// -----

// Tests composite tf.XlaCallModule is converted to func.call.

module {
  // CHECK-LABEL: func.func @main
  func.func @main(%arg0: tensor<1x1024xf32>) -> tensor<1x3xf32> {
    // CHECK: call @composite_dot_general_fn_1
    // CHECK-SAME: (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
    // CHECK-NOT: tf.XlaCallModule
    %0 = "tf.Const"() <{value = dense<0.5> : tensor<1024x3xf32>}> : () -> tensor<1024x3xf32>

}
func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
  %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
  return %0 : tensor<1x3xf32>
}
// CHECK-LABEL: func.func @main
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x1024xf32>) -> tensor<1x3xf32>

  func.return %1 : tensor<16x128xf32>
  // CHECK: return %0 : tensor<16x128xf32>
}

// CHECK-LABEL: FuseTransposeFCLhsToBatchMatmul
func.func @FuseTransposeFCLhsToBatchMatmul(%arg0: tensor<1024x4xf32>, %arg1: tensor<8x1024xf32>, %arg2: tensor<4x256xf32>) -> tensor<8x256xf32> {

  func.func @main(%arg0: tensor<128x10xf32>, %arg1: tensor<10x1024xf32>, %arg2: tensor<128x1024xf32>) -> (tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, tensor<128x1024xf32> {mhlo.sharding = ""}) {
    func.return %arg0, %arg1, %arg2 : tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>
  }
}

// The following xla::OpSharding protos are used:

  // CHECK: "tf.Identity"(%[[SHARDED_ARG1]])
  %0 = "tf.Identity"(%arg1) : (tensor<10x1024xf32>) -> tensor<10x1024xf32>

  // CHECK: "tf.Identity"(%arg2)
  %1 = "tf.Identity"(%arg2) : (tensor<128x1024xf32>) -> tensor<128x1024xf32>
  func.return %arg0, %0, %1 : tensor<128x10xf32>, tensor<10x1024xf32>, tensor<128x1024xf32>
}

// CHECK-LABEL: @RewriteHostComputeMlirOp

module attributes {tf.versions = {producer = 179 : i32}} {
  func.func @main(%arg0: tensor<128x10xf32> {mhlo.sharding = "\08\03\1A\02\01\02\22\02\00\01"}, %arg1: tensor<10x1024xf32> {mhlo.sharding = "\08\01\1A\01\01\22\01\00"}, %arg2: tensor<128x1024xf32> {mhlo.sharding = ""}) {
    func.return
  }
}

// The following xla::OpSharding protos are used:
//  Serialized string:
//   "\08\03\1A\02\01\02\22\02\00\01"

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x1x2xbf16>) -> tensor<1x1x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x1x2xf32>) -> tensor<1x1x2xf32>
  return %3 : tensor<1x1x2xf32>
}

// CHECK: func @cast_bf16_batch_matmul_v2_to_fp32
// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<10x2xf32>}> : () -> tensor<10x2xf32>

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_a", device = "", transpose_a = false, transpose_b = false} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    return %0 : tensor<*xf32>
  }

// CHECK: %[[cst:.*]] = "arith.constant"() <{value = dense<0.000000e+00> : tensor<2x1024xf32>}> : () -> tensor<2x1024xf32>

    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2x1024xf32>} : () -> tensor<2x1024xf32>
    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn} : (tensor<1x2x2x3xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }