Mamy Ratsimbazafy

Save/load from disk (serializing / marshalling)

23

Format to be defined: ### Non-binary (will certainly have size issues) - JSON via Nim [marshal](https://nim-lang.org/docs/marshal.html) module - YAML via [NimYAML](https://nimyaml.org/serialization.html) - ... ### Binary - MessagePack via [msgpack4nim](https://github.com/jangko/msgpack4nim) or...

[WIP] Laser forEach iterators

1

This replaces the use of `mapX_inline` and `applyX_inline` by the forEach / forEachContiguous / forEachParallel / forEachSerial laser iterators. This is particularly valuable for recurrent neural network like GRU because...

Fix https://github.com/mratsim/Arraymancer/issues/185 Pending https://github.com/nim-lang/Nim/pull/8715, upstream fix of https://github.com/nim-lang/Nim/issues/8706 This should avoid the regression https://github.com/mratsim/Arraymancer/issues/241

#510 removed Travis and Appveyor and introduces Github Actions but the windows CI is blocked by Openblas/Lapack not being found in msys2 for some reason.

https://github.com/mratsim/Arraymancer/pull/510/checks?check_run_id=2980291716#step:21:823  The first failure may be due to rounding issue with float32 and switching to float64 will solve the issue. 1e-11 is way over float32 precision. https://github.com/mratsim/Arraymancer/blob/8514e78993ea5852eb8c4d45f00cfbf13b3e193f/tests/linear_algebra/test_linear_algebra.nim#L130-L147 The second...

From IRC: https://irclogs.nim-lang.org/16-01-2021.html#15:08:07   | FromDiscord |  @mratsim QR decomposition test fails in arraymancer test suite. -- | -- | -- 15:08:34 | FromDiscord |  some of the tensor...

OpenMP barrier slowness in neural network examples with high core count.

1

The following bench, reduced to only call `linear` which is just a thin wrapper around BLAS, takes 1.6s without `-d:openmp` and 15s with `-d:openmp` ```Nim import ../src/arraymancer # Learning XOR...

Since Nim 1.2 Nim now adds -std:gnu++14 to the C++ target which we cannot avoid even after https://github.com/nim-lang/Nim/issues/13933 because the "override" only reorders the arguments passed to the compiler. Even...

Vulkan support

1

The most important blocker for Vulkan support was finding some AXPY example to understand how to use/allocate arbitrary sized buffers without using a texture hacks (like in OpenGL before Cuda/OpenCL...