Reland [mlir][test][gpu] Migrate CUDA tests to the TargetAttr compilation workflow (llvm#65768)

The revert happened due to a build bot failure that threw 'CUDA_ERROR_UNSUPPORTED_PTX_VERSION'.
The failure's root cause was a pass using "+ptx76" for compilation and an old CUDA driver
on the bot. This commit relands the patch with "+ptx60".

Original Gh PR: #65768
Original commit message:
    Migrate tests referencing `gpu-to-cubin` to the new compilation workflow
    using `TargetAttrs`. The `test-lower-to-nvvm` pass pipeline was modified
    to use the new compilation workflow to simplify the introduction of
    future tests.

    The `createLowerGpuOpsToNVVMOpsPass` function was removed, as it didn't
    allow for passing all options available in the `ConvertGpuOpsToNVVMOp`
    pass.

25 files changed, 63 insertions, 133 deletions

diff --git a/mlir/include/mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h b/mlir/include/mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h
index 46f29c6..e0f4c71 100644
--- a/mlir/include/mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h
+++ b/mlir/include/mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h
@@ -16,9 +16,7 @@ namespace mlir {
 class LLVMTypeConverter;
 class ConversionTarget;
 class RewritePatternSet;
-
-template <typename OpT>
-class OperationPass;
+class Pass;
 
 namespace gpu {
 class GPUModuleOp;
@@ -45,14 +43,6 @@ void populateGpuSubgroupReduceOpLoweringPattern(LLVMTypeConverter &converter,
 /// Collect a set of patterns to convert WMMA ops from GPU dialect to NVVM.
 void populateGpuWMMAToNVVMConversionPatterns(LLVMTypeConverter &converter,
                                              RewritePatternSet &patterns);
-
-/// Creates a pass that lowers GPU dialect operations to NVVM counterparts. The
-/// index bitwidth used for the lowering of the device side index computations
-/// is configurable.
-std::unique_ptr<OperationPass<gpu::GPUModuleOp>> createLowerGpuOpsToNVVMOpsPass(
-    unsigned indexBitwidth = kDeriveIndexBitwidthFromDataLayout,
-    bool hasRedux = false);
-
 } // namespace mlir
 
 #endif // MLIR_CONVERSION_GPUTONVVM_GPUTONVVMPASS_H_
diff --git a/mlir/include/mlir/Conversion/Passes.td b/mlir/include/mlir/Conversion/Passes.td
index ed37abf..3218760 100644
--- a/mlir/include/mlir/Conversion/Passes.td
+++ b/mlir/include/mlir/Conversion/Passes.td
@@ -486,7 +486,6 @@ def LowerHostCodeToLLVMPass : Pass<"lower-host-to-llvm", "ModuleOp"> {
 
 def ConvertGpuOpsToNVVMOps : Pass<"convert-gpu-to-nvvm", "gpu::GPUModuleOp"> {
   let summary = "Generate NVVM operations for gpu operations";
-  let constructor = "mlir::createLowerGpuOpsToNVVMOpsPass()";
   let dependentDialects = [
     "cf::ControlFlowDialect",
     "memref::MemRefDialect",
diff --git a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
index 06469dc..764b6a7 100644
--- a/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
+++ b/mlir/lib/Conversion/GPUToNVVM/LowerGpuOpsToNVVMOps.cpp
@@ -210,11 +210,7 @@ struct GPULaneIdOpToNVVM : ConvertOpToLLVMPattern<gpu::LaneIdOp> {
 /// code.
 struct LowerGpuOpsToNVVMOpsPass
     : public impl::ConvertGpuOpsToNVVMOpsBase<LowerGpuOpsToNVVMOpsPass> {
-  LowerGpuOpsToNVVMOpsPass() = default;
-  LowerGpuOpsToNVVMOpsPass(unsigned indexBitwidth, bool hasRedux = false) {
-    this->indexBitwidth = indexBitwidth;
-    this->hasRedux = hasRedux;
-  }
+  using Base::Base;
 
