def pack_uint8_r2_to_uint32(self, test_input):
num_rows, num_columns = test_input.get_shape().as_list()
num_output_columns = int(math.ceil(num_columns / 4.0))
padding_input = array_ops.pad(
math_ops.cast(test_input, dtype=dtypes.uint8),
constant_op.constant([[
0,
0,
], [0, num_output_columns * 4 - num_columns]]))
output = array_ops.zeros([num_rows, num_output_columns],
dtype=dtypes.uint32)
num_elements_per_pack = 4
shift_bits = 8
iota_r1 = math_ops.range(num_output_columns * num_elements_per_pack)
for p in range(num_elements_per_pack):
selected_index = math_ops.equal(
math_ops.mod(iota_r1, num_elements_per_pack), p)
gather_index = array_ops.boolean_mask(iota_r1, selected_index)
gathered_input = array_ops.gather(padding_input, gather_index, axis=1)
total_shift_bits = shift_bits * (num_elements_per_pack - p - 1)
left_shift_input = bitwise_ops.left_shift(
math_ops.cast(gathered_input, dtype=dtypes.uint32), total_shift_bits)
output = bitwise_ops.bitwise_or(output, left_shift_input)
return output
def testShapeInference(self):
dtype_list = [dtypes.int8, dtypes.int16, dtypes.int32, dtypes.int64,
dtypes.uint8, dtypes.uint16]
with self.test_session(use_gpu=True) as sess:
for dtype in dtype_list:
lhs = constant_op.constant([[0], [3], [5]], dtype=dtype)
rhs = constant_op.constant([[1, 2, 4]], dtype=dtype)
and_tensor = bitwise_ops.bitwise_and(lhs, rhs)
or_tensor = bitwise_ops.bitwise_or(lhs, rhs)
xor_tensor = bitwise_ops.bitwise_xor(lhs, rhs)
ls_tensor = bitwise_ops.left_shift(lhs, rhs)
rs_tensor = bitwise_ops.right_shift(lhs, rhs)
and_result, or_result, xor_result, ls_result, rs_result = sess.run(
[and_tensor, or_tensor, xor_tensor, ls_tensor, rs_tensor])
# Compare shape inference with result
self.assertAllEqual(and_tensor.get_shape().as_list(), and_result.shape)
self.assertAllEqual(and_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(or_tensor.get_shape().as_list(), or_result.shape)
self.assertAllEqual(or_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(xor_tensor.get_shape().as_list(), xor_result.shape)
self.assertAllEqual(xor_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(ls_tensor.get_shape().as_list(), ls_result.shape)
self.assertAllEqual(ls_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(rs_tensor.get_shape().as_list(), rs_result.shape)
self.assertAllEqual(rs_tensor.get_shape().as_list(), [3, 3])
def pack_uint8_r2_to_uint32(self, test_input):
num_rows, num_columns = test_input.get_shape().as_list()
num_output_columns = int(math.ceil(num_columns / 4.0))
padding_input = array_ops.pad(
math_ops.cast(test_input, dtype=dtypes.uint8),
constant_op.constant([[
0,
0,
], [0, num_output_columns * 4 - num_columns]]))
output = array_ops.zeros([num_rows, num_output_columns],
dtype=dtypes.uint32)
num_elements_per_pack = 4
shift_bits = 8
iota_r1 = math_ops.range(num_output_columns * num_elements_per_pack)
for p in range(num_elements_per_pack):
selected_index = math_ops.equal(
math_ops.mod(iota_r1, num_elements_per_pack), p)
gather_index = array_ops.boolean_mask(iota_r1, selected_index)
gathered_input = array_ops.gather(padding_input,
gather_index,
axis=1)
total_shift_bits = shift_bits * (num_elements_per_pack - p - 1)
left_shift_input = bitwise_ops.left_shift(
math_ops.cast(gathered_input, dtype=dtypes.uint32),
total_shift_bits)
output = bitwise_ops.bitwise_or(output, left_shift_input)
return output
def testShapeInference(self):
dtype_list = [dtypes.int8, dtypes.int16, dtypes.int32, dtypes.int64,
dtypes.uint8, dtypes.uint16]
with self.session() as sess:
for dtype in dtype_list:
lhs = constant_op.constant([[0], [3], [5]], dtype=dtype)
rhs = constant_op.constant([[1, 2, 4]], dtype=dtype)
and_tensor = bitwise_ops.bitwise_and(lhs, rhs)
or_tensor = bitwise_ops.bitwise_or(lhs, rhs)
xor_tensor = bitwise_ops.bitwise_xor(lhs, rhs)
ls_tensor = bitwise_ops.left_shift(lhs, rhs)
rs_tensor = bitwise_ops.right_shift(lhs, rhs)
and_result, or_result, xor_result, ls_result, rs_result = sess.run(
[and_tensor, or_tensor, xor_tensor, ls_tensor, rs_tensor])
# Compare shape inference with result
self.assertAllEqual(and_tensor.get_shape().as_list(), and_result.shape)
self.assertAllEqual(and_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(or_tensor.get_shape().as_list(), or_result.shape)
