[mlir][linalg][python] Add max operation in OpDSL

[gysit]

Summary

Add the max operation to the OpDSL and introduce a max pooling operation to test the implementation. As MLIR has no builtin max operation, the max function is lowered to a compare and select pair.

[stellaraccident]

Very nice - good to see this all coming together!

[gysit]

Rebase.

mlir/include/mlir/Dialect/Linalg/IR/LinalgNamedStructuredOps.yaml

-
 --- !LinalgOpConfig
 metadata: !LinalgOpMetadata
   name: matmul
   cpp_class_name: MatmulOp
   doc: |-
     Performs a matrix multiplication of two 2D inputs.
 
     Numeric casting is performed on the operands to the inner multiply, promoting
value: !ScalarExpression
         - !ScalarExpression
           symbolic_cast:
             type_var: U
             operands:
             - !ScalarExpression
               scalar_arg: I
 --- !LinalgOpConfig
 metadata: !LinalgOpMetadata
+  name: pooling_nhwc_max_poly
+  cpp_class_name: PoolingNhwcMaxPolyOp
+  doc: |-
+    Performs max pooling.
+
+    Numeric casting is performed on the input operand, promoting it to the same
+    data type as the accumulator/output.
+structured_op: !LinalgStructuredOpConfig
+  args:
+  - !LinalgOperandDefConfig
+    name: I
+    usage: InputOperand
+    type_var: T1
+    shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
+      (s0, s1, s2, s3)>
+  - !LinalgOperandDefConfig
+    name: K
+    usage: InputOperand
+    type_var: T2
+    shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
+      (s4, s5)>
+  - !LinalgOperandDefConfig
+    name: O
+    usage: OutputOperand
+    type_var: U
+    shape_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11] ->
+      (s0, s6, s7, s3)>
+  - !LinalgOperandDefConfig
+    name: strides
+    usage: IndexAttribute
+    type_var: I64
+    attribute_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11]
+      -> (s8, s9)>
+  - !LinalgOperandDefConfig
+    name: dilations
+    usage: IndexAttribute
+    type_var: I64
+    attribute_map: affine_map<()[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11]
+      -> (s10, s11)>
+  indexing_maps: !LinalgIndexingMapsConfig
+    static_indexing_maps:
+    - affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9,
+      s10, s11] -> (d0, d1 * s8 + d3 * s10, d2 * s9 + d4 * s11, d5)>
+    - affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9,
+      s10, s11] -> (d3, d4)>
+    - affine_map<(d0, d1, d2, d3, d4, d5)[s0, s1, s2, s3, s4, s5, s6, s7, s8, s9,
+      s10, s11] -> (d0, d1, d2, d5)>
+  iterator_types:
+  - parallel
+  - parallel
+  - parallel
+  - reduction
+  - reduction
+  - parallel
+  assignments:
+  - !ScalarAssign
+    arg: O
+    value: !ScalarExpression
+      scalar_apply:
+        fn_name: max
+        operands:
+        - !ScalarExpression
+          scalar_arg: O
+        - !ScalarExpression
+          symbolic_cast:
+            type_var: U
+            operands:
+            - !ScalarExpression
+              scalar_arg: I
+--- !LinalgOpConfig
+metadata: !LinalgOpMetadata
   name: fill_rng_2d
   cpp_class_name: FillRng2DOp
   doc: |-
     Fills the output tensor with pseudo random numbers.
 
     The operation generations pseudo random numbers using a linear congruential
     generator. It provides no guarantees regarding the distribution of the
     generated random numbers. Instead of generating the random numbers

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

Value applyfn__mul(Value lhs, Value rhs) {
     OpBuilder builder = getBuilder();
     if (isFloatingPoint(lhs))
       return builder.create<MulFOp>(lhs.getLoc(), lhs, rhs);
     if (isInteger(lhs))
       return builder.create<MulIOp>(lhs.getLoc(), lhs, rhs);
     llvm_unreachable("unsupported non numeric type");
   }
 
