[mlir][NVVM] Adds the NVVM target attribute.

[fmorac]

Summary

For an explanation of these patches see D154153.

Commit message:
This patch adds the NVVM target attribute for serializing GPU modules into
strings containing cubin.

Depends on D154113 and D154100 and D154097

[fmorac]

Update the commit message to include D154100

[fmorac]

Update commit message to include D154097

[fmorac]

Rebasing.

[fmorac]

Updated location of ModuleToObject.h

[mehdi_amini]

mlir/include/mlir/Dialect/GPU/IR/GPUCompilationAttr.td

23 ↗

(On Diff #543179)

Wondering why is this in the gpu dialect and not a target specific one? (like nvvm)

82 ↗

(On Diff #536066)

Nit: hasFastMath since it returns a bool. I would think that getFastMath would return a FastMathFlag attribute.

[mehdi_amini]

mlir/lib/Dialect/GPU/Targets/NVPTXTarget.cpp

213 ↗

(On Diff #543179)

We should have options to stop after emitting LLVMIR and PTX, seems like we have to go to cubin right now?

I would also think that going to PTX can be possible without requiring any dependency on the Cuda toolkit?

[mehdi_amini]

Do we have a way to test this without too much of the other code by the way?

[fmorac]

In D154117#4526902, @mehdi_amini wrote:

Do we have a way to test this without too much of the other code by the way?

Not really? I added some on the final patches were you have all the infra ready. The best test I can think of at this point, is an unit test for this serializer?

mlir/include/mlir/Dialect/GPU/IR/GPUCompilationAttr.td

23 ↗

(On Diff #543179)

I initially added these attributes (NVPTX & AMDGPU) to nvvm & rocdl, however I decided against that in the final patch because it added GPU dependencies to those dialects (includes & libs) and created more libs instead of a single (GPUTargets).

I'm open to change it.

82 ↗

(On Diff #536066)

I'll change it.

mlir/lib/Dialect/GPU/Targets/NVPTXTarget.cpp

213 ↗

(On Diff #543179)

These series of patches intend to be a full replacement of the current pipeline, upon approval we should intermediately deprecate the current pipeline and then remove it after a deprecation period.

The idea is that in future patches we change this behavior and stop at LLVM, but that requires having the LLVM Offload work done.

The benefit of this approach is that those future changes would be invisible to users and at the same time it makes them adopt this new mechanism now, that's why I go straight to cubin.

[mehdi_amini]

mlir/include/mlir/Dialect/GPU/IR/GPUCompilationAttr.td

23 ↗

(On Diff #543179)

Where will the dependency come from? The GPUTargetAttrInterface?

Maybe we should revisit the way the interface is specified. Worse case we're just one level of indirection away :)

Here if the issue is just the gpu::TargetOptions, maybe it can be declared with the interface definition, which makes it an isolated library?

mlir/lib/Dialect/GPU/Targets/NVPTXTarget.cpp

213 ↗

(On Diff #543179)

OK, thanks for explaining, can you add a TODO with this info in the code?

[fmorac]

Move the target attribute to NVVM and added it as an external model that it's promised.

[fmorac]

Fix typo in unit test.

[mehdi_amini]

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

56

Do we need to the init in the registration? Can we do the init when the serialization is called instead?

mlir/unittests/Target/LLVM/SerializeNVVMTarget.cpp

34

Not clear to me where we need the native target?

47

Can we tone this down? There is no reason to link all of MLIR into this unit-test.

[fmorac]

Addressed comments.

[mehdi_amini]

(Waiting on update to use ptxas)

[fmorac]

Swtiched to ptxas, added options for stopping compilation at LLVM IR & PTX, as well as unit tests testing the added functionalty.

[mehdi_amini]

Bring @tra here to get feedback on pros/cons of invoking ptxas through a temp file vs using the library APIs exposed by nvptxcompile.

[fmorac]

In D154117#4554942, @mehdi_amini wrote:

Bring @tra here to get feedback on pros/cons of invoking ptxas through a temp file vs using the library APIs exposed by nvptxcompile.

One immediate benefit is that we don't longer have a dependency on the toolkit to build this. You can detect ptxas by setting the env variable CUDA_ROOT to the location of the toolkit or by adding it to PATH.

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

56

Yes, it's possible I'll change it.

mlir/unittests/Target/LLVM/SerializeNVVMTarget.cpp

34

Not needed, removed.

47

Yes, I've removed it.

[guraypp]

In D154117#4554966, @fmorac wrote:

In D154117#4554942, @mehdi_amini wrote:

Bring @tra here to get feedback on pros/cons of invoking ptxas through a temp file vs using the library APIs exposed by nvptxcompile.

One immediate benefit is that we don't longer have a dependency on the toolkit to build this. You can detect ptxas by setting the env variable CUDA_ROOT to the location of the toolkit or by adding it to PATH.

I think compiling by ptxas has merits. One can different version of ptxas in case of performance regression.

[guraypp]

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

290

It would be nice to pass parameters to ptxas from MLIR.

325

What do you think about passsing -v always to ptxas and print the output (perhaps stdout file) using llvm::errs.
I found -v very useful to see compile-time register usage and local memory.

[guraypp]

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

286

Can we keep the PTX file? After codegen, it is very natural to look at the PTX file, if we keep the file via a flag, I think it would be great.

I implemented the dump-ptx flag to the gpu-to-cubin Pass earlier. I guess we cannot use that flag here.

[fmorac]

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

286

If you pass -debug-only=serialize-to-ptx (I'm changing it to serialize-to-isa) to mlir-opt then the PTX file will be printed to stdout. This mechanism also allows emitting PTX instead of binary, so you have multiple ways of obtaining the PTX.

290

You can, the args are passed through the targetOptions variable. Below you'll find that I add them to the PTX invocation.

325

We could consider adding a dedicated variable, but as it stands you can pass that variable through the cmdline and use --debug-only=serialize-to-binary to dump that output into stdout.

[mehdi_amini]

I think compiling by ptxas has merits. One can different version of ptxas in case of performance regression.

nvptxcompile provides the same facility I believe, doesn't it?

[guraypp]

In D154117#4558565, @mehdi_amini wrote:

I think compiling by ptxas has merits. One can different version of ptxas in case of performance regression.

nvptxcompile provides the same facility I believe, doesn't it?

