// RUN: odml-to-stablehlo-opt %s -tfl-parse-stablehlo-ops | FileCheck %s

module {
func.func @main(%arg0: tensor<2xi32>) -> tensor<2xi32> attributes {tf.entry_function = {inputs = "arg0", outputs = "tfl.custom1"}} {
  %0 = "tfl.custom"(%arg0, %arg0) {custom_code = "stablehlo.subtract", custom_option = #tfl<const_bytes : "0x00000100002401">} : (tensor<2xi32>, tensor<2xi32>) -> tensor<2xi32>
  func.return %0 : tensor<2xi32>
}
}

// CHECK:       module {

// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer --emit-stablehlo-ops=true %s -o - | flatbuffer_translate --tflite-flatbuffer-to-mlir --disable-vhlo-to-stablehlo=true - -o - | FileCheck %s
// test stablehlo roundtrip

module attributes {tfl.metadata = {"keep_stablehlo_constant" = "true"}} {
 func.func @main () -> tensor<1x1x1x96xf32> {
  %0 = "vhlo.constant_v1"() <{value = #vhlo.tensor_v1<dense<0.000000e+00> : tensor<f32>>}> : () -> tensor<1x1x1x96xf32>

// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer %s -o - | flatbuffer_translate --tflite-flatbuffer-to-mlir - -o - | FileCheck %s

// Confirm a wide array of attribute survives the round-trip
func.func @main(tensor<1x6x6x16xf32>) -> tensor<1x1x1x16xf32> {
^bb0(%arg0: tensor<1x6x6x16xf32>):

// RUN: tf-opt "--tfxla-device-specific-transforms=device-type=XLA_GPU_JIT" -verify-diagnostics -split-input-file %s | FileCheck -dump-input=fail %s

module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 1399 : i32}} {

// CHECK-LABEL: stateless_op
func.func @stateless_op() -> tensor<i32> {
  // CHECK: %cst = "tf.Const"() <{value = dense<1> : tensor<i32>}> : () -> tensor<i32>

// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer %s -o - | flatbuffer_translate --tflite-flatbuffer-to-mlir - -o - | FileCheck %s

func.func @main(%arg0: tensor<32x4x4x128xf32>, %arg1: tensor<1x32x42x128xf32>, %arg2: tensor<4xi32>) -> tensor<1x64x84x32xf32> {

// RUN: tf-mlir-translate -mlir-tf-mlir-to-str-attr -mlir-print-local-scope %s | FileCheck %s

module attributes {tf.versions = {producer = 888 : i32}} {
  func.func @main(%arg0: tensor<?xi32>) -> tensor<?xi32> {
    %0 = "tf.Identity"(%arg0) : (tensor<?xi32>) -> tensor<?xi32> loc(unknown)
    func.return %0 : tensor<?xi32> loc(unknown)
  } loc(unknown)
} loc(unknown)

// RUN: flatbuffer_translate -mlir-to-tflite-flatbuffer --use-buffer-offset %s -o - | flatbuffer_translate --tflite-flatbuffer-to-mlir - -o - | FileCheck %s

func.func @main(%arg0: tensor<32x4x4x128xf32>, %arg1: tensor<1x32x42x128xf32>, %arg2: tensor<4xi32>) -> tensor<1x64x84x32xf32> {

// RUN: stablehlo-quant-opt %s -stablehlo-remove-sharding-custom-call \
// RUN:   -split-input-file | FileCheck %s

// CHECK-LABEL: sharding_custom_call_removed
func.func @sharding_custom_call_removed(%arg0: tensor<3xf32>) -> tensor<3xf32> {
  %1 = stablehlo.custom_call @Sharding(%arg0) {mhlo.sharding = ""} : (tensor<3xf32>) -> tensor<3xf32>
  return %1 : tensor<3xf32>
}
// CHECK-NOT: custom_call

// -----

// RUN: tf-quant-opt %s -split-input-file -quant-preprocess-op='target-opset=XLA quantization-method=weight_only enable-per-channel-quantization=true' | FileCheck --check-prefix PerChannel %s

module {
  // For XLA weight-only per-channel depthwise convolution, tensor shape should have

// RUN: tf-opt %s -tfl-prepare-tf -tfl-legalize-tf='run-tfl-runtime-verification=false'  -tfl-optimize | FileCheck %s

func.func @broadcast_to_bf16(%arg0: tensor<3xbf16>, %arg1: tensor<2xi64>) -> tensor<3x3xbf16> {
  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<3xbf16>, tensor<2xi64>) -> tensor<3x3xbf16>
  func.return %0: tensor<3x3xbf16>

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>