Panfrost Gallium3D Driver Enables Working FP16 Support
The Panfrost Gallium3D driver providing open-source OpenGL driver support for Arm Mali Midgard and Bifrost hardware has another feature tacked on as of Friday night.
The Panfrost driver has been working steadily on supporting new driver functionality in the OpenGL space, especially improving the newer Bifrost support, albeit not yet any open-source Vulkan driver for Arm Mali hardware at this point.
The latest achievement for Panfrost is FP16 support. Half-precision floating-point support is now working for Panfrost and the change also enables lowering of medium temporaries to 16-bit and further use of 16-bit throughout the NIR intermediate representation. Previously the FP16 code path was disabled by default as it was "buggy experimental (don't use!) fp16." But now it's enabled at least to encourage further testing and hoping no nasty issues come up. The PAN_MESA_DEBUG=nofp16 option was added for disabling this FP16 code-path.
The FP16 support and other recent Panfrost improvements will be part of Mesa 20.2 due out in just over one month's time.
The Panfrost driver has been working steadily on supporting new driver functionality in the OpenGL space, especially improving the newer Bifrost support, albeit not yet any open-source Vulkan driver for Arm Mali hardware at this point.
The latest achievement for Panfrost is FP16 support. Half-precision floating-point support is now working for Panfrost and the change also enables lowering of medium temporaries to 16-bit and further use of 16-bit throughout the NIR intermediate representation. Previously the FP16 code path was disabled by default as it was "buggy experimental (don't use!) fp16." But now it's enabled at least to encourage further testing and hoping no nasty issues come up. The PAN_MESA_DEBUG=nofp16 option was added for disabling this FP16 code-path.
The FP16 support and other recent Panfrost improvements will be part of Mesa 20.2 due out in just over one month's time.
2 Comments