Though I haven't personally worked with nvptxcompile, it appears to be bundled with the toolkit. If one could use different toolkit version for PTX compilation, then yes it gives the same facility.

Recently, I used the driver (current MLIR compilation) and ptxas for PTX compilation. I noticed that even if I feed the same PTX code, the driver occasionally generates different SASS with ptxas. Maybe I hit a corner case but it was really hard to find the reason.

I believe it's crucial for us to know the potential disparities between nvptxcompile and ptxas if there is any.

[fmorac]

My 2 cents, as far as I know the nvptxcompiler and ptxas should produce the same code for a given release. However unlike ptxas the only way to change the version of nvptxcompiler is to recompile the NVVMTarget library with a different version.

Adding support for nvptxcompiler is like extra 20 lines of code in this patch, and a couple on CMake, how do we feel about having both?

If nvptxcompiler is detected during build then nvptxcompiler is used, if not, then ptxas is used, and we also give the user the ability to choose on CMake?

[mehdi_amini]

If nvptxcompiler is detected during build then nvptxcompiler is used, if not, then ptxas is used, and we also give the user the ability to choose on CMake?

Adding a CMake option looks good to me, I would pick a default and stick to it (that is fail the cmake configuration with a helpful message).

I don't like "magic" fallback because detection is fragile and it makes the behavior dependent on the environment of the user. It's harder to know what you build, and getting bug reports it makes it also harder to know what's happening.

[fmorac]

Added NVPTX compiler support & removed unnecessary fwd declarations. The CMake variable MLIR_ENABLE_NVPTXCOMPILER controls the usage of the NVPTXCompiler library and is disabled by default, as CMake 3.20 doesn't support CUDA::nvptxcompiler_static to find the library.

[tra]

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

124

A word of caution -- some linux distributions scatter CUDA SDK across the 'standard' linux filsystem paths, so a single getToolkitPath() would not be able to find all the necessary bits, as libdevice and the binaries will be in different places. You may need additional heuristics along the lines of what clang driver does.
https://github.com/llvm/llvm-project/blob/1b74459df8a6d960f7387f0c8379047e42811f58/clang/lib/Driver/ToolChains/Cuda.cpp#L180

[fmorac]

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

124

Thank you, that's good to know. I'll see how to rework it or add docs indicating how to make it work.
As it stands the user could specify the CUDA path to /usr/lib/cuda for libdevice, somewhat similar to --cuda-path.

For ptxas this mechanism searches several places in the following order:

1. Command line option.
2. PATH variable.
3. The toolkit path as specified by a couple of env variables or the one detected by CMake.

So it would work on Debian & Ubuntu, but the mechanism could be over burdening the user.

[fmorac]

Reapplied clang-format.
Adds a couple of syntax tests & fixes a bug not displaying the link option in the attribute.

[mehdi_amini]

Thanks Fabian!

mlir/CMakeLists.txt

122

The description isn't super clear, because it conflicts with what "the NVPTX backend" is in LLVM.
What about: "Statically link the nvptxlibrary instead of calling ptxas as a subprocess for compiling PTX to cubin" ?

[fmorac]

Changed the description of the CMake variable MLIR_ENABLE_NVPTXCOMPILER. Applied clang-format again.

[fmorac]

Fixed merge conflicts with D157183.

mlir/CMakeLists.txt

 if(${LLVM_NATIVE_ARCH} IN_LIST LLVM_TARGETS_TO_BUILD)
   set(MLIR_ENABLE_EXECUTION_ENGINE 1)
 else()
   set(MLIR_ENABLE_EXECUTION_ENGINE 0)
 endif()
 
 # Build the CUDA conversions and run according tests if the NVPTX backend
 # is available
-if ("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD AND MLIR_ENABLE_EXECUTION_ENGINE)
+if ("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
   set(MLIR_ENABLE_CUDA_CONVERSIONS 1)
 else()
   set(MLIR_ENABLE_CUDA_CONVERSIONS 0)
 endif()
 # TODO: we should use a config.h file like LLVM does
 add_definitions(-DMLIR_CUDA_CONVERSIONS_ENABLED=${MLIR_ENABLE_CUDA_CONVERSIONS})
 
 # Build the ROCm conversions and run according tests if the AMDGPU backend
 # is available.
 if ("AMDGPU" IN_LIST LLVM_TARGETS_TO_BUILD)
   set(MLIR_ENABLE_ROCM_CONVERSIONS 1)
 else()
   set(MLIR_ENABLE_ROCM_CONVERSIONS 0)
 endif()
 add_definitions(-DMLIR_ROCM_CONVERSIONS_ENABLED=${MLIR_ENABLE_ROCM_CONVERSIONS})
 
 set(MLIR_ENABLE_CUDA_RUNNER 0 CACHE BOOL "Enable building the mlir CUDA runner")
 set(MLIR_ENABLE_ROCM_RUNNER 0 CACHE BOOL "Enable building the mlir ROCm runner")
 set(MLIR_ENABLE_SPIRV_CPU_RUNNER 0 CACHE BOOL "Enable building the mlir SPIR-V cpu runner")
 set(MLIR_ENABLE_VULKAN_RUNNER 0 CACHE BOOL "Enable building the mlir Vulkan runner")
 
 option(MLIR_INCLUDE_TESTS
        "Generate build targets for the MLIR unit tests."
        ${LLVM_INCLUDE_TESTS})
 
 option(MLIR_INCLUDE_INTEGRATION_TESTS
        "Generate build targets for the MLIR integration tests.")
 
 set(MLIR_INSTALL_AGGREGATE_OBJECTS 1 CACHE BOOL
     "Installs object files needed for add_mlir_aggregate to work out of \

mlir/include/mlir/Dialect/LLVMIR/NVVMDialect.h

 // NVVM specific extensions to the LLVM type system.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef MLIR_DIALECT_LLVMIR_NVVMDIALECT_H_
 #define MLIR_DIALECT_LLVMIR_NVVMDIALECT_H_
 
 #include "mlir/Bytecode/BytecodeOpInterface.h"
+#include "mlir/Dialect/GPU/IR/CompilationInterfaces.h"
 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
 #include "mlir/IR/Dialect.h"
 #include "mlir/IR/OpDefinition.h"
 #include "mlir/Interfaces/SideEffectInterfaces.h"
 #include "llvm/IR/IntrinsicsNVPTX.h"
 
