[mlir][OpDSL] Support pointwise ops with rank zero inputs.

[gysit]

Summary

Allow pointwise operations to take rank zero input tensors similarly to scalar inputs. Use an empty indexing map to broadcast rank zero tensors to the iteration domain of the operation.

Depends On D120734

[gysit]

Fix warning.

mlir/python/mlir/dialects/linalg/opdsl/lang/emitter.py

if not iterator_types_attr:
     iterator_types_attr = ArrayAttr.get([StringAttr.get("parallel")] * rank)
     scalar_map = AffineMap.get(rank, 0, [])
     tensor_map = AffineMap.get_identity(rank)
     indexing_maps = []
     for arg_def in all_arg_defs:
       if arg_def.operand_def.kind == OperandKind.SCALAR:
         indexing_maps.append(scalar_map)
       if arg_def.operand_def.is_tensor():
+        idx = arg_def.operand_def.registered_index
+        if idx < len(ins) and ShapedType(ins[idx].type).rank == 0:
+          indexing_maps.append(scalar_map)
+        else:
           indexing_maps.append(tensor_map)
     indexing_maps_attr = ArrayAttr.get(
         [AffineMapAttr.get(am) for am in indexing_maps])
 
   generic_op = linalg.GenericOp(
       result_tensors=result_types,
       inputs=ins,
       outputs=outs,
       indexing_maps=indexing_maps_attr,

mlir/test/Dialect/Linalg/generalize-named-polymorphic-ops.mlir

func @generalize_elemwise_mul(%lhs : tensor<4x8xf32>, %rhs : tensor<4x8xf32>, %output : tensor<4x8xf32>) -> tensor<4x8xf32> {
   %0 = linalg.elemwise_binary {fun = #linalg.binary_fn<mul>}
                               ins(%lhs, %rhs: tensor<4x8xf32>, tensor<4x8xf32>)
                               outs(%output: tensor<4x8xf32>) -> tensor<4x8xf32>
   return %0: tensor<4x8xf32>
 }
 
 // CHECK-LABEL: @generalize_elemwise_mul
 // CHECK:        = arith.mulf
+
+// -----
+
+// Verifies pointwise ops support rank zero input tensors
+func @generalize_elemwise_rank_zero(%lhs : tensor<f32>, %rhs : tensor<f32>, %output : tensor<4x8xf32>) -> tensor<4x8xf32> {
+  %0 = linalg.elemwise_binary {fun = #linalg.binary_fn<sub>}
+                              ins(%lhs, %rhs: tensor<f32>, tensor<f32>)
+                              outs(%output: tensor<4x8xf32>) -> tensor<4x8xf32>
+  return %0: tensor<4x8xf32>
+}
+
+// CHECK-LABEL: @generalize_elemwise_rank_zero
+// CHECK:       linalg.generic
+// CHECK-SAME:  iterator_types = ["parallel", "parallel"]
+// CHECK:        = arith.subf

mlir/test/python/dialects/linalg/opdsl/emit_fill.py

 T1 = TV.T1
 T2 = TV.T2
 
 
 @linalg_structured_op
 def fill_poly(value=ScalarDef(T1), O=TensorDef(U, output=True)):
   O[None] = TypeFn.cast_signed(U, value)
 
+@linalg_structured_op
+def fill_rank_zero_poly(I=TensorDef(T1), O=TensorDef(U, output=True)):
+  O[None] = TypeFn.cast_signed(U, I[None])
 
 with Context() as ctx, Location.unknown():
   module = Module.create()
   f32 = F32Type.get()
   with InsertionPoint(module.body):
 
     # Fill indexing maps.
     # CHECK-DAG: #[[$MAP0:.+]] = affine_map<() -> ()>
     # CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0, d1) -> ()>
     # CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0, d1) -> (d0, d1)>
+    # CHECK-DAG: #[[$MAP3:.+]] = affine_map<(d0, d1, d2) -> ()>
+    # CHECK-DAG: #[[$MAP4:.+]] = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
 
     # CHECK-LABEL: @test_fill_0d
     # CHECK: linalg.generic
     # CHECK-SAME: indexing_maps = [#[[$MAP0]], #[[$MAP0]]
     # CHECK-SAME: iterator_types = []
     @builtin.FuncOp.from_py_func(f32, RankedTensorType.get([], f32))
     def test_fill_0d(value, init_result):
       return fill_poly(value, outs=[init_result])
 
     # CHECK-LABEL: @test_fill_2d
     # CHECK: linalg.generic
     # CHECK-SAME: indexing_maps = [#[[$MAP1]], #[[$MAP2]]]
     # CHECK-SAME: iterator_types = ["parallel", "parallel"]
     @builtin.FuncOp.from_py_func(f32, RankedTensorType.get([4, 16], f32))
     def test_fill_2d(value, init_result):
       return fill_poly(value, outs=[init_result])
 
+    # CHECK-LABEL: @test_fill_rank_zero_3d
+    # CHECK: linalg.generic
+    # CHECK-SAME: indexing_maps = [#[[$MAP3]], #[[$MAP4]]]
+    # CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel"]
+    @builtin.FuncOp.from_py_func(
+        RankedTensorType.get([], f32), RankedTensorType.get([4, 8, 16], f32))
+    def test_fill_rank_zero_3d(input, init_result):
+      return fill_rank_zero_poly(input, outs=[init_result])
 
 print(module)

mlir/test/python/integration/dialects/linalg/opsrun.py

 
 
 elemwise_boiler = """
 func @main() -> f32 attributes {llvm.emit_c_interface} {
   %v0 = arith.constant 0.0 : f32
   %v1 = arith.constant 1.0 : f32
   %v2 = arith.constant 2.0 : f32
 
