RADV Driver Introduces Limited Support For NVIDIA Cooperative Matrix Extension
([Mesa] 3 Hours Ago
VK_NV_cooperative_matrix2)
- Reference: 0001555962
- News link: https://www.phoronix.com/news/RADV-NV-Coperative-Matrix2
- Source link:
The Mesa Radeon Vulkan driver "RADV" has merged support for the VK_NV_cooperative_matrix2 NVIDIA Vulkan extension but it's hidden by default and only partially supported with a focus on helping FidelityFX Super Resolution 4 and VKD3D-Proton.
VK_NV_cooperative_matrix2 is a NVIDIA vendor extension for building on the cooperative matrix types introduced by VK_KHR_cooperative_matrix for expanding beyond just the scope of GEMM kernels. The extension was introduced last October with the [1]Vulkan 1.3.300 spec update.
The RADV driver in Mesa 25.2 now has limited support for the VK_NV_cooperative_matrix2 extension but it's disabled by default. This extension is only exposed when the "radv_cooperative_matrix2_nv" DriConf option is set.
Georg Lehmann who authored the support explained in [2]the merge request for RADV as well as touching common SPIR-V and NIR code within Mesa:
"VK_NV_cooperative_matrix2 has multiple different capabilities/features. This only implements CooperativeMatrixConversionsNV, as it's the most simple and vkd3d-proton wants to use it.
Note that if anyone wants to use this feature, the only operation that reasonably fast (free on gfx12) is ACC -> B conversion. ACC/B -> A conversion and ACC -> B transpose are pretty slow, and sadly the NV extension doesn't allow ACC -> A transpose, which would be fast for us."
After being under review the past month, the code is now merged for Mesa 25.2 with a focus on helping FSR4 and VKD3D-Proton for Steam Play (Proton) Linux gaming.
[1] https://www.phoronix.com/news/Vulkan-1.3.300-Released
[2] https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34793
VK_NV_cooperative_matrix2 is a NVIDIA vendor extension for building on the cooperative matrix types introduced by VK_KHR_cooperative_matrix for expanding beyond just the scope of GEMM kernels. The extension was introduced last October with the [1]Vulkan 1.3.300 spec update.
The RADV driver in Mesa 25.2 now has limited support for the VK_NV_cooperative_matrix2 extension but it's disabled by default. This extension is only exposed when the "radv_cooperative_matrix2_nv" DriConf option is set.
Georg Lehmann who authored the support explained in [2]the merge request for RADV as well as touching common SPIR-V and NIR code within Mesa:
"VK_NV_cooperative_matrix2 has multiple different capabilities/features. This only implements CooperativeMatrixConversionsNV, as it's the most simple and vkd3d-proton wants to use it.
Note that if anyone wants to use this feature, the only operation that reasonably fast (free on gfx12) is ACC -> B conversion. ACC/B -> A conversion and ACC -> B transpose are pretty slow, and sadly the NV extension doesn't allow ACC -> A transpose, which would be fast for us."
After being under review the past month, the code is now merged for Mesa 25.2 with a focus on helping FSR4 and VKD3D-Proton for Steam Play (Proton) Linux gaming.
[1] https://www.phoronix.com/news/Vulkan-1.3.300-Released
[2] https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/34793
phoronix