[MLIR][TOSA] Remove ReluN operator from TOSA dialect

[eric-k256]

Summary

ReluN has been removed from the TOSA specification. It can be replaced
in all instances with Clamp(0,N)

Signed-off-by: Eric Kunze <[e-mail]>

[jpienaar]

LG, do you have any pending PRs for downstream users?

[eric-k256]

I didn't find any uses of ReluN downstream, it must have been cleaned up already.

[jpienaar]

SGTM

[jpienaar]

Do you need help in landing?

[eric-k256]

Yes, help landing would be appreciated.

[eric-k256]

Rebased change

[jpienaar]

Sorry missed notification earlier (seems I have some filter configuration error as second one this week), will land once at desk

mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td

def Tosa_ClampOp : Tosa_Op<"clamp", [
   let results = (outs
     Tosa_Tensor:$output
   );
 
   let hasCanonicalizer = 1;
 }
 
 //===----------------------------------------------------------------------===//
-// Operator: reluN
-//===----------------------------------------------------------------------===//
-def Tosa_ReluNOp : Tosa_Op<"reluN", [
-    DeclareOpInterfaceMethods<InferShapedTypeOpInterface,
-                              ["inferReturnTypeComponents"]>,
-    NoSideEffect]> {
-  let summary = "Computes rectified linear: `max(features, N)`.";
-
-  let description = [{
-     ReLU with a scalar maximum value.
-  }];
-
-  let arguments = (ins
-    Tosa_Tensor:$input,
-    I64Attr:$max_int,
-    F32Attr:$max_fp
-  );
-
-  let results = (outs
-    Tosa_Tensor:$output
-  );
-}
-
-//===----------------------------------------------------------------------===//
 // Operator: sigmoid
 //===----------------------------------------------------------------------===//
 def Tosa_SigmoidOp : Tosa_Op<"sigmoid", [
     DeclareOpInterfaceMethods<InferShapedTypeOpInterface,
                               ["inferReturnTypeComponents"]>,
     NoSideEffect]> {
   let summary = "Computes elementwise sigmoid of input.";
 

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

if (isa<tosa::ClampOp>(op) && elementTy.isa<IntegerType>()) {
 
     auto minVal = rewriter.create<arith::ConstantIntOp>(
         loc, min, intTy.getIntOrFloatBitWidth());
     auto maxVal = rewriter.create<arith::ConstantIntOp>(
         loc, max, intTy.getIntOrFloatBitWidth());
     return clampIntHelper(loc, args[0], minVal, maxVal, rewriter);
   }
 
-  // tosa::ReluNOp
-  if (isa<tosa::ReluNOp>(op) && elementTy.isa<FloatType>()) {
-    auto zero =
-        rewriter.create<arith::ConstantOp>(loc, FloatAttr::get(elementTy, 0));
-    bool losesInfo = false;
-    APFloat max_apf = op->getAttr("max_fp").cast<FloatAttr>().getValue();
-    max_apf.convert(elementTy.cast<FloatType>().getFloatSemantics(),
-                    APFloat::rmNearestTiesToEven, &losesInfo);
-    auto n = rewriter.create<arith::ConstantOp>(
-        loc, elementTy, rewriter.getFloatAttr(elementTy, max_apf));
-    return clampFloatHelper(loc, args[0], zero, n, rewriter);
-  }
-
-  if (isa<tosa::ReluNOp>(op) && elementTy.isa<IntegerType>()) {
-    auto zero =
-        rewriter.create<arith::ConstantOp>(loc, IntegerAttr::get(elementTy, 0));
-    auto n = createConstFromIntAttribute<int32_t>(op, "max_int", elementTy,
-                                                  rewriter);
-    return clampIntHelper(loc, args[0], zero, n, rewriter);
-  }
-
   // tosa::SigmoidOp
   if (isa<tosa::SigmoidOp>(op) && elementTy.isa<FloatType>()) {
     auto one =
         rewriter.create<arith::ConstantOp>(loc, FloatAttr::get(elementTy, 1));
     auto negate = rewriter.create<arith::NegFOp>(loc, resultTypes, args[0]);
     auto exp = rewriter.create<mlir::math::ExpOp>(loc, resultTypes, negate);
     auto added = rewriter.create<arith::AddFOp>(loc, resultTypes, exp, one);
     return rewriter.create<arith::DivFOp>(loc, resultTypes, one, added);
patterns->add<
       PointwiseConverter<tosa::GreaterOp>,
       PointwiseConverter<tosa::GreaterEqualOp>,
       PointwiseConverter<tosa::EqualOp>,
       PointwiseConverter<tosa::MaximumOp>,
       PointwiseConverter<tosa::MinimumOp>,
       PointwiseConverter<tosa::CeilOp>,
       PointwiseConverter<tosa::FloorOp>,
       PointwiseConverter<tosa::ClampOp>,
-      PointwiseConverter<tosa::ReluNOp>,
       PointwiseConverter<tosa::SigmoidOp>,
       IdentityNConverter<tosa::IdentityOp>,
       ReduceConverter<tosa::ReduceAllOp>,
       ReduceConverter<tosa::ReduceAnyOp>,
       ReduceConverter<tosa::ReduceMinOp>,
       ReduceConverter<tosa::ReduceMaxOp>,
       ReduceConverter<tosa::ReduceSumOp>,
       ReduceConverter<tosa::ReduceProdOp>,

