature::OpenMP);
     disable_.set(LanguageFeature::CUDA); // !@cuf
+    disable_.set(LanguageFeature::GpuManaged);
+    disable_.set(LanguageFeature::GpuUnified);
     disable_.set(LanguageFeature::ImplicitNoneTypeNever);
     disable_.set(LanguageFeature::ImplicitNoneTypeAlways);
     disable_.set(LanguageFeature::DefaultSave);
diff --git a/flang/include/flang/Common/Fortran.h b/flang/include/flang/Common/Fortran.h
index 3b965fe60c2f0..0701e3e8b64cc 100644
--- a/flang/include/flang/Common/Fortran.h
+++ b/flang/include/flang/Common/Fortran.h
@@ -19,6 +19,7 @@
 #include <string>
 
 namespace Fortran::common {
+class LanguageFeatureControl;
 
 // Fortran has five kinds of intrinsic data types, plus the derived types.
 ENUM_CLASS(TypeCategory, Integer, Real, Complex, Character, Logical, Derived)
@@ -115,7 +116,8 @@ static constexpr IgnoreTKRSet ignoreTKRAll{IgnoreTKR::Type, IgnoreTKR::Kind,
 std::string AsFortran(IgnoreTKRSet);
 
 bool AreCompatibleCUDADataAttrs(std::optional<CUDADataAttr>,
-    std::optional<CUDADataAttr>, IgnoreTKRSet, bool allowUnifiedMatchingRule);
+    std::optional<CUDADataAttr>, IgnoreTKRSet, bool allowUnifiedMatchingRule,
+    const LanguageFeatureControl *features = nullptr);
 
 static constexpr char blankCommonObjectName[] = "__BLNK__";
 
diff --git a/flang/lib/Common/Fortran.cpp b/flang/lib/Common/Fortran.cpp
index 170ce8c225092..83ee68e3a62c0 100644
--- a/flang/lib/Common/Fortran.cpp
+++ b/flang/lib/Common/Fortran.cpp
@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "flang/Common/Fortran.h"
+#include "flang/Common/Fortran-features.h"
 
 namespace Fortran::common {
 
@@ -102,7 +103,13 @@ std::string AsFortran(IgnoreTKRSet tkr) {
 /// dummy argument attribute while `y` represents the actual argument attribute.
 bool AreCompatibleCUDADataAttrs(std::optional<CUDADataAttr> x,
     std::optional<CUDADataAttr> y, IgnoreTKRSet ignoreTKR,
-    bool allowUnifiedMatchingRule) {
+    bool allowUnifiedMatchingRule, const LanguageFeatureControl *features) {
+  bool isGpuManaged = features
+      ? features->IsEnabled(common::LanguageFeature::GpuManaged)
+      : false;
+  bool isGpuUnified = features
+      ? features->IsEnabled(common::LanguageFeature::GpuUnified)
+      : false;
   if (!x && !y) {
     return true;
   } else if (x && y && *x == *y) {
@@ -120,19 +127,27 @@ bool AreCompatibleCUDADataAttrs(std::optional<CUDADataAttr> x,
     return true;
   } else if (allowUnifiedMatchingRule) {
     if (!x) { // Dummy argument has no attribute -> host
-      if (y && (*y == CUDADataAttr::Managed || *y == CUDADataAttr::Unified)) {
+      if ((y && (*y == CUDADataAttr::Managed || *y == CUDADataAttr::Unified)) ||
+          (!y && (isGpuUnified || isGpuManaged))) {
         return true;
       }
     } else {
-      if (*x == CUDADataAttr::Device && y &&
-          (*y == CUDADataAttr::Managed || *y == CUDADataAttr::Unified)) {
-        return true;
-      } else if (*x == CUDADataAttr::Managed && y &&
-          *y == CUDADataAttr::Unified) {
-        return true;
-      } else if (*x == CUDADataAttr::Unified && y &&
-          *y == CUDADataAttr::Managed) {
-        return true;
+      if (*x == CUDADataAttr::Device) {
+        if ((y &&
+                (*y == CUDADataAttr::Managed || *y == CUDADataAttr::Unified)) ||
+            (!y && (isGpuUnified || isGpuManaged))) {
+          return true;
+        }
+      } else if (*x == CUDADataAttr::Managed) {
+        if ((y && *y == CUDADataAttr::Unified) ||
+            (!y && (isGpuUnified || isGpuManaged))) {
+          return true;
+        }
+      } else if (*x == CUDADataAttr::Unified) {
+        if ((y && *y == CUDADataAttr::Managed) ||
+            (!y && (isGpuUnified || isGpuManaged))) {
+          return true;
+        }
       }
     }
     return false;
diff --git a/flang/lib/Semantics/check-call.cpp b/flang/lib/Semantics/check-call.cpp
index 94afcbb68b349..8f51ef5ebeba3 100644
--- a/flang/lib/Semantics/check-call.cpp
+++ b/flang/lib/Semantics/check-call.cpp
@@ -914,7 +914,7 @@ static void CheckExplicitDataArg(const characteristics::DummyDataObject &dummy,
     }
     if (!common::AreCompatibleCUDADataAttrs(dummyDataAttr, actualDataAttr,
             dummy.ignoreTKR,
-            /*allowUnifiedMatchingRule=*/true)) {
+            /*allowUnifiedMatchingRule=*/true, &context.languageFeatures())) {
       auto toStr{[](std::optional<common::CUDADataAttr> x) {
         return x ? "ATTRIBUTES("s +
                 parser::ToUpperCaseLetters(common::EnumToString(*x)) + ")"s
diff --git a/flang/lib/Semantics/expression.cpp b/flang/lib/Semantics/expression.cpp
index c503ea3f0246f..b87370d9b3338 100644
--- a/flang/lib/Semantics/expression.cpp
+++ b/flang/lib/Semantics/expression.cpp
@@ -2501,8 +2501,13 @@ static constexpr int cudaInfMatchingValue{std::numeric_limits<int>::max()};
 
