Available versions of BLAS

Question

Hi, I am wondering what BLAS is available on JeVois? It would be a very useful information you might want to include in the document in case that we want build computation library or transfer frameworks (e.g. Tensorflow) from other platforms.

Answer 1

Yes, actually several versions of BLAS are already in there, see below. For deep networks, you may want to also look into lighter frameworks, line tiny-dnn or darknet, which may be less resource-intensive than tensorflow and also include optimizations for ARM processors using NEON (e.g., darknet-nnpack). We are implementing darknet YOLO now using darknet-nnpack and it is working quite well on JeVois (quite slow though, which hopefully we can improve using smaller networks).

itti@iLab1:~$ find /media/itti/LINUX/ -name *blas*

/media/itti/LINUX/usr/lib/python3.5/site-packages/numpy/core/tests/test_blasdot.py
/media/itti/LINUX/usr/lib/python3.5/site-packages/numpy/core/tests/test_blasdot.pyc
/media/itti/LINUX/usr/lib/python3.5/site-packages/numpy/core/_dotblas.cpython-35m-arm-linux-gnueabihf.so
/media/itti/LINUX/usr/lib/libblas.so.3.6.1
/media/itti/LINUX/usr/lib/libblas.so.3
/media/itti/LINUX/usr/lib/libopenblas.so.0
/media/itti/LINUX/usr/lib/libcblas.so
/media/itti/LINUX/usr/lib/libopenblas_armv7p-r0.2.19.dev.so
/media/itti/LINUX/usr/lib/libopenblas.so
/media/itti/LINUX/usr/lib/libgslcblas.so.0.0.0
/media/itti/LINUX/usr/lib/libblas.so
/media/itti/LINUX/usr/lib/libgslcblas.so.0
/media/itti/LINUX/usr/lib/libgslcblas.so
/media/itti/LINUX/usr/include/nnpack/blas.h

Comments

Thank you for your reply!

So with NNPack doing all the heavy jobs, we are safe to rely on that rather than implementing calculation library with scratch (e.g. with Neon or Ne10). Is that correct?