test_vectorize.py

import numpy as np

from collections import namedtuple
from itertools import product
from numba import vectorize
from numba import cuda, int32, float32, float64
from numba.cuda.cudadrv.driver import CudaAPIError, driver
from numba.cuda.testing import skip_on_cudasim
from numba.cuda.testing import CUDATestCase
import unittest


# Signatures to test with - these are all homogeneous in dtype, so the output
# dtype should match the input dtype - the output should not have been cast
# upwards, as reported in #8400: https://github.com/numba/numba/issues/8400
signatures = [int32(int32, int32),
              float32(float32, float32),
              float64(float64, float64)]

# The order here is chosen such that each subsequent dtype might have been
# casted to a previously-used dtype. This is unlikely to be an issue for CUDA,
# but there might be future circumstances in which it becomes relevant, perhaps
# if it supported Dynamic UFuncs, and we want to ensure that an implementation
# for a the given dtype is used rather than casting the input upwards.
dtypes = (np.float64, np.float32, np.int32)

# NumPy ndarray orders
orders = ('C', 'F')

# Input sizes corresponding to operations:
# - Less than one warp,
# - Less than one block,
# - Greater than one block (i.e. many blocks)
input_sizes = (8, 100, 2 ** 10 + 1)


@skip_on_cudasim('ufunc API unsupported in the simulator')
class TestCUDAVectorize(CUDATestCase):
    # Presumably chosen as an odd number unlikely to coincide with the total
    # thread count, and large enough to ensure a significant number of blocks
    # are used.
    N = 1000001

    def test_scalar(self):

        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        a = 1.2
        b = 2.3
        c = vector_add(a, b)
        self.assertEqual(c, a + b)

    def test_1d(self):

        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        for ty in dtypes:
            data = np.array(np.random.random(self.N), dtype=ty)
            expected = np.add(data, data)
            actual = vector_add(data, data)
            np.testing.assert_allclose(expected, actual)
            self.assertEqual(actual.dtype, ty)

    def test_1d_async(self):

        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        stream = cuda.stream()

        for ty in dtypes:
            data = np.array(np.random.random(self.N), dtype=ty)
            device_data = cuda.to_device(data, stream)

            dresult = vector_add(device_data, device_data, stream=stream)
            actual = dresult.copy_to_host()

            expected = np.add(data, data)

            np.testing.assert_allclose(expected, actual)
            self.assertEqual(actual.dtype, ty)

    def test_nd(self):

        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        for nd, dtype, order in product(range(1, 8), dtypes, orders):
            shape = (4,) * nd
            data = np.random.random(shape).astype(dtype)
            data2 = np.array(data.T, order=order)

            expected = data + data2
            actual = vector_add(data, data2)
            np.testing.assert_allclose(expected, actual)
            self.assertEqual(actual.dtype, dtype)

    def test_output_arg(self):
        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        A = np.arange(10, dtype=np.float32)
        B = np.arange(10, dtype=np.float32)

        expected = A + B
        actual = np.empty_like(A)
        vector_add(A, B, out=actual)

        np.testing.assert_allclose(expected, actual)
        self.assertEqual(expected.dtype, actual.dtype)

    def test_reduce(self):
        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        dtype = np.int32

        for n in input_sizes:
            x = np.arange(n, dtype=dtype)
            expected = np.add.reduce(x)
            actual = vector_add.reduce(x)
            np.testing.assert_allclose(expected, actual)
            # np.add.reduce is special-cased to return an int64 for any int
            # arguments, so we can't compare against its returned dtype when
            # we're checking the general reduce machinery (which just happens
            # to be using addition). Instead, compare against the input dtype.
            self.assertEqual(dtype, actual.dtype)

    def test_reduce_async(self):