+  Value applyfn__max(Value lhs, Value rhs) {
+    OpBuilder builder = getBuilder();
+    if (isFloatingPoint(lhs)) {
+      Value condition =
+          builder.create<CmpFOp>(lhs.getLoc(), CmpFPredicate::OGT, lhs, rhs);
+      return builder.create<SelectOp>(lhs.getLoc(), condition, lhs, rhs);
+    }
+    if (isInteger(lhs)) {
+      Value condition =
+          builder.create<CmpIOp>(lhs.getLoc(), CmpIPredicate::sgt, lhs, rhs);
+      return builder.create<SelectOp>(lhs.getLoc(), condition, lhs, rhs);
+    }
+    llvm_unreachable("unsupported non numeric type");
+  }
+
   void yieldOutputs(ValueRange values) {
     assert(!values.empty() && "linalg ops must yield outputs");
     if (values.empty())
       return;
     Value first = values.front();
     OpBuilder builder = getBuilder();
     builder.create<YieldOp>(first.getLoc(), values);
   }

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

class _BodyBuilder:
 
   def _eval_mul(self, lhs: Value, rhs: Value) -> Value:
     if _is_floating_point_type(lhs.type):
       return std.MulFOp(lhs.type, lhs, rhs).result
     if _is_integer_type(lhs.type) or _is_index_type(lhs.type):
       return std.MulIOp(lhs.type, lhs, rhs).result
     raise NotImplementedError("Unsupported 'mul' operand: {lhs}")
 
+  def _eval_max(self, lhs: Value, rhs: Value) -> Value:
+    i1 = IntegerType.get_signless(1)
+    if _is_floating_point_type(lhs.type):
+      ogt_attr = IntegerAttr.get(IntegerType.get_signless(64), 2)
+      cond = std.CmpFOp(i1, ogt_attr, lhs, rhs).result
+      return std.SelectOp(lhs.type, cond, lhs, rhs).result
+    if _is_integer_type(lhs.type) or _is_index_type(lhs.type):
+      sgt_attr = IntegerAttr.get(IntegerType.get_signless(64), 4)
+      cond = std.CmpIOp(i1, sgt_attr, lhs, rhs).result
+      return std.SelectOp(lhs.type, cond, lhs, rhs).result
+    raise NotImplementedError("Unsupported 'max' operand: {lhs}")
+
 
 def _infer_structured_outs(op_config: LinalgStructuredOpConfig,
                            in_arg_defs: Sequence[OperandDefConfig],
                            ins: Sequence[Value],
                            out_arg_defs: Sequence[OperandDefConfig],
                            outs: Sequence[Value]):
   """Infers implicit outs and output types.
 

mlir/python/mlir/dialects/linalg/opdsl/ops/core_named_ops.py

def pooling_nhwc_sum_poly(
   data type as the accumulator/output.
   """
   domain(D.n, D.oh, D.ow, D.kh, D.kw, D.c)
   O[D.n, D.oh, D.ow, D.c] += cast(
       U, I[D.n, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW, D.c])
 
 
 @linalg_structured_op
+def pooling_nhwc_max_poly(
+    I=TensorDef(T1, S.N, S.H, S.W, S.C),
+    K=TensorDef(T2, S.KH, S.KW, index_dims=[D.kh, D.kw]),
+    O=TensorDef(U, S.N, S.OH, S.OW, S.C, output=True),
+    strides=AttributeDef(S.SH, S.SW),
+    dilations=AttributeDef(S.DH, S.DW)):
+  """Performs max pooling.
+
+  Numeric casting is performed on the input operand, promoting it to the same
+  data type as the accumulator/output.
+  """
+  domain(D.n, D.oh, D.ow, D.kh, D.kw, D.c)
+  O[D.n, D.oh, D.ow, D.c] = ReduceFn.max(D.kh, D.kw)(
+      cast(U, I[D.n, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW,
+                D.c]))
+
+
+@linalg_structured_op
 def fill_rng_2d(
     min=ScalarDef(F64),
     max=ScalarDef(F64),
     seed=ScalarDef(I32),
     O=TensorDef(T, S.M, S.N, output=True)):
   """Fills the output tensor with pseudo random numbers.
 
