mgpusim icon indicating copy to clipboard operation
mgpusim copied to clipboard

Benchmark simpleconvolution Stops Mid-Execution

Open DDantalion opened this issue 10 months ago • 2 comments

I'm encountering an issue where the benchmark test for simpleconvolution fails to complete and stops running unexpectedly partway through.

Parameters Used in simpleconvolution.go:

var widthFlag = flag.Uint("width", 512, "The width of the input matrix.") var heightFlag = flag.Uint("height", 512, "The height of the input matrix.") var maskSizeFlag = flag.Uint("mask-size", 3, "The size of the mask.")

Command Executed: ./simpleconvolution -timing -unified-gpus 1,2,3,4 -use-unified-memory -report-all

If anyone has experienced similar issues or knows a possible fix, your insights would be greatly appreciated.

DDantalion avatar Jan 29 '25 07:01 DDantalion

I think it should be a bug, I will fix it as soon as possible. Thanks for your feedback!

DX990307 avatar Jan 30 '25 19:01 DX990307

I think the problem is that the driver seems unable to receive any page migration requests except for the first one. This problem also exists in conv2d with the parameter -W=1024, -H=1024 when enable page migration. Could you please share any strategies or solutions?

DDantalion avatar Feb 03 '25 22:02 DDantalion

Unified memory is known to have a hanging problem. We are in the process of reimplementing the unified memory system. Closed for now.

syifan avatar Jun 10 '25 23:06 syifan