 #include "mlir/Dialect/LLVMIR/NVVMOpsEnums.h.inc"
 

mlir/include/mlir/Dialect/LLVMIR/NVVMOps.td

let description = [{
     See for more information:
     https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#asynchronous-warpgroup-level-matrix-instructions-wgmma-wait-group
   }];
   let extraClassDefinition = [{
     std::string $cppClass::getPtx() { return std::string("wgmma.wait_group.sync.aligned %0;"); }
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// NVVM target attribute.
+//===----------------------------------------------------------------------===//
+
+def NVVM_TargettAttr : NVVM_Attr<"NVVMTarget", "target", []> {
+  let description = [{
+    GPU target attribute for controlling compilation of NVIDIA targets. All
+    parameters decay into default values if not present.
+
+    Examples:
+
+    1. Target with default values.
+    ```
+      gpu.module @mymodule [#nvvm.target] attributes {...} {
+        ...
+      }
+    ```
+
+    2. Target with `sm_90` chip and fast math.
+    ```
+      gpu.module @mymodule [#nvvm.target<chip = "sm_90", flags = {fast}>] {
+        ...
+      }
+    ```
+  }];
+  let parameters = (ins
+    DefaultValuedParameter<"int", "2", "Optimization level to apply.">:$O,
+    StringRefParameter<"Target triple.", "\"nvptx64-nvidia-cuda\"">:$triple,
+    StringRefParameter<"Target chip.", "\"sm_50\"">:$chip,
+    StringRefParameter<"Target chip features.", "\"+ptx60\"">:$features,
+    OptionalParameter<"DictionaryAttr", "Target specific flags.">:$flags,
+    OptionalParameter<"ArrayAttr", "Files to link to the LLVM module.">:$link
+  );
+  let assemblyFormat = [{
+    (`<` struct($O, $triple, $chip, $features, $flags)^ `>`)?
+  }];
+  let builders = [
+    AttrBuilder<(ins CArg<"int", "2">:$optLevel,
+                     CArg<"StringRef", "\"nvptx64-nvidia-cuda\"">:$triple,
+                     CArg<"StringRef", "\"sm_50\"">:$chip,
+                     CArg<"StringRef", "\"+ptx60\"">:$features,
+                     CArg<"DictionaryAttr", "nullptr">:$targetFlags,
+                     CArg<"ArrayAttr", "nullptr">:$linkFiles), [{
+      return Base::get($_ctxt, optLevel, triple, chip, features, targetFlags, linkFiles);
+    }]>
+  ];
+  let skipDefaultBuilders = 1;
+  let genVerifyDecl = 1;
+  let extraClassDeclaration = [{
+    bool hasFlag(StringRef flag) const;
+    bool hasFastMath() const;
+    bool hasFtz() const;
+  }];
+  let extraClassDefinition = [{
+    bool $cppClass::hasFlag(StringRef flag) const {
+      if (DictionaryAttr flags = getFlags())
+        return flags.get(flag) != nullptr;
+      return false;
+    }
+    bool $cppClass::hasFastMath() const {
+      return hasFlag("fast");
+    }
+    bool $cppClass::hasFtz() const {
+      return hasFlag("ftz");
+    }
+  }];
+}
+
 #endif // NVVMIR_OPS

mlir/include/mlir/InitAllExtensions.h

 // extensions to the system.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef MLIR_INITALLEXTENSIONS_H_
 #define MLIR_INITALLEXTENSIONS_H_
 
 #include "mlir/Dialect/Func/Extensions/AllExtensions.h"
+#include "mlir/Target/LLVM/NVVM/Target.h"
 
 #include <cstdlib>
 
 namespace mlir {
 
 /// This function may be called to register all MLIR dialect extensions with the
 /// provided registry.
 /// If you're building a compiler, you generally shouldn't use this: you would
 /// individually register the specific extensions that are useful for the
 /// pipelines and transformations you are using.
 inline void registerAllExtensions(DialectRegistry &registry) {
   func::registerAllExtensions(registry);
+  registerNVVMTarget(registry);
 }
 
 } // namespace mlir
 
 #endif // MLIR_INITALLEXTENSIONS_H_

mlir/include/mlir/Target/LLVM/NVVM/Target.h

+//===- Target.h - MLIR NVVM target registration -----------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This provides registration calls for attaching the NVVM target interface.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_TARGET_LLVM_NVVM_TARGET_H
+#define MLIR_TARGET_LLVM_NVVM_TARGET_H
+
+namespace mlir {
+class DialectRegistry;
+class MLIRContext;
+/// Registers the `TargetAttrInterface` for the `#nvvm.target` attribute in the
+/// given registry.
+void registerNVVMTarget(DialectRegistry &registry);
+
+/// Registers the `TargetAttrInterface` for the `#nvvm.target` attribute in the
+/// registry associated with the given context.
+void registerNVVMTarget(MLIRContext &context);
+} // namespace mlir
+
+#endif // MLIR_TARGET_LLVM_NVVM_TARGET_H