   The operation generations pseudo random numbers using a linear congruential

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

 // CHECK:      ^{{.*}}(%[[IN_ARG:.+]]: i32, %[[FILTER_ARG:.+]]: i32, %[[OUT_ARG:.+]]: i32)
 // CHECK-NEXT:   %[[MUL:.+]] = muli %[[IN_ARG]], %[[FILTER_ARG]] : i32
 // CHECK-NEXT:   %[[ADD:.+]] = addi %[[OUT_ARG]], %[[MUL]] : i32
 // CHECK-NEXT:   linalg.yield %[[ADD]] : i32
 // CHECK-NEXT: -> tensor<1x2x4x1xi32>
 
 // -----
 
+func @generalize_pooling_nhwc_max_poly_f32(%input : tensor<1x4x16x1xf32>, %shape: tensor<2x2xf32>, %output: tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32> {
+  %0 = linalg.pooling_nhwc_max_poly {dilations = dense<[1, 2]> : tensor<2xi64>, strides = dense<[2, 4]> : tensor<2xi64>}
+    ins(%input, %shape : tensor<1x4x16x1xf32>, tensor<2x2xf32>) outs(%output : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
+  return %0: tensor<1x2x4x1xf32>
+}
+
+// CHECK-LABEL: @generalize_pooling_nhwc_max_poly_f32
+// CHECK:      ^{{.*}}(%[[IN_ARG:.+]]: f32, %[[SHAPE_ARG:.+]]: f32, %[[OUT_ARG:.+]]: f32)
+// CHECK-NEXT:   %[[COND:.+]] = cmpf ogt, %[[OUT_ARG]], %[[IN_ARG]] : f32
+// CHECK-NEXT:   %[[MAX:.+]] = select %[[COND]], %[[OUT_ARG]], %[[IN_ARG]] : f32
+// CHECK-NEXT:   linalg.yield %[[MAX]] : f32
+// CHECK-NEXT: -> tensor<1x2x4x1xf32>
+
+// -----
+
+func @generalize_pooling_nhwc_max_poly_i32(%input : tensor<1x4x16x1xi32>, %shape: tensor<2x2xi32>, %output: tensor<1x2x4x1xi32>) -> tensor<1x2x4x1xi32> {
+  %0 = linalg.pooling_nhwc_max_poly {dilations = dense<[1, 2]> : tensor<2xi64>, strides = dense<[2, 4]> : tensor<2xi64>}
+    ins(%input, %shape : tensor<1x4x16x1xi32>, tensor<2x2xi32>) outs(%output : tensor<1x2x4x1xi32>) -> tensor<1x2x4x1xi32>
+  return %0: tensor<1x2x4x1xi32>
+}
+
+// CHECK-LABEL: @generalize_pooling_nhwc_max_poly_i32
+// CHECK:      ^{{.*}}(%[[IN_ARG:.+]]: i32, %[[SHAPE_ARG:.+]]: i32, %[[OUT_ARG:.+]]: i32)
+// CHECK-NEXT:   %[[COND:.+]] = cmpi sgt, %[[OUT_ARG]], %[[IN_ARG]] : i32
+// CHECK-NEXT:   %[[MAX:.+]] = select %[[COND]], %[[OUT_ARG]], %[[IN_ARG]] : i32
+// CHECK-NEXT:   linalg.yield %[[MAX]] : i32
+// CHECK-NEXT: -> tensor<1x2x4x1xi32>
+
+// -----
+
 func @generalize_pooling_nhwc_sum_poly_f32(%input : tensor<1x4x16x1xf32>, %shape: tensor<2x2xf32>, %output: tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32> {
   %0 = linalg.pooling_nhwc_sum_poly {dilations = dense<[1, 2]> : tensor<2xi64>, strides = dense<[2, 4]> : tensor<2xi64>}
     ins(%input, %shape : tensor<1x4x16x1xf32>, tensor<2x2xf32>) outs(%output : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
   return %0: tensor<1x2x4x1xf32>
 }
 
 // CHECK-LABEL: @generalize_pooling_nhwc_sum_poly_f32
 // CHECK:      ^{{.*}}(%[[IN_ARG:.+]]: f32, %[[SHAPE_ARG:.+]]: f32, %[[OUT_ARG:.+]]: f32)

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

 @linalg_structured_op
 def pooling_poly(
     I=TensorDef(T1, S.N, S.H, S.W, S.C),
     K=TensorDef(T2, S.KH, S.KW, index_dims=[D.kh, D.kw]),
     O=TensorDef(U, S.N, S.OH, S.OW, S.C, output=True),
     strides=AttributeDef(S.SH, S.SW),
     dilations=AttributeDef(S.DH, S.DW)):
   domain(D.n, D.oh, D.ow, D.kh, D.kw, D.c)
-  O[D.n, D.oh, D.ow, D.c] += cast(
-      U, I[D.n, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW, D.c])
+  O[D.n, D.oh, D.ow, D.c] = ReduceFn.max(D.kh, D.kw)(
+      cast(U, I[D.n, D.oh * S.SH + D.kh * S.DH, D.ow * S.SW + D.kw * S.DW,
+                D.c]))
 