   void runOnOperation() override {
     gpu::GPUModuleOp m = getOperation();
@@ -378,8 +374,3 @@ void mlir::populateGpuToNVVMConversionPatterns(LLVMTypeConverter &converter,
                                    "__nv_tanh");
   populateOpPatterns<math::TanOp>(converter, patterns, "__nv_tanf", "__nv_tan");
 }
-
-std::unique_ptr<OperationPass<gpu::GPUModuleOp>>
-mlir::createLowerGpuOpsToNVVMOpsPass(unsigned indexBitwidth, bool hasRedux) {
-  return std::make_unique<LowerGpuOpsToNVVMOpsPass>(indexBitwidth, hasRedux);
-}
diff --git a/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp b/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp
index a7fd5a2..24c4c4c4 100644
--- a/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp
+++ b/mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp
@@ -52,7 +52,7 @@ void mlir::sparse_tensor::buildSparseCompiler(
     pm.addPass(createSparseGPUCodegenPass());
     pm.addNestedPass<gpu::GPUModuleOp>(createStripDebugInfoPass());
     pm.addNestedPass<gpu::GPUModuleOp>(createConvertSCFToCFPass());
-    pm.addNestedPass<gpu::GPUModuleOp>(createLowerGpuOpsToNVVMOpsPass());
+    pm.addNestedPass<gpu::GPUModuleOp>(createConvertGpuOpsToNVVMOps());
   }
 
   // TODO(springerm): Add sparse support to the BufferDeallocation pass and add
diff --git a/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/dump-ptx.mlir b/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/dump-ptx.mlir
index 2c1ae3e..0cb06b7 100644
--- a/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/dump-ptx.mlir
+++ b/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/dump-ptx.mlir
@@ -1,6 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin{dump-ptx}))' \
+// RUN: | mlir-opt -test-lower-to-nvvm -debug-only=serialize-to-isa \
 // RUN: 2>&1 | FileCheck %s
 
