ricosjp / monolish Goto Github PK

Only the following two implementations are available:

• COO -> Dense
• COO -> CRS

Hi,

Just wondering whether you plan to provide support for asynchronous copy/compute with GPU calls such as dgemm.

I'm glad that your are using OpenMP offload. Since you say that the OpenMP offload is only currently available for NVIDIA. I was just wondering whether it is simply a question of time or if there were additional technical issues for implementing OpenMP offload for AMD.

Thanks,
Vijay

monolish_0.14.0-1+oss_amd64.deb and monolish_0.14.0-1+intel_amd64.deb is available in addition to monolish_0.14.0-1+oss+nvidia_amd64.deb and monolish_0.14.0-1+intel+nvidia_amd64.deb.

These new variants only contains CPU variant(libmonolish_cpu.so), but prevent from installing CUDA libraries which is very cumbersome. Installation document should reflect that.

ref: https://docs.nvidia.com/cuda/cublas/index.html#cublas-lt-t-gt-geam

The test/equateion in LinearOperator is not impl.

impl. test and update documents

impl. in github actions

• generate Doxyben (need to chenge version name)
• generate deb file
• generate monolish docker

need to get version number...

A.norm() ?

• get_device_num
• get_num_device
• set_num_device

In order to implement IC/ILU preconditioner

If it is not implemented, the equation class cannot be templated.

The solver class functions do nothing.

This is the same as creating a CCS format.

We will not implement CCS format, but will implement the transpose() of CRS.

• get/set device num. I/F
• enable option of mpicc in cmake : -DUSE_MPI
• Input MPI_COMM I/F
• implement BLAS Lv.1 #56
• implement global size of Dense
• implement BLAS Lv.2(Dense)
• Checking the working of CG in Dense (precon = none)
• implement comm_table of CRS
• implement BLAS Lv.2(CRS)
• Checking the working of CG in CRS (precon = none)

sample code can not link monolish

This is Pending. Because cusparse SpMM can not use Row major matrix in CUDA10.x

CUDA11.X can use Row major matrix. ( wait for #8 )

Now the operator[] function gives an error when eigs and eigvecs are already transferred to the GPU.
Change it to give a LOBPCG-specific error message.

./lobpcg_gpu.out
terminate called after throwing an instance of 'std::runtime_error'
  what():  Error, GPU vector cant use operator[]

• precond_apply(MATRIX, view1D, view1D)
• solve(MATRIX, view1D, view1D)

at For Developpers

https://i.riken.jp/en/supercom/documents/himenobmt/

Find out how to link and extract programs in Doxygen.

If you use Log Viewer v0.1.0, type conversion may not work when reading data, so you need to update to the new version.

not impl. test now

The equation test can be templated.

Current docker/ is not working.

Container specification

• monolish is installed with apt (only with tag)
• Put the samples in /opt/monolish/examples
• Make it downloadable from github container registry

Tasks:

• oss-cpu
• mkl-cpu
• oss-nvidia
• mkl-nvidia
• Enable gitlab-ci to deploy automatically (only with tag, How do I auto-specify a tag in gitlab-CI?)
• Document the usage.
• In the documentation, move the build method using allgebra to For developpers.

create subdir

currently uses cusparseDcsrmv() and cusparseScsrmv(), but these APIs are deprecated already on CUDA 10.1 and removed on CUDA 11.2.

We should use new cusparseSpMV() API (already available on CUDA 10.1)

need ARM CI Runner

monolish::matrix::CRS<double> A_crs(COO);
monolish::matrix::Dense<double> A_dense(COO);
moolish::vector<double> tmp(A.col());
monolish::blas::matvec(A_crs, tmp, tmp);   // tmp = A * tmp
monolish::blas::matvec(A_dense, tmp, tmp); // tmp = 0

If the pointers are the same, an error is thrown.

implemented in cusparse

https://docs.nvidia.com/cuda/cusparse/index.html#ic0

Currently, all CUDA-related calls are wrapped with check_CUDA() to handle error code enums.
These enums are different between CUDA, cuBLAS, cuSOLVER, and cuSPARSE; we need to use different wrappers.

in monolish::mpi

• MPI_Allgatherv();
• MPI_Alloc_mem();
• MPI_Allreduce();
• MPI_Alltoall();
• MPI_Barrier();
• MPI_Comm_rank();
• MPI_Comm_size();
• MPI_Finalize();
• MPI_Get_count();
• MPI_Init();
• MPI_Initialized();
• MPI_recv();
• MPI_send();
• MPI_Irecv();
• MPI_Isend();
• MPI_Scatterv();
• MPI_Waitall();
• MPI_Bcast()
• MPI_Gather()
• MPI_Scatter()
• MPI_Allgather();

monolish::matrix::COO A = monolish::util::2D_5p_laplace(M, N);

https://github.com/ricosjp/allgebra/releases/tag/21.05.0

• is_positive_definite()
• is_diag_dominant()
• is_symm()

make guide to build OSS and Intel

mumps or other?

It is necessary to consider the management method by tagging etc.
First, move everything.

Specify the method:

Float get_residual(vector &x);
https://github.com/ricosjp/monolish/blob/master/include/monolish_solver.hpp#L46

CUDA10.x and CUDA11.x have different APIs.

the resulting values are unreliable.