Utilities#

static constexpr std::size_t CUDA_ALLOCATION_ALIGNMENT = {256}#

Default alignment used for CUDA memory allocation.

bool is_pow2(std::size_t value) noexcept#

Returns whether or not value is a power of 2.

Parameters:

value – [in] value to check.

Returns:

True if the input is a power of two with non-negative integer exponent, false otherwise.

bool is_supported_alignment(std::size_t alignment) noexcept#

Returns whether or not alignment is a valid memory alignment.

Parameters:

alignment – [in] to check

Returns:

True if the alignment is valid, false otherwise.

bool is_supported_base_resource_alignment(
std::size_t alignment
) noexcept#

Returns whether or not alignment is supported by base CUDA memory resources.

Parameters:

alignment – [in] to check

Returns:

True if the alignment is a valid memory alignment and is no larger than CUDA_ALLOCATION_ALIGNMENT, false otherwise.

std::size_t align_up(
std::size_t value,
std::size_t alignment
) noexcept#

Align up to nearest multiple of specified power of 2.

Parameters:

  • value – [in] value to align

  • alignment – [in] amount, in bytes, must be a power of 2

Returns:

the aligned value

std::size_t align_down(
std::size_t value,
std::size_t alignment
) noexcept#

Align down to the nearest multiple of specified power of 2.

Parameters:

  • value – [in] value to align

  • alignment – [in] amount, in bytes, must be a power of 2

Returns:

the aligned value

bool is_aligned(std::size_t value, std::size_t alignment) noexcept#

Checks whether a value is aligned to a multiple of a specified power of 2.

Parameters:

  • value – [in] value to check for alignment

  • alignment – [in] amount, in bytes, must be a power of 2

Returns:

true if aligned

bool is_pointer_aligned(
void *ptr,
std::size_t alignment = rmm::CUDA_ALLOCATION_ALIGNMENT
) noexcept#

Checks whether the provided pointer is aligned to a specified alignment.

Parameters:

  • ptr – [in] pointer to check for alignment

  • alignment – [in] required alignment in bytes, must be a power of 2

Returns:

true if the pointer is aligned

void prefetch(
void const *ptr,
std::size_t size,
rmm::cuda_device_id device,
rmm::cuda_stream_view stream
)#

Prefetch memory to the specified device on the specified stream.

This function is a no-op if the pointer is not to CUDA managed memory or if concurrent managed access is not supported.

Throws:

rmm::cuda_error – if the prefetch fails.

Parameters:

  • ptr – The pointer to the memory to prefetch

  • size – The number of bytes to prefetch

  • device – The device to prefetch to

  • stream – The stream to use for the prefetch

template<typename T>
void prefetch(
cuda::std::span<T const> data,
rmm::cuda_device_id device,
rmm::cuda_stream_view stream
)#

Prefetch a span of memory to the specified device on the specified stream.

This function is a no-op if the buffer is not backed by CUDA managed memory.

Throws:

rmm::cuda_error – if the prefetch fails.

Parameters:

  • data – The span to prefetch

  • device – The device to prefetch to

  • stream – The stream to use for the prefetch