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

+//===- Utils.h - MLIR NVVM target compilation ------------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This files declares NVVM target related utility classes and functions.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_TARGET_LLVM_NVVM_UTILS_H
+#define MLIR_TARGET_LLVM_NVVM_UTILS_H
+
+#include "mlir/Dialect/GPU/IR/CompilationInterfaces.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
+#include "mlir/Target/LLVM/ModuleToObject.h"
+
+namespace mlir {
+class DialectRegistry;
+class MLIRContext;
+namespace NVVM {
+/// Searches & returns the path CUDA toolkit path, the search order is:
+/// 1. The `CUDA_ROOT` environment variable.
+/// 2. The `CUDA_HOME` environment variable.
+/// 3. The `CUDA_PATH` environment variable.
+/// 4. The CUDA toolkit path detected by CMake.
+/// 5. Returns an empty string.
+StringRef getCUDAToolkitPath();
+
+/// Base class for all NVVM serializations from GPU modules into binary strings.
+/// By default this class serializes into LLVM bitcode.
+class SerializeGPUModuleBase : public LLVM::ModuleToObject {
+public:
+  /// Initializes the `toolkitPath` with the path in `targetOptions` or if empty
+  /// with the path in `getCUDAToolkitPath`.
+  SerializeGPUModuleBase(Operation &module, NVVMTargetAttr target,
+                         const gpu::TargetOptions &targetOptions = {});
+
+  /// Initializes the LLVM NVPTX target by safely calling `LLVMInitializeNVPTX*`
+  /// methods if available.
+  static void init();
+
+  /// Returns the target attribute.
+  NVVMTargetAttr getTarget() const;
+
+  /// Returns the CUDA toolkit path.
+  StringRef getToolkitPath() const;
+
+  /// Returns the bitcode files to be loaded.
+  ArrayRef<std::string> getFileList() const;
+
+  /// Appends `nvvm/libdevice.bc` into `fileList`. Returns failure if the
+  /// library couldn't be found.
+  LogicalResult appendStandardLibs();
+
+  /// Loads the bitcode files in `fileList`.
+  virtual std::optional<SmallVector<std::unique_ptr<llvm::Module>>>
+  loadBitcodeFiles(llvm::Module &module,
+                   llvm::TargetMachine &targetMachine) override;
+
+protected:
+  /// NVVM target attribute.
+  NVVMTargetAttr target;
+
+  /// CUDA toolkit path.
+  std::string toolkitPath;
+
+  /// List of LLVM bitcode files to link to.
+  SmallVector<std::string> fileList;
+};
+} // namespace NVVM
+} // namespace mlir
+
+#endif // MLIR_TARGET_LLVM_NVVM_UTILS_H

mlir/lib/Dialect/LLVMIR/IR/NVVMDialect.cpp

#include "mlir/Dialect/LLVMIR/NVVMOps.cpp.inc"
   addAttributes<
 #define GET_ATTRDEF_LIST
 #include "mlir/Dialect/LLVMIR/NVVMOpsAttributes.cpp.inc"
       >();
 
   // Support unknown operations because not all NVVM operations are
   // registered.
   allowUnknownOperations();
+  declarePromisedInterface<gpu::TargetAttrInterface>();
 }
 
 LogicalResult NVVMDialect::verifyOperationAttribute(Operation *op,
                                                     NamedAttribute attr) {
   StringAttr attrName = attr.getName();
   // Kernel function attribute should be attached to functions.
   if (attrName == NVVMDialect::getKernelFuncAttrName()) {
     if (!isa<LLVM::LLVMFuncOp>(op)) {
if (attrName == NVVMDialect::getMinctasmAttrName() ||
     if (!llvm::dyn_cast<IntegerAttr>(attr.getValue()))
       return op->emitError()
              << "'" << attrName << "' attribute must be integer constant";
   }
 
   return success();
 }
 
+//===----------------------------------------------------------------------===//
+// NVVM target attribute.
+//===----------------------------------------------------------------------===//
+LogicalResult
+NVVMTargetAttr::verify(function_ref<InFlightDiagnostic()> emitError,
+                       int optLevel, StringRef triple, StringRef chip,
+                       StringRef features, DictionaryAttr flags,
+                       ArrayAttr files) {
+  if (optLevel < 0 || optLevel > 3) {
+    emitError() << "The optimization level must be a number between 0 and 3.";
+    return failure();
+  }
+  if (triple.empty()) {
+    emitError() << "The target triple cannot be empty.";
+    return failure();
+  }
+  if (chip.empty()) {
+    emitError() << "The target chip cannot be empty.";
+    return failure();
+  }
+  if (files && llvm::all_of(files, [](::mlir::Attribute attr) {
+        return attr && mlir::isa<StringAttr>(attr);
+      })) {
+    emitError() << "All the elements in the `link` array must be strings.";
+    return failure();
+  }
+  return success();
+}
+
 #define GET_OP_CLASSES
 #include "mlir/Dialect/LLVMIR/NVVMOps.cpp.inc"
 
 #define GET_ATTRDEF_CLASSES
 #include "mlir/Dialect/LLVMIR/NVVMOpsAttributes.cpp.inc"

mlir/lib/Target/LLVM/CMakeLists.txt

add_mlir_library(MLIRTargetLLVM
   Passes
   Support
   Target
   TargetParser
   LINK_LIBS PUBLIC
   MLIRExecutionEngineUtils
   MLIRTargetLLVMIRExport
 )
+
+if (MLIR_ENABLE_CUDA_CONVERSIONS)
+  set(NVPTX_LIBS
+    NVPTXCodeGen
+    NVPTXDesc
+    NVPTXInfo
+  )
+endif()
+
+add_mlir_dialect_library(MLIRNVVMTarget
+  NVVM/Target.cpp
+
+  ADDITIONAL_HEADER_DIRS
+  ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/LLVMIR
+
+  LINK_COMPONENTS
+  Core
+  MC
+  Target
+  ${NVPTX_LIBS}
+
+  LINK_LIBS PUBLIC
+  MLIRIR
+  MLIRExecutionEngineUtils
+  MLIRSupport
+  MLIRGPUDialect
+  MLIRTargetLLVM
+  MLIRNVVMToLLVMIRTranslation
+  )
+
+if(MLIR_ENABLE_CUDA_CONVERSIONS)
+  # Find the CUDA toolkit.
+  if (NOT DEFINED CUDAToolkit_ROOT)
+    find_package(CUDAToolkit)
+    get_filename_component(CUDAToolkit_ROOT ${CUDAToolkit_BIN_DIR} DIRECTORY ABSOLUTE)
+  endif()
+  message(VERBOSE "MLIR Default CUDA toolkit path: ${CUDAToolkit_ROOT}")
+
+  # Define the `CUDAToolkit` path.
+  target_compile_definitions(obj.MLIRNVVMTarget
+    PRIVATE
+    __DEFAULT_CUDATOOLKIT_PATH__="${CUDAToolkit_ROOT}"
+  )
+endif()

