Strange sparse matrix error on GPU

7 次查看（过去 30 天）
Matt J
Matt J 2018-12-13
编辑： Matt J 2018-12-17
I have a sparse matrix with a relatively small (16 MB) memory footprint
>> Whos A
Name Size Kilobytes Class Attributes
A 100000x100000 16406 double sparse
When I send it to the GPU and add the scalar 0, I get a puzzling error
>> gpuArray(A)+sparse(0);
Error using +
Maximum variable size allowed on the device is exceeded.
I really cannot see why this operation would cause any excessively large memory allocations. This is in Matlab version R2018a and the following GPU details. Does anyone have any insights as to why this occurs?
>> gpuDevice
ans =
CUDADevice with properties:
Name: 'GeForce GTX TITAN X'
Index: 1
ComputeCapability: '5.2'
SupportsDouble: 1
DriverVersion: 9.2000
ToolkitVersion: 9
MaxThreadsPerBlock: 1024
MaxShmemPerBlock: 49152
MaxThreadBlockSize: [1024 1024 64]
MaxGridSize: [2.1475e+09 65535 65535]
SIMDWidth: 32
TotalMemory: 1.2885e+10
AvailableMemory: 1.2259e+10
MultiprocessorCount: 24
ClockRateKHz: 1076000
ComputeMode: 'Default'
GPUOverlapsTransfers: 1
KernelExecutionTimeout: 0
CanMapHostMemory: 1
DeviceSupported: 1
DeviceSelected: 1
  1 个评论
Matt J
Matt J 2018-12-13
编辑：Matt J 2018-12-14
Here's a simple command to reproduce the problem,
>> gpuArray(speye(1e5))+sparse(0);
Error using +
Maximum variable size allowed on the device is exceeded.

请先登录，再进行评论。

请先登录，再回答此问题。

采纳的回答

Edric Ellis
Edric Ellis 2018-12-14
The maximum number of elements in a gpuArray is intmax('int32'). In this case, the scalar addition is forcing your sparse matrix to become a full array of size 1e5×1e5. (In this particular case, because the scalar you are adding is precisely zero, then it could theoretically become a no-op.)
>> (1e5*1e5) <= intmax('int32')
ans =
logical
0
  3 个评论
显示 1更早的评论
Edric Ellis
Edric Ellis 2018-12-17
编辑：Edric Ellis 2018-12-17
Ok, but even so, number of non-zeros required for the result exceeds intmax('int32') because unfortunately the gpuArray implementation of plus doesn't make a special case for adding zero - sparse gpuArray plus scalar always results in a "big sparse" array which is essentially a full array but with sparse storage.
Not terribly useful perhaps, but the following does work correctly:
>> gpuArray(speye(1e5)) + spalloc(1e5,1e5,0);
Matt J
Matt J 2018-12-17
编辑：Matt J 2018-12-17
OK. Well, I hope it's on the todo list of things to fix. :-)

请先登录，再进行评论。

更多回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 GPU Computing in MATLAB 的更多信息

标签

Community Treasure Hunt

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

Start Hunting!

Translated by