Extend mlir-gen to emit linalg named Ops (libxsmm#933).

Adds support to generate linalg named Ops for matmul, bias, relu.
This feature can be controlled using a new flag '--output'.

For example:
To generate generic linalg Ops use '--output=generic"
To generate named linalg Ops use '--output=named"

The default behaviour is to generate linalg generic Ops.

Adds named op test which pass out of the box.

-Adds another option "--keep-generic-matmul" to help generate generic
 matmul when linalg named ops output was chosen.

-Refactors the code.

test/Integration/mlir-gen.mlir

@@ -1,5 +1,8 @@
 // MLP with Softmax version
 // RUN: mlir-gen --kernel=const --bias --relu --seed=123 --batch=10 --layers=10,10,10 --softmax | tpp-run -e entry -entry-point-result=void
+// RUN: not --crash mlir-gen --output=named --kernel=const --bias --relu --seed=123 --batch=10 --layers=10,10,10 --softmax 2>&1 | FileCheck %s --check-prefix=SOFTMAX-TODO
+// SOFTMAX-TODO: Linalg named ops for softmax not implemented yet
+// SOFTMAX-TODO: UNREACHABLE executed
 
 // MLP without softmax
 // RUN: mlir-gen --kernel=const --bias --relu --seed=123 --batch=10 --layers=10,10,10 | tpp-run -e entry -entry-point-result=void

test/Passes/tile-and-fuse_named-op.mlir → test/Passes/tile-and-fuse-named-op.mlir

@@ -128,4 +128,4 @@ func.func @matmul_sequence_fusion_with_relu(%arg0: tensor<32x64xf32>, %arg1: ten
 // CHECK: scf.yield %{{.+}} : tensor<32x32xf32>
 // CHECK-NEXT: }
 
-// -----
+// -----

tools/mlir-gen/MLIRGen.cpp

@@ -517,7 +517,8 @@ Value MLIRGenerator::lowerNamedSoftmax(Value input, Value output) {
     return input;
 
   // TODO: Add lowering of softmax to sequence of named Ops
-
+  llvm_unreachable("Linalg named ops for softmax not implemented yet");
+
   auto outTy = cast<ShapedType>(input.getType());
   // Softmax flops = 4 * M * N = 4 * prod(outputDims)
   int64_t softmaxFlops = 1;