 // Compute the matching distance as described in section 3.2.3 of the CUDA
 // Fortran references.
-static int GetMatchingDistance(const characteristics::DummyArgument &dummy,
+static int GetMatchingDistance(const common::LanguageFeatureControl &features,
+    const characteristics::DummyArgument &dummy,
     const std::optional<ActualArgument> &actual) {
+  bool isGpuManaged = features.IsEnabled(common::LanguageFeature::GpuManaged);
+  bool isGpuUnified = features.IsEnabled(common::LanguageFeature::GpuUnified);
+  // assert((isGpuManaged != isGpuUnified) && "expect only one enabled.");
+
   std::optional<common::CUDADataAttr> actualDataAttr, dummyDataAttr;
   if (actual) {
     if (auto *expr{actual->UnwrapExpr()}) {
@@ -2529,6 +2534,9 @@ static int GetMatchingDistance(const characteristics::DummyArgument &dummy,
 
   if (!dummyDataAttr) {
     if (!actualDataAttr) {
+      if (isGpuUnified || isGpuManaged) {
+        return 3;
+      }
       return 0;
     } else if (*actualDataAttr == common::CUDADataAttr::Device) {
       return cudaInfMatchingValue;
@@ -2538,6 +2546,9 @@ static int GetMatchingDistance(const characteristics::DummyArgument &dummy,
     }
   } else if (*dummyDataAttr == common::CUDADataAttr::Device) {
     if (!actualDataAttr) {
+      if (isGpuUnified || isGpuManaged) {
+        return 2;
+      }
       return cudaInfMatchingValue;
     } else if (*actualDataAttr == common::CUDADataAttr::Device) {
       return 0;
@@ -2546,7 +2557,16 @@ static int GetMatchingDistance(const characteristics::DummyArgument &dummy,
       return 2;
     }
   } else if (*dummyDataAttr == common::CUDADataAttr::Managed) {
-    if (!actualDataAttr || *actualDataAttr == common::CUDADataAttr::Device) {
+    if (!actualDataAttr) {
+      if (isGpuUnified) {
+        return 1;
+      }
+      if (isGpuManaged) {
+        return 0;
+      }
+      return cudaInfMatchingValue;
+    }
+    if (*actualDataAttr == common::CUDADataAttr::Device) {
       return cudaInfMatchingValue;
     } else if (*actualDataAttr == common::CUDADataAttr::Managed) {
       return 0;
@@ -2554,7 +2574,16 @@ static int GetMatchingDistance(const characteristics::DummyArgument &dummy,
       return 1;
     }
   } else if (*dummyDataAttr == common::CUDADataAttr::Unified) {
-    if (!actualDataAttr || *actualDataAttr == common::CUDADataAttr::Device) {
+    if (!actualDataAttr) {
+      if (isGpuUnified) {
+        return 0;
+      }
+      if (isGpuManaged) {
+        return 1;
+      }
+      return cudaInfMatchingValue;
+    }
+    if (*actualDataAttr == common::CUDADataAttr::Device) {
       return cudaInfMatchingValue;
     } else if (*actualDataAttr == common::CUDADataAttr::Managed) {
       return 1;
@@ -2566,6 +2595,7 @@ static int GetMatchingDistance(const characteristics::DummyArgument &dummy,
 }
 