 
 @linalg_structured_op
 def fill_rng_poly(
     min=ScalarDef(F64),
     max=ScalarDef(F64),
     seed=ScalarDef(I32),
     O=TensorDef(T, S.M, S.N, output=True)):
def test_f32i32_conv(input, filter, init_result):
           input, filter, outs=[init_result], strides=[2, 4], dilations=[1, 2])
 
     # CHECK-LABEL: @test_f32i32_pooling
     # CHECK: linalg.generic
     # CHECK-SAME: indexing_maps = [#[[$CONV_MAP_I]], #[[$POOL_MAP_K]], #[[$CONV_MAP_O]]]
     # CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction", "parallel"]
     # CHECK:      ^{{.*}}(%[[IN:.+]]: f32, %[[SHAPE:.+]]: f32, %[[OUT:.+]]: i32)
     # CHECK-NEXT:   %[[IN_CAST:.+]] = fptosi %[[IN:.+]] : f32 to i32
-    # CHECK-NEXT:   %[[SUM:.+]] = addi %[[OUT]], %[[IN_CAST]] : i32
-    # CHECK-NEXT:   linalg.yield %[[SUM]] : i32
+    # CHECK-NEXT:   %[[COND:.+]] = cmpi sgt, %[[OUT]], %[[IN_CAST:.+]] : i32
+    # CHECK-NEXT:   %[[MAX:.+]] = select %[[COND]], %[[OUT]], %[[IN_CAST:.+]] : i32
+    # CHECK-NEXT:   linalg.yield %[[MAX]] : i32
     # CHECK-NEXT: -> tensor<2x4xi32>
     @builtin.FuncOp.from_py_func(
         RankedTensorType.get((4, 16), f32), RankedTensorType.get((2, 2), f32),
         RankedTensorType.get((2, 4), i32))
     def test_f32i32_pooling(input, shape, init_result):
       return pooling_poly(
           input, shape, outs=[init_result], strides=[2, 4], dilations=[1, 2])
 
+    # CHECK-LABEL: @test_f32f32_pooling
+    # CHECK: linalg.generic
+    # CHECK-SAME: indexing_maps = [#[[$CONV_MAP_I]], #[[$POOL_MAP_K]], #[[$CONV_MAP_O]]]
+    # CHECK-SAME: iterator_types = ["parallel", "parallel", "parallel", "reduction", "reduction", "parallel"]
+    # CHECK:      ^{{.*}}(%[[IN:.+]]: f32, %[[SHAPE:.+]]: f32, %[[OUT:.+]]: f32)
+    # CHECK-NEXT:   %[[COND:.+]] = cmpf ogt, %[[OUT]], %[[IN:.+]] : f32
+    # CHECK-NEXT:   %[[MAX:.+]] = select %[[COND]], %[[OUT]], %[[IN:.+]] : f32
+    # CHECK-NEXT:   linalg.yield %[[MAX]] : f32
+    # CHECK-NEXT: -> tensor<2x4xf32>
+    @builtin.FuncOp.from_py_func(
+        RankedTensorType.get((4, 16), f32), RankedTensorType.get((2, 2), f32),
+        RankedTensorType.get((2, 4), f32))
+    def test_f32f32_pooling(input, shape, init_result):
+      return pooling_poly(
+          input, shape, outs=[init_result], strides=[2, 4], dilations=[1, 2])
+
     # CHECK-LABEL: @test_i32_fill_rng
     # CHECK:      ^{{.*}}(%[[MIN:.+]]: f64, %[[MAX:.+]]: f64, %[[SEED:.+]]: i32, %{{.*}}
     # CHECK-DAG:    %[[IDX0:.+]] = linalg.index 0 : index
     # CHECK-DAG:    %[[IDX0_CAST:.+]] = index_cast %[[IDX0]] : index to i32
     # CHECK-DAG:    %[[RND0:.+]] = addi %[[IDX0_CAST]], %[[SEED]] : i32
     # CHECK-DAG:    %[[CST0:.+]] = constant 1103515245 : i64
     # CHECK-DAG:    %[[CST0_CAST:.+]] = trunci %[[CST0]] : i64 to i32
     # Skip the remaining random number computation and match the scaling logic.

