xla_ops_test.py

# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for XLA op wrappers."""

import functools

from absl.testing import parameterized
import numpy as np

from tensorflow.compiler.tests import xla_test
from tensorflow.compiler.tf2xla.ops import gen_xla_ops
from tensorflow.compiler.tf2xla.python import xla
from local_xla.xla import xla_data_pb2
from tensorflow.python.eager import def_function
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import errors
from tensorflow.python.framework import function
from tensorflow.python.framework import ops
from tensorflow.python.framework import tensor_shape
from tensorflow.python.framework import test_util
from tensorflow.python.ops import array_ops
from tensorflow.python.ops import array_ops_stack  # pylint: disable=g-direct-tensorflow-import
from tensorflow.python.ops import random_ops_util
from tensorflow.python.platform import googletest


class XlaOpsNumericalTest(xla_test.XLATestCase, parameterized.TestCase):

  def _assertOpOutputMatchesExpected(
      self, op, args, expected, equality_fn=None
  ):
    with self.session() as session:
      with self.test_scope():
        placeholders = [
            array_ops.placeholder(dtypes.as_dtype(arg.dtype), arg.shape)
            for arg in args
        ]
        feeds = {placeholders[i]: args[i] for i in range(0, len(args))}
        output = op(*placeholders)
      result = session.run(output, feeds)
      if not equality_fn:
        equality_fn = lambda x, y: self.assertAllClose(x, y, rtol=1e-3)
      equality_fn(result, expected)

  def testAdd(self):
    if xla_test.test.is_built_with_rocm():
      self.skipTest('Broken with rocm')
    for dtype in self.numeric_types:
      self._assertOpOutputMatchesExpected(
          xla.add,
          args=(
              np.array([1, 2, 3], dtype=dtype),
              np.array([4, 5, 6], dtype=dtype),
          ),
          expected=np.array([5, 7, 9], dtype=dtype),
      )

      self._assertOpOutputMatchesExpected(
          lambda x, y: xla.add(x, y, broadcast_dims=(0,)),
          args=(
              np.array([[1, 2], [3, 4]], dtype=dtype),
              np.array([7, 11], dtype=dtype),
          ),
          expected=np.array([[8, 9], [14, 15]], dtype=dtype),
      )

      self._assertOpOutputMatchesExpected(
          lambda x, y: xla.add(x, y, broadcast_dims=(1,)),
          args=(
              np.array([[1, 2], [3, 4]], dtype=dtype),
              np.array([7, 11], dtype=dtype),
          ),
          expected=np.array([[8, 13], [10, 15]], dtype=dtype),
      )

  def testBroadcast(self):
    for dtype in self.numeric_types:
      v = np.arange(4, dtype=np.int32).astype(dtype).reshape([2, 2])
      self._assertOpOutputMatchesExpected(
          lambda x: xla.broadcast(x, (7, 42)),
          args=(v,),
          expected=np.tile(v, (7, 42, 1, 1)),
      )

  @test_util.disable_mlir_bridge('Not supported yet')
  def testGather(self):
    operand = np.arange(10, dtype=np.int32).reshape([2, 5])
    start_indices = np.array([2], np.int32)
    slice_sizes = np.array([1, 3], np.int32)

    def gather(operand, start_indices):
      dimension_numbers = xla_data_pb2.GatherDimensionNumbers()
      dimension_numbers.offset_dims.extend([1])
      dimension_numbers.collapsed_slice_dims.extend([0])
      dimension_numbers.start_index_map.extend([0])
      dimension_numbers.index_vector_dim = 1
      return xla.gather(operand, start_indices, dimension_numbers, slice_sizes)

    self._assertOpOutputMatchesExpected(
        gather, args=(operand, start_indices), expected=np.array([[5, 6, 7]])
    )

  def testShiftRightLogical(self):
    self._assertOpOutputMatchesExpected(
        xla.shift_right_logical,
        args=(np.array([-1, 16], dtype=np.int32), np.int32(4)),
        expected=np.array([0x0FFFFFFF, 1], dtype=np.int32),
    )

    self._assertOpOutputMatchesExpected(
        xla.shift_right_logical,
        args=(np.array([0xFFFFFFFF, 16], dtype=np.uint32), np.uint32(4)),
        expected=np.array([0x0FFFFFFF, 1], dtype=np.uint32),
    )

  def testShiftRightArithmetic(self):
    self._assertOpOutputMatchesExpected(
        xla.shift_right_arithmetic,
        args=(np.array([-1, 16], dtype=np.int32), np.int32(4)),
        expected=np.array([-1, 1], dtype=np.int32),
    )

    self._assertOpOutputMatchesExpected(
        xla.shift_right_arithmetic,
        args=(np.array([0xFFFFFFFF, 16], dtype=np.uint32), np.uint32(4)),
        expected=np.array([0xFFFFFFFF, 1], dtype=np.uint32),
    )

  PRECISION_VALUES = (
      None,
      xla_data_pb2.PrecisionConfig.DEFAULT,
      xla_data_pb2.PrecisionConfig.HIGH,
      xla_data_pb2.PrecisionConfig.HIGHEST,
  )

