HelixSR 1.3.0
What's new
- Automatic wave size. HelixSR now picks the shader wave size that suits your GPU. This targets the blurry/ghosting image reported on RX 7000 cards and the ROG Ally under Windows.
- Ultra Performance about 30% faster at the same image quality, plus an optional faster Performance mode.
- Faster at several other common resolutions, with the same image quality.
- Clear message on Vega / Radeon VII. These GPUs are not supported in this version. Instead of a shaking low-quality picture, HelixSR now says so in
helixsr.logand leaves the network off. - Re-run the setup after updating (see Upgrading).
Faster Ultra Performance (network resolution)
New setting in helixsr.ini:
[Upscaling]
NetworkResolution = auto- auto (default): in Ultra Performance the network outputs twice the game's render size, and a final pass scales that to your screen. In our tests (compared with NVIDIA's Model E at full size) the image quality is the same, and it costs about 30% less GPU time.
- fast: also does this for Performance (network at 1.5x the render size). Faster, a little softer (foliage most).
- full: always the full output size, as in 1.2.0.
It is used only with render-resolution motion vectors (most games).
| Output | Mode | 1.2.0 | 1.3.0 |
|---|---|---|---|
| 3840x2160 | Ultra Performance | 3.40 ms | 2.16 ms |
| 2560x1440 | Ultra Performance | 1.63 ms | 1.15 ms |
| 1920x1080 | Ultra Performance | 1.09 ms | 0.75 ms |
| 3840x2160 | Performance, fast
| 3.50 ms | 2.71 ms |
| 2560x1440 | Performance, fast
| 1.67 ms | 1.35 ms |
| 1920x1080 | Performance, fast
| 1.11 ms | 0.94 ms |
Wave size (auto)
New setting in helixsr.ini:
[Compatibility]
WaveSize = autoauto (default), 32 or 64.
- auto on RDNA 1 and 2 (Steam Deck, RX 6000, BC-250): wave32, as before.
- auto on RDNA 3, 3.5 and 4 (RX 7000, RX 9000, ROG Ally / Z1, Radeon 780M / 890M, Strix): the network shaders run at wave64. HelixSR's own convolution kernels stay at wave32. The setup now builds both versions.
Why: with wave32, RDNA 3 under Windows showed a blurry, ghosting image. A community wave64 test build fixed it.
In our tests wave64 gives exactly the same output as wave32 (bit-identical) and is about 5% slower. The community wave64 build was 12-22% slower; ours reads lanes through shared memory (LDS), which is cheaper.
helixsr.log shows the choice in a line starting with wave size:.
Faster
Measured on the BC-250 at 2000 MHz, GPU time of the upscaling pass, same image quality:
| Output | Render size | 1.2.0 | 1.3.0 |
|---|---|---|---|
| 1920x1080 Quality | 1280x720 | 1.153 ms | 1.129 ms |
| 1920x804 Quality | 0.969 ms | 0.926 ms | |
| 2560x1071 Quality | 1.431 ms | 1.411 ms | |
| 2560x1440 | 1707x960 | 2.48 ms | 1.73 ms |
| 1920x1080 | 1280x721 | 1.56 ms | 1.16 ms |
| 3840x2160 | 2560x1441 | 5.45 ms | 3.70 ms |
- Two small network layers now split their work across more of the GPU (first three rows). 4K is unchanged.
- Render sizes just under a 1.5x ratio (a game rounding the render size up, e.g. 1707x960 for 1440p) now use the faster output kernel. Its output was checked equal to native Model E.
Not supported (Vega / Radeon VII)
Radeon Vega GPUs (Vega 56/64, Radeon VII and the Vega graphics in Ryzen APUs) have no wave32, which HelixSR still needs. They are not supported in 1.3.0. HelixSR now writes a clear message to helixsr.log and does not start the network, instead of showing the shaking placeholder picture from earlier versions. Vega support may come in a later version; there is no date.
Upgrading (re-run setup)
Run helixsr-setup again after updating, then copy the new helixsr_kernels.pak (and helixsr_weights.bin) next to the DLL in each game, as before.
The kernel file format changed. With a file from 1.2.0 the DLL still works, but without the new speedup, and it logs a hint to re-run the setup.
Thanks
Thanks to dawoodtogoo2020-code for building and testing a wave64 version on Windows and the ROG Ally X. That test showed that wave64 fixes the RDNA 3 image problem and led to this release's automatic wave size.