[Mlir-commits] [mlir] [MLIR][GPU] Lower remaining Math ops in the XeVM pipeline (PR #213874)

[llvmlistbot at llvm.org]

https://github.com/LouisLu060211 created https://github.com/llvm/llvm-project/pull/213874

`convert-math-to-xevm` only maps the Math ops that have a native or OpenCL counterpart. Anything else it doesn't cover, such as `math.isinf`, stays in the `gpu.module` through the rest of the device-side pipeline and reaches serialization, where translation to LLVM IR fails with a missing `LLVMTranslationDialectInterface` registration.

Add `convert-math-to-llvm` after the XeVM-specific conversions, mirroring what the NVVM pipeline already does. Running it last keeps the XeVM lowerings taking priority and leaves the generic path to handle the remainder.

Fixes #213583

>From 9a0508ed3b139d5a379f6f42bb344e6f144f5a69 Mon Sep 17 00:00:00 2001
From: LouisLu0602 <yaolu0602 at gmail.com>
Date: Tue, 4 Aug 2026 16:36:10 +0800
Subject: [PATCH] [MLIR][GPU] Lower remaining Math ops in the XeVM pipeline

`convert-math-to-xevm` only maps the Math ops that have a native or OpenCL
counterpart. Anything else it doesn't cover, such as `math.isinf`, stays in
the `gpu.module` through the rest of the device-side pipeline and reaches
serialization, where translation to LLVM IR fails with a missing
`LLVMTranslationDialectInterface` registration.

Add `convert-math-to-llvm` after the XeVM-specific conversions, mirroring what
the NVVM pipeline already does. Running it last keeps the XeVM lowerings taking
priority and leaves the generic path to handle the remainder.

Fixes #213583
---
 .../GPU/Pipelines/GPUToXeVMPipeline.cpp       |  6 +++++
 .../GPU/lower-to-xevm-pipeline-math.mlir      | 24 +++++++++++++++++++
 2 files changed, 30 insertions(+)
 create mode 100644 mlir/test/Dialect/GPU/lower-to-xevm-pipeline-math.mlir

diff --git a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
index 4dc9e2acfe235..3c035fb4a9ae4 100644
--- a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
+++ b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
@@ -145,6 +145,12 @@ void buildGPUPassPipeline(OpPassManager &pm,
     pm.addNestedPass<gpu::GPUModuleOp>(
         arith::createArithEmulateUnsupportedFloats(arithEmulateOptions));
   }
+  // Lower whatever Math ops are left. `convert-math-to-xevm` above only maps
+  // the ops that have a native or OpenCL counterpart; the rest (e.g.
+  // `math.isinf`) have no XeVM lowering and would otherwise survive into
+  // serialization, where translation to LLVM IR fails. This must run after the
+  // XeVM-specific conversions so those keep taking priority.
+  pm.addNestedPass<gpu::GPUModuleOp>(createConvertMathToLLVMPass());
   pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
   pm.addNestedPass<gpu::GPUModuleOp>(createReconcileUnrealizedCastsPass());
 }
diff --git a/mlir/test/Dialect/GPU/lower-to-xevm-pipeline-math.mlir b/mlir/test/Dialect/GPU/lower-to-xevm-pipeline-math.mlir
new file mode 100644
index 0000000000000..94917b398f1c0
--- /dev/null
+++ b/mlir/test/Dialect/GPU/lower-to-xevm-pipeline-math.mlir
@@ -0,0 +1,24 @@
+// RUN: mlir-opt %s --gpu-lower-to-xevm-pipeline="xegpu-op-level=lane" | FileCheck %s
+
+// `convert-math-to-xevm` only maps the Math ops that have a native or OpenCL
+// counterpart. The rest must still be lowered, or they survive into
+// serialization and translation to LLVM IR fails.
+// See https://github.com/llvm/llvm-project/issues/213583.
+
+// CHECK-LABEL: gpu.module @kernels
+// CHECK-NOT: math.
+module attributes {gpu.container_module} {
+  gpu.module @kernels {
+    gpu.func @classify_inf(%input : memref<1xf32>,
+                           %output : memref<1xi32>) kernel {
+      %c0 = arith.constant 0 : index
+      %zero = arith.constant 0 : i32
+      %one = arith.constant 1 : i32
+      %value = memref.load %input[%c0] : memref<1xf32>
+      %is_inf = math.isinf %value : f32
+      %result = arith.select %is_inf, %one, %zero : i32
+      memref.store %result, %output[%c0] : memref<1xi32>
+      gpu.return
+    }
+  }
+}