mlir/lib/Dialect/Tosa/IR/TosaOps.cpp

 NARY_SHAPE_INFER(tosa::LogicalRightShiftOp)
 NARY_SHAPE_INFER(tosa::LogicalXorOp)
 NARY_SHAPE_INFER(tosa::MaximumOp)
 NARY_SHAPE_INFER(tosa::MinimumOp)
 NARY_SHAPE_INFER(tosa::MulOp)
 NARY_SHAPE_INFER(tosa::NegateOp)
 NARY_SHAPE_INFER(tosa::PowOp)
 NARY_SHAPE_INFER(tosa::ReciprocalOp)
-NARY_SHAPE_INFER(tosa::ReluNOp)
 NARY_SHAPE_INFER(tosa::RescaleOp)
 NARY_SHAPE_INFER(tosa::ReverseOp)
 NARY_SHAPE_INFER(tosa::RsqrtOp)
 NARY_SHAPE_INFER(tosa::SelectOp)
 NARY_SHAPE_INFER(tosa::SubOp)
 NARY_SHAPE_INFER(tosa::TanhOp)
 NARY_SHAPE_INFER(tosa::SigmoidOp)
 #undef PRED_SHAPE_INFER

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

func.func @test_simple_f32(%arg0: tensor<1xf32>) -> () {
   %17 = "tosa.floor"(%0) : (tensor<1xf32>) -> tensor<1xf32>
 
   // CHECK: linalg.generic
   // CHECK: arith.minf
   // CHECK: arith.maxf
   %18 = "tosa.clamp"(%0) {min_int = 1 : i64, max_int = 5 : i64, min_fp = 1.0 : f32, max_fp = 5.0 : f32} : (tensor<1xf32>) -> tensor<1xf32>
 
   // CHECK: linalg.generic
-  // CHECK: arith.minf
-  // CHECK: arith.maxf
-  %19 = "tosa.reluN"(%0) {max_int = 5 : i64, max_fp = 5.0 : f32} : (tensor<1xf32>) -> tensor<1xf32>
-
-  // CHECK: linalg.generic
   // CHECK: arith.negf
   // CHECK: exp
   // CHECK: arith.addf
   // CHECK: arith.divf
-  %20 = "tosa.sigmoid"(%0) : (tensor<1xf32>) -> tensor<1xf32>
+  %19 = "tosa.sigmoid"(%0) : (tensor<1xf32>) -> tensor<1xf32>
 
