Solve merge conflicts and fixed tests after bump

Propagate the new TOSA mul shift operand into downstream test expectations and update onnx-mlir and torch-mlir to always supply the shift constant.

Author: mgehre-amd

mlir/test/Dialect/Tosa/canonicalize.mlir

@@ -607,8 +607,9 @@ func.func @mul_zero_broadcast_dynamic_result(%arg0: tensor<?x3xf32>) -> (tensor<
   // CHECK: tosa.mul
   // CHECK: tosa.mul
   %zeros = "tosa.const"() {value = dense<0.0> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
-  %1 = tosa.mul %arg0, %zeros {shift = 0 : i8} : (tensor<?x3xf32>, tensor<1x1xf32>) -> tensor<?x3xf32>
-  %2 = tosa.mul %zeros, %arg0 {shift = 0 : i8} : (tensor<1x1xf32>, tensor<?x3xf32>) -> tensor<?x3xf32>
+  %shift = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %1 = tosa.mul %arg0, %zeros, %shift : (tensor<?x3xf32>, tensor<1x1xf32>, tensor<1xi8>) -> tensor<?x3xf32>
+  %2 = tosa.mul %zeros, %arg0, %shift : (tensor<1x1xf32>, tensor<?x3xf32>, tensor<1xi8>) -> tensor<?x3xf32>
   return %1, %2 : tensor<?x3xf32>, tensor<?x3xf32>
 }
 
@@ -1581,7 +1582,8 @@ func.func @canonicalize_select_lrelu_zero_pattern(%arg0: tensor<13x21x3xf32>) ->
 // CHECK:           %[[VAL_1:.*]] = tosa.clamp %arg{{.*}} {max_fp = 0x7F800000 : f32, max_int = 9223372036854775807 : i64, min_fp = 0.000000e+00 : f32, min_int = -9223372036854775808 : i64} : (tensor<13x21x3xf32>) -> tensor<13x21x3xf32>
 // CHECK:           return %[[VAL_1]] : tensor<13x21x3xf32>
   %0 =  "tosa.const"() <{value = dense<0.000000e+00> : tensor<1x1x1xf32>}>: () -> tensor<1x1x1xf32>
-  %1 = tosa.mul %arg0, %0 {shift = 0 : i8}: (tensor<13x21x3xf32>, tensor<1x1x1xf32>) -> tensor<13x21x3xf32>
+  %shift = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %1 = tosa.mul %arg0, %0, %shift : (tensor<13x21x3xf32>, tensor<1x1x1xf32>, tensor<1xi8>) -> tensor<13x21x3xf32>
   %2 = tosa.greater_equal %arg0, %0: (tensor<13x21x3xf32>, tensor<1x1x1xf32>) -> tensor<13x21x3xi1>
   %3 = tosa.select %2, %arg0, %1: ( tensor<13x21x3xi1>, tensor<13x21x3xf32>, tensor<13x21x3xf32>) -> tensor<13x21x3xf32>
   return %3  :  tensor<13x21x3xf32>
@@ -1642,4 +1644,3 @@ func.func @canonicalize_select_to_clamp_i8_and_i64_pat2(%arg0: tensor<13x21x3xi8
   %3 = tosa.select %2, %1, %arg1: ( tensor<13x21x3xi1>, tensor<13x21x3xi64>, tensor<13x21x3xi64>) -> tensor<13x21x3xi64>
   return %3  :  tensor<13x21x3xi64>
 }
-

mlir/test/Dialect/Tosa/constant-mul-opt.mlir

@@ -17,7 +17,8 @@ func.func @mul_fold_float() -> tensor<4xf16> {
                         dense<[-132.7, -3.0, -0.0, 5.0]> :
                         tensor<4xf16>
                       } : () -> tensor<4xf16>
-  %2 = "tosa.mul"(%0, %1) : (tensor<4xf16>, tensor<4xf16>) -> tensor<4xf16>
+  %shift = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %2 = "tosa.mul"(%0, %1, %shift) : (tensor<4xf16>, tensor<4xf16>, tensor<1xi8>) -> tensor<4xf16>
   return %2 : tensor<4xf16>
 }
 
@@ -34,7 +35,8 @@ func.func @mul_fold_float_infinity_nan() -> tensor<7xf32> {
                         dense<[3.0, -3.0, -3.0, 3.0, 1.0, 0xFF800000, 0.0]> :
                         tensor<7xf32>
                       } : () -> tensor<7xf32>
-  %2 = "tosa.mul"(%0, %1) : (tensor<7xf32>, tensor<7xf32>) -> tensor<7xf32>
+  %shift = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %2 = "tosa.mul"(%0, %1, %shift) : (tensor<7xf32>, tensor<7xf32>, tensor<1xi8>) -> tensor<7xf32>
   return %2 : tensor<7xf32>
 }
 
@@ -51,7 +53,8 @@ func.func @add_fold_float_overflow() -> tensor<2xf32> {
                         dense<[2.1e+38, 1.1e+38]> :
                         tensor<2xf32>
                       } : () -> tensor<2xf32>
-  %2 = "tosa.mul"(%0, %1) : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xf32>
+  %shift = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
+  %2 = "tosa.mul"(%0, %1, %shift) : (tensor<2xf32>, tensor<2xf32>, tensor<1xi8>) -> tensor<2xf32>
   return %2 : tensor<2xf32>
 }
 

mlir/test/Dialect/Tosa/tosa-reduce-transposes.mlir

@@ -187,6 +187,7 @@ func.func @test_reshape_for_broadcast(%arg0: tensor<4x3x2xi32>) -> tensor<4x3x2x
 