self.assertAllEqual(or_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(xor_tensor.get_shape().as_list(), xor_result.shape)
self.assertAllEqual(xor_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(ls_tensor.get_shape().as_list(), ls_result.shape)
self.assertAllEqual(ls_tensor.get_shape().as_list(), [3, 3])
self.assertAllEqual(rs_tensor.get_shape().as_list(), rs_result.shape)
self.assertAllEqual(rs_tensor.get_shape().as_list(), [3, 3])
def testShiftsWithNegativeLHS(self):
dtype_list = [np.int8, np.int16, np.int32, np.int64]
with self.test_session(use_gpu=True) as sess:
for dtype in dtype_list:
lhs = np.array([-1, -5, -3, -14], dtype=dtype)
rhs = np.array([5, 0, 7, 11], dtype=dtype)
left_shift_result, right_shift_result = sess.run(
[bitwise_ops.left_shift(lhs, rhs),
bitwise_ops.right_shift(lhs, rhs)])
self.assertAllEqual(left_shift_result, np.left_shift(lhs, rhs))
self.assertAllEqual(right_shift_result, np.right_shift(lhs, rhs))
def testShiftsWithNegativeLHS(self):
dtype_list = [np.int8, np.int16, np.int32, np.int64]
with self.session(use_gpu=True) as sess:
for dtype in dtype_list:
lhs = np.array([-1, -5, -3, -14], dtype=dtype)
rhs = np.array([5, 0, 7, 11], dtype=dtype)
left_shift_result, right_shift_result = sess.run(
[bitwise_ops.left_shift(lhs, rhs),
bitwise_ops.right_shift(lhs, rhs)])
self.assertAllEqual(left_shift_result, np.left_shift(lhs, rhs))
self.assertAllEqual(right_shift_result, np.right_shift(lhs, rhs))
def testImplementationDefinedShiftsDoNotCrash(self):
dtype_list = [np.int8, np.int16, np.int32, np.int64]
with self.test_session(use_gpu=True) as sess:
for dtype in dtype_list:
lhs = np.array([-1, -5, -3, -14], dtype=dtype)
rhs = np.array([-2, 64, 101, 32], dtype=dtype)
# We intentionally do not test for specific values here since the exact
# outputs are implementation-defined. However, we should not crash or
# trigger an undefined-behavior error from tools such as
# AddressSanitizer.
sess.run([bitwise_ops.left_shift(lhs, rhs),
bitwise_ops.right_shift(lhs, rhs)])
def testImplementationDefinedShiftsDoNotCrash(self):
dtype_list = [np.int8, np.int16, np.int32, np.int64]
with self.session(use_gpu=True) as sess:
for dtype in dtype_list:
lhs = np.array([-1, -5, -3, -14], dtype=dtype)
rhs = np.array([-2, 64, 101, 32], dtype=dtype)
# We intentionally do not test for specific values here since the exact
# outputs are implementation-defined. However, we should not crash or
# trigger an undefined-behavior error from tools such as
# AddressSanitizer.
sess.run([bitwise_ops.left_shift(lhs, rhs),
bitwise_ops.right_shift(lhs, rhs)])
def testShiftsWithPositiveLHS(self):
dtype_list = [np.int8, np.int16, np.int32, np.int64,
np.uint8, np.uint16, np.uint32, np.uint64]
with self.session() as sess:
for dtype in dtype_list:
lhs = np.array([0, 5, 3, 14], dtype=dtype)
rhs = np.array([5, 0, 7, 3], dtype=dtype)
left_shift_result, right_shift_result = sess.run(
[bitwise_ops.left_shift(lhs, rhs),
bitwise_ops.right_shift(lhs, rhs)])
self.assertAllEqual(left_shift_result, np.left_shift(lhs, rhs))
self.assertAllEqual(right_shift_result, np.right_shift(lhs, rhs))
def binarize_dense_fast(x, transpose=False):
if transpose:
x = tf.transpose(x, [1, 0])
h, w = x.shape
num_bins = int(w / 64)
# Create shift tensor and apply it to binarized input.
shift_bits = tf.broadcast_to(tf.range(64, dtype=tf.int64), x.shape)
binary_x = tf.cast(x > 0, tf.int64)
binary_x = bitwise_ops.left_shift(binary_x, shift_bits)
# Split binarized x into chunks.
binary_chunks = tf.split(binary_x, num_bins, axis=-1)
# Combine chunks using bitwise or (equivalent to reduce sum).
packed_x = tf.reduce_sum(binary_chunks, axis=-1)
packed_x = tf.transpose(packed_x, [1, 0])
return packed_x
def binarize_dense(x, transpose=False):
if transpose:
x = tf.transpose(x, [1, 0])
h, w = x.shape
num_bins = int(w / 64)
binary_x = tf.cast(x > 0, tf.int64)
packed_x = []
for b in range(num_bins):
packed_x.append(tf.zeros(shape=[h], dtype=tf.int64))
for k in range(num_bins):
for b in range(64):
packed_x[k] = bitwise_ops.bitwise_or(
packed_x[k], bitwise_ops.left_shift(binary_x[:, 64 * k + b],
b))
packed_x = tf.stack(packed_x, axis=-1)
return packed_x
def uint32s_to_uint64(x):
return bitwise_ops.bitwise_or(
math_ops.cast(x[0], dtypes.uint64),
bitwise_ops.left_shift(math_ops.cast(x[1], dtypes.uint64),
constant_op.constant(32, dtypes.uint64)))
