Wrong result when using gpuArray with multiple implicit singleton expansions

조회 수: 1 (최근 30일)
Yunhui Zhou
Yunhui Zhou 2018년 10월 3일
답변: Wentao Du 2018년 10월 8일
Hi, I am using Matlab R2017b Update 8, and recently I find Matlab produce wrong result when calculating expressions using gpuArray with multiple implicit singleton expansion.
A minimal test case:
Matrixc = rand(2,2) * 3;
Matrixg = gpuArray(Matrixc);
wc = -2:1:2;
w3c = reshape(wc,1,1,length(wc));
w3g = gpuArray(w3c);
GPUc = gather(Matrixg(1,:).*w3g + 0.5*(Matrixg(1,:).^2+Matrixg.^2));
CPU = Matrixc(1,:).*w3c + 0.5*(Matrixc(1,:).^2+Matrixc.^2);
GPUtemp1 = Matrixg(1,:).*w3g;
GPUTemp2 = 0.5*(Matrixg(1,:).^2+Matrixg.^2);
GPUc2 = gather(GPUtemp1 + GPUTemp2);
disp(max(Matrixg(:) - Matrixc(:)));
disp(max(w3g(:) - w3c(:)));
disp(max(GPUc(:) - CPU(:)));
disp(max(GPUc2(:) - CPU(:)));
Produce results:
0
0
0.635767744700676
0
And as you can see here the problem can be by-passed by splitting up the expression to multiple steps. I want to know is it a Matlab bug or is it designed in this way?
UPDATE:
This is indeed a bug. See this bug report.

답변 (1개)

Wentao Du
Wentao Du 2018년 10월 8일
As described in the bug report , this bug was fixed as of R2018b and R2018a Update 4.
If you would like to use R2017b, you may try the workaround mentioned in the report.

카테고리

Help CenterFile Exchange에서 GPU Computing in MATLAB에 대해 자세히 알아보기

제품


릴리스

R2017b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by