libcu++ not present in CUDA 10.2 ?

Accelerated Computing CUDA CUDA Programming and Performance
1

The release notes for CUDA 10.2 mention “libcu++ (CUDA Standard C++ Library)”:

However, this does not seem to be included in at least 2 particular distributions I downloaded (Windows 10 x64, CentOS 7 runfile). Further, the documentation release notes in both cases in the respective doc folders (html and PDF) point to the Early Access version of ‘CUDA Tegra Release Notes for Development’ – presumably this was a mistake when packaging the documentation.

Was this feature not intended to be released w/ CUDA 10.2, or were the documentation and files mistakenly omitted? I searched for libcu++ on both installed distributions and came up empty.

2

The files are present in installdir/include/cuda/std . Some of them work already, some do not compile. I guess the folder structure is not correct, yet.

Here is an example using the new atomic type.

#include <cuda/std/std/atomic>

#include <cstdio>


__global__
void foo(cuda::std::atomic<int>* atomic){
    printf("thread %d from block %d, value %d\n",
           threadIdx.x, blockIdx.x, (*atomic)++);
}

int main(){

    cuda::std::atomic<int>* atomic;
    cudaMallocManaged(&atomic, sizeof(cuda::std::atomic<int>));

    *atomic = 0;

    foo<<<2,32>>>(atomic);

    cudaDeviceSynchronize();

    cudaFree(atomic);

    return 0;
}
3

I took a further look and see headers and even a makefile at least on the Windows distribution, but it’s hard to appreciate the existing functionality without some sort of documentation. Judging by the experimental folder, it would seem like some functionality is not completely tested/polished.