Merge topic 'cuda_architectures'

21131ca60c CUDA: Add CudaOnly.CompileFlags test
f0931b0790 CUDA: Convert tests to use CUDA_ARCHITECTURES
e98588aaba CUDA: Add CUDA_ARCHITECTURES target property

Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Patrick Stotko <stotko@cs.uni-bonn.de>
Merge-request: !4568

6 files changed, 103 insertions, 1 deletions

diff --git a/Source/cmCoreTryCompile.cxx b/Source/cmCoreTryCompile.cxx
index 9175dac..dc2df14 100644
--- a/Source/cmCoreTryCompile.cxx
+++ b/Source/cmCoreTryCompile.cxx
@@ -40,6 +40,8 @@ static std::string const kCMAKE_CXX_LINK_NO_PIE_SUPPORTED =
   "CMAKE_CXX_LINK_NO_PIE_SUPPORTED";
 static std::string const kCMAKE_CXX_LINK_PIE_SUPPORTED =
   "CMAKE_CXX_LINK_PIE_SUPPORTED";
+static std::string const kCMAKE_CUDA_ARCHITECTURES =
+  "CMAKE_CUDA_ARCHITECTURES";
 static std::string const kCMAKE_CUDA_COMPILER_TARGET =
   "CMAKE_CUDA_COMPILER_TARGET";
 static std::string const kCMAKE_ENABLE_EXPORTS = "CMAKE_ENABLE_EXPORTS";
@@ -713,6 +715,7 @@ int cmCoreTryCompile::TryCompileCode(std::vector<std::string> const& argv,
       vars.insert(kCMAKE_C_COMPILER_TARGET);
       vars.insert(kCMAKE_CXX_COMPILER_EXTERNAL_TOOLCHAIN);
       vars.insert(kCMAKE_CXX_COMPILER_TARGET);
+      vars.insert(kCMAKE_CUDA_ARCHITECTURES);
       vars.insert(kCMAKE_CUDA_COMPILER_TARGET);
       vars.insert(kCMAKE_ENABLE_EXPORTS);
       vars.insert(kCMAKE_LINK_SEARCH_END_STATIC);
diff --git a/Source/cmGeneratorTarget.cxx b/Source/cmGeneratorTarget.cxx
index 55157b4..b424b90 100644
--- a/Source/cmGeneratorTarget.cxx
+++ b/Source/cmGeneratorTarget.cxx
@@ -3085,6 +3085,95 @@ void cmGeneratorTarget::GetAppleArchs(const std::string& config,
   }
 }
 
+void cmGeneratorTarget::AddCUDAArchitectureFlags(std::string& flags) const
+{
+  struct CudaArchitecture
+  {
+    std::string name;
+    bool real{ true };
+    bool virtual_{ true };
+  };
+  std::vector<CudaArchitecture> architectures;
+
+  {
+    std::vector<std::string> options;
+    cmExpandList(this->GetSafeProperty("CUDA_ARCHITECTURES"), options);
+
+    if (options.empty()) {
+      switch (this->GetPolicyStatusCMP0104()) {
+        case cmPolicies::WARN:
+          if (!this->LocalGenerator->GetCMakeInstance()->GetIsInTryCompile()) {
+            this->Makefile->IssueMessage(
+              MessageType::AUTHOR_WARNING,
+              cmPolicies::GetPolicyWarning(cmPolicies::CMP0104) +
+                "\nCUDA_ARCHITECTURES is empty for target \"" +
+                this->GetName() + "\".");
+          }
+          CM_FALLTHROUGH;
+        case cmPolicies::OLD:
+          break;
+        default:
+          this->Makefile->IssueMessage(
+            MessageType::FATAL_ERROR,
+            "CUDA_ARCHITECTURES is empty for target \"" + this->GetName() +
+              "\".");
+      }
+    }
+
+    for (std::string& option : options) {
+      CudaArchitecture architecture;
+
+      // Architecture name is up to the first specifier.
+      std::size_t pos = option.find_first_of('-');
+      architecture.name = option.substr(0, pos);
+
+      if (pos != std::string::npos) {
+        cm::string_view specifier{ option.c_str() + pos + 1,
+                                   option.length() - pos - 1 };
+
+        if (specifier == "real") {
+          architecture.real = true;
+          architecture.virtual_ = false;
+        } else if (specifier == "virtual") {
+          architecture.real = false;
+          architecture.virtual_ = true;
+        } else {
+          this->Makefile->IssueMessage(
+            MessageType::FATAL_ERROR,
+            "Uknown CUDA architecture specifier \"" + std::string(specifier) +
+              "\".");
+        }
+      }
+
+      architectures.emplace_back(architecture);
+    }
+  }
+
+  std::string const& compiler =
+    this->Makefile->GetSafeDefinition("CMAKE_CUDA_COMPILER_ID");
+
+  if (compiler == "NVIDIA") {
+    for (CudaArchitecture& architecture : architectures) {
+      flags +=
+        " --generate-code=arch=compute_" + architecture.name + ",code=[";
+
+      if (architecture.virtual_) {
+        flags += "compute_" + architecture.name;
+
+        if (architecture.real) {
+          flags += ",";
+        }
+      }
+
+      if (architecture.real) {
+        flags += "sm_" + architecture.name;
+      }
+
+      flags += "]";
+    }
+  }
+}
+
 //----------------------------------------------------------------------------
 std::string cmGeneratorTarget::GetFeatureSpecificLinkRuleVariable(
   std::string const& var, std::string const& lang,
diff --git a/Source/cmGeneratorTarget.h b/Source/cmGeneratorTarget.h
index 9136928..7cd253d 100644
--- a/Source/cmGeneratorTarget.h
+++ b/Source/cmGeneratorTarget.h
@@ -421,6 +421,8 @@ public:
   void GetAppleArchs(const std::string& config,
                      std::vector<std::string>& archVec) const;
 
