ArrayFire out of memory doing convolution on NVIDIA Fermi?

Question

I'm trying to use ArrayFire to perform a convolution on a 9000x9000 pixel 3-channel image, which is ~75MB. My GPU is an NVIDIA GTX480 with 1536MB of RAM. I would expect ArrayFire to use 75MB for the input image and roughly 75MB for the output image. However, ArrayFire runs for a while and eventually says that it's out of memory:

Memory Usage: 1325 MB free (1536 MB total) //printed before calling convolutionTest()
warning: device memory is low //printed in convolutionTest()
src/gena/gi_mem.cpp:349: error: tried to allocate 309mb (45mb free / 1536mb total) //exception

When performing a convolution on a 75mb image on a GPU with 1536MB of memory, ArrayFire runs out of memory. Why does this happen, and what can I do about it?

Code:

#include <stdio.h>
#include <arrayfire.h>
using namespace af;

static const float h_sobel[] = {-2.0, -1.0,  0.0,
                                -1.0,  0.0,  1.0,
                                0.0,  1.0,  2.0}; // 3x3 sobel weights

static void convolutionTest() {
    array sobel_k = array(3, 3, h_sobel);
    array img_gray = loadimage("9k_x_9k.png", false); // 'false' makes it a 1 channel grayscale [0-255]
    array img_convolved = convolve(img_gray, sobel_k); // should I preallocate the output space?
}

int main(int argc, char** argv) {
    try {
        info();
        convolutionTest();
    } catch (af::exception& e) {
        fprintf(stderr, "%s\n", e.what()); //prints src/gena/gi_mem.cpp:349: error: tried to allocate 309mb (45mb free / 1536mb total)
    }
    return 0;
}

System configuration and notes:

ArrayFire 1.9
Ubuntu 10.04
CUDA 5.0
NVIDIA GTX480 (Fermi) GPU, which has 1536MB of RAM
helloworld and other ArrayFire examples work properly
ArrayFire's convolution has no trouble with a smaller image (e.g. 512x512 pixels)

ArrayFire crashes when doing `convolve()` on an image with 9000x9000 pixels. A 512x512 pixel image works fine. Feel free to download the 9000x9000 image [here](http://dl.dropbox.com/u/35993/9k_x_9k.png). — solvingPuzzles, Jan 08 '13 at 23:48
Is it failing with the code you provided here ? Or are you running convolutionTest in a for loop? — Pavan Yalamanchili, Jan 08 '13 at 23:50
keep in mind a 9000x9000 image has 81 million elements. This translates to 308MB for single channel (each pixel is a float value). — Pavan Yalamanchili, Jan 08 '13 at 23:56
Right, makes sense. So, if we have 308mb for the input and ~308mb for the output...perhaps there's a memory leak causing us to use up all the RAM? — solvingPuzzles, Jan 08 '13 at 23:57
Can we close this question here ? I figured out the problem, and this may be too localized for stack overflow. I will file a bug report internally. email support@accelereyes.com so I can keep you updated. Please provide the link to this question so we know it is you when you send it. — Pavan Yalamanchili, Jan 09 '13 at 00:01
Sounds good! Do you want to post an answer like "we're fixing it," and I'll accept the answer? — solvingPuzzles, Jan 09 '13 at 00:07

Pavan Yalamanchili · Accepted Answer · 2013-01-09T00:15:04.707

2

This is a bug in loadImage function. It is loading in all three channels causing it to use up more memory than necessary. AccelerEyes will fix it before the next nightly, and report back soon.

edited Jan 09 '13 at 00:15

answered Jan 09 '13 at 00:08

Pavan Yalamanchili

12,021
2
35
55

Could you display benchmarks for image convolution (Something like 1000 x 1000 with a 5x5 kernel)? Thanks. – Royi Oct 24 '13 at 18:54

ArrayFire out of memory doing convolution on NVIDIA Fermi?

1 Answers1