 static int ComputeCudaMatchingDistance(
+    const common::LanguageFeatureControl &features,
     const characteristics::Procedure &procedure,
     const ActualArguments &actuals) {
   const auto &dummies{procedure.dummyArguments};
@@ -2574,7 +2604,7 @@ static int ComputeCudaMatchingDistance(
   for (std::size_t i{0}; i < dummies.size(); ++i) {
     const characteristics::DummyArgument &dummy{dummies[i]};
     const std::optional<ActualArgument> &actual{actuals[i]};
-    int d{GetMatchingDistance(dummy, actual)};
+    int d{GetMatchingDistance(features, dummy, actual)};
     if (d == cudaInfMatchingValue)
       return d;
     distance += d;
@@ -2666,7 +2696,9 @@ std::pair<const Symbol *, bool> ExpressionAnalyzer::ResolveGeneric(
             CheckCompatibleArguments(*procedure, localActuals)) {
           if ((procedure->IsElemental() && elemental) ||
               (!procedure->IsElemental() && nonElemental)) {
-            int d{ComputeCudaMatchingDistance(*procedure, localActuals)};
+            int d{ComputeCudaMatchingDistance(
+                context_.languageFeatures(), *procedure, localActuals)};
+            llvm::errs() << "matching distance: " << d << "\n";
             if (d != crtMatchingDistance) {
               if (d > crtMatchingDistance) {
                 continue;
@@ -2688,8 +2720,8 @@ std::pair<const Symbol *, bool> ExpressionAnalyzer::ResolveGeneric(
           } else {
             elemental = &specific;
           }
-          crtMatchingDistance =
-              ComputeCudaMatchingDistance(*procedure, localActuals);
+          crtMatchingDistance = ComputeCudaMatchingDistance(
+              context_.languageFeatures(), *procedure, localActuals);
         }
       }
     }
diff --git a/flang/test/Semantics/cuf14.cuf b/flang/test/Semantics/cuf14.cuf
new file mode 100644
index 0000000000000..29c9ecf90677f
--- /dev/null
+++ b/flang/test/Semantics/cuf14.cuf
@@ -0,0 +1,55 @@
+! RUN: bbc -emit-hlfir -fcuda -gpu=unified %s -o - | FileCheck %s
+
+module matching
+  interface host_and_device
+    module procedure sub_host
+    module procedure sub_device
+  end interface
+
+  interface all
+    module procedure sub_host
+    module procedure sub_device
+    module procedure sub_managed
+    module procedure sub_unified
+  end interface
+
+  interface all_without_unified
+    module procedure sub_host
+    module procedure sub_device
+    module procedure sub_managed
+  end interface
+
+contains
+  subroutine sub_host(a)
+    integer :: a(:)
+  end
+
+  subroutine sub_device(a)
+    integer, device :: a(:)
+  end
+
+  subroutine sub_managed(a)
+    integer, managed :: a(:)
+  end
+
+  subroutine sub_unified(a)
+    integer, unified :: a(:)
+  end
+end module
+
+program m
+  use matching
+
+  integer, allocatable :: actual_host(:)
+
+  allocate(actual_host(10))
+
+  call host_and_device(actual_host)     ! Should resolve to sub_device
+  call all(actual_host)                 ! Should resolved to unified
+  call all_without_unified(actual_host) ! Should resolved to managed
+end
+
+! CHECK: fir.call @_QMmatchingPsub_device
+! CHECK: fir.call @_QMmatchingPsub_unified
+! CHECK: fir.call @_QMmatchingPsub_managed
+
diff --git a/flang/test/Semantics/cuf15.cuf b/flang/test/Semantics/cuf15.cuf
new file mode 100644
index 0000000000000..030dd6ff8ffe8
--- /dev/null
+++ b/flang/test/Semantics/cuf15.cuf
@@ -0,0 +1,55 @@
+! RUN: bbc -emit-hlfir -fcuda -gpu=managed %s -o - | FileCheck %s
+
+module matching
+  interface host_and_device
+    module procedure sub_host
+    module procedure sub_device
+  end interface
+
+  interface all
+    module procedure sub_host
+    module procedure sub_device
+    module procedure sub_managed
+    module procedure sub_unified
+  end interface
+
+  interface all_without_managed
+    module procedure sub_host
+    module procedure sub_device
+    module procedure sub_unified
+  end interface
+
+contains
+  subroutine sub_host(a)
+    integer :: a(:)
+  end
+
+  subroutine sub_device(a)
+    integer, device :: a(:)
+  end
+
+  subroutine sub_managed(a)
+    integer, managed :: a(:)
+  end
+
+  subroutine sub_unified(a)
+    integer, unified :: a(:)
+  end
+end module
+
+program m
+  use matching
+
+  integer, allocatable :: actual_host(:)
+
+  allocate(actual_host(10))
+
+  call host_and_device(actual_host)     ! Should resolve to sub_device
+  call all(actual_host)                 ! Should resolved to unified
+  call all_without_managed(actual_host) ! Should resolved to managed
+end
+
+! CHECK: fir.call @_QMmatchingPsub_device
+! CHECK: fir.call @_QMmatchingPsub_managed
+! CHECK: fir.call @_QMmatchingPsub_unified
+
diff --git a/flang/tools/bbc/bbc.cpp b/flang/tools/bbc/bbc.cpp
index ee2ff8562e9ff..085b988c6af5f 100644
--- a/flang/tools/bbc/bbc.cpp
+++ b/flang/tools/bbc/bbc.cpp
@@ -204,6 +204,10 @@ static llvm::cl::opt<bool> enableCUDA("fcuda",
                                       llvm::cl::desc("enable CUDA Fortran"),
                                       llvm::cl::init(false));
 
