RADV Driver Lands VK_EXT_shader_float8 Support
The Mesa Radeon Vulkan driver "RADV" is now the first Mesa in-tree driver supporting 8-bit floating point use within shaders via the new VK_EXT_shader_float8 extension.
VK_EXT_shader_float8 was introduced earlier this month with the Vulkan 1.4.317 spec update. VK_EXT_shader_float8 enables support for 8-bit floating point operations within shaders in order to support more AI / machine learning workloads with the Vulkan API.
With the open-source RADV driver, the 8-bit floating point support works with AMD RDNA4 / GFX12 GPUs and future hardware. This merge request is what wired up the VK_EXT_shader_float8 support within the RADV driver as well as associated SPIR-V and NIR plumbing.
This 8-bit FP support for the RADV driver will be part of next quarter's Mesa 25.2 release for helping out Vulkan with AI/ML software.
VK_EXT_shader_float8 was introduced earlier this month with the Vulkan 1.4.317 spec update. VK_EXT_shader_float8 enables support for 8-bit floating point operations within shaders in order to support more AI / machine learning workloads with the Vulkan API.
With the open-source RADV driver, the 8-bit floating point support works with AMD RDNA4 / GFX12 GPUs and future hardware. This merge request is what wired up the VK_EXT_shader_float8 support within the RADV driver as well as associated SPIR-V and NIR plumbing.
"This adds support for E4M3FN (4bit exponent, 3bit mantissa, no infinity) and E5M2 (5bit exponent, 2bit mantissa) float formats in vtn, nir and aco.
Only conversions and cmat muladd are supported. Like for bfloat16, all conversions go through fp32."
This 8-bit FP support for the RADV driver will be part of next quarter's Mesa 25.2 release for helping out Vulkan with AI/ML software.