16-bit Medium Precision Improvements Merged For AMD Radeon Mesa Code
([Mesa] 5 Hours Ago
RadeonSI mediump)
- Reference: 0001553116
- News link: https://www.phoronix.com/news/RadeonSI-Mesa-25.2-mediump
- Source link:
Prominent AMD Mesa developer Marek Olšák this week landed a number of fixes for the RadeonSI Gallium3D driver code in working to enable medium precision "mediump" support for this open-source graphics driver.
The 16-bit medium precision work for the RadeonSI driver isn't completely wrapped up yet but it's being worked on and Marek remarked that " it seems like it might happen ." With a set of 13 patches merged to Mesa 25.3 this week, there have been several mediump and 16-bit IO fixes. Marek explained in [1]that merge request :
"These are fixes and improvements for issues I discovered while trying to enable mediump IO for radeonsi.
#13298 is required for mediump XFB lowered to 16 bits, but this MR adds shader_info fields that allow not lowering mediump for certain cases of linked shaders like XFB, which allows drivers to support mediump IO for non-XFB cases of linked shaders while all other cases keep mediump IO as 32 bits. With that, mediump IO can be enabled only for a subset of cases of linked shaders that are known to work.
One commit also contains a workaround for RADV inside nir_opt_vectorize_io. I think this has been a known issue."
Those medium fixes were merged for the open-source AMD Radeon driver code. In a Mesa [2]issue , Marek also commented:
"While trying to enable mediump IO for radeonsi (and it seems like it might happen), I was able to fix a lot of issues with mediump in IO passes.
One of the remaining issues to fix is in gl_nir_lower_packed_varyings because of how mediump XFB works."
We'll see if RadeonSI medium precision support manages to wrapped up in time for the Mesa 25.2 release next quarter.
[1] https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35315
[2] https://gitlab.freedesktop.org/mesa/mesa/-/issues/13298
The 16-bit medium precision work for the RadeonSI driver isn't completely wrapped up yet but it's being worked on and Marek remarked that " it seems like it might happen ." With a set of 13 patches merged to Mesa 25.3 this week, there have been several mediump and 16-bit IO fixes. Marek explained in [1]that merge request :
"These are fixes and improvements for issues I discovered while trying to enable mediump IO for radeonsi.
#13298 is required for mediump XFB lowered to 16 bits, but this MR adds shader_info fields that allow not lowering mediump for certain cases of linked shaders like XFB, which allows drivers to support mediump IO for non-XFB cases of linked shaders while all other cases keep mediump IO as 32 bits. With that, mediump IO can be enabled only for a subset of cases of linked shaders that are known to work.
One commit also contains a workaround for RADV inside nir_opt_vectorize_io. I think this has been a known issue."
Those medium fixes were merged for the open-source AMD Radeon driver code. In a Mesa [2]issue , Marek also commented:
"While trying to enable mediump IO for radeonsi (and it seems like it might happen), I was able to fix a lot of issues with mediump in IO passes.
One of the remaining issues to fix is in gl_nir_lower_packed_varyings because of how mediump XFB works."
We'll see if RadeonSI medium precision support manages to wrapped up in time for the Mesa 25.2 release next quarter.
[1] https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/35315
[2] https://gitlab.freedesktop.org/mesa/mesa/-/issues/13298
phoronix