[mlir][sparse] Add another call to ConvertVectorToLLVMPass, to lower the vector operations added by ConvertMathToLLVMPass.

[bixia]

Summary

Run sparse_tanh with vectorization.

[aartbik]

mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp

82

we had this at L75, was that the wrong place, or do we really need it twice?

[bixia]

mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp

82

I can't move line75 to after line76, ConvertVectorToLLVM left some ops unconverted when I did that.
It seems that MemRefToLLVM relies on ConvertVectorToLLVM.

[aartbik]

mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp

82

Can we document that. Even a

pm.addPass(createConvertVectorToLLVMPass(options.lowerVectorToLLVMOptions())); // repeat

or so would flag that we at least thought about having the same pass twice

[bixia]

Add document for the repeated pass.

mlir/lib/Dialect/SparseTensor/Pipelines/SparseTensorPipelines.cpp

void mlir::sparse_tensor::buildSparseCompiler(
   pm.addPass(createLowerAffinePass());
   pm.addPass(createConvertVectorToLLVMPass(options.lowerVectorToLLVMOptions()));
   pm.addPass(createMemRefToLLVMConversionPass());
   pm.addNestedPass<func::FuncOp>(createConvertComplexToStandardPass());
   pm.addNestedPass<mlir::func::FuncOp>(mlir::arith::createArithExpandOpsPass());
   pm.addNestedPass<func::FuncOp>(createConvertMathToLLVMPass());
   pm.addPass(createConvertMathToLibmPass());
   pm.addPass(createConvertComplexToLibmPass());
+  // Repeat convert-vector-to-llvm.
+  pm.addPass(createConvertVectorToLLVMPass(options.lowerVectorToLLVMOptions()));
   pm.addPass(createConvertComplexToLLVMPass());
   pm.addPass(createConvertFuncToLLVMPass());
   pm.addPass(createReconcileUnrealizedCastsPass());
 }
 
 //===----------------------------------------------------------------------===//
 // Pipeline registration.
 //===----------------------------------------------------------------------===//

mlir/test/Integration/Dialect/SparseTensor/CPU/sparse_tanh.mlir

 // DEFINE: %{option} = enable-runtime-library=true
 // DEFINE: %{command} = mlir-opt %s --sparse-compiler=%{option} | \
 // DEFINE: mlir-cpu-runner \
 // DEFINE:  -e entry -entry-point-result=void  \
 // DEFINE:  -shared-libs=%mlir_lib_dir/libmlir_c_runner_utils%shlibext | \
 // DEFINE: FileCheck %s
 //
 // RUN: %{command}
 //
 // Do the same run, but now with direct IR generation.
 // REDEFINE: %{option} = enable-runtime-library=false
 // RUN: %{command}
+//
+// Do the same run, but now with direct IR generation and vectorization.
+// REDEFINE: %{option} = "enable-runtime-library=false vl=2 reassociate-fp-reductions=true enable-index-optimizations=true"
+// RUN: %{command}
 
 #SparseVector = #sparse_tensor.encoding<{ dimLevelType = [ "compressed" ] }>
 
 #trait_op = {
   indexing_maps = [
     affine_map<(i) -> (i)>   // X (out)
   ],
   iterator_types = ["parallel"],