Compute.DeviceAllocation Module Reference

Functions
std::size_t	Mila::Dnn::Compute::allocationGranularity (DeviceId device) noexcept
	The allocation granularity of device, or zero when it has none or cannot be asked.
constexpr std::size_t	Mila::Dnn::Compute::occupiedDeviceBytes (std::size_t bytes, std::size_t granularity) noexcept
	Bytes one allocation of bytes occupies on a device with the given granularity.
std::size_t	Mila::Dnn::Compute::occupiedTensorBytes (const auto &tensor)
	Bytes one tensor's allocation occupies on its own device.
std::size_t	Mila::Dnn::Compute::readFreeDeviceBytes (DeviceId device) noexcept
	The device's free memory, or zero when the device cannot say.

Variables
constexpr std::size_t	Mila::Dnn::Compute::kCudaAllocationGranularityBytes = std::size_t{ 2 } * 1024 * 1024
	The allocation granularity assumed for a CUDA device, the other half of the rule above.
constexpr std::size_t	Mila::Dnn::Compute::kPackedAllocationLimitBytes = std::size_t{ 1024 } * 1024
	Requests up to this size share blocks the driver packs; larger ones are rounded to the granularity.

Files
file	Mila/Src/Dnn/Compute/DeviceAllocation.ixx
	What one device allocation occupies once the driver has rounded it, and the device's free memory.