// RUN: tf-opt -split-input-file -verify-diagnostics -tf-resource-device-inference %s | FileCheck %s

!tf_res = tensor<*x!tf_type.resource<tensor<32xf32>>>

// Tests that the pass can correctly propagate device attributes inside the same
// function.

// CHECK-LABEL: func @propagate_in_function
func.func @propagate_in_function(
  %arg0: !tf_res {tf.device = "/TPU:0"},
  %arg1: !tf_res {tf.device = "/TPU:1"}) {
  tf_executor.graph {

// RUN: tac-translate -input-mlir -output-mlir -device-specs=NNAPI %s -o - 2>&1 | FileCheck %s

module {
  // CHECK-LABEL: main
  func.func @main(%arg0: tensor<4xf32>, %arg1: tensor<4xf32>) -> tensor<4xf32> {
    %0 = "tfl.squared_difference"(%arg0, %arg1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
    func.return %0 : tensor<4xf32>
    // CHECK:  [[VAL_0:%.*]] = tfl.sub %arg0, %arg1 {fused_activation_function = "NONE"} : tensor<4xf32>

// Verify arg/ret attributes are exported as device assignment for arg/retval
// nodes.

module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 121 : i32}} {
  func.func @main(%arg0: tensor<*xf32> {tf.device = "/CPU:0"}, %arg1: tensor<2x4x6x8xi32>) -> (tensor<*xf32>, tensor<2x4x6x8xi32> {tf.device = "/CPU:1"})

# Verify arg and ret devices are added as arg and ret attributes.

# CHECK-LABEL: func @main
# CHECK-SAME:  (%[[ARG_0:[a-z0-9]+]]: tensor<*xf32> {tf.device = "/CPU:0"}, %[[ARG_1:[a-z0-9]+]]: tensor<2x4x6x8xi32>) -> (tensor<*xf32>, tensor<*xi32> {tf.device = "/CPU:1"})

node {
  name: "args_0"
  op: "_Arg"
  device: "/CPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "index"
    value {

// RUN: tac-opt-all-backends -tfl-device-transform-gpu %s -split-input-file -verify-diagnostics | FileCheck %s

func.func @pack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<2x1xf32> {
  %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %0 : tensor<2x1xf32>
}

// CHECK:   func @pack(%[[VAL_0:.*]]: tensor<1xf32>, %[[VAL_1:.*]]: tensor<1xf32>) -> tensor<2x1xf32> {

// RUN: tac-opt-all-backends -tfl-device-transform-nnapi %s -split-input-file -verify-diagnostics | FileCheck %s

func.func @mean_4d_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x1x1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = true} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x1x1x512xf32>
  func.return %0 : tensor<1x1x1x512xf32>
}

A. Unique TensorFlower <******@****.***> 1605121757 -0800

  ASSERT_FALSE(arrived);
  TFE_TensorHandle* hdevice =
      TFE_TensorHandleCopyToDevice(hcpu, context, name, status.get());
  ASSERT_TRUE(arrived);
  ASSERT_FALSE(executed);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> matmul(
      MatMulOp(context, hcpu, hdevice), TFE_DeleteOp);
  TFE_OpSetDevice(matmul.get(), name, status.get());

// Creates a pass that extracts outside compilation (Host ops inside device
// cluster) at head/tail of Device cluster to run before/after XLA computation.
std::unique_ptr<mlir::OperationPass<mlir::ModuleOp>>
CreateExtractHeadTailOutsideCompilationPass();

// Creates a pass that extract outside compilation (Host ops inside cevice
// cluster) ops to a separate parallel_execute region to run on CPU.

    needed to apply space to depth optimization for a model that needs to load a
    pre-train checkpoint. This transform can be done on the host or TPU device
    based on the cost. As the size of the kernel is relatively small, this won't
    add additional cost to TPU device time. Below is the logic to transform the
    kernel of shape [7, 7, 3, 64] to [4, 4, 12, 84].

    ```python
    conv0 = tf.compat.v1.layers.Conv2D(