gpuSparse

Matlab mex wrappers to NVIDIA cuSPARSE (https://developer.nvidia.com/cusparse).

Uses int32 and single precision to save memory (Matlab sparse uses int64 and double).

Installation

1. Save in a folder called @gpuSparse on the Matlab path

2. A = gpuSparse('recompile') to trigger compilation of mex

3. <b>Recommended:</b> CUDA-11 for <i>much</i> faster transpose-multiply

Timings

<pre> <b>Due to memory layout (row/col-major) multiply and transpose-multiply differ in performance.</b> size(A) = 221,401 x 213,331 nnz(A) = 23,609,791 (0.05%) AT = precomputed transpose of A <b>CPU sparse</b> A*x (sparse) : Elapsed time is 1.370207 seconds. AT*y (sparse) : Elapsed time is 1.347447 seconds. A'*y (sparse) : Elapsed time is 0.267259 seconds. <b>GPU sparse</b> A*x (gpuArray) : Elapsed time is 0.137195 seconds. AT*y (gpuArray) : Elapsed time is 0.106331 seconds. A'*y (gpuArray) : Elapsed time is 0.232057 seconds. <i>(CUDA 11)</i> <s>A'*y (gpuArray) : Elapsed time is 16.733638 seconds.</s> <b>GPU gpuSparse</b> A*x (gpuSparse): Elapsed time is 0.068451 seconds. AT*y (gpuSparse): Elapsed time is 0.063651 seconds. A'*y (gpuSparse): Elapsed time is 0.059236 seconds. <i>(CUDA 11)</i> <s>A'*y (gpuSparse): Elapsed time is 3.094271 seconds.</s> </pre>