   // CHECK: linalg.generic
   // CHECK: arith.constant 0.000000e+00
   // CHECK: arith.constant 5.000000e-01
   // CHECK: arith.constant -2.14748365E+9
   // CHECK: arith.constant 2.14748365E+9
   // CHECK: arith.addf
   // CHECK: arith.subf
   // CHECK: arith.cmpf olt
   // CHECK: select
   // CHECK: arith.minf
   // CHECK: arith.maxf
   // CHECK: arith.fptosi
-  %21 = "tosa.cast"(%0) : (tensor<1xf32>) -> tensor<1xi32>
+  %20 = "tosa.cast"(%0) : (tensor<1xf32>) -> tensor<1xi32>
 
   // CHECK: linalg.generic
   // CHECK: arith.constant 0
   // CHECK: arith.cmpf
-  %22 = "tosa.cast"(%0) : (tensor<1xf32>) -> tensor<1xi1>
+  %21 = "tosa.cast"(%0) : (tensor<1xf32>) -> tensor<1xi1>
 
   // CHECK: linalg.generic
   // CHECK: arith.truncf
-  %23 = "tosa.cast"(%0) : (tensor<1xf32>) -> tensor<1xf16>
+  %22 = "tosa.cast"(%0) : (tensor<1xf32>) -> tensor<1xf16>
 
   // CHECK: linalg.generic
   // CHECK: arith.divf
-  %24 = "tosa.reciprocal"(%0) : (tensor<1xf32>) -> tensor<1xf32>
+  %23 = "tosa.reciprocal"(%0) : (tensor<1xf32>) -> tensor<1xf32>
 
   return
 }
 
 // -----
 
 // CHECK-LABEL: @test_simple_f16
 func.func @test_simple_f16(%arg0: tensor<1xf16>) -> () {
func.func @test_simple_i32(%arg0: tensor<1xi32>) -> () {
   %18 = "tosa.minimum"(%0, %1) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
 
   // CHECK: linalg.generic
   // CHECK: arith.cmpi
   // CHECK: select
   %19 = "tosa.clamp"(%0) {min_int = 1 : i64, max_int = 5 : i64, min_fp = 1.0 : f32, max_fp = 5.0 : f32} : (tensor<1xi32>) -> tensor<1xi32>
 
   // CHECK: linalg.generic
-  // CHECK: arith.cmpi
-  // CHECK: select
-  %20 = "tosa.reluN"(%0) {max_int = 5 : i64, max_fp = 5.0 : f32} : (tensor<1xi32>) -> tensor<1xi32>
-
-  // CHECK: linalg.generic
   // CHECK: arith.constant -32768
   // CHECK: arith.constant 32767
   // CHECK: arith.cmpi slt
   // CHECK: select
   // CHECK: arith.cmpi slt
   // CHECK: select
   // CHECK: arith.trunci
-  %21 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi16>
+  %20 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi16>
 
   // CHECK: linalg.generic
   // CHECK: arith.extsi
-  %22 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi64>
+  %21 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi64>
 
   // CHECK: linalg.generic
   // CHECK: arith.constant 0
   // CHECK: arith.cmpi
-  %23 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi1>
+  %22 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xi1>
 
   // CHECK: linalg.generic
   // CHECK: arith.sitofp
-  %24 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xf32>
+  %23 = "tosa.cast"(%0) : (tensor<1xi32>) -> tensor<1xf32>
 
   // CHECK: linalg.generic
   // CHECK: arith.constant 0
   // CHECK: arith.cmpi sgt
   // CHECK: arith.subi
   // CHECK: select
-  %25 = "tosa.abs"(%arg0) : (tensor<1xi32>) -> tensor<1xi32>
+  %24 = "tosa.abs"(%arg0) : (tensor<1xi32>) -> tensor<1xi32>
 
   return
 }
 
 // -----
 
 // CHECK-LABEL: @test_simple_ui8
 func.func @test_simple_ui8(%arg0: tensor<1xi8>) -> () {
func.func @test_i8(%arg0: tensor<1xi8>) -> () {
   return
 }
 
