Faiss GPU return different result than CPU

[mjack3]

Summary

I have a code that return higher performance running FAISS on CPU than GPU. The difference is huge.

Platform

Ubuntu 20.04 lTS

Faiss version: 1.7.2

Installed from: pip install faiss-gpu
https://pypi.org/project/faiss-gpu/

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 510.39.01    Driver Version: 510.39.01    CUDA Version: 11.6     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  NVIDIA GeForce ...  On   | 00000000:0A:00.0  On |                  N/A |
| 36%   29C    P8    39W / 370W |   1095MiB / 24576MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A      1325      G   /usr/lib/xorg/Xorg                102MiB |
|    0   N/A  N/A      1906      G   /usr/lib/xorg/Xorg                523MiB |
|    0   N/A  N/A      2043      G   /usr/bin/gnome-shell               89MiB |
|    0   N/A  N/A      2436      G   ...697963766717379249,131072      361MiB |
+-----------------------------------------------------------------------------+

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Fri_Dec_17_18:16:03_PST_2021
Cuda compilation tools, release 11.6, V11.6.55
Build cuda_11.6.r11.6/compiler.30794723_0

#define CUDNN_MAJOR 8
#define CUDNN_MINOR 3
#define CUDNN_PATCHLEVEL 3

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

I am using AUROC to meassure the performance of certain code. The "faiss part" of the code looks like this:

import faiss
from faiss.contrib import torch_utils


device = torch.device('cuda'):
bank_embeddings: torch.Tensor = torch.load(os.path.join(embedding_dir, 'bank_embeddings.pt'), map_location=device)
self.index = faiss.IndexFlatL2(bank_embeddings.shape[-1])
if device == torch.device('cuda'):
    provider = faiss.StandardGpuResources()  # use a single GPU
    self.index = faiss.index_cpu_to_gpu(provider, 0, self.index) # make it a flat GPU index
self.index.add(bank_embeddings)

Next, I do my search

D, I = self.index.search(n, 3)

If I use CPU my AUROC performance is 0.9754
If I use the GPU, my AUROC score is 0.4675

There is something wrong? The code is always the same :S

[mdouze]

No this is not normal.
The only possible discrepancy is when there are duplicate vectors in the dataset and their ordering in result lists is random.

[mjack3]

I have tested this:

import faiss
from faiss.contrib import torch_utils

device = torch.device('cuda') # GPU device
# device = torch.device('cpu') # CPU device

bank_embeddings: torch.Tensor = torch.load(os.path.join(embedding_dir, 'bank_embeddings.pt'), map_location=device)
 if device == torch.device('cuda'): # FIXME: the performance in GPU is decreased
       cfg = faiss.GpuIndexFlatConfig()
       resources = faiss.StandardGpuResources()
       self.index = faiss.GpuIndexFlatL2(resources, bank_embeddings.shape[-1], cfg)
else:
       self.index = faiss.IndexFlatL2(bank_embeddings.shape[-1])

self.index.add(bank_embeddings)

Next the same search

if not n.is_contiguous():
       n = n.contiguous() # n shape is: (784, 1536)
D, I = self.index.search(n, 3) # D shape is (784, 3)

D matrix if GPU is used

tensor([[15.5180, 18.6601, 19.2905],
        [18.4031, 21.5452, 22.1756],
        [20.4813, 23.6233, 24.2538],
        ...,
        [23.9322, 27.0743, 27.7047],
        [21.9629, 25.1050, 25.7354],
        [17.7017, 20.8438, 21.4742]], device='cuda:0')

D matrix if CPU is used:

tensor([[4.4300, 6.7436, 6.9746],
        [4.1022, 4.9629, 6.6677],
        [4.2402, 6.0263, 7.2301],
        ...,
        [4.8605, 5.8493, 6.3380],
        [4.7552, 5.8216, 5.8592],
        [4.7130, 6.0770, 6.9383]])

Values are quite differents.

Do you think this could be solve manually?

[mjack3]

Did you any progress?

[mdouze]

Could you repro with self-contained code?

[mjack3]

Could you repro with self-contained code?

Sure. Give me until tomorrow :).

Thanks!

[mjack3]

Hello @mdouze

I updated a self-contained code. There you just need to execute the main.py for running the tests. Let me know if you can access properly.

https://github.com/mjack3/FAISS_GPU_TEST

Note: I am using the torch+CUDA==11.3 although requeriments.txt has the torch+CUDA==10.2. I think that won't make difference.

Thank you so much

[mdouze]

Running your test on Faiss 1.7.2 and torch 1.10.1 yields https://gist.github.com/mdouze/df42bd8ad8c74c42fedd3fd1043e83b7

NB that the Cuda version does matter and that pip is not a supported way of installing Faiss (use conda).

[mjack3]

pip is not a supported way of installing
sure there is the problem. I have my project environment configured with pip. Do you plan to support pip ?

[roomo7time]

For my cace, I was experiencing the cpu-gpu difference issue with faiss-gpu==1.7.2 that I installed with pip. After uninstalling it, and reintalling faiss-gpu by conda using conda install -c pytorch faiss-gpu==1.7.3 the issue has been resolved.