In the driver version after nvidia 410, the Qt test program, when the Qimage size is 816x600, the GPU memory usage is 13M, but when the Qimage size is 817x600, the GPU memory usage continues to increase to 431M

[zhangboin]

In the driver version after nvidia 410, the Qt test program, when the Qimage size is 816x600, the GPU memory usage is 13M, but when the Qimage size is 817x600, the GPU memory usage continues to increase to 431M：
Qt source as follows：
#include
#include
#include
#include
#include

class Widget : public QOpenGLWidget
{
public:
Widget(int width, int height) : QOpenGLWidget(0), image(width, height, QImage::Format_RGB32) {}
//Widget(int width, int height) : QOpenGLWidget(0), image(width, height, QImage::Format_RGB16) {}
protected:
void paintGL() override
{
glFinish();
QPainter p(this);
static unsigned int index = 100;
image.fill(index++);
p.drawImage(QPoint(0, 0), image);
update();
}
private:
QImage image;
};

int main(int argc, char** argv)
{
QApplication app(argc, argv);

int width = 1024;
int height = 768;

if (argc == 3) {
    width = atoi(argv[1]);
    height = atoi(argv[2]);
}

Widget w(width, height);
w.resize(width, height);
w.show();

return app.exec();

}
Open the video card driver debug switch, 817x600 will have more ioctl, as follows:

[6714.002807] NVRM: ioctl(0x2a, 0x9870760, 0x20)

Some of the ioctl additions are as follows:

[6714.004671] NVRM: ioctl(0x4a, 0x9872690, 0xb8)

[6714.004870] NVRM: ioctl(0x4a, 0x9872520, 0xb8)

[6714.004921] NVRM: ioctl(0x57, 0x9872400, 0x38)

+++++++++++++++++

[6714.005669] NVRM: ioctl(0x52, 0x9872a20, 0x10)

[mtijanic]

The OpenGL driver allocates larger chunks of memory and then sub-allocates from there for texture memory. I guess that this is where you hit a threshold and it decides to allocate another chunk. I imagine if you keep increasing the size after that, the memory usage will stay constant for a while before it makes another jump.

[mtijanic]

Heh, there is some weird effect where 2048x2048 is less than 1024x1025, could be due to no power-of-two that triggers some edge case. I'll forward this to the OGL devs to see if there's anything out of the ordinary, but I wouldn't read too much into that info.

By the way, how are you measuring the GPU memory usage? Is this just that app's usage or the total system usage (e.g. as reported by nvidia-smi)

[zhangboin]

it reports by nvidia-smi

[zhangboin]

However, when the GPU memory is not enough, when the new program is used to apply for video memory, the video memory occupied by the qt test will be dynamically released, so I think it will not be nvidia driver allocated video memory has a lot of cache。

[zhangboin]

@mtijanic Hi，are there any other ideas or experimental verification measures for this problem. Thank you!

[ahuillet]

This is tracked as NVIDIA bug 5003509 and being worked on.