jeffwong/cudaglmnet

A lasso solver on the GPU

cudaglmnet's Issues

cudaglmnet fails to compile

When I try to compile cudaglmnet, I get the following error sequence. I include outputs of ./configure and system information after the error log.

What gives? What's going wrong, and how to fix? I'm not an expert on configuration issues with CUDA/nvcc, so I haven't much of a clue here, and any pointers would be helpful.

Thanks!

Paul Deane

Starting after cloning sudaglmnet and cding to the installation directory:

sudo R CMD INSTALL ./

• installing to library ‘/usr/local/lib/R/site-library’
• installing source package ‘cudaglmnet’ ...
  ** using staged installation
  configure: checking for critical tools used by configure
  checking "environment variable COMPILER_BINDIR"... "COMPILER_BINDIR not set; using user's default path"
  checking "environment variable CUDA_HOME"... "CUDA_HOME not set; using default /usr/local/cuda"
  checking "environment variable CULA_HOME"... "CULA_HOME not set; using default /usr/local/cula"
  checking for /usr/local/cuda/bin/nvcc... yes
  "nvcc found"
  checking "whether this is the 64 bit linux version of CUDA"... checking for /usr/local/cuda/lib64/libcublas.so... yes
  "yes -- using /usr/local/cuda/lib64 for CUDA libs"
  "using -I/usr/share/R/include for R header files"
  "using -Wl,--export-dynamic -fopenmp -Wl,-Bsymbolic-functions -Wl,-z,relro -L/usr/lib/R/lib -lR -lpcre2-8 -llzma -lbz2 -lz -lrt -ldl -lm -licuuc -licui18n for R shared libraries"
  checking "whether this is the 64 bit version of CULA"... checking for /usr/local/cula/lib64/libcula.so... no
  checking for /usr/local/cula/lib64/libcula.dylib... no
  "no"
  checking "checking whether cula is present at all"... checking for /usr/local/cula/lib/libcula.so... no
  checking for /usr/local/cula/lib/libcula.dylib... no
  "cula is not installed"
  configure: determining host platform
  checking build system type... x86_64-unknown-linux-gnu
  checking host system type... x86_64-unknown-linux-gnu
  configure: creating ./config.status
  config.status: creating src/Makefile
  ** libs
  ** arch -
  /usr/local/cuda/bin/nvcc -O2 -shared -Xlinker -Wl,--export-dynamic -fopenmp -Wl,-Bsymbolic-functions -Wl,-z,relro -L/usr/lib/R/lib -lR -lpcre2-8 -llzma -lbz2 -lz -lrt -ldl -lm -licuuc -licui18n -L/usr/local/cuda/lib64 -lcublas main.o -o cudaglmnet.so
  nvcc fatal : Unknown option '-fopenmp'
  make: *** [Makefile:24: cudaglmnet.so] Error 1
  ERROR: compilation failed for package ‘cudaglmnet’
• removing ‘/usr/local/lib/R/site-library/cudaglmnet’
• restoring previous ‘/usr/local/lib/R/site-library/cudaglmnet’

SYSTEM INFORMATION

Description: Ubuntu 20.04 LTS
Release: 20.04
Codename: focal

GPU: NVIDIA GeForce RTX 3070
R version: 4.2.2 Patched (2022-11-10 r83330)

./configure
configure: checking for critical tools used by configure
checking "environment variable COMPILER_BINDIR"... "COMPILER_BINDIR not set; using user's default path"
checking "environment variable CUDA_HOME"... "using CUDA_HOME=/usr/local/cuda-12.0"
checking "environment variable CULA_HOME"... "CULA_HOME not set; using default /usr/local/cula"
checking for /usr/local/cuda-12.0/bin/nvcc... yes
"nvcc found"
checking "whether this is the 64 bit linux version of CUDA"... checking for /usr/local/cuda-12.0/lib64/libcublas.so... yes
"yes -- using /usr/local/cuda-12.0/lib64 for CUDA libs"
"using -I/usr/share/R/include for R header files"
"using -L/usr/local/R/lib -lR for R shared libraries"
checking "whether this is the 64 bit version of CULA"... checking for /usr/local/cula/lib64/libcula.so... no
checking for /usr/local/cula/lib64/libcula.dylib... no
"no"
checking "checking whether cula is present at all"... checking for /usr/local/cula/lib/libcula.so... no
checking for /usr/local/cula/lib/libcula.dylib... no
"cula is not installed"
configure: determining host platform
checking build system type... x86_64-unknown-linux-gnu
checking host system type... x86_64-unknown-linux-gnu
configure: creating ./config.status
config.status: creating src/Makefile