  @parameterized.parameters(*PRECISION_VALUES)
  def testConv(self, precision):
    for dtype in set(self.float_types).intersection(
        set([dtypes.bfloat16.as_numpy_dtype, np.float32])
    ):
      def conv_1d_fn(lhs, rhs):
        dnums = xla_data_pb2.ConvolutionDimensionNumbers()
        num_spatial_dims = 1
        dnums.input_batch_dimension = 0
        dnums.input_feature_dimension = 1
        dnums.output_batch_dimension = 0
        dnums.output_feature_dimension = 1
        dnums.kernel_output_feature_dimension = 0
        dnums.kernel_input_feature_dimension = 1
        dnums.input_spatial_dimensions.extend(range(2, 2 + num_spatial_dims))
        dnums.kernel_spatial_dimensions.extend(range(2, 2 + num_spatial_dims))
        dnums.output_spatial_dimensions.extend(range(2, 2 + num_spatial_dims))
        precision_config = None
        if precision:
          precision_config = xla_data_pb2.PrecisionConfig()
          precision_config.operand_precision.extend([precision, precision])
        return xla.conv(
            lhs,
            rhs,
            window_strides=(1,),
            padding=((2, 1),),
            lhs_dilation=(1,),
            rhs_dilation=(2,),
            dimension_numbers=dnums,
            precision_config=precision_config,
        )

      self._assertOpOutputMatchesExpected(
          conv_1d_fn,
          args=(
              np.array([[[3, 4, 5, 6]]], dtype=dtype),
              np.array([[[-2, -3]]], dtype=dtype),
          ),
          expected=np.array([[[-9, -12, -21, -26, -10]]], dtype=dtype),
      )

  def testConvPreferredElementType(self):
    dtype = np.float16
    preferred_element_type = np.float32

    def conv_1d_fn(lhs, rhs):
      dnums = xla_data_pb2.ConvolutionDimensionNumbers()
      num_spatial_dims = 1
      dnums.input_batch_dimension = 0
      dnums.input_feature_dimension = 1
      dnums.output_batch_dimension = 0
      dnums.output_feature_dimension = 1
      dnums.kernel_output_feature_dimension = 0
      dnums.kernel_input_feature_dimension = 1
      dnums.input_spatial_dimensions.extend(range(2, 2 + num_spatial_dims))
      dnums.kernel_spatial_dimensions.extend(range(2, 2 + num_spatial_dims))
      dnums.output_spatial_dimensions.extend(range(2, 2 + num_spatial_dims))
      precision_config = None
      return xla.conv(
          lhs,
          rhs,
          window_strides=(1,),
          padding=((2, 1),),
          lhs_dilation=(1,),
          rhs_dilation=(2,),
          dimension_numbers=dnums,
          precision_config=precision_config,
          preferred_element_type=preferred_element_type,
      )

    self._assertOpOutputMatchesExpected(
        conv_1d_fn,
        args=(
            np.array([[[3, 4, 5, 6]]], dtype=dtype),
            np.array([[[-2, -3]]], dtype=dtype),
        ),
        expected=np.array(
            [[[-9, -12, -21, -26, -10]]], dtype=preferred_element_type
        ),
    )

  @parameterized.parameters(*PRECISION_VALUES)
  def testDotGeneral(self, precision):
    for dtype in self.float_types:

      def dot_fn(lhs, rhs):
        dnums = xla_data_pb2.DotDimensionNumbers()
        dnums.lhs_contracting_dimensions.append(2)
        dnums.rhs_contracting_dimensions.append(1)
        dnums.lhs_batch_dimensions.append(0)
        dnums.rhs_batch_dimensions.append(0)
        precision_config = None
        if precision:
          precision_config = xla_data_pb2.PrecisionConfig()
          precision_config.operand_precision.extend([precision, precision])
        return xla.dot_general(
            lhs, rhs, dimension_numbers=dnums, precision_config=precision_config
        )

      lhs = np.array(
          [
              [[1, 2], [3, 4]],
              [[5, 6], [7, 8]],
          ],
          dtype=dtype,
      )
      rhs = np.array(
          [
              [[1, 2, 3], [4, 5, 6]],
              [[7, 8, 9], [10, 11, 12]],
          ],
          dtype=dtype,
      )
      self._assertOpOutputMatchesExpected(
          dot_fn,
          args=(lhs, rhs),
          expected=np.array(
              [
                  [[9, 12, 15], [19, 26, 33]],
                  [[95, 106, 117], [129, 144, 159]],
              ],
              dtype=dtype,
          ),
      )

  def testDotGeneralInt8xInt8ToInt32(self):
    def dot_fn(lhs, rhs):
      dnums = xla_data_pb2.DotDimensionNumbers()
      dnums.lhs_contracting_dimensions.append(2)
      dnums.rhs_contracting_dimensions.append(1)
      dnums.lhs_batch_dimensions.append(0)
      dnums.rhs_batch_dimensions.append(0)
      return xla.dot_general(
          lhs, rhs, dimension_numbers=dnums, preferred_element_type=np.int32
      )

    lhs = np.array(
        [
            [[1, 2], [3, 4]],
            [[5, 6], [7, 8]],
        ],
        dtype=np.int8,
    )
    rhs = np.array(
        [
            [[1, 2, 3], [4, 5, 6]],
            [[7, 8, 9], [10, 11, 12]],
        ],
        dtype=np.int8,
    )
    self._assertOpOutputMatchesExpected(
        dot_fn,
        args=(lhs, rhs),
        expected=np.array(
            [
                [[9, 12, 15], [19, 26, 33]],
                [[95, 106, 117], [129, 144, 159]],
            ],
            dtype=np.int32,
        ),
    )

  def testNeg(self):
    for dtype in self.numeric_types - {np.uint8, np.int8}:
      self._assertOpOutputMatchesExpected(
          xla.neg,
          args=(np.array([1, 2, 3], dtype=dtype),),
          expected=np.array([-1, -2, -3], dtype=dtype),
      )

  def testPad(self):
    for dtype in self.numeric_types:

      def pad_fn(x):
        return xla.pad(
            x,
            padding_value=7,
            padding_low=[2, 1],
            padding_high=[1, 2],
            padding_interior=[1, 0],
        )

      self._assertOpOutputMatchesExpected(
          pad_fn,
          args=(np.arange(4, dtype=np.int32).astype(dtype).reshape([2, 2]),),
          expected=np.array(
              [
                  [7, 7, 7, 7, 7],
                  [7, 7, 7, 7, 7],
                  [7, 0, 1, 7, 7],
                  [7, 7, 7, 7, 7],
                  [7, 2, 3, 7, 7],
                  [7, 7, 7, 7, 7],
              ],
              dtype=dtype,
          ),
      )

