vello_gpu: Optimize full-sized texture sampling by laurenz-canva · Pull Request #1964 · linebender/vello

laurenz-canva · 2026-09-29T14:55:41Z

This PR adds a fast path for images where we want to sample the whole texture. When this is the case, we can make use of WebGL sampler objects to define whether to use NN/bilinear sampling as well as which extend mode to use, without us having to reimplement these operations in software! This is much faster, see this video and compare to the FPS numbers from #1963 instead. We now fully saturate the 90FPS of this device 🎉!

IMG_1338.MOV

The only two downsides are that

It makes our shaders a bit bigger. In a follow-up, I will expand the image-bicubic feature to instead allow compiling out all non-native image operations.
We need to reject non-filterable textures, which I don't think is a big deal for now. In the future, we can refine this so non-filterable textures are accepted again, as long as the user only uses ImageQuality::Low.

Once this (and the preceding PR) have been approved, I will make sure to rerun tests on other devices to ensure we are not creating any regressions by making the shader bigger.

LaurenzV · 2026-09-30T08:13:48Z

+    /// **It is important that the provided width and height match the actual dimensions of the
+    /// texture, otherwise, rendering might be corrupted!**
+    Full {
+        /// Width of the bound texture view.
+        width: u16,
+        /// Height of the bound texture view.
+        height: u16,
+    },


In theory, we could derive that information ourselves using textureDimensions in the shader. However, I faintly remember this being a performance footgun. Hence why I think it's better to require that information to be passed along.

LaurenzV · 2026-09-30T08:34:50Z

+            );
+            out.sample_xy = get_native_image_translate(image_texel1)
+                + get_native_image_transform(image_texel0) * pos;
+            out.payload = get_native_image_opacity(image_texel1);


Bit annoying to abuse the payload like this. 😅 But it avoids having to do another texture sample in the fragment shader, as all the metadata we need is passed directly.

LaurenzV · 2026-10-01T06:38:17Z

+                (paint_and_rect_flag >> EXTERNAL_TEXTURE_SLOT_SHIFT) & 0x3u;
+
+            if external_texture_slot == 0u {
+                final_color = paint_alpha * textureSampleLevel(external_texture_0, external_sampler_0, sample_xy, 0.0);


TODO: Use textureSample?

I think it's better to leave this as is, and revisit once we properly support mip-chains.

LaurenzV · 2026-10-09T07:17:19Z

+    if image.tint.is_some_and(|tint| {
+        tint.mode != TintMode::Multiply || tint.color.components[..3] != [1.0; 3]


In the future, we cna likely support Multiply tinting as well, by storing the full rgba instead of just the opacity.

…ender#1963) This PR makes the following changes: - It adds a shim in the test suite so that we can run most external texture tests (the once that sample the full texture) with Vello CPU as well. By doing so, Vello CPU can serve as the reference, and we increase the coverage of tests that compare against Vello CPU and GPU. - It fixes the issue where, instead of extending each tap of bilinear/bicubic samples, we only extended the original sample location. This is wrong, and is the reason why a couple of image tests were inconsistent between CPU and GPU. In addition to that, it also ports some changes to the extend logic that we recently applied to Vello CPU, to make the results more numerically stable and the code more comparable to Vello CPU. **This does unfortunately regress performance: Rendering a full-sized image on my Android phone goes from around 73FPS to 66FPS.** See the below videos. Before: https://github.com/user-attachments/assets/0660b553-9076-4a1d-88ec-88328b034501 After: https://github.com/user-attachments/assets/66fd5fcd-f266-4eb1-b739-29bb78ee8783 But the previous approach was just fundamentally wrong. However, the good news is that with linebender#1964, I will introduce a fast path that uses GPU-native bilinear sampling for images that are sampled from the whole external texture, **which will not only undo this slowdown, but in fact make image rendering even faster compared to current main, when using external textures and sampling the whole texture! See that PR for more information.** Since rendering whole images is the most common operation (except for glyph caching, but this is experimental right now, anyway. And nearest-neighbor sampling is less affected than bilinear sampling), in my opinion this is a trade-off worth taking. Using the image atlas will unfortunately stay slower for now, even with linebender#1964, but I think that's something we have to accept for now. Once we revisit glyph caching, we can figure out how to best fix this. We could also improve this in the future by restricting the possible image sampling modes for atlas images (for example, only allowing extend mode `Pad`). But the problem is simply that when we sample from an arbitrary subregion of an atlas, we have to emulate correct extension ourselves, which is much slower than letting the hardware do it, so this should be avoided. --------- Co-authored-by: Laurenz Stampfl <laurenz.stampfl+github@gmail.com>

laurenz-canva marked this pull request as draft September 29, 2026 14:58

laurenz-canva added this pull request to stack #1965 September 29, 2026 14:59

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from 048b162 to 42ba050 Compare September 29, 2026 15:08

