// CHECK-DAG: %[[input_18:.*]] = "quantfork.stats"({{.*}}) <{layerStats = dense<[-8.000000e-01, 1.600000e+00]> : tensor<2xf32>}> : (tensor<1x4xf32>) -> tensor<1x4xf32>
// CHECK-DAG: %[[input_19:.*]] = "quantfork.stats"({{.*}}) <{layerStats = dense<[-2.000000e+00, 4.000000e+00]> : tensor<2xf32>}> : (tensor<1x2xf32>) -> tensor<1x2xf32>

// CHECK: "tfl.unidirectional_sequence_lstm"({{.*}}, %[[input_18]], %[[input_19]], %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}})

// RUN:   -split-input-file | FileCheck %s --check-prefix=STATIC-RANGE-PTQ-TO-COMPUTE-HEAVY

// STATIC-RANGE-PTQ-TO-COMPUTE-HEAVY: @main
func.func @main(%arg0: tensor<1x2xf32>) -> tensor<1x2xf32> {
  %0 = stablehlo.add %arg0, %arg0 : tensor<1x2xf32>
  return %0 : tensor<1x2xf32>
}
// Tests that `composite_add_fn_1` does not quantize when quantizing
// only compute-heavy ops.

// CHECK-LABEL: reduce_window_max_activation_transpose
// CHECK-SAME: (%[[ARG:.+]]: tensor<16x8xf32>) -> tensor<4x8xf32>
func.func @reduce_window_max_activation_transpose_rank2(%arg0: tensor<16x8xf32>) -> tensor<4x8xf32> {
  %0 = stablehlo.constant dense<0xFF800000> : tensor<f32>  // -inf
  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<16x8xf32>) -> tensor<8x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({

  %split_dim_2 = arith.constant dense<1> : tensor<1xi32>
  %4, %5 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x2xf32>)
  %6:2 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x?xf32>)

}

// CHECK-LABEL: QuantizeConcat
func.func @QuantizeConcat(tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.000000e-01:128>> {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>):
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2xf32>

      %2 = "tf.ReadVariableOp"(%arg15) : (tensor<*x!tf_type.resource<tensor<7x7x3x64xf32>>>) -> tensor<7x7x3x64xf32>
      %3 = "tf.ReadVariableOp"(%arg16) : (tensor<*x!tf_type.resource<tensor<1001xf32>>>) -> tensor<1001xf32>
      %4 = "tf.ReadVariableOp"(%arg17) : (tensor<*x!tf_type.resource<tensor<64x1001xf32>>>) -> tensor<64x1001xf32>
      %5 = "tf.ReadVariableOp"(%arg18) : (tensor<*x!tf_type.resource<tensor<f32>>>) -> tensor<f32>

  %0 = "tf_device.cluster_func"(%arg0, %arg1) {func = @_func, use_spmd_for_xla_partitioning = true, use_tpu = true, num_cores_per_replica = 2 : i64} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1:2 = "tf.TPUPartitionedOutputV2"(%0) {device = "", partition_dims = [2, 1]} : (tensor<2x2xf32>) -> (tensor<1x2xf32>, tensor<1x2xf32>)

  %0 = "tf.Const"() {value = dense<0.0> : tensor<1x2xf32>} : () -> tensor<1x2xf32>

  // Operand and identity shapes are broadcastable. However, we cannot fold
  // because the operand does not match the result shape.
  %1 = "tf.AddV2"(%arg0, %0) : (tensor<4x1xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>

func.func @float_matmul(
  %arg0: tensor<1x10xf32>, %arg1: tensor<10x10xf32>) -> (tensor<*xf32>, tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %0 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = false
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>

// support dynamic shaped operands like the native lowering. Verify that
// fallback lowering is preferred for static shaped operands when available.

// CHECK-LABEL: batchmatmulv2
func.func @batchmatmulv2(%arg0: tensor<1x4x2xf32>, %arg1: tensor<3x2x4xf32>) -> tensor<3x4x4xf32> {
  // NO_FALLBACK: mhlo.dynamic_broadcast_in_dim
  // NO_FALLBACK: mhlo.dot_general

  // SUPPORTED_FALLBACK_DEVICE: mhlo.reduce