Unify pipeline for XeGPU tests

intel · Oct 10, 2024 · 97bd8aa · 97bd8aa
1 parent 9d73c3c
commit 97bd8aa
Show file tree

Hide file tree

Showing 6 changed files with 6 additions and 64 deletions.
diff --git a/lib/Transforms/VnniTransformation.cpp b/lib/Transforms/VnniTransformation.cpp
@@ -386,7 +386,7 @@ static void updateDpasOp(mlir::OpBuilder &builder, mlir::xegpu::DpasOp &op,
   auto rhs = op.getRhs();
   // B operand of DPAS has multiple uses and
   // the rest uses don't use vnni.
-  if (!analysis.getLayout(rhs)) {
+  if (!analysis.getLayout(rhs) && isVNNIApplicable(rhs.getType())) {
     builder.setInsertionPoint(op);
     auto cast = mlir::cast<mlir::TypedValue<mlir::VectorType>>(rhs);
     auto &&[newRhs, root] = applyVnniTransform(builder, cast);

diff --git a/.../VnniTransform/gemm_with_extract_e2e.mlir → .../Dialect/XeGPU/gemm_with_extract_e2e.mlir b/.../VnniTransform/gemm_with_extract_e2e.mlir → .../Dialect/XeGPU/gemm_with_extract_e2e.mlir
diff --git a/test/Integration/Dialect/XeGPU/optimize_transpose.mlir b/test/Integration/Dialect/XeGPU/optimize_transpose.mlir
@@ -1,8 +1,8 @@
-// RUN: %python_executable %imex_runner --requires=l0-runtime -i %s --pass-pipeline-file=%p/xegpu-to-func-vc-optimize-transpose.pp \
+// RUN: %python_executable %imex_runner --requires=l0-runtime -i %s --pass-pipeline-file=%p/xegpu-to-func-vc.pp \
 // RUN:                                       --runner imex-cpu-runner -e main \
 // RUN:                                       --entry-point-result=void \
 // RUN:                                       --shared-libs=%irunner_utils,%mlir_runner_utils,%mlir_c_runner_utils,%levelzero_runtime --filecheck
-// RUN: %python_executable %imex_runner --requires=sycl-runtime -i %s --pass-pipeline-file=%p/xegpu-to-func-vc-optimize-transpose.pp \
+// RUN: %python_executable %imex_runner --requires=sycl-runtime -i %s --pass-pipeline-file=%p/xegpu-to-func-vc.pp \
 // RUN:                                        --runner imex-cpu-runner -e main \
 // RUN:                                        --entry-point-result=void \
 // RUN:                                        --shared-libs=%irunner_utils,%mlir_runner_utils,%mlir_c_runner_utils,%sycl_runtime --filecheck

diff --git a/test/Integration/Dialect/XeGPU/xegpu-to-func-vc-optimize-transpose.pp b/test/Integration/Dialect/XeGPU/xegpu-to-func-vc-optimize-transpose.pp
diff --git a/test/Integration/Dialect/XeGPU/xegpu-to-func-vc.pp b/test/Integration/Dialect/XeGPU/xegpu-to-func-vc.pp
@@ -3,6 +3,9 @@
 // spirv dialect (for device code) lowering pipeline.
 // Ready for imex runner starting from GPU dialect.
 builtin.module(
+    gpu.module(imex-xegpu-hoist-transpose,
+        imex-xegpu-apply-vnni-transformation,
+        imex-xegpu-optimize-transpose)
     imex-vector-linearize
     gpu.module(convert-xegpu-to-vc)
     cse

diff --git a/test/Transforms/VnniTransform/xegpu-to-func-vc.pp b/test/Transforms/VnniTransform/xegpu-to-func-vc.pp