 // CHECK: Generated by LLVM NVPTX Back-End
diff --git a/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-mma-2-4-f16.mlir b/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-mma-2-4-f16.mlir
index 8eb90fd..80972f2 100644
--- a/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-mma-2-4-f16.mlir
+++ b/mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-mma-2-4-f16.mlir
@@ -2,10 +2,9 @@
 // NOTE: this test requires gpu-sm80
 //
 // RUN: mlir-opt \
-// RUN: --pass-pipeline="builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,convert-nvgpu-to-nvvm,affine-expand-index-ops,lower-affine,convert-arith-to-llvm),convert-vector-to-llvm,canonicalize,cse,gpu.module(gpu-to-cubin{chip=sm_80 features=+ptx71}))" \
+// RUN: --pass-pipeline="builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,convert-nvgpu-to-nvvm,affine-expand-index-ops,lower-affine,convert-arith-to-llvm),convert-vector-to-llvm,canonicalize,cse)" \
 // RUN: %s \
-// RUN: | mlir-opt --convert-vector-to-scf --convert-scf-to-cf -convert-cf-to-llvm --convert-vector-to-llvm \
-// RUN:            --convert-arith-to-llvm --gpu-to-llvm --reconcile-unrealized-casts \
+// RUN: | mlir-opt --test-lower-to-nvvm="cubin-chip=sm_80 cubin-features=+ptx71" \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_c_runner_utils \
diff --git a/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-reduction-distribute.mlir b/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-reduction-distribute.mlir
index 8571c5ca..8c99149 100644
--- a/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-reduction-distribute.mlir
+++ b/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-reduction-distribute.mlir
@@ -1,9 +1,7 @@
 // RUN: mlir-opt %s -test-vector-warp-distribute="hoist-uniform distribute-transfer-write propagate-distribution" -canonicalize |\
 // RUN: mlir-opt -test-vector-warp-distribute=rewrite-warp-ops-to-scf-if |\
 // RUN: mlir-opt -lower-affine -convert-vector-to-scf -convert-scf-to-cf -convert-vector-to-llvm \
-// RUN:  -convert-arith-to-llvm -gpu-kernel-outlining |\
-// RUN: mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,reconcile-unrealized-casts,gpu-to-cubin))' |\
-// RUN: mlir-opt -gpu-to-llvm -reconcile-unrealized-casts |\
+// RUN:  -convert-arith-to-llvm -test-lower-to-nvvm | \
 // RUN: mlir-cpu-runner -e main -entry-point-result=void \
 // RUN:   -shared-libs=%mlir_cuda_runtime \
 // RUN:   -shared-libs=%mlir_c_runner_utils \
diff --git a/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-warp-distribute.mlir b/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-warp-distribute.mlir
index c671c18..f26c18c 100644
--- a/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-warp-distribute.mlir
+++ b/mlir/test/Integration/Dialect/Vector/GPU/CUDA/test-warp-distribute.mlir
@@ -2,9 +2,7 @@
 // everything on the same thread.
 // RUN: mlir-opt %s -test-vector-warp-distribute=rewrite-warp-ops-to-scf-if -canonicalize | \
 // RUN: mlir-opt -convert-vector-to-scf -convert-scf-to-cf -convert-cf-to-llvm -convert-vector-to-llvm -convert-arith-to-llvm \
-// RUN:  -gpu-kernel-outlining |\
-// RUN: mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,reconcile-unrealized-casts,gpu-to-cubin))' |\
-// RUN: mlir-opt -gpu-to-llvm -reconcile-unrealized-casts |\
+// RUN:  -test-lower-to-nvvm | \
 // RUN: mlir-cpu-runner -e main -entry-point-result=void \
 // RUN:   -shared-libs=%mlir_cuda_runtime \
 // RUN:   -shared-libs=%mlir_c_runner_utils \
@@ -15,9 +13,7 @@
 // RUN: mlir-opt %s  -test-vector-warp-distribute="hoist-uniform distribute-transfer-write" \
