mkn.gpu

CUDA/HIP C++20 convenience wrappers

README

Example

The same lambda run over host data and over managed data, see tst/any/readme.cpp.

#include "mkn/gpu.hpp"

using namespace mkn::gpu;

static constexpr std::size_t N = 1'000'000;

// host data: plain host memory, the device cannot read it, so copy in and out
std::vector<float> host(N, 1);
DeviceMem<float> dev{host};  // allocate on the device and copy host -> device
DLauncher{N}([] __device__(float* d) { d[idx()] += 1; }, dev).sync();  // dev is passed as float*
host = dev();  // copy device -> host

// managed data: the same pointer is valid on host and device, pages migrate on access
std::vector<float, ManagedAllocator<float>> managed(N, 1);
DLauncher{managed}([] __device__(float* d) { d[idx()] += 1; }, managed).sync();  // no copies

What happens:

	host data (std::vector + DeviceMem)	managed data (ManagedAllocator)
memory	separate host and device allocations	one allocation, visible to both
copies	explicit: DeviceMem{host} in, dev() out	none written, the driver migrates pages
argument to the lambda	DeviceMem becomes its device float*	the vector becomes its float*
host access after sync()	only after copying back	direct, managed[i]
cost	two full copies over PCIe	page faults on first touch each side, see doc/tracing

A plain std::vector cannot be passed to a launch directly, its memory is on the host only.

DLauncher{N} guesses the launch from the size: blocks of the device maximum, reduced to the kernel's own limit when its register or shared memory use is high, and enough blocks to cover N. Each element runs once, idx() is its index.

On the CPU backend the same code runs on the host: DeviceMem and managed memory are host memory and the lambda runs once per element in a loop.

Backend

Resolved at compile time from the headers available, see inc/mkn/gpu/def.hpp:

headers found	backend
cuda_runtime.h only	CUDA
hip/hip_runtime.h only	ROCm
neither	CPU
both	compile error, define one

Defines override the detection:

define	use
-DMKN_GPU_CUDA=1	CUDA, e.g. when both headers are found
-DMKN_GPU_ROCM=1	ROCm, e.g. when both headers are found
-DMKN_GPU_CPU=1	CPU while GPU headers are available, e.g. to test without a device