sparse in cuda matlab shows bad performance

This question is simple

I understand that MATLAB solves sparse linear equations by multi wavefront method on CPU

like sparse_a Is a sparse b is a full vector then x can be computed by x=sparse_a\b

is there any method to let matlab compute sparse in gpu?

gpu_sparse_a=gpuArray(sparse_a ) b_gpu=gpuArray(b)

and then x=gpu_sparse_a\b_gpu

result shows that gpu compute sparse slower than cpu....why?is sparse in cpu transfer into gpu actually become a full matrix?

[俊凯 王]

C=sparse(hang_combine,lie_combine,value_combine);

C_gpu=gpuArray(C);

f_gpu=gpuArray(f);

u_gpu=C_gpu\f_gpu;

code like this ,why is this slower than cpu ? in cpu,just u=C\f

[Matt J]

Is it slower? What proof/demonstration of that do you have for us?

[俊凯 王]

clc;clear;
cpp_parallel_mine=[];
matlab_serial=[];
time=0;
nonzero_per_row=20;
ii=[];
jj=[];
M=40*40;
N=M;
nonzero=nonzero_per_row*M;
for i=1:1:M
    jj_apend=randperm(N,nonzero_per_row);
    ii_apend=repmat((i), 1, nonzero_per_row);
    ii=[ii,ii_apend];
    jj=[jj,jj_apend];
end
row_max=max(ii);
col_max=max(jj);
sr=jj;
cpu_sparse=sparse(ii,jj,sr,row_max,col_max);
b=randi([100,200],M,1);
tic
x=cpu_sparse\b;
disp("cpu sparse solver")
toc
gpu_sparse=gpuArray(cpu_sparse);
b_gpu=gpuArray(b);
tic
x_gpu=gpu_sparse\b_gpu;
disp("gpu sparse solver");
toc

[俊凯 王]

easy code，randi sparse in cpu and gpu，if you run in matlab ，performance shows different，if M is larger ，Greater efficiency gap

[俊凯 王]

and i wonder if the sparse solver on the CPU uses a multi-core CPU

Answer 1

[Joss Knight]

It is often slower. The problem is that sparse factorizations create dense matrices...basically, it's hard to parallelize.

We generally advise to use the sparse iterative solvers, generally with preconditioners, instead. These are typically faster on GPU and CPU. Look for gmres, cgs, pcg and so on.