Mesa 21.2 Begins Seeing Intel Xe-HP Graphics Driver Changes
Xe-HP is Intel's forthcoming data center / high performance Xe GPU. Xe-HP is said to already be sampling with customers while the formal launch is anticipated for later in the calendar year. On the Linux kernel side, Intel discrete graphics are nearing a working state at least for DG1 and thus we are beginning to see more work flow into the user-space driver code now with Mesa and presumably their Compute Runtime / IGC stack soon as well.
What hit Mesa 21.2-devel this morning is the initial back of graphics compiler changes for supporting the Xe-HP ISA. The merge request offers some convenient insight into some of the Xe-HP changes from the architecture perspective compared to existing Gen12 Xe, "the most invasive compiler changes relative to previous generations are the result of the preexisting FPU pipeline being split into multiple asynchronous pipelines (a floating-point, integer and long AKA double-precision pipeline), which is highly visible to software because the hardware is not able to guarantee data coherency across instructions (already since TGL), so the compiler is now responsible for keeping track of which pipeline will be executing which instruction and specifying synchronization primitives in order to resolve any cross-pipeline dependencies."
The Xe-HP graphics compiler support on top of the existing Xe support is at the moment just over 700 lines of code changed. Obviously the shader compiler support is necessary before the Intel open-source engineers can move on with the rest of the OpenGL and Vulkan bring-up. We are also tracking work on the Compute Runtime / IGC side for their OpenCL and oneAPI Level Zero bring-up to come for Xe-HP.
This initial Xe-HP compiler support is now in Mesa Git and expect more Intel Xe-HP enablement to continue this cycle.