 // RUN:   -test-vector-warp-distribute=rewrite-warp-ops-to-scf-if -canonicalize | \
 // RUN: mlir-opt -convert-vector-to-scf -convert-scf-to-cf -convert-cf-to-llvm -convert-vector-to-llvm -convert-arith-to-llvm \
-// RUN:  -gpu-kernel-outlining |\
-// RUN: mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,reconcile-unrealized-casts,gpu-to-cubin))' |\
-// RUN: mlir-opt -gpu-to-llvm -reconcile-unrealized-casts |\
+// RUN:  -test-lower-to-nvvm | \
 // RUN: mlir-cpu-runner -e main -entry-point-result=void \
 // RUN:   -shared-libs=%mlir_cuda_runtime \
 // RUN:   -shared-libs=%mlir_c_runner_utils \
@@ -27,9 +23,7 @@
 // RUN: mlir-opt %s  -test-vector-warp-distribute="hoist-uniform distribute-transfer-write propagate-distribution" \
 // RUN:   -test-vector-warp-distribute=rewrite-warp-ops-to-scf-if -canonicalize | \
 // RUN: mlir-opt -convert-vector-to-scf -convert-scf-to-cf -convert-cf-to-llvm -convert-vector-to-llvm -convert-arith-to-llvm \
-// RUN:  -gpu-kernel-outlining |\
-// RUN: mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,reconcile-unrealized-casts,gpu-to-cubin))' |\
-// RUN: mlir-opt -gpu-to-llvm -reconcile-unrealized-casts |\
+// RUN:  -test-lower-to-nvvm | \
 // RUN: mlir-cpu-runner -e main -entry-point-result=void \
 // RUN:   -shared-libs=%mlir_cuda_runtime \
 // RUN:   -shared-libs=%mlir_c_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f16.mlir b/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f16.mlir
index 535ba52..591bf1b 100644
--- a/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f16.mlir
+++ b/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f16.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin{chip=sm_70}))' \
-// RUN: | mlir-opt --convert-scf-to-cf -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm="cubin-chip=sm_70" \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32-bare-ptr.mlir b/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32-bare-ptr.mlir
index c4ca465..51bd23f 100644
--- a/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32-bare-ptr.mlir
+++ b/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32-bare-ptr.mlir
@@ -3,9 +3,7 @@
 // Similar to the wmma-matmul-f32 but but with the memref bare pointer lowering convention.
 // This test also uses gpu.memcpy operations (instead of gpu.host_register).
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm{use-bare-ptr-memref-call-conv=1},gpu-to-cubin{chip=sm_70}))' \
-// RUN: | mlir-opt --convert-scf-to-cf -gpu-to-llvm="use-bare-pointers-for-host=1 use-bare-pointers-for-kernels=1" \
+// RUN: | mlir-opt -test-lower-to-nvvm="host-bare-ptr-calling-convention=1 kernel-bare-ptr-calling-convention=1 cubin-chip=sm_70" \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --entry-point-result=void \
diff --git a/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32.mlir b/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32.mlir
index ae410dc..0307b3d 100644
--- a/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32.mlir
+++ b/mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f32.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin{chip=sm_70}))' \
-// RUN: | mlir-opt --convert-scf-to-cf -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm="cubin-chip=sm_70" \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/all-reduce-and.mlir b/mlir/test/Integration/GPU/CUDA/all-reduce-and.mlir
index f4324a1..b131b86 100644
--- a/mlir/test/Integration/GPU/CUDA/all-reduce-and.mlir
+++ b/mlir/test/Integration/GPU/CUDA/all-reduce-and.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
@@ -10,9 +8,7 @@
 
