What is the fastest way to convert image-rs::Image into slint::Image?

[qarmin]

Currently I'm using code similar to

pub type ImageBufferRgba = image::ImageBuffer<image::Rgba<u8>, Vec<u8>>;

fn convert_into_slint_image(img: DynamicImage) -> slint::Image {
    let image_buffer: ImageBufferRgba = img.to_rgba8();
    let buffer = slint::SharedPixelBuffer::<slint::Rgba8Pixel>::clone_from_slice(image_buffer.as_raw(), image_buffer.width(), image_buffer.height());
    slint::Image::from_rgba8(buffer)
}

let img = image::open(image_name).unwrap();
let img2 = convert_into_slint_image(img);
app.global::<GuiState>().set_preview_image(slint_image);

In debug build, jpg image 5184x3495(~18 million pixels) which weight is 19.1 MB loads in ~1.1 second

Loading image took: 374.248528ms, converting image took: 782.257233ms, setting image took: 4.586µs

In release build it takes lower amount of time

Loading image took: 367.533911ms, converting image took: 85.82906ms, setting image took: 1.47µs

Is there any way to decrease convert time between formats especially in debug builds?

[ogoffart]

The best would be to avoid conversion altogether. (ie: trying to always operate with a SharedPixelBuffer.

But there might be optimizations to do in clone_from_slice() The current implementation:

slint/internal/core/graphics/image.rs

Lines 129 to 139 in df7657d

	pub fn clone_from_slice<SourcePixelType>(
	pixel_slice: &[SourcePixelType],
	width: u32,
	height: u32,
	) -> Self
	where
	[SourcePixelType]: rgb::AsPixels<Pixel>,
	{
	use rgb::AsPixels;
	Self { width, height, data: pixel_slice.as_pixels().iter().cloned().collect() }
	}

Maybe .iter().cloned().collect() is not ideal especially debug mode. Maybe something like this would work

Vec::from(pixel_slice.as_pixels())

and would be more optimized in debug mode.

[ogoffart]

I pushed that change in 1a0e065 which should in principle improve the debug build

[qarmin]

Before

Loading image took: 374.248528ms, converting image took: 782.257233ms, setting image took: 4.586µs

Now

Loading image took: 380.358729ms, converting image took: 488.832646ms, setting image took: 7.982µs

[tronical]

convert_into_slint_image() does two things:

(1) convert into a rgba8 image
(2) creating a copy via clone_from_slice

It would be interesting to know how long step (1) takes. Perhaps there's a way to make it so that the decoder creates rgba8 right away and so that this step vanishes.

For (2) - if it turns out to be costly - we would need a way adopt a Vec's data. That would require changing the SharedVector internals to allow for an indirection that we typically don't need.

Perhaps another option that might work with today's APIs would be to use image-rs lower level APIs to create a Reader, determine dimensions, allocate a SharedPixelBuffer and decode directly into it. Not sure if that's always possible or at what other cost this comes.

[qarmin]

DEBUG:
Converting image to rgba8 took: 43.705888ms
Converting slice into shared pixelbuffer took: 748.677265ms
Converting sharedpixelbuffer to slint image took: 416ns
Loading image took: 390.485237ms, converting image took: 800.394429ms, setting image took: 5.231µs

RELEASE:
Converting image to rgba8 took: 44.366056ms
Converting slice into shared pixelbuffer took: 35.693432ms
Converting sharedpixelbuffer to slint image took: 202ns
Loading image took: 352.566009ms, converting image took: 88.694365ms, setting image took: 1.305µs