  def testSetDynamicDimensionSize(self):
    dynamic_size = 7

    # XLA doesn't support this for bfloat16.
    for dtype in set(self.numeric_types).intersection(
        set([np.int32, np.float32, np.float64, np.complex64])
    ):
      def xla_set_dynamic_dimension_size_fn(x):
        # Tell XLA to cut the array to size=dynamic_size.
        return gen_xla_ops.xla_set_dynamic_dimension_size(
            x, dim_index=0, size=dynamic_size
        )

      a = np.arange(10, dtype=np.int32).astype(dtype)
      expected = a[:dynamic_size]

      self._assertOpOutputMatchesExpected(
          xla_set_dynamic_dimension_size_fn, args=(a,), expected=expected
      )

  def testPadNegative(self):
    for dtype in self.numeric_types:

      def pad_fn(x):
        return xla.pad(
            x,
            padding_value=7,
            padding_low=[0, -1],
            padding_high=[1, -2],
            padding_interior=[1, 2],
        )

      self._assertOpOutputMatchesExpected(
          pad_fn,
          args=(np.arange(6, dtype=np.int32).astype(dtype).reshape([2, 3]),),
          expected=np.array(
              [[7, 7, 1, 7], [7, 7, 7, 7], [7, 7, 4, 7], [7, 7, 7, 7]],
              dtype=dtype,
          ),
      )

  @parameterized.parameters(
      random_ops_util.Algorithm.THREEFRY,
      random_ops_util.Algorithm.PHILOX,
      random_ops_util.Algorithm.AUTO_SELECT,
  )
  def testRngBitGeneratorIsDeterministic(self, algorithm):
    dtype = np.uint32
    key = np.array([1, 2], dtype=np.uint64)
    shape = (10, 12)

    def rng_fun_is_deterministic(k):
      res1 = xla.rng_bit_generator(algorithm, k, shape, dtype=dtype)
      res2 = xla.rng_bit_generator(algorithm, k, shape, dtype=dtype)
      return (res1[0] - res2[0], res1[1] - res2[1])

    self._assertOpOutputMatchesExpected(
        rng_fun_is_deterministic,
        args=(key,),
        expected=(
            np.zeros(key.shape, dtype=key.dtype),
            np.zeros(shape, dtype=dtype),
        ),
    )

  def testReduce(self):
    for dtype in set(self.numeric_types).intersection(
        set([dtypes.bfloat16.as_numpy_dtype, np.float32])
    ):

      @function.Defun(dtype, dtype)  # pylint: disable=cell-var-from-loop
      def sum_reducer(x, y):
        return x + y

      def sum_reduction(dims):
        def fn(x):
          return xla.reduce(
              x, init_value=0, dimensions_to_reduce=dims, reducer=sum_reducer  # pylint: disable=cell-var-from-loop
          )

        return fn

      self._assertOpOutputMatchesExpected(
          sum_reduction(dims=[]),
          args=(np.arange(12, dtype=np.int32).astype(dtype).reshape([3, 4]),),
          expected=np.arange(12, dtype=np.int32).astype(dtype).reshape([3, 4]),
      )
      self._assertOpOutputMatchesExpected(
          sum_reduction(dims=[0]),
          args=(np.arange(12, dtype=np.int32).astype(dtype).reshape([3, 4]),),
          expected=np.array([12, 15, 18, 21], dtype=dtype),
      )
      self._assertOpOutputMatchesExpected(
          sum_reduction(dims=[1]),
          args=(np.arange(12, dtype=np.int32).astype(dtype).reshape([3, 4]),),
          expected=np.array([6, 22, 38], dtype=dtype),
      )
      self._assertOpOutputMatchesExpected(
          sum_reduction(dims=[0, 1]),
          args=(np.arange(12, dtype=np.int32).astype(dtype).reshape([3, 4]),),
          expected=dtype(66),
      )

      @function.Defun(dtype, dtype)  # pylint: disable=cell-var-from-loop
      def mul_reducer(x, y):
        return x * y

      def mul_reduction(dims):
        def fn(x):
          return xla.reduce(
              x, init_value=1, dimensions_to_reduce=dims, reducer=mul_reducer  # pylint: disable=cell-var-from-loop
          )

        return fn

      self._assertOpOutputMatchesExpected(
          mul_reduction(dims=[0]),
          args=(np.arange(12, dtype=np.int32).astype(dtype).reshape([3, 4]),),
          expected=np.array([0, 45, 120, 231], dtype=dtype),
      )

  IS_XLA_VARIADIC_REDUCE_V2 = [True, False]

  @parameterized.parameters(IS_XLA_VARIADIC_REDUCE_V2)
  def testVariadicReduceKahanSum(self, is_v2):
    for dtype in set(self.numeric_types).intersection(
        set([np.float32, np.complex64])
    ):
      @def_function.function
      def kahan_sum_reducer(t0, t1):
        (s0, c0), (s1, c1) = t0, t1
        s0minusc = s0 - (c0 + c1)
        t = s1 + s0minusc
        c = (t - s1) - s0minusc
        s = t
        return s, c

      def kahan_sum_reduction(dims, output_idx):
        def fn(x):
          arg = array_ops.zeros([], dtype)  # pylint: disable=cell-var-from-loop
          reducer = kahan_sum_reducer.get_concrete_function(  # pylint: disable=cell-var-from-loop
              (arg, arg), (arg, arg)
          )

          if is_v2:
            return xla.variadic_reduce(
                (x, array_ops.zeros_like(x)),
                init_values=(arg, arg),
                dimensions_to_reduce=dims,
                reducer=reducer,
            )[output_idx]
          else:
            return gen_xla_ops.xla_variadic_reduce(
                (x, array_ops.zeros_like(x)),
                init_value=(arg, arg),
                dimensions_to_reduce=dims,
                reducer=reducer,
            )[output_idx]

