func.return %1 : tensor<16x128xf32>
  // CHECK: return %0 : tensor<16x128xf32>
}

// CHECK-LABEL: FuseTransposeFCLhsToBatchMatmul
func.func @FuseTransposeFCLhsToBatchMatmul(%arg0: tensor<1024x4xf32>, %arg1: tensor<8x1024xf32>, %arg2: tensor<4x256xf32>) -> tensor<8x256xf32> {

  %cst = arith.constant dense<1> : tensor<1xi32>
  %0 = "tfl.exp"(%arg0) : (tensor<8x128xf32>) -> tensor<8x128xf32>
  %1 = "tfl.sum"(%0, %cst) {keep_dims = true} : (tensor<8x128xf32>, tensor<1xi32>) -> tensor<8x1xf32>
  %2 = "tfl.div"(%0, %1) {fused_activation_function = "NONE"} : (tensor<8x128xf32>, tensor<8x1xf32>) -> tensor<8x128xf32>
  func.return %2 : tensor<8x128xf32>

      %17 = "tf.Concat"(%cst_1, %16#5, %16#15, %16#25, %16#35) : (tensor<i32>, tensor<32x1024xf32>, tensor<32x1024xf32>, tensor<32x1024xf32>, tensor<32x1024xf32>) -> tensor<128x1024xf32>

  func.return %0 : tensor<1024xf32>

  // 1.0 * 2.0 * 512 + 4.0 = 1028.0
  // CHECK: %[[CST:.*]] = arith.constant dense<1.028000e+03> : tensor<1024xf32>
  // CHECK:  return %[[CST]]
}

// CHECK-LABEL: @ConstantFoldFullyConnectedNoBias
func.func @ConstantFoldFullyConnectedNoBias() -> tensor<1024xf32> {

constexpr absl::string_view kModuleStatic = R"mlir(
  module {
    func.func @main(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      return %0 : tensor<1x3xf32>
    }
  }
)mlir";

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_a", device = "", transpose_a = false, transpose_b = false} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    return %0 : tensor<*xf32>
  }

// CHECK: %[[cst:.*]] = "arith.constant"() <{value = dense<0.000000e+00> : tensor<2x1024xf32>}> : () -> tensor<2x1024xf32>

}
func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
  %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
  return %0 : tensor<1x3xf32>
}
// CHECK-LABEL: func.func @main
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x1024xf32>) -> tensor<1x3xf32>

    %1 = stablehlo.constant dense<-1.280000e+02> : tensor<1x1024xf32>
    %2 = stablehlo.constant dense<0.003921567> : tensor<1x1024xf32>
    %3 = stablehlo.divide %arg0, %2 : tensor<1x1024xf32>
    %4 = stablehlo.add %3, %1 : tensor<1x1024xf32>
    %5 = "tf.Identity"(%4) {device = ""} : (tensor<1x1024xf32>) -> tensor<1x1024xf32>
    return %5 : tensor<1x1024xf32>
  }

}

    %19 = stablehlo.convert %18 : (tensor<8x16x4xi32>) -> tensor<8x16x4xf32>
    %20 = stablehlo.broadcast_in_dim %9, dims = [0, 1, 2] : (tensor<1x1x1xf32>) -> tensor<8x16x4xf32>
    %21 = stablehlo.multiply %19, %20 : tensor<8x16x4xf32>  // * s1 s2
    %22 = call @uniform_quantize_1(%21, %7, %8) : (tensor<8x16x4xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<8x16x4xi8>