Reland [mlir][Target] Improve ROCDL gpu serialization API (#96198)

Reland: #95456

This patch improves the ROCDL gpu serialization API by:
- Introducing the enum `AMDGCNLibraries` for specifying the AMD GCN
device code libraries to use during linking.
- Removing `getCommonBitcodeLibs` in favor of `AMDGCNLibraries`.
Previously `getCommonBitcodeLibs` would try to load all AMD GCN bitcode
librariesm now it will only load the requested libraries.
- Exposing the `compileToBinary` method and making it virtual, allowing
downstream users to re-use this method.
- Exposing `moduleToObjectImpl`, this method provides a prototype flow
for compiling to binary, allowing downstream users to re-use this
method.
- It also avoids constructing the control variables if no device
libraries are being used.
- Changes the style of the error messages to be composable, ie no full
stops.
- Adds an error message for when the ROCm toolkit can't be found but it
was required.

Author: fabianmcg

mlir/include/mlir/Target/LLVM/ROCDL/Utils.h

@@ -27,6 +27,19 @@ namespace ROCDL {
 /// 5. Returns an empty string.
 StringRef getROCMPath();
 
+/// Helper enum for specifying the AMD GCN device libraries required for
+/// compilation.
+enum class AMDGCNLibraries : uint32_t {
+  None = 0,
+  Ockl = 1,
+  Ocml = 2,
+  OpenCL = 4,
+  Hip = 8,
+  LastLib = Hip,
+  LLVM_MARK_AS_BITMASK_ENUM(LastLib),
+  All = (LastLib << 1) - 1
+};
+
 /// Base class for all ROCDL serializations from GPU modules into binary
 /// strings. By default this class serializes into LLVM bitcode.
 class SerializeGPUModuleBase : public LLVM::ModuleToObject {
@@ -49,8 +62,8 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
   /// Returns the bitcode files to be loaded.
   ArrayRef<std::string> getFileList() const;
 
-  /// Appends standard ROCm device libraries like `ocml.bc`, `ockl.bc`, etc.
-  LogicalResult appendStandardLibs();
+  /// Appends standard ROCm device libraries to `fileList`.
+  LogicalResult appendStandardLibs(AMDGCNLibraries libs);
 
   /// Loads the bitcode files in `fileList`.
   virtual std::optional<SmallVector<std::unique_ptr<llvm::Module>>>
@@ -63,15 +76,20 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
   LogicalResult handleBitcodeFile(llvm::Module &module) override;
 
 protected:
-  /// Appends the paths of common ROCm device libraries to `libs`.
-  LogicalResult getCommonBitcodeLibs(llvm::SmallVector<std::string> &libs,
-                                     SmallVector<char, 256> &libPath,
-                                     StringRef isaVersion);
-
   /// Adds `oclc` control variables to the LLVM module.
-  void addControlVariables(llvm::Module &module, bool wave64, bool daz,
-                           bool finiteOnly, bool unsafeMath, bool fastMath,
-                           bool correctSqrt, StringRef abiVer);
+  void addControlVariables(llvm::Module &module, AMDGCNLibraries libs,
+                           bool wave64, bool daz, bool finiteOnly,
+                           bool unsafeMath, bool fastMath, bool correctSqrt,
+                           StringRef abiVer);
+
+  /// Compiles assembly to a binary.
+  virtual std::optional<SmallVector<char, 0>>
+  compileToBinary(const std::string &serializedISA);
+
+  /// Default implementation of `ModuleToObject::moduleToObject`.
+  std::optional<SmallVector<char, 0>>
+  moduleToObjectImpl(const gpu::TargetOptions &targetOptions,
+                     llvm::Module &llvmModule);
 
   /// Returns the assembled ISA.
   std::optional<SmallVector<char, 0>> assembleIsa(StringRef isa);
@@ -84,6 +102,9 @@ class SerializeGPUModuleBase : public LLVM::ModuleToObject {
 
   /// List of LLVM bitcode files to link to.
   SmallVector<std::string> fileList;
+
+  /// AMD GCN libraries to use when linking, the default is using none.
+  AMDGCNLibraries deviceLibs = AMDGCNLibraries::None;
 };
 } // namespace ROCDL
 } // namespace mlir

mlir/lib/Target/LLVM/CMakeLists.txt

@@ -127,17 +127,18 @@ add_mlir_dialect_library(MLIRROCDLTarget
   )
 
 if(MLIR_ENABLE_ROCM_CONVERSIONS)
-  if (NOT ("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD))
-    message(SEND_ERROR
-      "Building mlir with ROCm support requires the AMDGPU backend")
-  endif()
-
   if (DEFINED ROCM_PATH)
     set(DEFAULT_ROCM_PATH "${ROCM_PATH}" CACHE PATH "Fallback path to search for ROCm installs")
   elseif(DEFINED ENV{ROCM_PATH})
     set(DEFAULT_ROCM_PATH "$ENV{ROCM_PATH}" CACHE PATH "Fallback path to search for ROCm installs")
   else()
-    set(DEFAULT_ROCM_PATH "/opt/rocm" CACHE PATH "Fallback path to search for ROCm installs")
+    if (WIN32)
+      # Avoid setting an UNIX path for Windows.
+      # TODO: Eventually migrate to FindHIP once it becomes a part of CMake.
+      set(DEFAULT_ROCM_PATH "" CACHE PATH "Fallback path to search for ROCm installs")
+    else()
+      set(DEFAULT_ROCM_PATH "/opt/rocm" CACHE PATH "Fallback path to search for ROCm installs")
+    endif()
   endif()
   message(VERBOSE "MLIR Default ROCM toolkit path: ${DEFAULT_ROCM_PATH}")