Intel Opens Up "IMF LA" As A GPU Compute Speed Boost To Better Compete With Windows
Released today was IGC 1.0.5761. This routine update to the Intel Graphics Compiler has a number of low-level compiler additions and other changes as usual. All quite low level but then there was the mention of "IMF LA open-sourcing."
What the heck is this "IMF LA"? Hitting the Intel Graphics Compiler were a bunch of commits referencing IMF LA and changing around the common math functions to use the new code... There has been no public documentation on IMF LA but was able to get clarification that it's the Intel Math Functions "Low Accuracy" mode.
The Intel Math Functions Low Accuracy provides enough mathematical precision to meet the requirements mandated by the OpenCL specification but is "low accuracy" compared to the existing implementation. Reducing the precision is done in the name of performance while still abiding by OpenCL requirements around precision.
Switching over to IMF LA the hope is to help close the current performance gap between their Linux and Windows drivers. There still are additional optimizations being pursued but this is part of that work in aiming to make their Windows and Linux drivers more at feature parity. So it sounds like the Windows driver mode up to now was already this IMF "Low Accuracy" mode and now the Linux driver has this same implementation now that it's been open-sourced.
With oneAPI Level Zero and now Intel offering discrete graphics cards for the data center, Intel has been ramping up their open-source Linux graphics driver support particularly around ensuring optimal performance is being achieved. Obviously in cases like this with IMF LA it's not only their data center / discrete GPUs that stand to benefit but all supported Intel graphics hardware. They have made much progress already while looking ahead to 2021 they are still striving for more aggressive performance, a possible GUI control panel, and also employing IGC by the Intel Mesa drivers for juicing out more performance.
Needless to say, I'll be firing up some fresh Intel OpenCL compute benchmarks in the coming days to see exactly how much IMF LA helps out.