[Mlir-commits] [mlir] [GPUToXeVMPipeline][Pipeline] Modify pipeline to convert GPU module separately to Host code. (PR #166204)

[Md Abdullah Shahneous Bari]

https://github.com/mshahneo updated https://github.com/llvm/llvm-project/pull/166204

>From 411ded5110a1b6ec5c8246dde17a02479bd1d075 Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Wed, 29 Oct 2025 20:27:44 +0000
Subject: [PATCH 1/3] Add CSE and canonicalize before final conversion to
 binary.

---
 mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
index 1a1485ba2e02c..38ddd5b42124e 100644
--- a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
+++ b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
@@ -106,6 +106,8 @@ void buildPostGPUCommonPassPipeline(
   pm.addPass(createLowerAffinePass());
   pm.addPass(createConvertToLLVMPass());
   pm.addPass(createReconcileUnrealizedCastsPass());
+  pm.addNestedPass<gpu::GPUModuleOp>(createCanonicalizerPass());
+  pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
   // gpu-module-to-binary
   {
     GpuModuleToBinaryPassOptions gpuToModuleBinOptions;

>From 1f8c24bf3f30168c21a3c81c890cc1972905905b Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Mon, 3 Nov 2025 17:48:30 +0000
Subject: [PATCH 2/3] Convert GPU code to LLVM seperately.

---
 mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp | 1 +
 1 file changed, 1 insertion(+)

diff --git a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
index 38ddd5b42124e..1c46fa518dd99 100644
--- a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
+++ b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
@@ -85,6 +85,7 @@ void buildGPUPassPipeline(OpPassManager &pm,
     pm.addNestedPass<gpu::GPUModuleOp>(
         createConvertGpuOpsToLLVMSPVOps(gpuToLLVMSPVOptions));
   }
+  pm.addNestedPass<gpu::GPUModuleOp>(createConvertToLLVMPass());
   pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
   pm.addNestedPass<gpu::GPUModuleOp>(createReconcileUnrealizedCastsPass());
 }

>From 50bb0456b5f6be3ceb00c21122c5d5e56653546a Mon Sep 17 00:00:00 2001
From: "Shahneous Bari, Md Abdullah" <md.abdullah.shahneous.bari at intel.com>
Date: Tue, 25 Nov 2025 02:56:47 +0000
Subject: [PATCH 3/3] Add `convert-vector-to-llvm` pass to the pipeline.

`convert-vector-to-llvm` pass uses some `vector` transform passes
that is not added by `convert-to-llvm`, hence this pass is necessary
for proper lowering of vector ops. As not all vector ops are directly
converted to llvm, some ops are gradually lowered to other vector ops
and those ops are in turn covnerted to llvm.
---
 mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
index 1c46fa518dd99..9cec2df14c246 100644
--- a/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
+++ b/mlir/lib/Dialect/GPU/Pipelines/GPUToXeVMPipeline.cpp
@@ -85,7 +85,6 @@ void buildGPUPassPipeline(OpPassManager &pm,
     pm.addNestedPass<gpu::GPUModuleOp>(
         createConvertGpuOpsToLLVMSPVOps(gpuToLLVMSPVOptions));
   }
-  pm.addNestedPass<gpu::GPUModuleOp>(createConvertToLLVMPass());
   pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
   pm.addNestedPass<gpu::GPUModuleOp>(createReconcileUnrealizedCastsPass());
 }
@@ -105,6 +104,7 @@ void buildPostGPUCommonPassPipeline(
     pm.addPass(createGpuToLLVMConversionPass(gpuToLLVMOptions));
   }
   pm.addPass(createLowerAffinePass());
+  pm.addPass(createConvertVectorToLLVMPass());
   pm.addPass(createConvertToLLVMPass());
   pm.addPass(createReconcileUnrealizedCastsPass());
   pm.addNestedPass<gpu::GPUModuleOp>(createCanonicalizerPass());