        return fn

      xs = np.array([1e5, np.pi, -1e5, np.exp(1.0)])
      xs = np.array([xs, xs[::-1] / 3, xs / 7], dtype)
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[], output_idx=0), args=(xs,), expected=xs
      )
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[], output_idx=1),
          args=(xs,),
          expected=np.zeros_like(xs),
      )
      shuffle_indices = np.argsort(np.random.randn(xs.shape[0]))
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[0], output_idx=0),
          args=(xs[shuffle_indices],),
          expected=np.array(
              [
                  np.exp(1) / 3 + 1e5 * 8 / 7,
                  np.pi * 8 / 7 - 1e5 / 3,
                  -1e5 * 8 / 7 + np.pi / 3,
                  np.exp(1) * 8 / 7 + 1e5 / 3,
              ],
              dtype=dtype,
          ),
      )
      error_term_equality = functools.partial(
          self.assertAllClose, rtol=1e-3, atol=0.005
      )
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[0], output_idx=1),
          args=(xs[shuffle_indices],),
          expected=np.zeros_like(xs[0]),
          equality_fn=error_term_equality,
      )
      shuffle_indices = np.argsort(np.random.randn(xs.shape[1]))
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[1], output_idx=0),
          args=(xs[:, shuffle_indices],),
          expected=np.array(
              [
                  np.pi + np.exp(1.0),
                  (np.pi + np.exp(1.0)) / 3,
                  (np.pi + np.exp(1.0)) / 7,
              ],
              dtype=dtype,
          ),
      )
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[1], output_idx=1),
          args=(xs[:, shuffle_indices],),
          expected=np.zeros_like(xs[:, 0]),
          equality_fn=error_term_equality,
      )
      # Now, shuffle both dims.
      xs = xs[np.argsort(np.random.randn(xs.shape[0]))]
      xs = xs[:, np.argsort(np.random.randn(xs.shape[1]))]
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[0, 1], output_idx=0),
          args=(xs,),
          expected=dtype((np.pi + np.exp(1.0)) * 31 / 21),
      )
      self._assertOpOutputMatchesExpected(
          kahan_sum_reduction(dims=[0, 1], output_idx=1),
          args=(xs,),
          expected=dtype(0),
          equality_fn=error_term_equality,
      )

  @parameterized.parameters(IS_XLA_VARIADIC_REDUCE_V2)
  def testVariadicReduceSingleOp(self, is_v2):
    @def_function.function
    def reducer_add(op_element, acc_val):
      return (op_element + acc_val,)

    for dtype in set(self.numeric_types):
      values = np.array([[1, 3, 5], [4, 6, 8]], dtype=dtype)
      init_val = np.array(0, dtype=dtype)
      arg_spec = array_ops.zeros([], dtype)  # pylint: disable=cell-var-from-loop
      reducer_func = reducer_add.get_concrete_function(arg_spec, arg_spec)

      def reduce(values, *, dimensions_to_reduce):
        if is_v2:
          return xla.variadic_reduce(
              (values,),
              (init_val,),  # pylint: disable=cell-var-from-loop
              dimensions_to_reduce=dimensions_to_reduce,
              reducer=reducer_func,  # pylint: disable=cell-var-from-loop
          )[0]
        else:
          return gen_xla_ops.xla_variadic_reduce(
              (values,),
              (init_val,),  # pylint: disable=cell-var-from-loop
              dimensions_to_reduce=dimensions_to_reduce,
              reducer=reducer_func,  # pylint: disable=cell-var-from-loop
          )[0]

      # Reduce dimension 0
      self._assertOpOutputMatchesExpected(
          functools.partial(reduce, dimensions_to_reduce=(0,)),
          args=(values,),
          expected=np.array([5, 9, 13], dtype=dtype),
      )

      # Reduce dimension 1
      self._assertOpOutputMatchesExpected(
          functools.partial(reduce, dimensions_to_reduce=(1,)),
          args=(values,),
          expected=np.array([9, 18], dtype=dtype),
      )

      # Reduce dimensions 0 and 1
      self._assertOpOutputMatchesExpected(
          functools.partial(reduce, dimensions_to_reduce=(0, 1)),
          args=(values,),
          expected=np.array(27, dtype=dtype),
      )

  def testVariadicReduceV2DifferentTypes(self):
    # Two ops, with different dtypes
    @def_function.function
    def reducer_add(op_element_1, op_element_2, acc_val_1, acc_val_2):
      return (op_element_1 + acc_val_1, op_element_2 + acc_val_2)

    for dtype in set(self.numeric_types):
      values_1 = np.array([[1, 3, 5], [4, 6, 8]], dtype=dtype)
      values_2 = values_1.astype(np.int32)

      init_val_1 = np.array(0, dtype=dtype)  # pylint: disable=cell-var-from-loop
      init_val_2 = init_val_1.astype(np.int32)

      arg_spec_1 = array_ops.zeros([], dtype)  # pylint: disable=cell-var-from-loop
      arg_spec_2 = array_ops.zeros([], np.int32)
      reducer_func = reducer_add.get_concrete_function(
          arg_spec_1, arg_spec_2, arg_spec_1, arg_spec_2
      )  # pylint: disable=cell-var-from-loop

      def reduce(*values, dimensions_to_reduce):
        return xla.variadic_reduce(
            values,
            (
                init_val_1,  # pylint: disable=cell-var-from-loop
                init_val_2,  # pylint: disable=cell-var-from-loop
            ),
            dimensions_to_reduce=dimensions_to_reduce,
            reducer=reducer_func,  # pylint: disable=cell-var-from-loop
        )

      # Reduce dimension 0
      self._assertOpOutputMatchesExpected(
          functools.partial(reduce, dimensions_to_reduce=(0,)),
          args=(values_1, values_2),
          expected=(
              np.array([5, 9, 13], dtype=dtype),
              np.array([5, 9, 13], dtype=np.int32),
          ),
      )