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

+//===- Target.cpp - MLIR LLVM NVVM target compilation -----------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This files defines NVVM target related functions including registration
+// calls for the `#nvvm.target` compilation attribute.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Target/LLVM/NVVM/Target.h"
+
+#include "mlir/Dialect/GPU/IR/GPUDialect.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
+#include "mlir/Target/LLVM/NVVM/Utils.h"
+#include "mlir/Target/LLVMIR/Dialect/GPU/GPUToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/LLVMIR/LLVMToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/NVVM/NVVMToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Export.h"
+
+#include "llvm/Support/FileSystem.h"
+#include "llvm/Support/FileUtilities.h"
+#include "llvm/Support/FormatVariadic.h"
+#include "llvm/Support/MemoryBuffer.h"
+#include "llvm/Support/Path.h"
+#include "llvm/Support/Process.h"
+#include "llvm/Support/Program.h"
+#include "llvm/Support/TargetSelect.h"
+
+#include <cstdlib>
+
+using namespace mlir;
+using namespace mlir::NVVM;
+
+#ifndef __DEFAULT_CUDATOOLKIT_PATH__
+#define __DEFAULT_CUDATOOLKIT_PATH__ ""
+#endif
+
+namespace {
+// Implementation of the `TargetAttrInterface` model.
+class NVVMTargetAttrImpl
+    : public gpu::TargetAttrInterface::FallbackModel<NVVMTargetAttrImpl> {
+public:
+  std::optional<SmallVector<char, 0>>
+  serializeToObject(Attribute attribute, Operation *module,
+                    const gpu::TargetOptions &options) const;
+};
+} // namespace
+
+// Register the NVVM dialect, the NVVM translation & the target interface.
+void mlir::registerNVVMTarget(DialectRegistry &registry) {
+  registerNVVMDialectTranslation(registry);
+  registry.addExtension(+[](MLIRContext *ctx, NVVM::NVVMDialect *dialect) {
+    NVVMTargetAttr::attachInterface<NVVMTargetAttrImpl>(*ctx);
+  });
+}
+
+void mlir::registerNVVMTarget(MLIRContext &context) {
+  DialectRegistry registry;
+  registerNVVMTarget(registry);
+  context.appendDialectRegistry(registry);
+}
+
+// Search for the CUDA toolkit path.
+StringRef mlir::NVVM::getCUDAToolkitPath() {
+  if (const char *var = std::getenv("CUDA_ROOT"))
+    return var;
+  if (const char *var = std::getenv("CUDA_HOME"))
+    return var;
+  if (const char *var = std::getenv("CUDA_PATH"))
+    return var;
+  return __DEFAULT_CUDATOOLKIT_PATH__;
+}
+
+SerializeGPUModuleBase::SerializeGPUModuleBase(
+    Operation &module, NVVMTargetAttr target,
+    const gpu::TargetOptions &targetOptions)
+    : ModuleToObject(module, target.getTriple(), target.getChip(),
+                     target.getFeatures(), target.getO()),
+      target(target), toolkitPath(targetOptions.getToolkitPath()),
+      fileList(targetOptions.getLinkFiles()) {
+
+  // If `targetOptions` have an empty toolkitPath use `getCUDAToolkitPath`
+  if (toolkitPath.empty())
+    toolkitPath = getCUDAToolkitPath();
+
+  // Append the files in the target attribute.
+  if (ArrayAttr files = target.getLink())
+    for (Attribute attr : files.getValue())
+      if (auto file = dyn_cast<StringAttr>(attr))
+        fileList.push_back(file.str());
+
+  // Append libdevice to the files to be loaded.
+  (void)appendStandardLibs();
+}
+
+void SerializeGPUModuleBase::init() {
+  static llvm::once_flag initializeBackendOnce;
+  llvm::call_once(initializeBackendOnce, []() {
+  // If the `NVPTX` LLVM target was built, initialize it.
+#if MLIR_CUDA_CONVERSIONS_ENABLED == 1
+    LLVMInitializeNVPTXTarget();
+    LLVMInitializeNVPTXTargetInfo();
+    LLVMInitializeNVPTXTargetMC();
+    LLVMInitializeNVPTXAsmPrinter();
+#endif
+  });
+}
+
+NVVMTargetAttr SerializeGPUModuleBase::getTarget() const { return target; }
+
+StringRef SerializeGPUModuleBase::getToolkitPath() const { return toolkitPath; }
+
+ArrayRef<std::string> SerializeGPUModuleBase::getFileList() const {
+  return fileList;
+}
+
+// Try to append `libdevice` from a CUDA toolkit installation.
+LogicalResult SerializeGPUModuleBase::appendStandardLibs() {
+  StringRef pathRef = getToolkitPath();
+  if (pathRef.size()) {
+    SmallVector<char, 256> path;
+    path.insert(path.begin(), pathRef.begin(), pathRef.end());
+    pathRef = StringRef(path.data(), path.size());
+    if (!llvm::sys::fs::is_directory(pathRef)) {
+      getOperation().emitError() << "CUDA path: " << pathRef
+                                 << " does not exist or is not a directory.\n";
+      return failure();
+    }
+    llvm::sys::path::append(path, "nvvm", "libdevice", "libdevice.10.bc");
+    pathRef = StringRef(path.data(), path.size());
+    if (!llvm::sys::fs::is_regular_file(pathRef)) {
+      getOperation().emitError() << "LibDevice path: " << pathRef
+                                 << " does not exist or is not a file.\n";
+      return failure();
+    }
+    fileList.push_back(pathRef.str());
+  }
+  return success();
+}
+
+std::optional<SmallVector<std::unique_ptr<llvm::Module>>>
+SerializeGPUModuleBase::loadBitcodeFiles(llvm::Module &module,