/usr/local/cuda-12.0/bin/nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2022 NVIDIA Corporation
Built on Mon_Oct_24_19:12:58_PDT_2022
Cuda compilation tools, release 12.0, V12.0.76
Build cuda_12.0.r12.0/compiler.31968024_0

cat /usr/local/cuda-12.0/version.json
{
"cuda" : {
"name" : "CUDA SDK",
"version" : "12.0.20221203"
},
"cuda_cccl" : {
"name" : "CUDA C++ Core Compute Libraries",
"version" : "12.0.90"
},
"cuda_cudart" : {
"name" : "CUDA Runtime (cudart)",
"version" : "12.0.107"
},
"cuda_cuobjdump" : {
"name" : "cuobjdump",
"version" : "12.0.76"
},
"cuda_cupti" : {
"name" : "CUPTI",
"version" : "12.0.90"
},
"cuda_cuxxfilt" : {
"name" : "CUDA cu++ filt",
"version" : "12.0.76"
},
"cuda_demo_suite" : {
"name" : "CUDA Demo Suite",
"version" : "12.0.76"
},
"cuda_gdb" : {
"name" : "CUDA GDB",
"version" : "12.0.90"
},
"cuda_nsight" : {
"name" : "Nsight Eclipse Plugins",
"version" : "12.0.78"
},
"cuda_nvcc" : {
"name" : "CUDA NVCC",
"version" : "12.0.76"
},
"cuda_nvdisasm" : {
"name" : "CUDA nvdisasm",
"version" : "12.0.76"
},
"cuda_nvml_dev" : {
"name" : "CUDA NVML Headers",
"version" : "12.0.76"
},
"cuda_nvprof" : {
"name" : "CUDA nvprof",
"version" : "12.0.90"
},
"cuda_nvprune" : {
"name" : "CUDA nvprune",
"version" : "12.0.76"
},
"cuda_nvrtc" : {
"name" : "CUDA NVRTC",
"version" : "12.0.76"
},
"cuda_nvtx" : {
"name" : "CUDA NVTX",
"version" : "12.0.76"
},
"cuda_nvvp" : {
"name" : "CUDA NVVP",
"version" : "12.0.90"
},
"cuda_opencl" : {
"name" : "CUDA OpenCL",
"version" : "12.0.76"
},
"cuda_sanitizer_api" : {
"name" : "CUDA Compute Sanitizer API",
"version" : "12.0.90"
},
"libcublas" : {
"name" : "CUDA cuBLAS",
"version" : "12.0.1.189"
},
"libcufft" : {
"name" : "CUDA cuFFT",
"version" : "11.0.0.21"
},
"libcufile" : {
"name" : "GPUDirect Storage (cufile)",
"version" : "1.5.0.59"
},
"libcurand" : {
"name" : "CUDA cuRAND",
"version" : "10.3.1.50"
},
"libcusolver" : {
"name" : "CUDA cuSOLVER",
"version" : "11.4.2.57"
},
"libcusparse" : {
"name" : "CUDA cuSPARSE",
"version" : "12.0.0.76"
},
"libnpp" : {
"name" : "CUDA NPP",
"version" : "12.0.0.30"
},
"libnvjitlink" : {
"name" : "JIT Linker Library",
"version" : "12.0.76"
},
"libnvjpeg" : {
"name" : "CUDA nvJPEG",
"version" : "12.0.0.28"
},
"libnvvm_samples" : {
"name" : "NVVM Samples",
"version" : "12.0.94"
},
"nsight_compute" : {
"name" : "Nsight Compute",
"version" : "2022.4.0.15"
},
"nsight_systems" : {
"name" : "Nsight Systems",
"version" : "2022.4.2.18"
},
"nvidia_driver" : {
"name" : "NVIDIA Linux Driver",
"version" : "525.60.13"
},
"nvidia_fs" : {
"name" : "NVIDIA file-system",
"version" : "2.14.12"
}
}

some device vectors do not need to be init to 0

thrust::device_vector(n, init) initializes device memory than does a memcpy to initialize the values. It is only necessary to init X and y, the other vectors can be left alone with random values, saving many memcpys

