Namespaces
namespace	metal

Classes
struct	DeviceInfo
struct	FieldProblem
struct	GpuStatus

Enumerations
enum class	Backend { Cpu , Cuda , Metal }

Functions
Backend	compiled_backend ()
	Backend compiled into this build. CPU references are always available.
std::vector< DeviceInfo >	list_devices ()
	All visible devices with current memory budget. Empty without an available GPU.
std::optional< DeviceInfo >	find_device (std::string_view uuid_or_name)
	Finds a device by exact UUID or by a case-sensitive substring of its name ("V100", "P100").
bool	has_free_memory (int device_index, std::size_t bytes)
	True when bytes plus a safety margin fit in the device's free memory right now.
int	field_cpu (const FieldProblem &p, std::vector< std::int32_t > &out)
	CPU reference. out is resized to layershw. Returns the number of sweep rounds.
GpuStatus	field_gpu (int device_index, const FieldProblem &p, std::vector< std::int32_t > &out)
	Compiled GPU backend; on failure (no device, no memory) returns !ok and leaves out empty.
std::vector< std::uint64_t >	philox_fill_cpu (std::uint64_t seed, std::uint32_t stage, std::uint64_t item, std::size_t n)
GpuStatus	philox_fill_cuda (int cuda_index, std::uint64_t seed, std::uint32_t stage, std::uint64_t item, std::size_t n, std::vector< std::uint64_t > &out)
	GPU version. On failure (no CUDA, not enough free memory, launch error) returns !ok and leaves out empty.

Variables
constexpr std::size_t	kDeviceMemoryMargin = std::size_t{256} << 20
	Safety margin kept free on every device for other processes and runtime overheads.
constexpr std::int32_t	kFieldInf = std::numeric_limits<std::int32_t>::max() / 4

Enumeration Type Documentation

◆ Backend

enum class tmk::gpu::Backend

strong

Enumerator
Cpu
Cuda
Metal

Definition at line 16 of file device.hpp.

Function Documentation

◆ compiled_backend()

Backend tmk::gpu::compiled_backend ( )

Backend compiled into this build. CPU references are always available.

◆ field_cpu()

int tmk::gpu::field_cpu	(	const FieldProblem &	p,
		std::vector< std::int32_t > &	out )

CPU reference. out is resized to layers*h*w. Returns the number of sweep rounds.

◆ field_gpu()

GpuStatus tmk::gpu::field_gpu	(	int	device_index,
		const FieldProblem &	p,
		std::vector< std::int32_t > &	out )

Compiled GPU backend; on failure (no device, no memory) returns !ok and leaves out empty.

◆ find_device()

std::optional< DeviceInfo > tmk::gpu::find_device ( std::string_view uuid_or_name )

Finds a device by exact UUID or by a case-sensitive substring of its name ("V100", "P100").

◆ has_free_memory()

bool tmk::gpu::has_free_memory	(	int	device_index,
		std::size_t	bytes )

True when bytes plus a safety margin fit in the device's free memory right now.

◆ list_devices()

std::vector< DeviceInfo > tmk::gpu::list_devices ( )

All visible devices with current memory budget. Empty without an available GPU.

◆ philox_fill_cpu()

std::vector< std::uint64_t > tmk::gpu::philox_fill_cpu	(	std::uint64_t	seed,
		std::uint32_t	stage,
		std::uint64_t	item,
		std::size_t	n )

◆ philox_fill_cuda()

GpuStatus tmk::gpu::philox_fill_cuda	(	int	cuda_index,
		std::uint64_t	seed,
		std::uint32_t	stage,
		std::uint64_t	item,
		std::size_t	n,
		std::vector< std::uint64_t > &	out )

GPU version. On failure (no CUDA, not enough free memory, launch error) returns !ok and leaves out empty.

Variable Documentation

◆ kDeviceMemoryMargin

std::size_t tmk::gpu::kDeviceMemoryMargin = std::size_t{256} << 20

inlineconstexpr

Safety margin kept free on every device for other processes and runtime overheads.

Definition at line 40 of file device.hpp.

◆ kFieldInf

std::int32_t tmk::gpu::kFieldInf = std::numeric_limits<std::int32_t>::max() / 4