+                                         llvm::TargetMachine &targetMachine) {
+  SmallVector<std::unique_ptr<llvm::Module>> bcFiles;
+  if (failed(loadBitcodeFilesFromList(module.getContext(), targetMachine,
+                                      fileList, bcFiles, true)))
+    return std::nullopt;
+  return bcFiles;
+}
+
+#if MLIR_CUDA_CONVERSIONS_ENABLED == 1
+namespace {
+class NVPTXSerializer : public SerializeGPUModuleBase {
+public:
+  NVPTXSerializer(Operation &module, NVVMTargetAttr target,
+                  const gpu::TargetOptions &targetOptions);
+
+  // Compile PTX to cubin.
+  std::optional<SmallVector<char, 0>>
+  compileToBinary(const std::string &ptxCode);
+
+  std::optional<SmallVector<char, 0>>
+  moduleToObject(llvm::Module &llvmModule,
+                 llvm::TargetMachine &targetMachine) override;
+
+private:
+  using TmpFile = std::pair<llvm::SmallString<128>, llvm::FileRemover>;
+
+  // Create a temp file.
+  std::optional<TmpFile> createTemp(StringRef name, StringRef suffix);
+
+  // Find the PTXAS compiler. The search order is:
+  // 1. The toolkit path in `targetOptions`.
+  // 2. In the system PATH.
+  // 3. The path from `getCUDAToolkitPath()`.
+  std::optional<std::string> findPtxas() const;
+
+  // Target options.
+  gpu::TargetOptions targetOptions;
+};
+} // namespace
+
+NVPTXSerializer::NVPTXSerializer(Operation &module, NVVMTargetAttr target,
+                                 const gpu::TargetOptions &targetOptions)
+    : SerializeGPUModuleBase(module, target, targetOptions),
+      targetOptions(targetOptions) {}
+
+std::optional<NVPTXSerializer::TmpFile>
+NVPTXSerializer::createTemp(StringRef name, StringRef suffix) {
+  llvm::SmallString<128> filename;
+  std::error_code ec =
+      llvm::sys::fs::createTemporaryFile(name, suffix, filename);
+  if (ec) {
+    getOperation().emitError() << "Couldn't create the temp file: `" << filename
+                               << "`, error message: " << ec.message();
+    return std::nullopt;
+  }
+  return TmpFile(filename, llvm::FileRemover(filename.c_str()));
+}
+
+std::optional<std::string> NVPTXSerializer::findPtxas() const {
+  // Find the `ptxas` compiler.
+  // 1. Check the toolkit path given in the command line.
+  StringRef pathRef = targetOptions.getToolkitPath();
+  SmallVector<char, 256> path;
+  if (pathRef.size()) {
+    path.insert(path.begin(), pathRef.begin(), pathRef.end());
+    llvm::sys::path::append(path, "bin", "ptxas");
+    if (llvm::sys::fs::can_execute(path))
+      return StringRef(path.data(), path.size()).str();
+  }
+
+  // 2. Check PATH.
+  if (std::optional<std::string> ptxasCompiler =
+          llvm::sys::Process::FindInEnvPath("PATH", "ptxas"))
+    return *ptxasCompiler;
+
+  // 3. Check `getCUDAToolkitPath()`.
+  pathRef = getCUDAToolkitPath();
+  path.clear();
+  if (pathRef.size()) {
+    path.insert(path.begin(), pathRef.begin(), pathRef.end());
+    llvm::sys::path::append(path, "bin", "ptxas");
+    if (llvm::sys::fs::can_execute(path))
+      return StringRef(path.data(), path.size()).str();
+  }
+  return std::nullopt;
+}
+
+// TODO: clean this method & have a generic tool driver, switch to the
+// `nvptxcompiler` library or never emit binaries with this mechanism and let
+// another stage take care of it.
+std::optional<SmallVector<char, 0>>
+NVPTXSerializer::compileToBinary(const std::string &ptxCode) {
+  // Find the PTXAS compiler.
+  std::optional<std::string> ptxasCompiler = findPtxas();
+  if (!ptxasCompiler) {
+    getOperation().emitError()
+        << "Couldn't find the `ptxas` compiler. Please specify the toolkit "
+           "path, add the compiler to $PATH, or set one of the environment "
+           "variables in `NVVM::getCUDAToolkitPath()`.";
+    return std::nullopt;
+  }
+
+  // Base name for all temp files: mlir-<module name>-<target triple>-<chip>.
+  std::string basename = llvm::formatv(
+      "mlir-{0}-{1}-{2}",
+      dyn_cast<gpu::GPUModuleOp>(&getOperation()).getNameAttr().getValue(),
+      getTarget().getTriple(), getTarget().getChip());
+
+  // Create temp files:
+  std::optional<TmpFile> ptxFile = createTemp(basename, "%%.ptx");
+  if (!ptxFile)
+    return std::nullopt;
+  std::optional<TmpFile> stdoutFile = createTemp(basename, "%%.stdout");
+  if (!stdoutFile)
+    return std::nullopt;
+  std::optional<TmpFile> stderrFile = createTemp(basename, "%%.stderr");
+  if (!stderrFile)
+    return std::nullopt;
+  std::optional<TmpFile> cubinFile = createTemp(basename, "%%.cubin");
+  if (!cubinFile)
+    return std::nullopt;
+
+  std::error_code ec;
+  // Dump the PTX to a temp file.
+  {
+    llvm::raw_fd_ostream ptxStream(ptxFile->first, ec);
+    if (ec) {
+      getOperation().emitError()
+          << "Couldn't open the file: `" << ptxFile->first
+          << "`, error message: " << ec.message();
+      return std::nullopt;
+    }
+    ptxStream << ptxCode;
+    if (ptxStream.has_error()) {
+      getOperation().emitError()
+          << "An error occurred while writing the PTX to: `" << ptxFile->first