      # Reduce dimension 1
      self._assertOpOutputMatchesExpected(
          functools.partial(reduce, dimensions_to_reduce=(1,)),
          args=(values_1, values_2),
          expected=(
              np.array([9, 18], dtype=dtype),
              np.array([9, 18], dtype=np.int32),
          ),
      )

      # Reduce dimensions 0 and 1
      self._assertOpOutputMatchesExpected(
          functools.partial(reduce, dimensions_to_reduce=(0, 1)),
          args=(values_1, values_2),
          expected=(np.array(27, dtype=dtype), np.array(27, dtype=np.int32)),
      )

      # Reduce not dimensions
      self._assertOpOutputMatchesExpected(
          functools.partial(reduce, dimensions_to_reduce=()),
          args=(values_1, values_2),
          expected=(values_1, values_2),
      )

  @test_util.disable_mlir_bridge('Not supported yet')
  def testScatter(self):
    test_array = np.arange(9).astype(np.int32).reshape((3, 3))
    scatter_indices = np.array([0, 2], dtype=np.int32)
    updates = np.array([[10, 20, 30], [70, 80, 90]], dtype=np.int32)

    dnums = xla_data_pb2.ScatterDimensionNumbers()
    dnums.update_window_dims.append(1)
    dnums.inserted_window_dims.append(0)
    dnums.scatter_dims_to_operand_dims.append(0)
    dnums.index_vector_dim = 1

    add_numbers = function.Defun(np.int32, np.int32)(lambda x, y: x + y)

    def test_fn(
        scatter_input,
        scatter_indices,
        scatter_updates,
    ):
      return gen_xla_ops.xla_scatter(
          scatter_input,
          scatter_indices,
          scatter_updates,
          add_numbers,
          dnums.SerializeToString(),
          indices_are_sorted=False,
      )

    expected = np.array([[10, 21, 32], [3, 4, 5], [76, 87, 98]], dtype=np.int32)
    self._assertOpOutputMatchesExpected(
        test_fn,
        args=(test_array, scatter_indices, updates),
        expected=expected,
    )

  def testSelectAndScatter(self):
    for dtype in set(self.numeric_types).intersection(
        set([dtypes.bfloat16.as_numpy_dtype, np.float32])
    ):

      @function.Defun(dtype, dtype)  # pylint: disable=cell-var-from-loop
      def add_scatter(x, y):
        return x + y

      @function.Defun(dtype, dtype)  # pylint: disable=cell-var-from-loop
      def ge_select(x, y):
        return x >= y

      def test_fn(operand, source):
        return xla.select_and_scatter(
            operand,
            window_dimensions=[2, 3, 1, 1],
            window_strides=[2, 2, 1, 1],
            padding=[[0, 0]] * 4,
            source=source,
            init_value=0,
            select=ge_select,  # pylint: disable=cell-var-from-loop
            scatter=add_scatter,  # pylint: disable=cell-var-from-loop
        )

      self._assertOpOutputMatchesExpected(
          test_fn,
          args=(
              np.array(
                  [
                      [7, 2, 5, 3, 8],
                      [3, 8, 9, 3, 4],
                      [1, 5, 7, 5, 6],
                      [0, 6, 2, 10, 2],
                  ],
                  dtype=dtype,
              ).reshape((4, 5, 1, 1)),
              np.array([[2, 6], [3, 1]], dtype=dtype).reshape((2, 2, 1, 1)),
          ),
          expected=np.array(
              [
                  [0, 0, 0, 0, 0],
                  [0, 0, 8, 0, 0],
                  [0, 0, 3, 0, 0],
                  [0, 0, 0, 1, 0],
              ],
              dtype=dtype,
          ).reshape((4, 5, 1, 1)),
      )

  def testTranspose(self):
    for dtype in self.numeric_types:
      v = np.arange(4, dtype=np.int32).astype(dtype).reshape([2, 2])
      self._assertOpOutputMatchesExpected(
          lambda x: xla.transpose(x, [1, 0]), args=(v,), expected=v.T
      )

  def testDynamicSlice(self):
    for dtype in self.numeric_types:
      self._assertOpOutputMatchesExpected(
          xla.dynamic_slice,
          args=(
              np.arange(1000, dtype=np.int32)
              .astype(dtype)
              .reshape([10, 10, 10]),
              np.array([5, 7, 3]),
              np.array([2, 3, 2]),
          ),
          expected=np.array(
              np.array([
                  [[573, 574], [583, 584], [593, 594]],
                  [[673, 674], [683, 684], [693, 694]],
              ]),
              dtype=dtype,
          ),
      )

  def testDynamicSliceWithIncorrectStartIndicesShape(self):
    with self.session() as session:
      with self.test_scope():
        output = xla.dynamic_slice(
            np.arange(1000, dtype=np.int32).reshape([10, 10, 10]),
            np.array([5, 7]),
            np.array([2, 3, 4]),
        )
      with self.assertRaises(errors.InvalidArgumentError) as invalid_arg_error:
        session.run(output)
      self.assertRegex(
          invalid_arg_error.exception.message,
          (
              r'has mismatched number of slice sizes \(3\) and number of start'
              r' indices \(2\)'
          ),
      )

  def testDynamicSliceWithIncorrectSizeIndicesShape(self):
    with self.session() as session:
      with self.test_scope():
        output = xla.dynamic_slice(
            np.arange(1000, dtype=np.int32).reshape([10, 10, 10]),
            np.array([5, 7, 3]),
            np.array([2, 3]),
        )
      with self.assertRaises(errors.InvalidArgumentError) as invalid_arg_error:
        session.run(output)
      self.assertRegex(
          invalid_arg_error.exception.message,
          (
              r'has mismatched number of slice sizes \(2\) and number of start'
              r' indices \(3\)'
          ),
      )

  def test_optimization_barrier(self):
    args = (
        np.array([[5, 6, 7]], dtype=np.float32),
        np.array([[1, 2, 3]], dtype=int),
    )

    self._assertOpOutputMatchesExpected(
        xla.optimization_barrier, args=args, expected=args
    )

