multi gpu acceleration

[ppeterpp]

according to the multi-gpu code given in the readme, i used below to do interpolation

streams = []
cycle=2
engine_path="/workspace/tensorrt/ckpts_no_avx512/rife419_beta_v2_ensembleTrue_op20_fp16_clamp_sim_2160*3840.engine"
for i in range(cycle):
    streams.append(core.std.SelectEvery(rife_trt(clip, multi=3, engine_path=engine_path, num_streams=4, device_id=i), cycle=cycle, offsets=i))
clip = core.std.Interleave(streams)

it seems like each stream is taking as input the whole clip, which might be redundant? And I also found that increasing the num of cycle does no help to make it faster.
using cycle=2:

Output 750 frames in 18.43 seconds (40.70 fps)me=00:00:03.99 bitrate=49365.4kbits/s speed=0.228x     
Filtername           Filter mode   Time (%)   Time (s)
Model                parallel       227.65      41.95
Model                parallel       227.23      41.87
Bicubic              parallel       166.12      30.61
Bicubic              parallel        53.94       9.94
VideoSource          unordered       18.00       3.32
BlankClip            parallel         0.09       0.02
BlankClip            parallel         0.07       0.01
BlankClip            parallel         0.07       0.01
BlankClip            parallel         0.05       0.01
Interleave           parallel         0.02       0.00
Interleave           parallel         0.02       0.00
SelectEvery          parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
DuplicateFrames      parallel         0.00       0.00
DuplicateFrames      parallel         0.00       0.00
Trim                 parallel         0.00       0.00
Loop                 parallel         0.00       0.00
Trim                 parallel         0.00       0.00
Loop                 parallel         0.00       0.00
Interleave           parallel         0.00       0.00
Interleave           parallel         0.00       0.00
<Destroyed>          <unknown>        0.00       0.00

using cycle=4:

Output 750 frames in 20.00 seconds (37.49 fps)me=00:00:03.96 bitrate=49697.3kbits/s speed=0.209x     
Filtername           Filter mode   Time (%)   Time (s)
Bicubic              parallel       170.90      34.19
Model                parallel        83.73      16.75
Model                parallel        83.39      16.68
Model                parallel        79.77      15.96
Model                parallel        78.67      15.74
Bicubic              parallel        54.29      10.86
VideoSource          unordered       15.86       3.17
BlankClip            parallel         0.06       0.01
BlankClip            parallel         0.06       0.01
BlankClip            parallel         0.06       0.01
BlankClip            parallel         0.05       0.01
BlankClip            parallel         0.05       0.01
BlankClip            parallel         0.05       0.01
BlankClip            parallel         0.05       0.01
BlankClip            parallel         0.05       0.01
SelectEvery          parallel         0.04       0.01
Interleave           parallel         0.02       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
Interleave           parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
Interleave           parallel         0.01       0.00
SelectEvery          parallel         0.01       0.00
SelectEvery          parallel         0.00       0.00
SelectEvery          parallel         0.00       0.00
SelectEvery          parallel         0.00       0.00
SelectEvery          parallel         0.00       0.00
DuplicateFrames      parallel         0.00       0.00
DuplicateFrames      parallel         0.00       0.00
DuplicateFrames      parallel         0.00       0.00
DuplicateFrames      parallel         0.00       0.00
Trim                 parallel         0.00       0.00
Trim                 parallel         0.00       0.00
Trim                 parallel         0.00       0.00
Loop                 parallel         0.00       0.00
Trim                 parallel         0.00       0.00
Loop                 parallel         0.00       0.00
Loop                 parallel         0.00       0.00
Loop                 parallel         0.00       0.00
Interleave           parallel         0.00       0.00
Interleave           parallel         0.00       0.00
Interleave           parallel         0.00       0.00
Interleave           parallel         0.00       0.00
<Destroyed>          <unknown>        0.00       0.00

time used per 'Model' drop half, but the total time even becomes longer. do you have any suggestion?

[styler00dollar]

I added this example because tepete told me that this works for upscaling. I don't have access to a multi-gpu system, so I can't test. You may have to find different code or it may not even work properly for interpolation, but rife is so fast that you don't really need multi-gpu and are encode limited. It makes more sense for slow workloads like upscaling.