 // -----
 
 // CHECK-LABEL: @test_clamp_f16
 func.func @test_clamp_f16(%arg0: tensor<1xf16>) -> () {
   // CHECK: linalg.generic
-  // CHECK: ^bb0(%[[ARG1:.+]]: f16, 
+  // CHECK: ^bb0(%[[ARG1:.+]]: f16,
   // CHECK-DAG: %[[C0:.+]] = arith.constant 0.0
   // CHECK-DAG: %[[C6:.+]] = arith.constant 6.0
   // CHECK-DAG: %[[MIN:.+]] = arith.minf %[[ARG1]], %[[C0]]
   // CHECK-DAG: %[[MAX:.+]] = arith.maxf %[[MIN]], %[[C6]]
   %0 = "tosa.clamp"(%arg0) {min_int = 0 : i64, max_int = 0 : i64, min_fp = 0.0 : f32, max_fp = 6.0 : f32} : (tensor<1xf16>) -> tensor<1xf16>
 
   return
 }

mlir/test/Dialect/Tosa/ops.mlir

 // -----
 // CHECK-LABEL: clamp
 func.func @test_clamp(%arg0: tensor<13x21x3xf32>) -> tensor<13x21x3xf32> {
   %0 = "tosa.clamp"(%arg0) {min_fp = 0.0 : f32, max_fp = 1.0: f32, min_int = 0 : i64, max_int = 1 : i64} : (tensor<13x21x3xf32>) -> tensor<13x21x3xf32>
   return %0 : tensor<13x21x3xf32>
 }
 
 // -----
-// CHECK-LABEL: relu
-func.func @test_relu(%arg0: tensor<13x21x3xf32>) -> tensor<13x21x3xf32> {
-  %0 = "tosa.reluN"(%arg0) {max_fp = 3.40282347E+38 : f32, max_int = 0 : i64} : (tensor<13x21x3xf32>) -> tensor<13x21x3xf32>
-  return %0 : tensor<13x21x3xf32>
-}
-
-
-// -----
 // CHECK-LABEL: sigmoid
 func.func @test_sigmoid(%arg0: tensor<13x21x3xf32>) -> tensor<13x21x3xf32> {
   %0 = "tosa.sigmoid"(%arg0) : (tensor<13x21x3xf32>) -> tensor<13x21x3xf32>
   return %0 : tensor<13x21x3xf32>
 }
 
 // -----
 // CHECK-LABEL: tanh

mlir/test/Dialect/Tosa/tosa-infer-shapes.mlir

func.func @test_unary_f32(%arg0 : tensor<4xf32>) -> () {
   %5 = "tosa.log"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
 
   // CHECK: "tosa.negate"(%arg0) : (tensor<4xf32>) -> tensor<4xf32>
   %6 = "tosa.negate"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
 
   // CHECK: "tosa.reciprocal"(%arg0) : (tensor<4xf32>) -> tensor<4xf32>
   %7 = "tosa.reciprocal"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
 
-  // CHECK: "tosa.reluN"(%arg0) {{.+}} : (tensor<4xf32>) -> tensor<4xf32>
-  %8 = "tosa.reluN"(%arg0) { max_int = 10 : i64, min_int = 0 : i64, min_fp = 0.0 : f32, max_fp = 10.0 : f32 } : (tensor<4xf32>) -> tensor<*xf32>
-
   // CHECK: "tosa.reverse"(%arg0) {axis = 0 : i64} : (tensor<4xf32>) -> tensor<4xf32>
-  %9 = "tosa.reverse"(%arg0) { axis = 0 : i64 } : (tensor<4xf32>) -> tensor<?xf32>
+  %8 = "tosa.reverse"(%arg0) { axis = 0 : i64 } : (tensor<4xf32>) -> tensor<?xf32>
 
   // CHECK: "tosa.rsqrt"(%arg0) : (tensor<4xf32>) -> tensor<4xf32>
-  %10 = "tosa.rsqrt"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
+  %9 = "tosa.rsqrt"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
 
   // CHECK: "tosa.tanh"(%arg0) : (tensor<4xf32>) -> tensor<4xf32>
-  %11 = "tosa.tanh"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
+  %10 = "tosa.tanh"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
 
   // CHECK: "tosa.sigmoid"(%arg0) : (tensor<4xf32>) -> tensor<4xf32>
-  %12 = "tosa.sigmoid"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
+  %11 = "tosa.sigmoid"(%arg0) : (tensor<4xf32>) -> tensor<*xf32>
 
   // CHECK: "tosa.cast"(%arg0) : (tensor<4xf32>) -> tensor<4xi32>
-  %13 = "tosa.cast"(%arg0) : (tensor<4xf32>) -> tensor<*xi32>
+  %12 = "tosa.cast"(%arg0) : (tensor<4xf32>) -> tensor<*xi32>
   return
 }
 
 // -----
 
 // CHECK-LABEL: @test_unary_i32
 func.func @test_unary_i32(%arg0 : tensor<4xi32>) -> () {
   // CHECK: "tosa.abs"(%arg0) : (tensor<4xi32>) -> tensor<4xi32>
   %0 = "tosa.abs"(%arg0) : (tensor<4xi32>) -> tensor<*xi32>
 
   // CHECK: "tosa.bitwise_not"(%arg0) : (tensor<4xi32>) -> tensor<4xi32>
   %1 = "tosa.bitwise_not"(%arg0) : (tensor<4xi32>) -> tensor<*xi32>
 
   // CHECK: "tosa.clamp"(%arg0) {{.+}} : (tensor<4xi32>) -> tensor<4xi32>
   %2 = "tosa.clamp"(%arg0) { max_int = 10 : i64, min_int = 0 : i64, min_fp = 0.0 : f32, max_fp = 10.0 : f32 } : (tensor<4xi32>) -> tensor<*xi32>
 
   // CHECK: "tosa.clz"(%arg0) : (tensor<4xi32>) -> tensor<4xi32>
   %3 = "tosa.clz"(%arg0) : (tensor<4xi32>) -> tensor<*xi32>
 
   // CHECK: "tosa.negate"(%arg0) : (tensor<4xi32>) -> tensor<4xi32>
   %4 = "tosa.negate"(%arg0) : (tensor<4xi32>) -> tensor<*xi32>
 
-  // CHECK: "tosa.reluN"(%arg0) {{.+}} : (tensor<4xi32>) -> tensor<4xi32>
-  %5 = "tosa.reluN"(%arg0) { max_int = 10 : i64, min_int = 0 : i64, min_fp = 0.0 : f32, max_fp = 10.0 : f32 } : (tensor<4xi32>) -> tensor<*xi32>
-
   // CHECK: "tosa.reverse"(%arg0) {axis = 0 : i64} : (tensor<4xi32>) -> tensor<4xi32>
-  %6 = "tosa.reverse"(%arg0) { axis = 0 : i64 } : (tensor<4xi32>) -> tensor<?xi32>
+  %5 = "tosa.reverse"(%arg0) { axis = 0 : i64 } : (tensor<4xi32>) -> tensor<?xi32>
 
   // CHECK: "tosa.rescale"(%arg0) {{.+}} : (tensor<4xi32>) -> tensor<4xi16>
-  %7 = "tosa.rescale"(%arg0) {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = [42 : i32, 43 : i32], shift = [14 : i32, 15 : i32], scale32 = false, double_round = false, per_channel = false} : (tensor<4xi32>)  -> (tensor<*xi16>)
+  %6 = "tosa.rescale"(%arg0) {input_zp = 243 : i32, output_zp = 252 : i32, multiplier = [42 : i32, 43 : i32], shift = [14 : i32, 15 : i32], scale32 = false, double_round = false, per_channel = false} : (tensor<4xi32>)  -> (tensor<*xi16>)
 
   // CHECK: "tosa.identity"(%arg0) : (tensor<4xi32>) -> tensor<4xi32>
-  %8 = "tosa.identity"(%arg0) : (tensor<4xi32>) -> tensor<?xi32>
+  %7 = "tosa.identity"(%arg0) : (tensor<4xi32>) -> tensor<?xi32>
   return
 }
 
 // -----
 
 // CHECK-LABEL: @test_binary_scalar_f32
 func.func @test_binary_scalar_f32(%arg0 : tensor<4xf32>, %arg1 : tensor<f32>) -> () {
   // CHECK: "tosa.add"(%arg0, %arg1) : (tensor<4xf32>, tensor<f32>) -> tensor<4xf32>