  def test_reduce_precision(self):
    arg = np.array([1 + 2**-2 + 2**-4, 128, 256], dtype=np.float32)
    expected = np.array([1 + 2**-2, 128, float('Inf')], dtype=np.float32)
    exponent_bits = 4
    mantissa_bits = 2
    self._assertOpOutputMatchesExpected(
        lambda x: xla.reduce_precision(x, exponent_bits, mantissa_bits),
        args=(arg,),
        expected=expected,
        equality_fn=self.assertAllEqual,
    )

    arg = np.array([4], dtype=np.float32)
    expected = np.array([4], dtype=np.float32)
    # Test passing numbers that cannot fit in a 32-bit integer.
    exponent_bits = 2**33
    mantissa_bits = 2**33
    self._assertOpOutputMatchesExpected(
        lambda x: xla.reduce_precision(x, exponent_bits, mantissa_bits),
        args=(arg,),
        expected=expected,
        equality_fn=self.assertAllEqual,
    )


class XlaOpsShapeInferenceTest(xla_test.XLATestCase, parameterized.TestCase):

  def testDotShapeInference(self):
    a = array_ops.placeholder(np.float32, shape=(1, 2, 3, 4))
    b = array_ops.placeholder(np.float32, shape=(4, 5, 2, 6))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(1)
    dim_nums.rhs_contracting_dimensions.append(2)
    dim_nums.lhs_batch_dimensions.append(3)
    dim_nums.rhs_batch_dimensions.append(0)

    c = xla.dot_general(a, b, dim_nums)
    self.assertEqual(c.shape, tensor_shape.TensorShape([4, 1, 3, 5, 6]))

  def testDotDifferentNumberOfContractingDimensions(self):
    a = array_ops.placeholder(np.float32, shape=(4, 4, 4, 4))
    b = array_ops.placeholder(np.float32, shape=(4, 4, 4, 4))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(2)
    dim_nums.rhs_contracting_dimensions.append(2)
    dim_nums.rhs_contracting_dimensions.append(3)

    with self.assertRaisesRegex(
        ValueError,
        'Must specify the same number of contracting '
        'dimensions for lhs and rhs. Got: 1 and 2',
    ):
      xla.dot_general(a, b, dim_nums)

  def testDotDifferentContractingDimensionsSizes(self):
    a = array_ops.placeholder(np.float32, shape=(2, 2, 2, 2))
    b = array_ops.placeholder(np.float32, shape=(4, 4, 4, 4))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(2)
    dim_nums.rhs_contracting_dimensions.append(3)

    with self.assertRaisesRegex(
        ValueError, 'Dimensions must be equal, but are 2 and 4'
    ):
      xla.dot_general(a, b, dim_nums)

  def testDotDifferentNumberOfBatchDimensions(self):
    a = array_ops.placeholder(np.float32, shape=(4, 4, 4, 4))
    b = array_ops.placeholder(np.float32, shape=(4, 4, 4, 4))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_batch_dimensions.append(2)
    dim_nums.rhs_batch_dimensions.append(2)
    dim_nums.rhs_batch_dimensions.append(3)

    with self.assertRaisesRegex(
        ValueError,
        'Must specify the same number of batch '
        'dimensions for lhs and rhs. Got: 1 and 2',
    ):
      xla.dot_general(a, b, dim_nums)

  def testDotDifferentBatchDimensionsSizes(self):
    a = array_ops.placeholder(np.float32, shape=(2, 2, 2, 2))
    b = array_ops.placeholder(np.float32, shape=(4, 4, 4, 2))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(2)
    dim_nums.rhs_contracting_dimensions.append(3)
    dim_nums.lhs_batch_dimensions.append(0)
    dim_nums.rhs_batch_dimensions.append(0)

    with self.assertRaisesRegex(
        ValueError, 'Dimensions must be equal, but are 2 and 4'
    ):
      xla.dot_general(a, b, dim_nums)

  def testDotUnknownNonContractingDimension(self):
    a = array_ops.placeholder(np.float32, shape=(None, 16))
    b = array_ops.placeholder(np.float32, shape=(16, 2))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(1)
    dim_nums.rhs_contracting_dimensions.append(0)

    c = xla.dot_general(a, b, dim_nums)
    self.assertEqual(c.shape.as_list(), [None, 2])

  def testDotUnknownContractingDimension(self):
    a = array_ops.placeholder(np.float32, shape=(3, None))
    b = array_ops.placeholder(np.float32, shape=(None, 2))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(1)
    dim_nums.rhs_contracting_dimensions.append(0)

    c = xla.dot_general(a, b, dim_nums)
    self.assertEqual(c.shape.as_list(), [3, 2])

  def testDotUnknownAndKnownContractingDimension(self):
    a = array_ops.placeholder(np.float32, shape=(3, 4))
    b = array_ops.placeholder(np.float32, shape=(None, 2))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(1)
    dim_nums.rhs_contracting_dimensions.append(0)

    c = xla.dot_general(a, b, dim_nums)
    self.assertEqual(c.shape.as_list(), [3, 2])

  def testDotUnknownBatchDimension(self):
    a = array_ops.placeholder(np.float32, shape=(None, 3, 4))
    b = array_ops.placeholder(np.float32, shape=(None, 4))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(2)
    dim_nums.rhs_contracting_dimensions.append(1)
    dim_nums.lhs_batch_dimensions.append(0)
    dim_nums.rhs_batch_dimensions.append(0)

    c = xla.dot_general(a, b, dim_nums)
    self.assertEqual(c.shape.as_list(), [None, 3])

  def testDotUnknownAndKnownBatchDimension(self):
    a = array_ops.placeholder(np.float32, shape=(2, 3, 4))
    b = array_ops.placeholder(np.float32, shape=(None, 4))

    dim_nums = xla_data_pb2.DotDimensionNumbers()
    dim_nums.lhs_contracting_dimensions.append(2)
    dim_nums.rhs_contracting_dimensions.append(1)
    dim_nums.lhs_batch_dimensions.append(0)
    dim_nums.rhs_batch_dimensions.append(0)

    c = xla.dot_general(a, b, dim_nums)
    self.assertEqual(c.shape.as_list(), [2, 3])

  def testDynamicSlice(self):
    start = array_ops.placeholder(np.int32, shape=(2, 3, 4))

