How to get the communication volume and how to evaluate individual DNN layer? about spu HOT 6 CLOSED

Hi @warpoons

Close to the end of profiling information, there is a line with total bytes send
[2024-04-03 17:19:25.930] [info] [api.cc:204] Link details: total send bytes 2609750688, send actions 4112

from spu.

Hi @warpoons

Close to the end of profiling information, there is a line with total bytes send [2024-04-03 17:19:25.930] [info] [api.cc:204] Link details: total send bytes 2609750688, send actions 4112

Thanks for your prompt reply. Sorry for my carelessness.
Would you mind taking a look at my question about how to evaluate individual each DNN layer? Thanks.

from spu.

Consider we don't have a good way to identify boundaries between layers in IR after compilation, collecting data that belongs to a specific layer is almost mission impossible at this point.

from spu.

Consider we don't have a good way to identify boundaries between layers in IR after compilation, collecting data that belongs to a specific layer is almost mission impossible at this point.

What about if I only create a single-layer network that is equivalent to the target layer I want to evaluate, then use SPU to evaluate it. Can it be possible? Please refer to "#599 (comment)". Thanks.

from spu.

Consider we don't have a good way to identify boundaries between layers in IR after compilation, collecting data that belongs to a specific layer is almost mission impossible at this point.

What about if I only create a single-layer network that is equivalent to the target layer I want to evaluate, then use SPU to evaluate it. Can it be possible? Please refer to "#599 (comment)". Thanks.

Yes, that is definitely a feasible solution.

Also consider some optimizations might happen across layers, you may get slightly worse performance by just running a single layer. The difference should be marginal most of time, but just keep this in mind.

from spu.

Consider we don't have a good way to identify boundaries between layers in IR after compilation, collecting data that belongs to a specific layer is almost mission impossible at this point.

What about if I only create a single-layer network that is equivalent to the target layer I want to evaluate, then use SPU to evaluate it. Can it be possible? Please refer to "#599 (comment)". Thanks.

Yes, that is definitely a feasible solution.

Also consider some optimizations might happen across layers, you may get slightly worse performance by just running a single layer. The difference should be marginal most of time, but just keep this in mind.

Hi @anakinxc. Thanks for your kind reminder. I'll try this and report back here if it is feasible. Thanks!

from spu.