+  void AddCUDAArchitectureFlags(std::string& flags) const;
+
   std::string GetFeatureSpecificLinkRuleVariable(
     std::string const& var, std::string const& lang,
     std::string const& config) const;
diff --git a/Source/cmLocalGenerator.cxx b/Source/cmLocalGenerator.cxx
index a55e094..ca84f1e 100644
--- a/Source/cmLocalGenerator.cxx
+++ b/Source/cmLocalGenerator.cxx
@@ -1945,6 +1945,8 @@ void cmLocalGenerator::AddLanguageFlags(std::string& flags,
         this->AppendFlags(flags, "-swift-version " + std::string(v));
       }
     }
+  } else if (lang == "CUDA") {
+    target->AddCUDAArchitectureFlags(flags);
   }
 
   // Add MSVC runtime library flags.  This is activated by the presence
diff --git a/Source/cmPolicies.h b/Source/cmPolicies.h
index d3daad8..0eb42bf 100644
--- a/Source/cmPolicies.h
+++ b/Source/cmPolicies.h
@@ -308,6 +308,10 @@ class cmMakefile;
          3, 17, 0, cmPolicies::WARN)                                          \
   SELECT(POLICY, CMP0103,                                                     \
          "multiple export() with same FILE without APPEND is not allowed.",   \
+         3, 18, 0, cmPolicies::WARN)                                          \
+  SELECT(POLICY, CMP0104,                                                     \
+         "CMAKE_CUDA_ARCHITECTURES now detected for NVCC, empty "             \
+         "CUDA_ARCHITECTURES not allowed.",                                   \
          3, 18, 0, cmPolicies::WARN)
 
 #define CM_SELECT_ID(F, A1, A2, A3, A4, A5, A6) F(A1)
@@ -338,7 +342,8 @@ class cmMakefile;
   F(CMP0081)                                                                  \
   F(CMP0083)                                                                  \
   F(CMP0095)                                                                  \
-  F(CMP0099)
+  F(CMP0099)                                                                  \
+  F(CMP0104)
 
 /** \class cmPolicies
  * \brief Handles changes in CMake behavior and policies
diff --git a/Source/cmTarget.cxx b/Source/cmTarget.cxx
index 0c4311d..955a5cc 100644
--- a/Source/cmTarget.cxx
+++ b/Source/cmTarget.cxx
@@ -363,6 +363,7 @@ cmTarget::cmTarget(std::string const& name, cmStateEnums::TargetType type,
     initProp("CUDA_SEPARABLE_COMPILATION");
     initProp("CUDA_RESOLVE_DEVICE_SYMBOLS");
     initProp("CUDA_RUNTIME_LIBRARY");
+    initProp("CUDA_ARCHITECTURES");
     initProp("LINK_SEARCH_START_STATIC");
     initProp("LINK_SEARCH_END_STATIC");
     initProp("Swift_LANGUAGE_VERSION");