LaurenzV force-pushed the laurenz/test_image_fast_path branch from 42ba050 to bfec171 Compare September 30, 2026 06:18

laurenz-canva force-pushed the laurenz/test_image_fast_path branch 2 times, most recently from 5b66f7a to 9eaef9e Compare September 30, 2026 07:09

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from 9eaef9e to c31e48a Compare September 30, 2026 07:47

LaurenzV mentioned this pull request Sep 30, 2026

vello_gpu: Fix methods for bilinear and bicubic image sampling #1963

Merged

laurenz-canva force-pushed the laurenz/test_image_fast_path branch 2 times, most recently from fa12eb3 to 597aa52 Compare September 30, 2026 08:42

LaurenzV reviewed Sep 30, 2026

View reviewed changes

LaurenzV marked this pull request as ready for review September 30, 2026 08:58

LaurenzV requested a review from grebmeg September 30, 2026 08:58

LaurenzV reviewed Oct 1, 2026

View reviewed changes

LaurenzV removed the request for review from grebmeg October 1, 2026 09:29

LaurenzV marked this pull request as draft October 1, 2026 09:30

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from 597aa52 to 340a2ca Compare October 1, 2026 10:58

laurenz-canva removed this pull request from stack #1965 October 1, 2026 10:59

laurenz-canva added this pull request to stack #1969 October 1, 2026 11:00

LaurenzV force-pushed the laurenz/test_image_fast_path branch from 340a2ca to f9e3ed5 Compare October 2, 2026 06:21

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from f9e3ed5 to 3df9f0b Compare October 5, 2026 09:05

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from 3df9f0b to 5fde82a Compare October 5, 2026 14:05

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from 5fde82a to c54ee4d Compare October 5, 2026 14:35

LaurenzV mentioned this pull request Oct 8, 2026

vello_gpu: Add initial support for excluding certain paints from shaders at compile-time #1978

Merged

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from c54ee4d to bf7f46d Compare October 8, 2026 08:04

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from bf7f46d to b835478 Compare October 8, 2026 08:08

laurenz-canva marked this pull request as ready for review October 8, 2026 08:45

LaurenzV requested a review from grebmeg October 8, 2026 08:50

LaurenzV mentioned this pull request Oct 9, 2026

vello_gpu: Add NV12 YCbCr external texture support #1995

Open

laurenz-canva force-pushed the laurenz/test_image_fast_path branch from b835478 to fa4fcb7 Compare October 9, 2026 06:17

Base automatically changed from laurenz/fix_bilinear_image to main October 9, 2026 07:00

vello_gpu: Optimize full-sized texture sampling

3e7e39c

LaurenzV force-pushed the laurenz/test_image_fast_path branch from fa4fcb7 to 3e7e39c Compare October 9, 2026 07:00

LaurenzV reviewed Oct 9, 2026

View reviewed changes

Navigation Menu

Repository navigation

vello_gpu: Optimize full-sized texture sampling - #1964

vello_gpu: Optimize full-sized texture sampling#1964
laurenz-canva wants to merge 1 commit into
mainfrom
laurenz/test_image_fast_path

laurenz-canva commented Sep 29, 2026 •

edited

Loading

Uh oh!

LaurenzV Sep 30, 2026

Uh oh!

LaurenzV Sep 30, 2026

Uh oh!

LaurenzV Oct 1, 2026

Uh oh!

laurenz-canva Oct 8, 2026

Uh oh!

LaurenzV Oct 9, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		if image.tint.is_some_and(\|tint\| {
		tint.mode != TintMode::Multiply \|\| tint.color.components[..3] != [1.0; 3]

Conversation

laurenz-canva commented Sep 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

LaurenzV Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

LaurenzV Sep 30, 2026

Choose a reason for hiding this comment

Uh oh!

LaurenzV Oct 1, 2026

Choose a reason for hiding this comment

Uh oh!

laurenz-canva Oct 8, 2026

Choose a reason for hiding this comment

Uh oh!

LaurenzV Oct 9, 2026

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

laurenz-canva commented Sep 29, 2026 •

edited

Loading