    # If slice_sizes are known, the operand shape does not matter.
    # The shape of the output is equal to slice_sizes.
    slice_sizes = np.array([1, 2, 4], dtype=np.int32)
    for a_shape in [(2, 3, 4), (None, 3, 4), None]:
      a = array_ops.placeholder(np.float32, shape=a_shape)
      res = xla.dynamic_slice(a, start, slice_sizes)
      self.assertEqual(res.shape.as_list(), [1, 2, 4])

    # The first two dimension slice sizes are known
    slice_sizes = array_ops_stack.stack(
        [1, 2, array_ops.placeholder(np.int32, [])]
    )
    for a_shape in [(2, 3, 4), (None, 3, 4), None]:
      a = array_ops.placeholder(np.float32, shape=a_shape)
      res = xla.dynamic_slice(a, start, slice_sizes)
      self.assertEqual(res.shape.as_list(), [1, 2, None])

    # If slice_sizes has known rank and dimension, but is not a constant
    # then output has the same rank, but with unknown dimensions.
    slice_sizes = array_ops.placeholder(np.int32, [3])
    for a_shape in [(2, 3, 4), (None, 3, 4), None]:
      a = array_ops.placeholder(np.float32, shape=a_shape)
      res = xla.dynamic_slice(a, start, slice_sizes)
      self.assertEqual(res.shape.as_list(), [None, None, None])

    # slice sizes has known rank, but unknown dimensions.
    # then the output has the same rank as the operand, but with unknown
    # dimensions.
    slice_sizes = array_ops.placeholder(np.int32, [None])
    for a_shape in [(2, 3, 4), (None, 3, 4)]:
      a = array_ops.placeholder(np.float32, shape=a_shape)
      res = xla.dynamic_slice(a, start, slice_sizes)
      self.assertEqual(res.shape.as_list(), [None, None, None])

    a = array_ops.placeholder(np.float32, shape=None)
    slice_sizes = array_ops.placeholder(np.int32, [None])
    res = xla.dynamic_slice(a, start, slice_sizes)
    self.assertIsNone(res.shape.rank)

  def testDynamicUpdateSlice(self):
    a = array_ops.placeholder(np.float32, shape=(2, 3, 4))
    upd = array_ops.placeholder(np.float32, shape=(1, 2, 3))
    start_indices = array_ops.placeholder(np.int32, shape=(3,))

    res = xla.dynamic_update_slice(a, upd, start_indices)
    self.assertEqual(res.shape.as_list(), [2, 3, 4])

    a = array_ops.placeholder(np.float32, shape=(None, 3, None))
    res = xla.dynamic_update_slice(a, upd, start_indices)
    self.assertEqual(res.shape.as_list(), [None, 3, None])

  def testPadShapeInference(self):
    a = array_ops.placeholder(np.float32, shape=(2, 3))

    c = xla.pad(
        a,
        padding_value=7,
        padding_low=[2, 1],
        padding_high=[1, 2],
        padding_interior=[1, 4],
    )

    self.assertEqual(c.shape, tensor_shape.TensorShape([6, 14]))

    c = xla.pad(
        a,
        padding_value=7,
        padding_low=[2, -2],
        padding_high=[1, -2],
        padding_interior=[1, 2],
    )

    self.assertEqual(c.shape, tensor_shape.TensorShape([6, 3]))

    c = xla.pad(
        array_ops.placeholder(np.float32, shape=(None, 2)),
        padding_value=7,
        padding_low=[0, 1],
        padding_high=[0, 2],
        padding_interior=[0, 4],
    )
    self.assertEqual(c.shape.as_list(), [None, 9])

    # 0-sized input dimension and interior padding
    c = xla.pad(
        array_ops.placeholder(np.float32, shape=(2, 0)),
        padding_value=7,
        padding_low=[2, 1],
        padding_high=[1, 1],
        padding_interior=[1, 2],
    )

    self.assertEqual(c.shape, tensor_shape.TensorShape([6, 2]))

    with self.assertRaisesRegex(
        ValueError, 'padding_value input must be scalar, found rank 1 '
    ):
      xla.pad(
          a,
          padding_value=[0, 1],
          padding_low=[0, 0],
          padding_high=[0, 0],
          padding_interior=[0, 0],
      )

    with self.assertRaisesRegex(
        ValueError, 'padding_low must be a 1D tensor of size 2 '
    ):
      xla.pad(
          a,
          padding_value=7,
          padding_low=[0, 0, 0],
          padding_high=[0, 0],
          padding_interior=[0, 0],
      )

    with self.assertRaisesRegex(
        ValueError, 'padding_high must be a 1D tensor of size 2 '
    ):
      xla.pad(
          a,
          padding_value=7,
          padding_low=[0, 0],
          padding_high=[0, 0, 0],
          padding_interior=[0, 0],
      )

    with self.assertRaisesRegex(
        ValueError, 'padding_interior must be a 1D tensor of size 2 '
    ):
      xla.pad(
          a,
          padding_value=7,
          padding_low=[0, 0],
          padding_high=[0, 0],
          padding_interior=[0],
      )

    with self.assertRaisesRegex(
        ValueError,
        'padding_interior must contain only non-negative values, found -2 ',
    ):
      xla.pad(
          a,
          padding_value=7,
          padding_low=[0, 0],
          padding_high=[0, 0],
          padding_interior=[-2, 0],
      )

    with self.assertRaisesRegex(
        ValueError, 'resulting padded dimension has negative size -1 '
    ):
      xla.pad(
          a,
          padding_value=7,
          padding_low=[-3, 0],
          padding_high=[0, 0],
          padding_interior=[0, 0],
      )