 // CHECK-LABEL: @test_resnet18_common_case
 // COM: note that %74 is now represented by %arg2
+// CHECK-DAG: %[[CONST0:.+]] = "tosa.const"() <{value = dense<0> : tensor<1xi8>}> : () -> tensor<1xi8>
 // CHECK-DAG: %[[VAL_2:.*]] = "tosa.const"() <{value = dense_resource<torch_tensor_64_torch.float32_1> : tensor<64xf32>}> : () -> tensor<64xf32>
 // CHECK-DAG: %[[VAL_3:.*]] = "tosa.const"() <{value = dense_resource<torch_tensor_64_torch.float32> : tensor<64xf32>}> : () -> tensor<64xf32>
 // CHECK-DAG: %[[VAL_4:.*]] = "tosa.const"() <{value = dense<9.99999974E-6> : tensor<1xf32>}> : () -> tensor<1xf32>
@@ -197,9 +198,9 @@ func.func @test_reshape_for_broadcast(%arg0: tensor<4x3x2xi32>) -> tensor<4x3x2x
 // CHECK-DAG: %[[VAL_9:.*]] = tosa.reshape %arg0 {new_shape = array<i64: 1, 1, 1, 64>} : (tensor<64xf32>) -> tensor<1x1x1x64xf32>
 // CHECK-DAG: %[[VAL_10:.*]] = tosa.sub %arg2, %[[VAL_9]] : (tensor<1x112x112x64xf32>, tensor<1x1x1x64xf32>) -> tensor<1x112x112x64xf32>
 // CHECK-DAG: %[[VAL_11:.*]] = tosa.reshape %[[VAL_8]] {new_shape = array<i64: 1, 1, 1, 64>} : (tensor<64xf32>) -> tensor<1x1x1x64xf32>
-// CHECK-DAG: %[[VAL_12:.*]] = tosa.mul %[[VAL_10]], %[[VAL_11]] : (tensor<1x112x112x64xf32>, tensor<1x1x1x64xf32>) -> tensor<1x112x112x64xf32>
+// CHECK-DAG: %[[VAL_12:.*]] = tosa.mul %[[VAL_10]], %[[VAL_11]], %[[CONST0]] : (tensor<1x112x112x64xf32>, tensor<1x1x1x64xf32>, tensor<1xi8>) -> tensor<1x112x112x64xf32>
 // CHECK-DAG: %[[VAL_13:.*]] = tosa.reshape %[[VAL_3]] {new_shape = array<i64: 1, 1, 1, 64>} : (tensor<64xf32>) -> tensor<1x1x1x64xf32>
-// CHECK-DAG: %[[VAL_14:.*]] = tosa.mul %[[VAL_12]], %[[VAL_13]] : (tensor<1x112x112x64xf32>, tensor<1x1x1x64xf32>) -> tensor<1x112x112x64xf32>
+// CHECK-DAG: %[[VAL_14:.*]] = tosa.mul %[[VAL_12]], %[[VAL_13]], %[[CONST0]] : (tensor<1x112x112x64xf32>, tensor<1x1x1x64xf32>, tensor<1xi8>) -> tensor<1x112x112x64xf32>
 // CHECK-DAG: %[[VAL_15:.*]] = tosa.reshape %[[VAL_2]] {new_shape = array<i64: 1, 1, 1, 64>} : (tensor<64xf32>) -> tensor<1x1x1x64xf32>
 // CHECK-DAG: %[[VAL_16:.*]] = tosa.add %[[VAL_14]], %[[VAL_15]] : (tensor<1x112x112x64xf32>, tensor<1x1x1x64xf32>) -> tensor<1x112x112x64xf32>
 // CHECK-DAG: %[[VAL_17:.*]] = tosa.clamp %[[VAL_16]] {max_fp = 3.40282347E+38 : f32, max_int = 2147483647 : i64, min_fp = 0.000000e+00 : f32, min_int = 0 : i64} : (tensor<1x112x112x64xf32>) -> tensor<1x112x112x64xf32>
@@ -220,9 +221,9 @@ func.func @test_resnet18_common_case(%arg0: tensor<64xf32>, %arg1: tensor<64xf32
     %79 = tosa.reshape %arg0 {new_shape = array<i64: 1, 64, 1, 1>} : (tensor<64xf32>) -> tensor<1x64x1x1xf32>
     %80 = tosa.sub %75, %79 : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x112x112xf32>
     %81 = tosa.reshape %78 {new_shape = array<i64: 1, 64, 1, 1>} : (tensor<64xf32>) -> tensor<1x64x1x1xf32>
-    %82 = tosa.mul %80, %81 : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x112x112xf32>
+    %82 = tosa.mul %80, %81, %shift : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>, tensor<1xi8>) -> tensor<1x64x112x112xf32>
     %83 = tosa.reshape %60 {new_shape = array<i64: 1, 64, 1, 1>} : (tensor<64xf32>) -> tensor<1x64x1x1xf32>
-    %84 = tosa.mul %82, %83 : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x112x112xf32>
+    %84 = tosa.mul %82, %83, %shift : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>, tensor<1xi8>) -> tensor<1x64x112x112xf32>
     %85 = tosa.reshape %59 {new_shape = array<i64: 1, 64, 1, 1>} : (tensor<64xf32>) -> tensor<1x64x1x1xf32>
     %86 = tosa.add %84, %85 : (tensor<1x64x112x112xf32>, tensor<1x64x1x1xf32>) -> tensor<1x64x112x112xf32>
     %87 = tosa.clamp %86 {max_fp = 3.40282347E+38 : f32, max_int = 2147483647 : i64, min_fp = 0.000000e+00 : f32, min_int = 0 : i64} : (tensor<1x64x112x112xf32>) -> tensor<1x64x112x112xf32>