Getting started

Setup:

make -C enc_cuda install PREFIX="$(realpath install)"
make -C app install PREFIX="$(realpath install)"

## OR ##
./install.sh

## OR (no logging) ##
CPPFLAGS=-DNDEBUG ./install.sh

Run:

sudo LD_LIBRARY_PATH="./install/lib:/usr/local/gdev/lib64:/usr/local/lib64" ./install/bin/cuda_enc_app

Content

"Encrypted CUDA" library

The library libenccuda overrides functions of the CUDA driver API to make memory transfers between host and GPU encrypted with AES-256-CTR:

CUresult cuMemAlloc(CUdeviceptr *dev_ptr, unsigned int bytesize);
CUresult cuMemFree(CUdeviceptr dev_ptr);
CUresult cuMemcpyDtoH(void *dstHost, CUdeviceptr srcDevice, unsigned int ByteCount);
CUresult cuMemcpyHtoD(CUdeviceptr dstDevice, const void *srcHost, unsigned int ByteCount);

In addition, it exposes a function used to setup the symmetric key, initial counter value, and prepare the GPU for AES encryption:

CUresult cuda_enc_setup(char * key, char * iv);

See enc_cuda/enc_cuda.h for a description of the function.

The AES routines used are:

• On the host, the OpenSSL AES-CTR functions (libcrypto).
• On the GPU, the AES-CTR functions implemented by Romain Dolbeau for a WIP paper. Even though it is a WIP paper, it appears to be the most complete open source implementation of an AES cipher available online.
  • At the time of writing, his webpage seems to be down, but is still available on <archive.org>:
    • project presentation and code
    • the WIP paper
  • More specifically, we only tried using the function that is reported as most high-performing in the paper (aes_ctr_cuda_BTB32SRDIAGKEY0_PRMT_8nocoalnocoal).

Test app

app contains an example that simply copies memory to the device, and back to the host.

Limitations

• The counter value is NOT incremented between encryptions
• This is not AES-GCM