+          << "`.";
+      return std::nullopt;
+    }
+    ptxStream.flush();
+  }
+
+  // Create PTX args.
+  SmallVector<StringRef, 12> ptxasArgs(
+      {StringRef("ptxas"), StringRef("-arch"), getTarget().getChip(),
+       StringRef(ptxFile->first), StringRef("-o"),
+       StringRef(cubinFile->first)});
+
+  std::pair<llvm::BumpPtrAllocator, SmallVector<const char *>> cmdOpts =
+      targetOptions.tokenizeCmdOptions();
+  for (auto arg : cmdOpts.second)
+    ptxasArgs.push_back(arg);
+
+  std::optional<StringRef> redirects[] = {
+      std::nullopt,
+      stdoutFile->first,
+      stderrFile->first,
+  };
+
+  // Invoke PTXAS.
+  std::string message;
+  if (llvm::sys::ExecuteAndWait(ptxasCompiler.value(), ptxasArgs,
+                                /*Env=*/std::nullopt,
+                                /*Redirects=*/redirects,
+                                /*SecondsToWait=*/0,
+                                /*MemoryLimit=*/0,
+                                /*ErrMsg=*/&message)) {
+    if (message.empty()) {
+      llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> ptxasStderr =
+          llvm::MemoryBuffer::getFile(stderrFile->first);
+      if (ptxasStderr)
+        getOperation().emitError()
+            << "PTXAS invocation failed. PTXAS error log:\n"
+            << ptxasStderr->get()->getBuffer();
+      else
+        getOperation().emitError() << "PTXAS invocation failed.";
+      return std::nullopt;
+    }
+    getOperation().emitError()
+        << "PTXAS invocation failed, error message: " << message;
+    return std::nullopt;
+  }
+
+// Dump the output of PTXAS, helpful if the verbose flag was passed.
+#define DEBUG_TYPE "serialize-to-binary"
+  LLVM_DEBUG({
+    llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> ptxasStdout =
+        llvm::MemoryBuffer::getFile(stdoutFile->first);
+    if (ptxasStdout && (*ptxasStdout)->getBuffer().size()) {
+      llvm::dbgs() << "PTXAS output for module: "
+                   << dyn_cast<gpu::GPUModuleOp>(&getOperation()).getNameAttr()
+                   << "\n";
+      llvm::dbgs() << (*ptxasStdout)->getBuffer() << "\n";
+      llvm::dbgs().flush();
+    }
+  });
+#undef DEBUG_TYPE
+
+  // Read the cubin file.
+  llvm::ErrorOr<std::unique_ptr<llvm::MemoryBuffer>> cubinBuffer =
+      llvm::MemoryBuffer::getFile(cubinFile->first);
+  if (!cubinBuffer) {
+    getOperation().emitError()
+        << "Couldn't open the file: `" << cubinFile->first
+        << "`, error message: " << cubinBuffer.getError().message();
+    return std::nullopt;
+  }
+  StringRef cubinStr = (*cubinBuffer)->getBuffer();
+  return SmallVector<char, 0>(cubinStr.begin(), cubinStr.end());
+}
+
+std::optional<SmallVector<char, 0>>
+NVPTXSerializer::moduleToObject(llvm::Module &llvmModule,
+                                llvm::TargetMachine &targetMachine) {
+  // Return LLVM IR if the compilation target is offload.
+#define DEBUG_TYPE "serialize-to-llvm"
+  LLVM_DEBUG({
+    llvm::dbgs() << "LLVM IR for module: "
+                 << dyn_cast<gpu::GPUModuleOp>(&getOperation()).getNameAttr()
+                 << "\n";
+    llvm::dbgs() << llvmModule << "\n";
+    llvm::dbgs().flush();
+  });
+#undef DEBUG_TYPE
+  if (targetOptions.getCompilationTarget() == gpu::TargetOptions::offload)
+    return SerializeGPUModuleBase::moduleToObject(llvmModule, targetMachine);
+
+  // Emit PTX code.
+  std::optional<std::string> serializedISA =
+      translateToISA(llvmModule, targetMachine);
+  if (!serializedISA) {
+    getOperation().emitError() << "Failed translating the module to ISA.";
+    return std::nullopt;
+  }
+#define DEBUG_TYPE "serialize-to-ptx"
+  LLVM_DEBUG({
+    llvm::dbgs() << "PTX for module: "
+                 << dyn_cast<gpu::GPUModuleOp>(&getOperation()).getNameAttr()
+                 << "\n";
+    llvm::dbgs() << *serializedISA << "\n";
+    llvm::dbgs().flush();
+  });
+#undef DEBUG_TYPE
+
+  // Return PTX if the compilation target is assembly.
+  if (targetOptions.getCompilationTarget() == gpu::TargetOptions::assembly)
+    return SmallVector<char, 0>(serializedISA->begin(), serializedISA->end());
+
+  // Compile to binary.
+  return compileToBinary(*serializedISA);
+}
+#endif // MLIR_CUDA_CONVERSIONS_ENABLED == 1
+
+std::optional<SmallVector<char, 0>>
+NVVMTargetAttrImpl::serializeToObject(Attribute attribute, Operation *module,
+                                      const gpu::TargetOptions &options) const {
+  assert(module && "The module must be non null.");
+  if (!module)
+    return std::nullopt;
+  if (!mlir::isa<gpu::GPUModuleOp>(module)) {
+    module->emitError("Module must be a GPU module.");
+    return std::nullopt;
+  }
+#if MLIR_CUDA_CONVERSIONS_ENABLED == 1
+  NVPTXSerializer serializer(*module, cast<NVVMTargetAttr>(attribute), options);
+  serializer.init();
+  return serializer.run();
+#else
+  module->emitError(
+      "The `NVPTX` target was not built. Please enable it when building LLVM.");
+  return std::nullopt;
+#endif // MLIR_CUDA_CONVERSIONS_ENABLED == 1
+}