-  %lhs = memref.alloc() : memref<4x8xf32>
+  %lhs = memref.alloc() : memref<f32>
   %rhs = memref.alloc() : memref<4x8xf32>
   %O0 = memref.alloc() : memref<4x8xf32>
   %O1 = memref.alloc() : memref<4x8xf32>
-  linalg.fill(%v1, %lhs) : f32, memref<4x8xf32>
+  linalg.fill(%v1, %lhs) : f32, memref<f32>
   linalg.fill(%v2, %rhs) : f32, memref<4x8xf32>
   linalg.fill(%v0, %O0) : f32, memref<4x8xf32>
   linalg.fill(%v0, %O1) : f32, memref<4x8xf32>
 
   call @elemwise_exp_add_on_buffers(%lhs, %rhs, %O0) :
-    (memref<4x8xf32>, memref<4x8xf32>, memref<4x8xf32>) -> ()
+    (memref<f32>, memref<4x8xf32>, memref<4x8xf32>) -> ()
   call @elemwise_log_mul_on_buffers(%lhs, %rhs, %O1) :
-    (memref<4x8xf32>, memref<4x8xf32>, memref<4x8xf32>) -> ()
+    (memref<f32>, memref<4x8xf32>, memref<4x8xf32>) -> ()
 
   %c0 = arith.constant 0 : index
   %res0 = memref.load %O0[%c0, %c0] : memref<4x8xf32>
   %res1 = memref.load %O1[%c0, %c0] : memref<4x8xf32>
 
   %0 = arith.addf %res0, %res1 : f32
 
   // TODO: FFI-based solution to allow testing and printing with python code.

 def test_elemwise_builtin():
   with Context() as ctx, Location.unknown():
     module = Module.create()
     f32 = F32Type.get()
     i8 = IntegerType.get_signless(8)
     with InsertionPoint(module.body):
 
       @builtin.FuncOp.from_py_func(
-          MemRefType.get((4, 8), f32), MemRefType.get((4, 8), f32),
+          MemRefType.get((), f32), MemRefType.get((4, 8), f32),
           MemRefType.get((4, 8), f32))
       def elemwise_exp_add_on_buffers(lhs, rhs, out):
         linalg.elemwise_unary(lhs, outs=[out])
         linalg.elemwise_binary(out, rhs, outs=[out])
 
       @builtin.FuncOp.from_py_func(
-          MemRefType.get((4, 8), f32), MemRefType.get((4, 8), f32),
+          MemRefType.get((), f32), MemRefType.get((4, 8), f32),
           MemRefType.get((4, 8), f32))
       def elemwise_log_mul_on_buffers(lhs, rhs, out):
         linalg.elemwise_unary(lhs, outs=[out], fun=UnaryFn.log)
         linalg.elemwise_binary(out, rhs, outs=[out], fun=BinaryFn.mul)
 
     execution_engine = ExecutionEngine(transform(module, elemwise_boiler))
 
     # TODO: FFI-based solution to allow testing and printing with python code.

 def test_elemwise_generic():
   with Context() as ctx, Location.unknown():
     module = Module.create()
     f32 = F32Type.get()
     i8 = IntegerType.get_signless(8)
     with InsertionPoint(module.body):
 
       @builtin.FuncOp.from_py_func(
-          MemRefType.get((4, 8), f32), MemRefType.get((4, 8), f32),
+          MemRefType.get((), f32), MemRefType.get((4, 8), f32),
           MemRefType.get((4, 8), f32))
       def elemwise_exp_add_on_buffers(lhs, rhs, out):
         linalg.elemwise_unary(lhs, outs=[out], emit_generic=True)
         linalg.elemwise_binary(out, rhs, outs=[out], emit_generic=True)
 
       @builtin.FuncOp.from_py_func(
-          MemRefType.get((4, 8), f32), MemRefType.get((4, 8), f32),
+          MemRefType.get((), f32), MemRefType.get((4, 8), f32),
           MemRefType.get((4, 8), f32))
       def elemwise_log_mul_on_buffers(lhs, rhs, out):
         linalg.elemwise_unary(
             lhs, outs=[out], fun=UnaryFn.log, emit_generic=True)
         linalg.elemwise_binary(
             out, rhs, outs=[out], fun=BinaryFn.mul, emit_generic=True)
 
     execution_engine = ExecutionEngine(transform(module, elemwise_boiler))

mlir/tools/mlir-linalg-ods-gen/mlir-linalg-ods-yaml-gen.cpp

 // {0}: Class name
 static const char rankPolyStructuredOpIndexingMapsFormat[] = R"FMT(
 ArrayAttr {0}::indexing_maps() {{
   MLIRContext *context = getContext();
   AffineMap scalarMap = AffineMap::get(getNumParallelLoops(), 0, context);
   AffineMap tensorMap = AffineMap::getMultiDimIdentityMap(
     getNumParallelLoops(), context);
   SmallVector<AffineMap> indexingMaps;
-  for (OpOperand *opOperand : getInputAndOutputOperands())
-    indexingMaps.push_back(isScalar(opOperand) ? scalarMap : tensorMap);
+  for (OpOperand *opOperand : getInputOperands())
+    indexingMaps.push_back(getRank(opOperand) == 0 ? scalarMap : tensorMap);
+  for (OpOperand *opOperand : getOutputOperands())
+    indexingMaps.push_back(tensorMap);
   return Builder(getContext()).getAffineMapArrayAttr(indexingMaps);
 }
 )FMT";
 
 // Implementations of fold and getEffects.
 // Parameters:
 // {0}: Class name
 const char structuredOpFoldersFormat[] = R"FMT(