The website says the latest release is 1.1, but the Makefile says the current version is 1.0.3.

Also, there are no tags on the repo. Any chance you could tag releases? Thanks!

With modern compilers, the lines -((int32_t)(t1 & 1)) & random->tmat in tinymt32.h/c throw warnings of the form

-((int32_t)(t1 & 1)) & random->tmat
^----------------------------------
warning: conversion to ‘uint32_t {aka unsigned int}’ from ‘int’ may change the sign of the result [-Wsign-conversion]

when compiled with -Wsign-conversion.

The same warning is probably also present in tinmy64.h/c.

Maybe an intermediate const variable and an explicit cast can solve this.

Used in: alpaka-group/alpaka#579
Diff: alpaka-group/alpaka@783b8f7

Found in alpaka-group/alpaka#994 by @krzikalla

These lines

throw a -Wconversion warning:

tinymt32.h: warning: conversion to ‘float’ from ‘uint32_t {aka unsigned int}’ may alter its value [-Wconversion]
     return (tinymt32_temper(random) >> 8) * TINYMT32_MUL;
        ~~~~~~~~~~~~~~~~~~~~~~~~~^~~~~

Possible fix via explicit cast:

 return float(tinymt32_temper(random) >> 8) * TINYMT32_MUL;

Do you have any plans to create TinyMT Java repository?

Because CHANGE-LOG.txt mention "ver. 1.1.2" while the README files say "ver. 1.1.1" and there is no tag "v1.1.2", I was wondering if the version 1.1.2 is an ongoing project. Or should the current codebase be considered as version 1.1.2?

To avoid possible confusion, maybe CHANGE-LOG.txt needs release date information for each version.

(By the way, there is a typo in README.md, s/TitinyMT/TinyMT/.)

Hi!

I would like to humbly bring forth a feature request, namely having a C++ STL compatible version satisfying UniformRandomBitGenerator, something along the lines of std::mersenne_twister_engine.

The problem with most STL implementations is that they do not satisfy the StandardLayoutType concept, which precludes its usage in ComputeCpp, the only true, OpenCL accelerated SYCL implementation thus far. SYCL is the CUDA of OpenCL, to put it shortly, a single-source, language extension-free abstraction over OpenCL. The next best thing after ice cream. :)

ComputeCpp because is not a single-pass compiler and has to interact with other host compiler, mandates that classes used in device code be standard layout. Compilers have a limited amount of freedom in how to layout types in memory and if the host and device compilers do not agree on their layout, one will read garbled things originating from the other, hence the least common denominator is standard layout types.

I have created a stub implementation based on cppreference and started tweaking it based on MSVCs implementation and tinymt64.h. Ideally, the implementations of tinymt64 and tinymt32 should not differ. Some compile-time trickery (TMP) might be needed, but mostly flat constexpr functions. MSVCs implementation has no specializations as far as I saw, so it's fairly C code with minimal inheritance (the source of non-stanard layout) and a few static constexpr helper variables.

TinyMT.zip

I believe that someone with deep knowledge and mediocre C++ knowledge can put together this class in a fairly short amount of time. It might just be that std::mersenne_twister_engine is actually TineMT compatible, I just couldn't identify all the template params inside tinymt64.h.

Again, I would like to kindly ask for a standard layout tiny_mersenne_twister_engine implementation that satisfies UniformRandomBitGenerator to be used either with only the STL or even inside SYCL device code for parallel Monte-Carlo simulations.