        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        stream = cuda.stream()
        dtype = np.int32

        for n in input_sizes:
            x = np.arange(n, dtype=dtype)
            expected = np.add.reduce(x)
            dx = cuda.to_device(x, stream)
            actual = vector_add.reduce(dx, stream=stream)
            np.testing.assert_allclose(expected, actual)
            # Compare against the input dtype as in test_reduce().
            self.assertEqual(dtype, actual.dtype)

    def test_manual_transfer(self):
        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        n = 10
        x = np.arange(n, dtype=np.int32)
        dx = cuda.to_device(x)
        expected = x + x
        actual = vector_add(x, dx).copy_to_host()
        np.testing.assert_equal(expected, actual)
        self.assertEqual(expected.dtype, actual.dtype)

    def test_ufunc_output_2d(self):
        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        n = 10
        x = np.arange(n, dtype=np.int32).reshape(2, 5)
        dx = cuda.to_device(x)
        vector_add(dx, dx, out=dx)

        expected = x + x
        actual = dx.copy_to_host()
        np.testing.assert_equal(expected, actual)
        self.assertEqual(expected.dtype, actual.dtype)

    def check_tuple_arg(self, a, b):
        @vectorize(signatures, target='cuda')
        def vector_add(a, b):
            return a + b

        r = vector_add(a, b)
        np.testing.assert_equal(np.asarray(a) + np.asarray(b), r)

    def test_tuple_arg(self):
        a = (1.0, 2.0, 3.0)
        b = (4.0, 5.0, 6.0)
        self.check_tuple_arg(a, b)

    def test_namedtuple_arg(self):
        Point = namedtuple('Point', ('x', 'y', 'z'))
        a = Point(x=1.0, y=2.0, z=3.0)
        b = Point(x=4.0, y=5.0, z=6.0)
        self.check_tuple_arg(a, b)

    def test_tuple_of_array_arg(self):
        arr = np.arange(10, dtype=np.int32)
        a = (arr, arr + 1)
        b = (arr + 2, arr + 2)
        self.check_tuple_arg(a, b)

    def test_tuple_of_namedtuple_arg(self):
        Point = namedtuple('Point', ('x', 'y', 'z'))
        a = (Point(x=1.0, y=2.0, z=3.0), Point(x=1.5, y=2.5, z=3.5))
        b = (Point(x=4.0, y=5.0, z=6.0), Point(x=4.5, y=5.5, z=6.5))
        self.check_tuple_arg(a, b)

    def test_namedtuple_of_array_arg(self):
        xs1 = np.arange(10, dtype=np.int32)
        ys1 = xs1 + 2
        xs2 = np.arange(10, dtype=np.int32) * 2
        ys2 = xs2 + 1
        Points = namedtuple('Points', ('xs', 'ys'))
        a = Points(xs=xs1, ys=ys1)
        b = Points(xs=xs2, ys=ys2)
        self.check_tuple_arg(a, b)

    def test_name_attribute(self):
        @vectorize('f8(f8)', target='cuda')
        def bar(x):
            return x ** 2

        self.assertEqual(bar.__name__, 'bar')

    def test_no_transfer_for_device_data(self):
        # Initialize test data on the device prior to banning host <-> device
        # transfer

        noise = np.random.randn(1, 3, 64, 64).astype(np.float32)
        noise = cuda.to_device(noise)

        # A mock of a CUDA function that always raises a CudaAPIError

        def raising_transfer(*args, **kwargs):
            raise CudaAPIError(999, 'Transfer not allowed')

        # Use the mock for transfers between the host and device

        old_HtoD = getattr(driver, 'cuMemcpyHtoD', None)
        old_DtoH = getattr(driver, 'cuMemcpyDtoH', None)

        setattr(driver, 'cuMemcpyHtoD', raising_transfer)
        setattr(driver, 'cuMemcpyDtoH', raising_transfer)

        # Ensure that the mock functions are working as expected

        with self.assertRaisesRegex(CudaAPIError, "Transfer not allowed"):
            noise.copy_to_host()

        with self.assertRaisesRegex(CudaAPIError, "Transfer not allowed"):
            cuda.to_device([1])

        try:
            # Check that defining and calling a ufunc with data on the device
            # induces no transfers

            @vectorize(['float32(float32)'], target='cuda')
            def func(noise):
                return noise + 1.0

            func(noise)
        finally:
            # Replace our mocks with the original implementations. If there was
            # no original implementation, simply remove ours.

            if old_HtoD is not None:
                setattr(driver, 'cuMemcpyHtoD', old_HtoD)
            else:
                del driver.cuMemcpyHtoD
            if old_DtoH is not None:
                setattr(driver, 'cuMemcpyDtoH', old_DtoH)
            else:
                del driver.cuMemcpyDtoH


if __name__ == '__main__':
    unittest.main()