 // Same as above but with the memref bare pointer lowering convention.
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm{use-bare-ptr-memref-call-conv=1},gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm="use-bare-pointers-for-kernels=1" \
+// RUN: | mlir-opt -test-lower-to-nvvm="kernel-bare-ptr-calling-convention=1" \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/all-reduce-max.mlir b/mlir/test/Integration/GPU/CUDA/all-reduce-max.mlir
index 0a8d38f..155423db 100644
--- a/mlir/test/Integration/GPU/CUDA/all-reduce-max.mlir
+++ b/mlir/test/Integration/GPU/CUDA/all-reduce-max.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/all-reduce-min.mlir b/mlir/test/Integration/GPU/CUDA/all-reduce-min.mlir
index bcd785d..e5047b6 100644
--- a/mlir/test/Integration/GPU/CUDA/all-reduce-min.mlir
+++ b/mlir/test/Integration/GPU/CUDA/all-reduce-min.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/all-reduce-op.mlir b/mlir/test/Integration/GPU/CUDA/all-reduce-op.mlir
index aa4b0e8..163e9fd 100644
--- a/mlir/test/Integration/GPU/CUDA/all-reduce-op.mlir
+++ b/mlir/test/Integration/GPU/CUDA/all-reduce-op.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/all-reduce-or.mlir b/mlir/test/Integration/GPU/CUDA/all-reduce-or.mlir
index 2e7d046..381db26 100644
--- a/mlir/test/Integration/GPU/CUDA/all-reduce-or.mlir
+++ b/mlir/test/Integration/GPU/CUDA/all-reduce-or.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/all-reduce-region.mlir b/mlir/test/Integration/GPU/CUDA/all-reduce-region.mlir
index 32cfa27..23c6c11 100644
--- a/mlir/test/Integration/GPU/CUDA/all-reduce-region.mlir
+++ b/mlir/test/Integration/GPU/CUDA/all-reduce-region.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/all-reduce-xor.mlir b/mlir/test/Integration/GPU/CUDA/all-reduce-xor.mlir
index 30767b9..3c5a100 100644
--- a/mlir/test/Integration/GPU/CUDA/all-reduce-xor.mlir
+++ b/mlir/test/Integration/GPU/CUDA/all-reduce-xor.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/async.mlir b/mlir/test/Integration/GPU/CUDA/async.mlir
index e6dd91a..d2a5127 100644
--- a/mlir/test/Integration/GPU/CUDA/async.mlir
+++ b/mlir/test/Integration/GPU/CUDA/async.mlir
@@ -1,7 +1,7 @@
 // RUN: mlir-opt %s \
 // RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-async-region -gpu-to-llvm \
+// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm),nvvm-attach-target)' \
+// RUN: | mlir-opt -gpu-async-region -gpu-to-llvm -gpu-module-to-binary \
 // RUN: | mlir-opt -async-to-async-runtime -async-runtime-ref-counting \
 // RUN: | mlir-opt -convert-async-to-llvm -convert-func-to-llvm \
 // RUN: | mlir-cpu-runner \
diff --git a/mlir/test/Integration/GPU/CUDA/gpu-to-cubin.mlir b/mlir/test/Integration/GPU/CUDA/gpu-to-cubin.mlir
index afcb674..a5d04f73 100644
--- a/mlir/test/Integration/GPU/CUDA/gpu-to-cubin.mlir
+++ b/mlir/test/Integration/GPU/CUDA/gpu-to-cubin.mlir
@@ -1,8 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -convert-vector-to-scf -convert-scf-to-cf -convert-cf-to-llvm -convert-vector-to-llvm -convert-arith-to-llvm \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm -reconcile-unrealized-casts \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/multiple-all-reduce.mlir b/mlir/test/Integration/GPU/CUDA/multiple-all-reduce.mlir
index 444e287..7657bf4 100644
--- a/mlir/test/Integration/GPU/CUDA/multiple-all-reduce.mlir
+++ b/mlir/test/Integration/GPU/CUDA/multiple-all-reduce.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/printf.mlir b/mlir/test/Integration/GPU/CUDA/printf.mlir
index fce7739..1a35d1e 100644
--- a/mlir/test/Integration/GPU/CUDA/printf.mlir
+++ b/mlir/test/Integration/GPU/CUDA/printf.mlir
@@ -1,6 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/shuffle.mlir b/mlir/test/Integration/GPU/CUDA/shuffle.mlir
index 6a784ca..40fcea8 100644
--- a/mlir/test/Integration/GPU/CUDA/shuffle.mlir
+++ b/mlir/test/Integration/GPU/CUDA/shuffle.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/Integration/GPU/CUDA/two-modules.mlir b/mlir/test/Integration/GPU/CUDA/two-modules.mlir
index 5f6e5d7..5a9acdf 100644
--- a/mlir/test/Integration/GPU/CUDA/two-modules.mlir
+++ b/mlir/test/Integration/GPU/CUDA/two-modules.mlir
@@ -1,7 +1,5 @@
 // RUN: mlir-opt %s \
-// RUN: | mlir-opt -gpu-kernel-outlining \
-// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm,gpu-to-cubin))' \
-// RUN: | mlir-opt -gpu-to-llvm \
+// RUN: | mlir-opt -test-lower-to-nvvm \
 // RUN: | mlir-cpu-runner \
 // RUN:   --shared-libs=%mlir_cuda_runtime \
 // RUN:   --shared-libs=%mlir_runner_utils \
diff --git a/mlir/test/lib/Dialect/GPU/TestLowerToNVVM.cpp b/mlir/test/lib/Dialect/GPU/TestLowerToNVVM.cpp
index 5db6f56..5d0c420 100644
--- a/mlir/test/lib/Dialect/GPU/TestLowerToNVVM.cpp
+++ b/mlir/test/lib/Dialect/GPU/TestLowerToNVVM.cpp
@@ -65,11 +65,11 @@ struct TestLowerToNVVMOptions
       llvm::cl::init("nvptx64-nvidia-cuda")};
   PassOptions::Option<std::string> cubinChip{
       *this, "cubin-chip", llvm::cl::desc("Chip to use to serialize to cubin."),
-      llvm::cl::init("sm_80")};
+      llvm::cl::init("sm_50")};
   PassOptions::Option<std::string> cubinFeatures{
       *this, "cubin-features",
       llvm::cl::desc("Features to use to serialize to cubin."),
-      llvm::cl::init("+ptx76")};
+      llvm::cl::init("+ptx60")};
 };
 
