dlm is a library for optimised memory transfers between heterogeneous devices and main memory, that includes GPUs, NICs, etc. It chooses the most efficient way of data copying considering DMA, zero-copy technique (e.g. DMA-BUF) and peer-to-peer. It supports various HPC frameworks and provides distinct memory entity with a common abstract interface.
- virtual memory storage, vms (RAM)
- OpenCL
- Infiniband (ibverbs)
- gcc
- OpenCL 1.2+ SDK
- rdma-core
- CMake 2.6
- cmocka
mkdir build && cd ./build
cmake ../ && make
# run tests
ctest#include <dlm/providers/vms.h>
#include <dlm/providers/opencl.h>
// initialise
struct dlm_mem_vms *vms = dlm_vms_allocate_memory(size);
struct dlm_mem_cl *cl = dlm_cl_allocate_memory(ctx, size, CL_MEM_READ_WRITE);
check(vms, cl);
sz = vms->mem.size; // equal or greater
fill_array((char *)vms->va, sz);
// request memory_copy
struct dlm_mem *dvms = dlm_to_mem(vms);
struct dlm_mem *dcl = dlm_to_mem(cl);
ret = dlm_mem_copy(dvms, dcl);
check(ret);
// synchronise (standard poll)
// e.g. use <mem>->fd to integrate it in libevent
wait_with_epoll(cl->mem.fd); // or vms->mem.fd to use vms
check(cl->err);
// use the copyied memory
cl_mem opencl_mem = cl->clmem;
run_opencl(opencl_mem);
// deallocate
// dlm uses ref counting via retain/release.
dlm_mem_release(&vms->mem);
dlm_mem_release(&cl->mem);