Intel Vulkan Driver Gets Patches For New KHR_shader_float16_int8 Functionality

Yesterday saw the release of Vulkan 1.1.95 that introduced the new VK_KHR_shader_float18_int8 extension for supporting 16-bit floating-point types and 8-bit integer types within shader code for arithmetic operations, compared to earlier extensions limiting these data types to load/store operations. NVIDIA released a same-day driver update for the new 1.1.95 extensions while now Intel's "ANV" open-source Vulkan code is the second Linux driver seeing this support (or first if just looking at the open-source drivers).

A set of 59 patches dealing with around one thousand lines of code was sent out this morning for wiring up VK_KHR_shader_float16_int8 into the Intel ANV driver. The Float16 and Int8 support in shader code can be handled on Intel Broadwell graphics hardware and newer.

Bringing up this support included changes to their SPIR-V to NIR translator for this intermediate representation used by the ANV driver. Some of the compiler back-end bits are useful too for more than just the Intel Vulkan driver, such as the Intel OpenGL ES support.

The patch series for now can be found on the Mesa-dev list. We'll see if the big patch series gets reviewed in time for the upcoming Mesa 19.0 cycle.

Iago Toral Quiroga of Igalia who spearheaded these driver changes also wrote a blog post going into more technical details about the new extension and driver plumbing for it.