%fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%r1, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x48xf32>, tensor<48xf32>, tensor<48xf32>) -> tensor<3x3x48xf32>
  %r2 = "tf.Reshape"(%fq, %s2) {T = f32, Tshape = i32, device = ""} : (tensor<3x3x48xf32>, tensor<4xi32>) -> tensor<3x3x48x1xf32>

  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%r1, %mini, %maxi) {num_bits = 3, narrow_range = false} : (tensor<3x3x48xf32>, tensor<48xf32>, tensor<48xf32>) -> tensor<3x3x48xf32>
  %r2 = "tf.Reshape"(%fq, %s2) {T = f32, Tshape = i32, device = ""} : (tensor<3x3x48xf32>, tensor<4xi32>) -> tensor<3x3x48x1xf32>

func.func @batchmatmulv2_basic(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
// CHECK-LABEL:   func @batchmatmulv2_basic
// CHECK-SAME:        ([[LHS:%.*]]: tensor<1x4x2xf32>, [[RHS:%.*]]: tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
// CHECK:           [[LHSSHAPE:%.*]] = shape.shape_of [[LHS]] : tensor<1x4x2xf32>
// CHECK:           [[RHSSHAPE:%.*]] = shape.shape_of [[RHS]] : tensor<3x2x4xf32>
// CHECK:           [[CM2:%.*]] = arith.constant -2 : index

  %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
  func.return %0 : tensor<3x4x4xf32>
}

func.func @unary_einsum_reduce_sum_transpose(%arg0: tensor<3x4x5x6xf32>) -> tensor<3x5x4xf32> {
  %0 = "tf.Einsum"(%arg0) {T = "tfdtype$DT_FLOAT", equation = "...gse->...sg"}: (tensor<3x4x5x6xf32>) -> tensor<3x5x4xf32>
  func.return %0 : tensor<3x5x4xf32>
  // CHECK-LABEL: unary_einsum_reduce_sum_transpose
  // CHECK-DAG: %[[cst:.*]] = arith.constant dense<3> : tensor<1xi32>

func.func @conv_fn(%arg0: tensor<1x3x3x4xf32>) -> tensor<1x3x3x4xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<3x3x4x4xf32>
  %1 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x4xf32>) -> tensor<1x3x3x4xf32>
  func.return %1: tensor<1x3x3x4xf32>
}

        %0 = stablehlo.constant dense<2.000000e+00> : tensor<3x3x4x4xf32>
        %1 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x4xf32>) -> tensor<1x3x3x4xf32>

// Contains the `stablehlo.transpose` op of the arg (e.g. [b, f, 0, 1] to
// [b, 0, 1, f]). The weight constant is folded into [0, 1, i, o] format.
// CHECK-DAG: %[[CST:.+]] = stablehlo.constant dense<3.000000e+00> : tensor<3x3x8x8xf32>
// CHECK: %[[TRANSPOSE_1:.+]] = stablehlo.transpose %arg0, dims = [0, 2, 3, 1] : (tensor<1x8x4x4xf32>) -> tensor<1x4x4x8xf32>

    // CHECK: mhlo.reduce
    // CHECK: mhlo.dot_general
    // CHECK: mhlo.transpose
    %0 = "tf.BatchMatMulV2"(%arg0, %arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<3x2x4xf32>) -> tensor<3x4x4xf32>
    func.return %0 : tensor<3x4x4xf32>
  }

  // CHECK-LABEL: approx_topk

    %9 = stablehlo.convert %3 : (tensor<3x3x4x4xi8>) -> tensor<3x3x4x4xf32>
    %10 = stablehlo.convolution(%8, %9) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x4xf32>) -> tensor<1x3x3x4xf32>
    %11 = stablehlo.reshape %2 : (tensor<1x1x1x1xi8>) -> tensor<1xi8>