16-bit Support Being Worked On For Intel ANV, Gallium3D TGSI
Marek Olšák of AMD has proposed adding TGSI 16-bit support, with TGSI being the intermediate representation (IR) used by Gallium3D. That's evolved into a lengthy discussion by upstream developers for supporting 16-bit floats and integers within TGSI. No patches yet, but that will probably be coming soon.
Meanwhile, Alejandro Piñeiro of Igalia has pushed his latest patches for supporting Vulkan's 16-bit extensions in the Intel ANV driver. Those extensions are VK_KHR_16bit_storage for Vulkan and the associated SPV_KHR_16bit_storage extension in SPIR-V. These extensions will work with Intel Broadwell "Gen 8" graphics hardware and newer. The Intel ANV 16-bit support comes in at 47 patches at just over one thousand new lines of code.
VK_KHR_16bit_storage was introduced this summer with Vulkan 1.0.54 and this half-float push is partially driven by lessening memory bandwidth requirements if not needing 32-bit or 64-bit precision/range.
1 Comment
