Skip to content

moinakg/salsa20_core_cuda

Repository files navigation

salsa20_core_cuda

Port of Salsa20 core crypto function to CUDA.

Running the code

You can simply run "make" within the directory to build the code if the CUDA 5.0 toolkit is installed in the default location. If CUDA is installed in some other place or you are using an older CUDA toolkit then you can run make like this:

make CUDA_PATH=/path/to/cuda/toolkit

The following executable binaries will be generated:

vecCrypt - Standard single-stream GPU based Salsa20 encryption code.

vecCrypt_strm - CUDA Streams based overlapped copy and execute version. It uses 16 streams by default. This is typically faster than the previous version.

vecCrypt_strm_cpuxor - CUDA Streams based overlapped copy and execute version with final CTR mode XOR done on the CPU. It involves PCI data transfer of only the keystream back to the host. Each CUDA stream gets a corresponding host thread that does the XOR once the stream finishes. The XOR is done using optimized SSE2 intrinsics. All buffers are aligned at 16-byte boundary so normal SSE loads are used. This version is by far the fastest of the three.

Tweaking the code

Some of the constants in the file "common.h" are suited to CUDA Capability 1.x devices and can be tweaked for CUDA Capability 2.x or later for better performance.

The constant THREADS_PER_BLOCK is dependent on the amount of registers and shared memory in the device. It can be increased for devices with CUDA Capability 2.x and later.

The constant NUM_STREAMS determines the number of CUDA streams to be used. This can be tweaked to check results with different number of streams. The test buffer (approx 244MB) is split up into these overlapped streams to hide PCI transfer latency.

About

Port of Salsa20 core crypto function to CUDA

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published