search

ARM-software / ComputeLibrary

The Compute Library is a set of computer vision and machine learning functions optimised for both Arm CPUs and GPUs using SIMD technologies.
MIT License
2.82k stars 774 forks source link

How to Set the the num_threads when calling the runtime function? #1025

Closed GGGGxxxxxxxxr closed 1 year ago

GGGGxxxxxxxxr commented 1 year ago

H! I am building a WDSR model based on Arm Compute Library. As there are some operators (such as SkipLink) not supported by ArmComputeLibrary currently, I just manually call the runtime functions instead of using the GraphAPI.

I wonder how to set the num_threads during layer.run() when I don't utilize the GraphAPI?

Here is an example part of the code:

NEConvolutionLayer conv3_1;
conv3_1.configure(&layer2_out, &conv1_weight, &conv1_bias, &layer3_out, PadStrideInfo(1, 1, 1, 1));

conv3_1.run();

For example, I would like to use NUM_THREADS=1 in this conv3_1.run(), how can I set it?

Thanks!

GGGGxxxxxxxxr commented 1 year ago

Some update, I somehow managed to manually set the num_threads in CPPScheduler.cpp, but I am wondering whether I could have a function call to set it instead of modifying the backend source code too much.

morgolock commented 1 year ago

Hi @GGGGxxxxxxxxr

You can do this at runtime with the code below:

Scheduler::get().set_num_threads(threads_num);

For more details see: https://github.com/ARM-software/ComputeLibrary/blob/main/arm_compute/runtime/IScheduler.h#L146

Hope this helps