  def testVariadicReduceV2SingleArg(self):
    @def_function.function
    def reducer_add(op_element, acc_val):
      return (op_element + acc_val,)

    dtype = np.float32
    arg_spec = array_ops.zeros([], dtype)  # pylint: disable=cell-var-from-loop
    reducer_func = reducer_add.get_concrete_function(arg_spec, arg_spec)

    res = xla.variadic_reduce(
        (array_ops.placeholder(np.float32, shape=(3, 4, 5)),),
        (array_ops.placeholder(np.float32, shape=()),),
        dimensions_to_reduce=(1,),
        reducer=reducer_func,
    )
    self.assertLen(res, 1)
    self.assertEqual(res[0].shape, tensor_shape.TensorShape([3, 5]))

  def testVariadicReduceV2MultipleArgs(self):
    @def_function.function
    def reducer_adds(
        op_element_1,
        op_element_2,
        op_element_3,
        acc_val_1,
        acc_val_2,
        acc_val_3,
    ):
      return (
          op_element_1 + acc_val_1,
          op_element_2 + acc_val_2,
          op_element_3 + acc_val_3,
      )

    dtype = np.float32
    arg1_spec = array_ops.zeros([], dtype)  # pylint: disable=cell-var-from-loop
    arg2_spec = array_ops.zeros([], np.int32)
    arg3_spec = array_ops.zeros([], np.int32)
    reducer_func = reducer_adds.get_concrete_function(
        arg1_spec, arg2_spec, arg3_spec, arg1_spec, arg2_spec, arg3_spec
    )

    def reduce_with_shapes(shape1, shape2, shape3, dimensions_to_reduce=(1,)):
      inputs = (
          array_ops.placeholder(np.float32, shape=shape1),
          array_ops.placeholder(np.int32, shape=shape2),
          array_ops.placeholder(np.int32, shape=shape3),
      )
      init_values = (
          array_ops.placeholder(np.float32, shape=()),
          array_ops.placeholder(np.int32, shape=()),
          array_ops.placeholder(np.int32, shape=()),
      )

      return xla.variadic_reduce(
          inputs,
          init_values,
          dimensions_to_reduce=dimensions_to_reduce,
          reducer=reducer_func,
      )

    def assert_output_shapes(output, expected_shape):
      self.assertLen(output, 3)
      self.assertEqual(output[0].shape.as_list(), list(expected_shape))
      self.assertEqual(output[1].shape.as_list(), list(expected_shape))
      self.assertEqual(output[2].shape.as_list(), list(expected_shape))

    output = reduce_with_shapes((3, 4, 5), (3, 4, 5), (3, 4, 5))
    assert_output_shapes(output, (3, 5))

    output = reduce_with_shapes(
        (3, 4, 5), (3, 4, 5), (3, 4, 5), dimensions_to_reduce=()
    )
    assert_output_shapes(output, (3, 4, 5))

    output = reduce_with_shapes(None, (3, None, 5), (None, 4, 5))
    assert_output_shapes(output, (3, 5))

    output = reduce_with_shapes(None, (3, None, 5), None)
    assert_output_shapes(output, (3, 5))

    output = reduce_with_shapes(None, (None, None, 5), None)
    assert_output_shapes(output, (None, 5))

    output = reduce_with_shapes(None, None, None)
    self.assertLen(output, 3)
    self.assertIsNone(output[0].shape.rank)
    self.assertIsNone(output[1].shape.rank)
    self.assertIsNone(output[2].shape.rank)

    with self.assertRaisesRegex(
        ValueError, 'All inputs must have the same shape'
    ):
      reduce_with_shapes((3, 4, 5), (13, 4, 5), (3, 4, 5))

    with self.assertRaisesRegex(
        ValueError, 'All inputs must have the same shape'
    ):
      reduce_with_shapes((None, 4, 5), (3, None, 5), (13, 4, 5))

    with self.assertRaisesRegex(
        ValueError, 'All inputs must have the same shape'
    ):
      reduce_with_shapes((None, 4, 5), (3, None, 5), (13, 4, 5))

  @parameterized.product(
      algorithm=[random_ops_util.Algorithm.THREEFRY,
                 random_ops_util.Algorithm.PHILOX,
                 random_ops_util.Algorithm.AUTO_SELECT],
      dtype=[np.uint8, np.uint64],
  )
  def testRngBitGenerator(self, algorithm, dtype):
    initial_state = array_ops.placeholder(np.uint64, shape=(2,))
    shape = (2, 3)
    res = xla.rng_bit_generator(algorithm, initial_state, shape, dtype=dtype)

    self.assertEqual(res[0].shape, initial_state.shape)
    self.assertEqual(res[1].shape, shape)

    # The initial_state has unknown dimension size
    initial_state = array_ops.placeholder(np.uint64, shape=(None,))
    shape = (2, 3)
    res = xla.rng_bit_generator(algorithm, initial_state, shape, dtype=dtype)

    self.assertEqual(res[0].shape.as_list(), initial_state.shape.as_list())
    self.assertEqual(res[1].shape, shape)

    # The initial_state has unknown rank
    initial_state = array_ops.placeholder(np.uint64, shape=None)
    shape = (2, 3)
    res = xla.rng_bit_generator(algorithm, initial_state, shape, dtype=dtype)

    self.assertEqual(res[0].shape.as_list(), [None])
    self.assertEqual(res[1].shape, shape)

    # The output shape has unknown dimension
    initial_state = array_ops.placeholder(np.uint64, shape=(None,))
    shape = (None, 3)
    with self.assertRaisesRegex(
        TypeError, 'Failed to convert elements .* to Tensor'
    ):
      res = xla.rng_bit_generator(algorithm, initial_state, shape, dtype=dtype)


if __name__ == '__main__':
  # This test is using Tensorflow sessions which are not compatible with eager
  # mode.
  ops.disable_eager_execution()
  googletest.main()