Reduce host to device transfer overhead

Do memcpy from host to device in parallel

https://developer.nvidia.com/content/how-overlap-data-transfers-cuda-cc

copy functions can be optimized using scatter?

http://stackoverflow.com/questions/1585755/cuda-what-is-scattered-write

copy functions might perform better using coalescing

may be better to use cudaMemCopy and take advantage of memory coalescing instead of using a kernel to copy elemnts one by one

CopySubMatrix should use shared memory

gpu_indices[j] is loaded multiple times

active variable set redundant computations

isActive array is copied from device to host, then host counts the number of 1's.
This can be done by summing the isActive array
numActive is set and then copied from host to device

sudo R CMD INSTALL ./ fails

$ sudo R CMD INSTALL ./

• installing to library ‘/usr/local/lib/R/site-library’
• installing source package ‘cudaglmnet’ ...
  configure: checking for critical tools used by configure
  checking "environment variable COMPILER_BINDIR"... "COMPILER_BINDIR not set; using user's default path"
  checking "environment variable CUDA_HOME"... "CUDA_HOME not set; using default /usr/local/cuda"
  checking "environment variable CULA_HOME"... "CULA_HOME not set; using default /usr/local/cula"
  checking for /usr/local/cuda/bin/nvcc... yes
  "nvcc found"
  checking "whether this is the 64 bit linux version of CUDA"... checking for /usr/local/cuda/lib64/libcublas.so... yes
  "yes -- using /usr/local/cuda/lib64 for CUDA libs"
  "using -I/usr/share/R/include for R header files"
  "using -L/usr/lib/R/lib -lR for R shared libraries"
  checking "whether this is the 64 bit version of CULA"... checking for /usr/local/cula/lib64/libcula.so... no
  checking for /usr/local/cula/lib64/libcula.dylib... no
  "no"
  checking "checking whether cula is present at all"... checking for /usr/local/cula/lib/libcula.so... no
  checking for /usr/local/cula/lib/libcula.dylib... no
  "cula is not installed"
  configure: determining host platform
  checking build system type... x86_64-unknown-linux-gnu
  checking host system type... x86_64-unknown-linux-gnu
  configure: creating ./config.status
  config.status: creating src/Makefile
  ** libs
  ** arch -
  make: Nothing to be done for `all'.
  installing to /usr/local/lib/R/site-library/cudaglmnet/libs
  ** R
  ** preparing package for lazy loading
  ** help
  *** installing help indices
  ** building package indices
  ** testing if installed package can be loaded
  Error in dyn.load(file, DLLpath = DLLpath, ...) :
  unable to load shared object '/usr/local/lib/R/site-library/cudaglmnet/libs/cudaglmnet.so':
  libcublas.so.7.5: cannot open shared object file: No such file or directory
  Error: loading failed
  Execution halted
  ERROR: loading failed
• removing ‘/usr/local/lib/R/site-library/cudaglmnet’

Support for other families in glmnet other than Gaussian?

Am I incorrect in stating that this only supports Guassian family from glmnet? Can you please clearly state whether this also supports cox, binomial, etc that glmnet currently supports?

Positivity constraints on fitted coefficients?

In glmnet one can add positivity constraints on the fitted coefficients using argument lower.limits. I was just wondering if it would be possible to add this capability in cudaglmnet also? (Support for additional families like poisson, multinomial and mgaussian would also be nice of course)

install package

Hello,
Do I need to install a package to use 'cudaglmnet' function?
It seems to be the case..
If so, how install that package ?
Thank you

Most functions can be implemented using thrust

Most functions used here can be rewritten using a transform functor, reduce, or zip