[mlir][Target] Support Fatbin target for static nvptxcompiler

Author: MikaOvO

mlir/lib/Target/LLVM/CMakeLists.txt

@@ -66,27 +66,39 @@ if ("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
       set(MLIR_CUDAToolkit_ROOT ${CUDAToolkit_LIBRARY_ROOT})
     endif()
 
-    # Add the `nvptxcompiler` library.
+    # Add the `nvptxcompiler` library and the `nvfatbin` library.
     if(MLIR_ENABLE_NVPTXCOMPILER)
       # Find the `nvptxcompiler` library.
       # TODO: Bump the MLIR CMake version to 3.25 and use `CUDA::nvptxcompiler_static`.
       find_library(MLIR_NVPTXCOMPILER_LIB_PATH nvptxcompiler_static
                   PATHS ${CUDAToolkit_LIBRARY_DIR} NO_DEFAULT_PATH)
+      find_library(MLIR_NVFATBIN_LIB_PATH nvfatbin_static
+                  PATHS ${CUDAToolkit_LIBRARY_DIR} NO_DEFAULT_PATH)
 
       # Fail if `nvptxcompiler_static` couldn't be found.
       if(MLIR_NVPTXCOMPILER_LIB_PATH STREQUAL "MLIR_NVPTXCOMPILER_LIB_PATH-NOTFOUND")
         message(FATAL_ERROR
                 "Requested using the `nvptxcompiler` library backend but it couldn't be found.")
       endif()
 
+      # Fail if `nvfatbin_static` couldn't be found.
+      if(MLIR_NVFATBIN_LIB_PATH STREQUAL "MLIR_NVFATBIN_LIB_PATH-NOTFOUND")
+        message(FATAL_ERROR
+                "Requested using the `nvfatbin` library backend but it couldn't be found.")
+      endif()    
+
       add_library(MLIR_NVPTXCOMPILER_LIB STATIC IMPORTED GLOBAL)
+      add_library(MLIR_NVFATBIN_LIB STATIC IMPORTED GLOBAL)
       # Downstream projects can modify this path and use it in CMake. For example:
       # add_library(MLIR_NVPTXCOMPILER_LIB STATIC IMPORTED GLOBAL)
       # set_property(TARGET MLIR_NVPTXCOMPILER_LIB PROPERTY IMPORTED_LOCATION ${...})  
       # where `...` is to be replaced with the path to the library.
       set_property(TARGET MLIR_NVPTXCOMPILER_LIB PROPERTY IMPORTED_LOCATION ${MLIR_NVPTXCOMPILER_LIB_PATH})
-      # Link against `nvptxcompiler_static`. TODO: use `CUDA::nvptxcompiler_static`.
+      set_property(TARGET MLIR_NVFATBIN_LIB PROPERTY IMPORTED_LOCATION ${MLIR_NVFATBIN_LIB_PATH})  
+      # Link against `nvptxcompiler_static` and `nvfatbin_static`. 
+      # TODO: use `CUDA::nvptxcompiler_static` and `CUDA::nvfatbin_static`.
       target_link_libraries(MLIRNVVMTarget PRIVATE MLIR_NVPTXCOMPILER_LIB)
+      target_link_libraries(MLIRNVVMTarget PRIVATE MLIR_NVFATBIN_LIB)
       target_include_directories(obj.MLIRNVVMTarget PUBLIC ${CUDAToolkit_INCLUDE_DIRS})
     endif()
   else()

mlir/lib/Target/LLVM/NVVM/Target.cpp

@@ -473,6 +473,18 @@ NVPTXSerializer::compileToBinary(const std::string &ptxCode) {
     }                                                                          \
   } while (false)
 
+#include <nvFatbin.h>
+
+#define RETURN_ON_NVFATBIN_ERROR(expr)                                         \
+  do {                                                                         \
+    auto result = (expr);                                                      \
+    if (result != nvFatbinResult::NVFATBIN_SUCCESS) {                          \
+      emitError(loc) << llvm::Twine(#expr).concat(" failed with error: ")      \
+                     << nvFatbinGetErrorString(result);                        \
+      return std::nullopt;                                                     \
+    }                                                                          \
+  } while (false)
+
 std::optional<SmallVector<char, 0>>
 NVPTXSerializer::compileToBinaryNVPTX(const std::string &ptxCode) {
   Location loc = getOperation().getLoc();
@@ -486,6 +498,11 @@ NVPTXSerializer::compileToBinaryNVPTX(const std::string &ptxCode) {
       targetOptions.tokenizeCmdOptions();
   cmdOpts.second.append(
       {"-arch", getTarget().getChip().data(), "--opt-level", optLevel.c_str()});
+  bool useFatbin32 = false;
+  for (const char *option : cmdOpts.second) {
+    if (StringRef(option) == "-32")
+      useFatbin32 = true;
+  }
 
   // Create the compiler handle.
   RETURN_ON_NVPTXCOMPILER_ERROR(
@@ -538,6 +555,30 @@ NVPTXSerializer::compileToBinaryNVPTX(const std::string &ptxCode) {
   });
 #undef DEBUG_TYPE
   RETURN_ON_NVPTXCOMPILER_ERROR(nvPTXCompilerDestroy(&compiler));
+
+  if (targetOptions.getCompilationTarget() == gpu::CompilationTarget::Fatbin) {
+    const char *cubinOpts[1] = {"-64"};
+    if (useFatbin32) {
+      cubinOpts[0] = {"-32"};
+    }
+    nvFatbinHandle handle;
+
+    auto chip = getTarget().getChip();
+    chip.consume_front("sm_");
+
+    RETURN_ON_NVFATBIN_ERROR(nvFatbinCreate(&handle, cubinOpts, 1));
+    RETURN_ON_NVFATBIN_ERROR(nvFatbinAddCubin(
+        handle, binary.data(), binary.size(), chip.data(), nullptr));
+    RETURN_ON_NVFATBIN_ERROR(nvFatbinAddPTX(
+        handle, ptxCode.data(), ptxCode.size(), chip.data(), nullptr, nullptr));
+
+    size_t fatbinSize;
+    RETURN_ON_NVFATBIN_ERROR(nvFatbinSize(handle, &fatbinSize));
+    SmallVector<char, 0> fatbin(fatbinSize, 0);
+    RETURN_ON_NVFATBIN_ERROR(nvFatbinGet(handle, (void *)fatbin.data()));
+    RETURN_ON_NVFATBIN_ERROR(nvFatbinDestroy(&handle));
+    return fatbin;
+  }
   return binary;
 }
 #endif // MLIR_ENABLE_NVPTXCOMPILER