+static llvm::cl::opt<std::string>
+    enableGPUMode("gpu", llvm::cl::desc("Enable GPU Mode managed|unified"),
+                  llvm::cl::init(""));
+
 static llvm::cl::opt<bool> fixedForm("ffixed-form",
                                      llvm::cl::desc("enable fixed form"),
                                      llvm::cl::init(false));
@@ -495,6 +499,12 @@ int main(int argc, char **argv) {
     options.features.Enable(Fortran::common::LanguageFeature::CUDA);
   }
 
+  if (enableGPUMode == "managed") {
+    options.features.Enable(Fortran::common::LanguageFeature::GpuManaged);
+  } else if (enableGPUMode == "unified") {
+    options.features.Enable(Fortran::common::LanguageFeature::GpuUnified);
+  }
+
   if (fixedForm) {
     options.isFixedForm = fixedForm;
   }

flang/include/flang/Common/Fortran-features.h

flang/lib/Common/Fortran.cpp

flang/lib/Semantics/expression.cpp

flang/lib/Common/Fortran.cpp

[flang][cuda] Extends matching distance computation

c916853

clementval requested review from klausler and wangzpgi May 10, 2024 21:07

llvmbot added flang Flang issues not falling into any other category flang:semantics labels May 10, 2024

klausler reviewed May 10, 2024

View reviewed changes

Address comments

6cf2d7d

klausler reviewed May 10, 2024

View reviewed changes

flang/lib/Common/Fortran.cpp Outdated Show resolved Hide resolved

klausler approved these changes May 10, 2024

View reviewed changes

braces

26d79d3

clementval merged commit e8eb52d into llvm:main May 13, 2024
4 checks passed

clementval deleted the cuda_managed_unified branch May 13, 2024 15:34

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[flang][cuda] Extends matching distance computation #91810

[flang][cuda] Extends matching distance computation #91810

Uh oh!

clementval commented May 10, 2024

Uh oh!

llvmbot commented May 10, 2024

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

[flang][cuda] Extends matching distance computation #91810

[flang][cuda] Extends matching distance computation #91810

Uh oh!

Conversation

clementval commented May 10, 2024

Uh oh!

llvmbot commented May 10, 2024

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!