def test_host_alloc_mapped(self):
ary = cuda.mapped_array(10, dtype=np.uint32)
ary.fill(123)
self.assertTrue(all(ary == 123))
driver.device_memset(ary, 0, driver.device_memory_size(ary))
self.assertTrue(all(ary == 0))
self.assertTrue(sum(ary != 0) == 0)
def test_host_alloc_pinned(self):
ary = cuda.pinned_array(10, dtype=np.uint32)
ary.fill(123)
self.assertTrue(all(ary == 123))
devary = cuda.to_device(ary)
driver.device_memset(devary, 0, driver.device_memory_size(devary))
self.assertTrue(all(ary == 123))
devary.copy_to_host(ary)
self.assertTrue(all(ary == 0))
def test_host_alloc_pinned(self):
ary = cuda.pinned_array(10, dtype=np.uint32)
ary.fill(123)
self.assertTrue(all(ary == 123))
devary = cuda.to_device(ary)
driver.device_memset(devary, 0, driver.device_memory_size(devary))
self.assertTrue(all(ary == 123))
devary.copy_to_host(ary)
self.assertTrue(all(ary == 0))
def test_memset(self):
dtype = np.dtype('uint32')
n = 10
sz = dtype.itemsize * 10
devary = self.context.memalloc(sz)
driver.device_memset(devary, 0xab, sz)
hstary = np.empty(n, dtype=dtype)
driver.device_to_host(hstary, devary, sz)
hstary2 = np.array([0xabababab] * n, dtype=np.dtype('uint32'))
self.assertTrue(np.all(hstary == hstary2))
def test_memset(self):
dtype = np.dtype('uint32')
n = 10
sz = dtype.itemsize * 10
devary = self.context.memalloc(sz)
driver.device_memset(devary, 0xab, sz)
hstary = np.empty(n, dtype=dtype)
driver.device_to_host(hstary, devary, sz)
hstary2 = np.array([0xabababab] * n, dtype=np.dtype('uint32'))
self.assertTrue(np.all(hstary == hstary2))
def test_host_alloc_driver(self):
n = 32
mem = cuda.current_context().memhostalloc(n, mapped=True)
dtype = np.dtype(np.uint8)
ary = np.ndarray(shape=n // dtype.itemsize, dtype=dtype, buffer=mem)
magic = 0xab
driver.device_memset(mem, magic, n)
self.assertTrue(np.all(ary == magic))
ary.fill(n)
recv = np.empty_like(ary)
driver.device_to_host(recv, mem, ary.size)
self.assertTrue(np.all(ary == recv))
self.assertTrue(np.all(recv == n))
def test_host_alloc_driver(self):
n = 32
mem = cuda.current_context().memhostalloc(n, mapped=True)
dtype = np.dtype(np.uint8)
ary = np.ndarray(shape=n // dtype.itemsize, dtype=dtype, buffer=mem)
magic = 0xAB
driver.device_memset(mem, magic, n)
self.assertTrue(np.all(ary == magic))
ary.fill(n)
recv = np.empty_like(ary)
driver.device_to_host(recv, mem, ary.size)
self.assertTrue(np.all(ary == recv))
self.assertTrue(np.all(recv == n))
def _test_managed_alloc_driver(self, memory_factor, attach_global=True):
# Verify that we can allocate and operate on managed
# memory through the CUDA driver interface.
total_mem_size = self.get_total_gpu_memory()
n_bytes = int(memory_factor * total_mem_size)
ctx = cuda.current_context()
mem = ctx.memallocmanaged(n_bytes, attach_global=attach_global)
dtype = np.dtype(np.uint8)
n_elems = n_bytes // dtype.itemsize
ary = np.ndarray(shape=n_elems, dtype=dtype, buffer=mem)
magic = 0xab
device_memset(mem, magic, n_bytes)
ctx.synchronize()
# Note that this assertion operates on the CPU, so this
# test effectively drives both the CPU and the GPU on
# managed memory.
self.assertTrue(np.all(ary == magic))
def test_host_alloc_mapped(self):
ary = cuda.mapped_array(10, dtype=np.uint32)
ary.fill(123)
self.assertTrue(all(ary == 123))
driver.device_memset(ary, 0, driver.device_memory_size(ary))
self.assertTrue(all(ary == 0))