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

func @main() -> i32 attributes {llvm.emit_c_interface} {
   // TODO: FFI-based solution to allow testing and printing with python code.
   return %0 : i32
 }
 """
 
 pooling_boiler = """
 func @main() -> i32 attributes {llvm.emit_c_interface} {
   %v0 = constant 0 : i32
+  %v42 = constant 42.0 : f64
   %v1 = constant 1.0 : f64
 
   %input = memref.alloc() : memref<1x4x16x1xf64>
   %shape = memref.alloc() : memref<2x2xf64>
   %output = memref.alloc() : memref<1x2x4x1xi32>
   linalg.fill(%v1, %input) : f64, memref<1x4x16x1xf64>
   linalg.fill(%v1, %shape) : f64, memref<2x2xf64>
   linalg.fill(%v0, %output) : i32, memref<1x2x4x1xi32>
 
+  %c0 = constant 0 : index
+  memref.store %v42, %input[%c0, %c0, %c0, %c0] : memref<1x4x16x1xf64>
+
   call @pooling_on_buffers(%input, %shape, %output) :
     (memref<1x4x16x1xf64>, memref<2x2xf64>, memref<1x2x4x1xi32>) -> ()
 
-  %c0 = constant 0 : index
   %0 = memref.load %output[%c0, %c0, %c0, %c0] : memref<1x2x4x1xi32>
 
   // TODO: FFI-based solution to allow testing and printing with python code.
   return %0 : i32
 }
 """
 
+
 def transform(module, boilerplate):
   import mlir.conversions
   import mlir.dialects.linalg.passes
   import mlir.transforms
 
   # TODO: Allow cloning functions from one module to another.
   # Atm we have to resort to string concatenation.
   mod = Module.parse(
with Context() as ctx, Location.unknown():
     f64 = F64Type.get()
     i32 = IntegerType.get_signless(32)
     with InsertionPoint(module.body):
 
       @builtin.FuncOp.from_py_func(
           MemRefType.get((1, 4, 16, 1), f64), MemRefType.get((2, 2), f64),
           MemRefType.get((1, 2, 4, 1), i32))
       def pooling_on_buffers(input, shape, output):
-        linalg.pooling_nhwc_sum_poly(
-            input,
-            shape,
-            outs=[output],
-            strides=[2, 4],
-            dilations=[1, 2])
+        linalg.pooling_nhwc_max_poly(
+            input, shape, outs=[output], strides=[2, 4], dilations=[1, 2])
 
     execution_engine = ExecutionEngine(transform(module, pooling_boiler))
 
     # TODO: FFI-based solution to allow testing and printing with python code.
     # Prepare arguments: one result i32.
     # Arguments must be passed as pointers.
     c_int_p = ctypes.c_int * 1
     res = c_int_p(-1)
     execution_engine.invoke("main", res)
 
     log("RESULT: ", res[0])
-    # CHECK: RESULT: 4
+    # CHECK: RESULT: 42
 
 
 test_pooling_builtin()
 
 
 def test_pooling_generic():
   with Context() as ctx, Location.unknown():
     module = Module.create()
     f64 = F64Type.get()
     i32 = IntegerType.get_signless(32)
     with InsertionPoint(module.body):
 
       @builtin.FuncOp.from_py_func(
           MemRefType.get((1, 4, 16, 1), f64), MemRefType.get((2, 2), f64),
           MemRefType.get((1, 2, 4, 1), i32))
       def pooling_on_buffers(input, shape, output):
-        linalg.pooling_nhwc_sum_poly(
+        linalg.pooling_nhwc_max_poly(
             input,
             shape,
             outs=[output],
             strides=[2, 4],
             dilations=[1, 2],
             emit_generic=True)
 
     execution_engine = ExecutionEngine(transform(module, pooling_boiler))
 
     # TODO: FFI-based solution to allow testing and printing with python code.
     # Prepare arguments: one result i32.
     # Arguments must be passed as pointers.
     c_int_p = ctypes.c_int * 1
     res = c_int_p(-1)
     execution_engine.invoke("main", res)
 
     log("RESULT: ", res[0])
-    # CHECK: RESULT: 4
+    # CHECK: RESULT: 42
 
 
 test_pooling_generic()