speakspeak / cuzk Goto Github PK

In my case, the code won't compile with error like:

/usr/lib/gcc/x86_64-pc-linux-gnu/7.5.0/include/c++/type_traits:83:83: error: redefinition of ‘constexpr const _Tp   std::integral_constant<_Tp, __v>::value’
   template<typename _Tp, _Tp __v>

Fix it by appending flag: --std=c++14.

如过有多张GPU那么您的代码时自动适配多GPU的吗。还是只能运行在一张GPU上。

文件位置：
depends/libstl-cuda/algorithm.cu
报错函数：
cub::DeviceRadixSort::SortPairs()
报错日志：
./../depends/libmatrix-cuda/transpose/../depends/libstl-cuda/algorithm.cu(131): error: no instance of overloaded function "cub::DeviceRadixSort::SortPairs" matches the argument list argument types are: (void *, size_t, size_t *, libff::bls12_381_pp::G1_type **, size_t *, libff::bls12_381_pp::G1_type **, size_t, size_t, size_t)
备注：sort_pair_host中的地址指针key_out_addr和output_addr的声明方式好像也是错的，get_host模板初始化有问题

Makefile, as provided, builds only one exec file: msmtesta.
The reason is that it is missing the compilation scope in its header, such as:

all: msma msmb  # limit the compilation scope if you like

There may be some undocumented dependencies, on a Linux VM (with Debian as OS) the following happens:

(.detectron2) root@2d29a0a64e2d:/home/zkp/cuZK/test# make
nvcc -arch=sm_35 -rdc=true --expt-extended-lambda ./MSMtestbn.cu libgmp.a ../depends/libff-cuda/curves/alt_bn128/alt_bn128_pp_host.cu  ../depends/libff-cuda/curves/alt_bn128/alt_bn128_init_host.cu  ../depends/libff-cuda/curves/alt_bn128/alt_bn128_g1_host.cu ../depends/libff-cuda/curves/alt_bn128/alt_bn128_g2_host.cu  ../depends/libstl-cuda/memory.cu ../depends/libff-cuda/common/utils.cu ../depends/libff-cuda/mini-mp-cuda/mini-mp-cuda.cu -o msmtesta
nvcc warning : The 'compute_35', 'compute_37', 'sm_35', and 'sm_37' architectures are deprecated, and may be removed in a future release (Use -Wno-deprecated-gpu-targets to suppress warning).
In file included from ./MSMtestbn.cu:45:
./../depends/libff-cuda/fields/bigint_host.cuh:4:10: fatal error: gmp.h: No such file or directory
    4 | #include <gmp.h>

I think a Docker image showing the minimal setup could solve the problem and improve reproducibility.
In this specific case it is the package libgmp3-dev.

While the README says that NVIDIA V100 has been used (a data-center product), the compilation nvcc -arch=sm_35 uses a much lower capability of 3.5 (typical for a workstation).

I wonder whether this setup is really compatible and could the code benefit of upgrading the computing capabilities?