 //===----------------------------------------------------------------------===//
@@ -126,13 +126,14 @@ void buildGpuPassPipeline(OpPassManager &pm,
 
   // TODO: C++20 designated initializers.
   // The following pass is inconsistent.
-  // ConvertGpuOpsToNVVMOpsOptions convertGpuOpsToNVVMOpsOptions;
-  // convertGpuOpsToNVVMOpsOptions.indexBitwidth =
-  //   options.kernelIndexBitWidth;
+  // TODO: fix inconsistence.
+  ConvertGpuOpsToNVVMOpsOptions convertGpuOpsToNVVMOpsOptions;
+  convertGpuOpsToNVVMOpsOptions.useBarePtrCallConv =
+      options.kernelUseBarePtrCallConv;
+  convertGpuOpsToNVVMOpsOptions.indexBitwidth = options.kernelIndexBitWidth;
+  convertGpuOpsToNVVMOpsOptions.useOpaquePointers = true;
   pm.addNestedPass<gpu::GPUModuleOp>(
-      // TODO: fix inconsistence.
-      createLowerGpuOpsToNVVMOpsPass(/*indexBitWidth=*/
-                                     options.kernelIndexBitWidth));
+      createConvertGpuOpsToNVVMOps(convertGpuOpsToNVVMOpsOptions));
 
   // TODO: C++20 designated initializers.
   ConvertNVGPUToNVVMPassOptions convertNVGPUToNVVMPassOptions;
@@ -141,22 +142,6 @@ void buildGpuPassPipeline(OpPassManager &pm,
       createConvertNVGPUToNVVMPass(convertNVGPUToNVVMPassOptions));
   pm.addNestedPass<gpu::GPUModuleOp>(createConvertSCFToCFPass());
 
-  // TODO: C++20 designated initializers.
-  GpuToLLVMConversionPassOptions gpuToLLVMConversionOptions;
-  // Note: hostBarePtrCallConv must be false for now otherwise
-  // gpu::HostRegister is ill-defined: it wants unranked memrefs but can't
-  // lower the to bare ptr.
-  gpuToLLVMConversionOptions.hostBarePtrCallConv =
-      options.hostUseBarePtrCallConv;
-  gpuToLLVMConversionOptions.kernelBarePtrCallConv =
-      options.kernelUseBarePtrCallConv;
-  gpuToLLVMConversionOptions.useOpaquePointers = true;
-
-  // TODO: something useful here.
-  // gpuToLLVMConversionOptions.gpuBinaryAnnotation = "";
-  pm.addNestedPass<gpu::GPUModuleOp>(
-      createGpuToLLVMConversionPass(gpuToLLVMConversionOptions));
-
   // Convert vector to LLVM (always needed).
   // TODO: C++20 designated initializers.
   ConvertVectorToLLVMPassOptions convertVectorToLLVMPassOptions;
@@ -170,11 +155,6 @@ void buildGpuPassPipeline(OpPassManager &pm,
 