mlir/unittests/Target/LLVM/CMakeLists.txt

 add_mlir_unittest(MLIRTargetLLVMTests
+  SerializeNVVMTarget.cpp
   SerializeToLLVMBitcode.cpp
 )
 
 llvm_map_components_to_libnames(llvm_libs nativecodegen)
 
 target_link_libraries(MLIRTargetLLVMTests
   PRIVATE
   MLIRTargetLLVM
+  MLIRNVVMTarget
+  MLIRGPUDialect
+  MLIRNVVMDialect
   MLIRLLVMDialect
   MLIRLLVMToLLVMIRTranslation
   MLIRBuiltinToLLVMIRTranslation
+  MLIRNVVMToLLVMIRTranslation
+  MLIRGPUToLLVMIRTranslation
   ${llvm_libs}
 )
 
 if (DEFINED LLVM_NATIVE_TARGET)
   target_compile_definitions(MLIRTargetLLVMTests
     PRIVATE
     -DLLVM_NATIVE_TARGET_TEST_ENABLED=1
   )
 else()
   target_compile_definitions(MLIRTargetLLVMTests
     PRIVATE
     -DLLVM_NATIVE_TARGET_TEST_ENABLED=0
   )
 endif()

mlir/unittests/Target/LLVM/SerializeNVVMTarget.cpp

+//===- SerializeNVVMTarget.cpp ----------------------------------*- C++ -*-===//
+//
+// This file is licensed under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/GPU/IR/GPUDialect.h"
+#include "mlir/Dialect/LLVMIR/NVVMDialect.h"
+#include "mlir/IR/MLIRContext.h"
+#include "mlir/InitAllDialects.h"
+#include "mlir/Parser/Parser.h"
+#include "mlir/Target/LLVM/NVVM/Target.h"
+#include "mlir/Target/LLVMIR/Dialect/Builtin/BuiltinToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/GPU/GPUToLLVMIRTranslation.h"
+#include "mlir/Target/LLVMIR/Dialect/LLVMIR/LLVMToLLVMIRTranslation.h"
+
+#include "llvm/IRReader/IRReader.h"
+#include "llvm/Support/MemoryBufferRef.h"
+#include "llvm/Support/Process.h"
+#include "llvm/Support/TargetSelect.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/TargetParser/Host.h"
+
+#include "gmock/gmock.h"
+
+using namespace mlir;
+
+// Skip the test if the NVPTX target was not built.
+#if MLIR_CUDA_CONVERSIONS_ENABLED == 0
+#define SKIP_WITHOUT_NVPTX(x) DISABLED_##x
+#else
+#define SKIP_WITHOUT_NVPTX(x) x
+#endif
+
+class MLIRTargetLLVMNVVM : public ::testing::Test {
+protected:
+  virtual void SetUp() {
+    registerBuiltinDialectTranslation(registry);
+    registerLLVMDialectTranslation(registry);
+    registerGPUDialectTranslation(registry);
+    registerNVVMTarget(registry);
+  }
+
+  // Checks if PTXAS is in PATH.
+  bool hasPtxas() {
+    // Find the `ptxas` compiler.
+    std::optional<std::string> ptxasCompiler =
+        llvm::sys::Process::FindInEnvPath("PATH", "ptxas");
+    return ptxasCompiler.has_value();
+  }
+
+  // Dialect registry.
+  DialectRegistry registry;
+
+  // MLIR module used for the tests.
+  const std::string moduleStr = R"mlir(
+      gpu.module @nvvm_test {
+        llvm.func @nvvm_kernel(%arg0: f32) attributes {gpu.kernel, nvvm.kernel} {
+        llvm.return
+      }
+    })mlir";
+};
+
+// Test NVVM serialization to LLVM.
+TEST_F(MLIRTargetLLVMNVVM, SKIP_WITHOUT_NVPTX(SerializeNVVMMToLLVM)) {
+  MLIRContext context(registry);
+
+  OwningOpRef<ModuleOp> module =
+      parseSourceString<ModuleOp>(moduleStr, &context);
+  ASSERT_TRUE(!!module);
+
+  // Create an NVVM target.
+  NVVM::NVVMTargetAttr target = NVVM::NVVMTargetAttr::get(&context);
+
+  // Serialize the module.
+  auto serializer = dyn_cast<gpu::TargetAttrInterface>(target);
+  ASSERT_TRUE(!!serializer);
+  gpu::TargetOptions options("", {}, "", gpu::TargetOptions::offload);
+  for (auto gpuModule : (*module).getBody()->getOps<gpu::GPUModuleOp>()) {
+    std::optional<SmallVector<char, 0>> object =
+        serializer.serializeToObject(gpuModule, options);
+    // Check that the serializer was successful.
+    ASSERT_TRUE(object != std::nullopt);
+    ASSERT_TRUE(object->size() > 0);
+
+    // Read the serialized module.
+    llvm::MemoryBufferRef buffer(StringRef(object->data(), object->size()),
+                                 "module");
+    llvm::LLVMContext llvmContext;
+    llvm::Expected<std::unique_ptr<llvm::Module>> llvmModule =
+        llvm::getLazyBitcodeModule(buffer, llvmContext);
+    ASSERT_TRUE(!!llvmModule);
+    ASSERT_TRUE(!!*llvmModule);
+
+    // Check that it has a function named `foo`.
+    ASSERT_TRUE((*llvmModule)->getFunction("nvvm_kernel") != nullptr);
+  }
+}
+
+// Test NVVM serialization to PTX.
+TEST_F(MLIRTargetLLVMNVVM, SKIP_WITHOUT_NVPTX(SerializeNVVMToPTX)) {
+  MLIRContext context(registry);
+
+  OwningOpRef<ModuleOp> module =
+      parseSourceString<ModuleOp>(moduleStr, &context);
+  ASSERT_TRUE(!!module);
+
+  // Create an NVVM target.
+  NVVM::NVVMTargetAttr target = NVVM::NVVMTargetAttr::get(&context);
+
+  // Serialize the module.
+  auto serializer = dyn_cast<gpu::TargetAttrInterface>(target);
+  ASSERT_TRUE(!!serializer);
+  gpu::TargetOptions options("", {}, "", gpu::TargetOptions::assembly);
+  for (auto gpuModule : (*module).getBody()->getOps<gpu::GPUModuleOp>()) {
+    std::optional<SmallVector<char, 0>> object =
+        serializer.serializeToObject(gpuModule, options);
+    // Check that the serializer was successful.
+    ASSERT_TRUE(object != std::nullopt);
+    ASSERT_TRUE(object->size() > 0);
+
+    ASSERT_TRUE(
+        StringRef(object->data(), object->size()).contains("nvvm_kernel"));
+  }
+}
+
+// Test NVVM serialization to Binary.
+TEST_F(MLIRTargetLLVMNVVM, SKIP_WITHOUT_NVPTX(SerializeNVVMToBinary)) {
+  if (!hasPtxas())
+    GTEST_SKIP() << "PTXAS compiler not found, skipping test.";
+
+  MLIRContext context(registry);
+
+  OwningOpRef<ModuleOp> module =
+      parseSourceString<ModuleOp>(moduleStr, &context);
+  ASSERT_TRUE(!!module);
+
+  // Create an NVVM target.
+  NVVM::NVVMTargetAttr target = NVVM::NVVMTargetAttr::get(&context);
+
+  // Serialize the module.
+  auto serializer = dyn_cast<gpu::TargetAttrInterface>(target);
+  ASSERT_TRUE(!!serializer);
+  gpu::TargetOptions options("", {}, "", gpu::TargetOptions::binary);
+  for (auto gpuModule : (*module).getBody()->getOps<gpu::GPUModuleOp>()) {
+    std::optional<SmallVector<char, 0>> object =
+        serializer.serializeToObject(gpuModule, options);
+    // Check that the serializer was successful.
+    ASSERT_TRUE(object != std::nullopt);
+    ASSERT_TRUE(object->size() > 0);
+  }
+}