Intel Iris Gallium3D Driver Adds Compute Kernel Support In Mesa 20.3
While Mesa 20.2 isn't even releasing for a few weeks, Mesa 20.3 is already seeing new feature work that will debut next quarter.
Intel's Jason Ekstrand has landed a set of patches for handling of kernels within Iris, Intel's modern Gallium3D driver. He commented, "This MR contains most of the patches required to handle kernels in iris. I've had them lying around in a branch in some form or another for a while. We should upstream what we can."
In particular this is about handling the MESA_SHADER_KERNEL type. This is what was added to Mesa back in early 2019 by Karol Herbst of Red Hat as part of their OpenCL work on Nouveau. MESA_SHADER_KERNEL is for representing OpenCL kernels.
As part of this kernel bring-up is also related work like handling of 64-bit intrinsics, support for serialized NIR, and supporting the on-disk shader cache with these kernels. The actual bring-up isn't too much given that OpenGL compute shaders are already supported and all of the common work in place with the NIR compiler back-end.
Interesting work although for those focused on GPU computing the Intel Compute Runtime stack already provides open-source OpenCL 2.1 support (and OpenCL 3.0 already for Gen12, will be enabled for existing hardware when CL 3.0 is ratified) along with Intel oneAPI Level Zero support.
Intel's Jason Ekstrand has landed a set of patches for handling of kernels within Iris, Intel's modern Gallium3D driver. He commented, "This MR contains most of the patches required to handle kernels in iris. I've had them lying around in a branch in some form or another for a while. We should upstream what we can."
In particular this is about handling the MESA_SHADER_KERNEL type. This is what was added to Mesa back in early 2019 by Karol Herbst of Red Hat as part of their OpenCL work on Nouveau. MESA_SHADER_KERNEL is for representing OpenCL kernels.
As part of this kernel bring-up is also related work like handling of 64-bit intrinsics, support for serialized NIR, and supporting the on-disk shader cache with these kernels. The actual bring-up isn't too much given that OpenGL compute shaders are already supported and all of the common work in place with the NIR compiler back-end.
Interesting work although for those focused on GPU computing the Intel Compute Runtime stack already provides open-source OpenCL 2.1 support (and OpenCL 3.0 already for Gen12, will be enabled for existing hardware when CL 3.0 is ratified) along with Intel oneAPI Level Zero support.
Add A Comment