I am able to reproduce this issue with OCV CPU but not with CUDA. The CUDA backend gives identical timings for both models.

I have replaced the timing code with this:

using std::chrono::steady_clock;
auto start = steady_clock::now();
auto preds = net.forward();
auto end = steady_clock::now();

CUDA backend:

timings for model 1

Total time:303ms
Total time:14ms
Total time:16ms
Total time:15ms
Total time:15ms
Total time:16ms
Total time:16ms
Total time:15ms
Total time:15ms
Total time:16ms
Total time:15ms
Total time:16ms
Total time:17ms
Total time:13ms
Total time:15ms
Total time:15ms
Total time:15ms
Total time:14ms
Total time:15ms
Total time:14ms
test

timings for model 2

Total time:271ms
Total time:15ms
Total time:14ms
Total time:14ms
Total time:14ms
Total time:14ms
Total time:15ms
Total time:13ms
Total time:15ms
Total time:15ms
Total time:15ms
Total time:14ms
Total time:14ms
Total time:15ms
Total time:16ms
Total time:14ms
Total time:14ms
Total time:22ms
Total time:13ms
Total time:15ms
test

OCV CPU:

timings for model 1

Total time:219ms
Total time:69ms
Total time:69ms
Total time:68ms
Total time:69ms
Total time:68ms
Total time:70ms
Total time:70ms
Total time:71ms
Total time:73ms
Total time:72ms
Total time:72ms
Total time:73ms
Total time:72ms
Total time:72ms
Total time:73ms
Total time:72ms
Total time:72ms
Total time:72ms
Total time:72ms
test

timings for model 2

Total time:505ms
Total time:357ms
Total time:355ms
Total time:352ms
Total time:352ms
Total time:353ms
Total time:353ms
Total time:352ms
Total time:351ms
Total time:350ms
Total time:350ms
Total time:351ms
Total time:350ms
Total time:350ms
Total time:350ms
Total time:350ms
Total time:350ms
Total time:350ms
Total time:349ms
Total time:350ms
test

Uh oh!

same dnn model with different params, the speed is different #17259

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions