Skip to content

Why cuda is slower than cpu #536

Open
@Hukongtao

Description

@Hukongtao

🐛 Describe the bug

Minimal reproducible code

import time

import torch
from torchcodec.decoders import VideoDecoder

device = "cpu"
# device = "cuda"


video_path = "NASAs_Most_Scientifically_Complex_Space_Observatory_Requires_Precision-MP4_small.mp4"
decoder = VideoDecoder(
    video_path, 
    device=device,
    # dimension_order="NHWC",
    # seek_mode="approximate",
    num_ffmpeg_threads=8,
)

index_list = [106, 125, 127, 130, 132, 144, 146, 171, 180, 181, 189, 194, 195, 199, 203, 204, 204, 214, 227, 242, 259, 263, 266, 296, 303, 314, 320, 323, 325, 328, 333, 338, 338, 350, 370, 373, 381, 384, 384, 384, 396, 400, 444, 448, 452, 463, 467, 470, 473, 479, 487, 489, 489, 529, 532, 532, 559, 564, 570, 617, 649, 658, 658, 665, 674, 691, 703, 704, 716, 718, 733, 743, 750, 754, 765, 777, 786, 792, 814, 818, 818, 821, 833, 847, 854, 858, 859, 877, 881, 891, 925, 926, 928, 949, 954, 967, 975, 982, 987, 990]

data = decoder.get_frames_at(indices=index_list)

loop_number = 20
start_time = time.time()
for i in range(loop_number):
    data = decoder.get_frames_at(indices=index_list)
    print(data.data.shape)
end_time = time.time()
# print(data.data)
# print(data.data.shape)

print(f"spend time:{(end_time - start_time) / loop_number}")

cpu spend time: 0.9169630885124207
gpu spend time:3.559278666973114
Use different devices, and you will find that cuda takes more time than cpu. I don't understand why.

Versions

As mentioned above

Metadata

Metadata

Assignees

No one assigned

    Labels

    questionFurther information is requested

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions