[mlir] Lower math dialect later in gpu-lower-to-nvvm-pipeline

[grypp]

This PR moves lowering of math dialect later in the pipeline. Because math dialect is lowered correctly by createConvertGpuOpsToNVVMOps for GPU target, and it needs to run it first.

Reland #78556

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-gpu

Author: Guray Ozen (grypp)

Changes

This PR moves lowering of math dialect later in the pipeline. Because math dialect is lowered correctly by createConvertGpuOpsToNVVMOps for GPU target, and it needs to run it first.

Reland #78556

---

Full diff: https://github.com/llvm/llvm-project/pull/81489.diff

2 Files Affected:

• (modified) mlir/lib/Dialect/GPU/Pipelines/GPUToNVVMPipeline.cpp (+1-1)
• (added) mlir/test/Dialect/GPU/test-nvvm-pipeline.mlir (+30)

diff --git a/mlir/lib/Dialect/GPU/Pipelines/GPUToNVVMPipeline.cpp b/mlir/lib/Dialect/GPU/Pipelines/GPUToNVVMPipeline.cpp
index 0b4739214bf2f1..935f0deaf9c8a6 100644
--- a/mlir/lib/Dialect/GPU/Pipelines/GPUToNVVMPipeline.cpp
+++ b/mlir/lib/Dialect/GPU/Pipelines/GPUToNVVMPipeline.cpp
@@ -51,7 +51,6 @@ void buildCommonPassPipeline(
   pm.addPass(createConvertVectorToSCFPass());
   pm.addPass(createConvertSCFToCFPass());
   pm.addPass(createConvertNVVMToLLVMPass());
-  pm.addPass(createConvertMathToLLVMPass());
   pm.addPass(createConvertFuncToLLVMPass());
   pm.addPass(memref::createExpandStridedMetadataPass());
 
@@ -98,6 +97,7 @@ void buildHostPostPipeline(OpPassManager &pm,
   GpuModuleToBinaryPassOptions gpuModuleToBinaryPassOptions;
   gpuModuleToBinaryPassOptions.compilationTarget = options.cubinFormat;
   pm.addPass(createGpuModuleToBinaryPass(gpuModuleToBinaryPassOptions));
+  pm.addPass(createConvertMathToLLVMPass());
   pm.addPass(createCanonicalizerPass());
   pm.addPass(createCSEPass());
   pm.addPass(createReconcileUnrealizedCastsPass());
diff --git a/mlir/test/Dialect/GPU/test-nvvm-pipeline.mlir b/mlir/test/Dialect/GPU/test-nvvm-pipeline.mlir
new file mode 100644
index 00000000000000..07e719798b8511
--- /dev/null
+++ b/mlir/test/Dialect/GPU/test-nvvm-pipeline.mlir
@@ -0,0 +1,30 @@
+// REQUIRES: host-supports-nvptx
+// RUN: mlir-opt %s \
+// RUN:  | mlir-opt -gpu-lower-to-nvvm-pipeline="cubin-format=isa" \
+// RUN:   | FileCheck %s
+
+// RUN: mlir-opt %s \
+// RUN:  | mlir-opt -gpu-lower-to-nvvm-pipeline="cubin-format=isa" \
+// RUN:    --mlir-print-ir-after=convert-gpu-to-nvvm 2>&1 \
+// RUN:  | FileCheck %s --check-prefixes=CHECK-NVVM
+
+// This test checks whether the GPU region is compiled correctly to PTX by 
+// pipeline. It doesn't test IR for GPU side, but it can test Host IR and 
+// generated PTX.
+
+// CHECK-LABEL: llvm.func @test_math(%arg0: f32) {
+func.func @test_math(%arg0 : f32) {
+    %c2 = arith.constant 2 : index
+    %c1 = arith.constant 1 : index
+    // CHECK: gpu.launch_func  @test_math_kernel::@test_math_kernel
+    // CHECK: gpu.binary @test_math_kernel  [#gpu.object<#nvvm.target
+    gpu.launch 
+        blocks(%0, %1, %2) in (%3 = %c1, %4 = %c1, %5 = %c1) 
+        threads(%6, %7, %8) in (%9 = %c2, %10 = %c1, %11 = %c1) { 
+        // CHECK-NVVM: __nv_expf 
+        %s1 = math.exp %arg0 : f32
+        gpu.printf "%f" %s1 : f32
+        gpu.terminator
+    }
+    return
+}
\ No newline at end of file