// CHECK-LABEL: test_batch_matmul_adj_to_einsum
  // CHECK: %[[RES_EINSUM:[0-9]*]] = "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...nk->...mn"}> : (tensor<1x2x3xf32>, tensor<4x3xf32>) -> tensor<1x2x4xf32>
  // CHECK: return %[[RES_EINSUM]] : tensor<1x2x4xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = true} : (tensor<1x2x3xf32>, tensor<4x3xf32>) -> tensor<1x2x4xf32>

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

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

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
  %1 = "tfl.pseudo_const"() {value = dense<[[0.1]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kModuleXlaCallModuleOpWithDefaultQuantizationMethod);
  ASSERT_TRUE(module_op);

// CHECK-NEXT:      %22 = mhlo.dynamic_reshape %2, %21 : (tensor<2x?x3x4xf32>, tensor<4xi32>) -> tensor<2x?x3x4xf32>
// CHECK-NEXT:      %23 = "tfl.batch_matmul"(%12, %22) <{adj_x = false, adj_y = false, asymmetric_quantize_inputs = false}> : (tensor<2x?x2x3xf32>, tensor<2x?x3x4xf32>) -> tensor<2x?x2x4xf32>
// CHECK-NEXT:      %24 = "tfl.shape"(%arg0) : (tensor<2x?x2x3xf32>) -> tensor<4xi32>

      return %0 : tensor<1x1x4xf32>
    }
  )mlir";

  const OwningOpRef<ModuleOp> module_op =
      ParseModuleOpString(kXlaCallModuleOpWithQuantizationMethodAttr);
  ASSERT_TRUE(module_op);

    %14 = stablehlo.multiply %12, %13 : tensor<1x4x3xf32>  // s1 * s2
    %15 = call @uniform_quantize_1(%14, %5, %6) : (tensor<1x4x3xf32>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xi8>
    %16 = call @uniform_dequantize_0(%15, %5, %6) : (tensor<1x4x3xi8>, tensor<1x1x1xf32>, tensor<1x1x1xi8>) -> tensor<1x4x3xf32>
    return %16 : tensor<1x4x3xf32>
  }

  // Use a four dimension sharding (devices=[1,1,1,1]0)
  // Since the input tensor only has three dimensions, we expect this to fail.
  %0 = "tf.XlaSharding"(%arg0) { _XlaSharding = "\08\03\1A\04\01\01\01\01\22\01\00" } : (tensor<1x2x3xi32>) -> tensor<1x2x3xi32>
  %1 = "tf.A"(%0) : (tensor<1x2x3xi32>) -> (tensor<1x2x3xi32>)
  func.return %1: tensor<1x2x3xi32>
}

// -----

// CHECK-LABEL: func @check_retval_sharding_errors

// CHECK:           %[[MUL:.*]] = "tf.Mul"(%[[SUB]], %[[CAST0]]) : (tensor<1x24xi32>, tensor<1x24xi32>) -> tensor<1x24xi32>
// CHECK:           %[[SCATTER1:.*]] = "tf.TensorScatterAdd"(%[[CST1]], %arg1, %[[CAST1]]) : (tensor<1x24xi32>, tensor<1x2x2xi32>, tensor<1x2xi32>) -> tensor<1x24xi32>

  func.return %17 : tensor<1x2x3xf32>

  // CHECK: %[[NONE:.*]] = "tfl.no_value"() <{value}> : () -> none
  // CHECK: %[[DQ_1:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_2:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>
  // CHECK: %[[DQ_3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x1xf16>) -> tensor<1x1xf32>