   // Finally we can reconcile unrealized casts.
   pm.addNestedPass<gpu::GPUModuleOp>(createReconcileUnrealizedCastsPass());
-
-#if MLIR_GPU_TO_CUBIN_PASS_ENABLE
-  pm.addNestedPass<gpu::GPUModuleOp>(createGpuSerializeToCubinPass(
-      options.cubinTriple, options.cubinChip, options.cubinFeatures));
-#endif // MLIR_GPU_TO_CUBIN_PASS_ENABLE
 }
 
 void buildLowerToNVVMPassPipeline(OpPassManager &pm,
@@ -251,22 +231,16 @@ void buildLowerToNVVMPassPipeline(OpPassManager &pm,
   //===----------------------------------------------------------------------===//
   // Host post-GPUModule-specific stuff.
   //===----------------------------------------------------------------------===//
-  // Convert vector to LLVM (always needed).
+  // Attach an NVVM target to all the GPU modules with the provided target
+  // options.
   // TODO: C++20 designated initializers.
-  ConvertVectorToLLVMPassOptions convertVectorToLLVMPassOptions;
-  convertVectorToLLVMPassOptions.reassociateFPReductions = true;
-  pm.addNestedPass<func::FuncOp>(
-      createConvertVectorToLLVMPass(convertVectorToLLVMPassOptions));
+  GpuNVVMAttachTargetOptions nvvmTargetOptions;
+  nvvmTargetOptions.triple = options.cubinTriple;
+  nvvmTargetOptions.chip = options.cubinChip;
+  nvvmTargetOptions.features = options.cubinFeatures;
+  pm.addPass(createGpuNVVMAttachTarget(nvvmTargetOptions));
 
-  ConvertIndexToLLVMPassOptions convertIndexToLLVMPassOpt3;
-  // Must be 64b on the host, things don't compose properly around
-  // gpu::LaunchOp and gpu::HostRegisterOp.
-  // TODO: fix GPU layering.
-  convertIndexToLLVMPassOpt3.indexBitwidth = options.hostIndexBitWidth;
-  pm.addPass(createConvertIndexToLLVMPass(convertIndexToLLVMPassOpt3));
-
-  // This must happen after cubin translation otherwise gpu.launch_func is
-  // illegal if no cubin annotation is present.
+  // Convert GPU to LLVM.
   // TODO: C++20 designated initializers.
   GpuToLLVMConversionPassOptions gpuToLLVMConversionOptions;
   // Note: hostBarePtrCallConv must be false for now otherwise
@@ -277,10 +251,28 @@ void buildLowerToNVVMPassPipeline(OpPassManager &pm,
   gpuToLLVMConversionOptions.kernelBarePtrCallConv =
       options.kernelUseBarePtrCallConv;
   gpuToLLVMConversionOptions.useOpaquePointers = true;
+
   // TODO: something useful here.
   // gpuToLLVMConversionOptions.gpuBinaryAnnotation = "";
   pm.addPass(createGpuToLLVMConversionPass(gpuToLLVMConversionOptions));
 
+  // Serialize all GPU modules to binaries.
+  pm.addPass(createGpuModuleToBinaryPass());
+
+  // Convert vector to LLVM (always needed).
+  // TODO: C++20 designated initializers.
+  ConvertVectorToLLVMPassOptions convertVectorToLLVMPassOptions;
+  convertVectorToLLVMPassOptions.reassociateFPReductions = true;
+  pm.addNestedPass<func::FuncOp>(
+      createConvertVectorToLLVMPass(convertVectorToLLVMPassOptions));
+
+  ConvertIndexToLLVMPassOptions convertIndexToLLVMPassOpt3;
+  // Must be 64b on the host, things don't compose properly around
+  // gpu::LaunchOp and gpu::HostRegisterOp.
+  // TODO: fix GPU layering.
+  convertIndexToLLVMPassOpt3.indexBitwidth = options.hostIndexBitWidth;
+  pm.addPass(createConvertIndexToLLVMPass(convertIndexToLLVMPassOpt3));
+
   // Convert Func to LLVM (always needed).
   // TODO: C++20 designated initializers.
   ConvertFuncToLLVMPassOptions convertFuncToLLVMPassOptions2;