Scrying the AMD GFX1250 LLVM Tea Leaves
chipsandcheese.comI'm a heavy user of NVIDIA LOP3 instruction (uint32). I wonder when AMD will finally support it well.
They already do, it's called v_bitop3_b32. Added in CDNA4, present in CDNA5 and will be making its way in RDNA5, if the LLVM code is to be trusted.
https://www.amd.com/content/dam/amd/en/documents/instinct-te...
I'm very interested in your use case.
I looked into bit-slicing techniques, but they increase throughput at the cost of massive latency. This is not acceptable for realtime audio.
Fusing multiple boolean logic functions into one.
Where does someone start on kernel development for an RDNA4? Books, resources, whatever?
8 year developer just getting into to AI side of the house
...not just implementing barriers in hardware but also full monitors, allowing the wave to be notified if a specific cache line is evicted from the L2 cache.
What is this feature for? When do you need it?
The explanation is missing that the hardware also notifies you of modifications, which makes it useful for implementing cross-cluster synchronization primitives.
Compare to MONITOR/MWAIT on x86: https://www.felixcloutier.com/x86/mwait#mwait-for-address-ra...
GPUs are absurdly SMT. Like 8 threads or 10 threads (or tracked instruction pointers) per hardware instruction execution unit (for AMD, the Workgroup Processor, WGP)
I have to imagine that some kind of queue or data structure would benefit from this information. Especially with server GPU tasks staying resident inside of a WGP for literally hours at a time.
sure, but what kind?
wait/arrive seems obvious to me. So I'll talk about that... As a synchronization primitive, its very useful for things like:
I don't know what a "cluster" is, but its clearly some new subdivision of threads (aside from wavefront, group, block, grid, etc. etc.). Because locking is so inefficient on GPUs, you probably want to use barriers in the typical case.parallel_memset(data, 0); // The 1024 threads go off and, in parallel + some efficient manner, initialize the datastructure barrier(); // wait for all 1024 threads to reach here before continuing. do_stuff(); // This probably relies on all default data being set before startingCluster arrive + Cluster wait allow for you to split the barrier into two units. Maybe the first half of a loop must be synchronized, but there's "spare extra work" you can do before calling wait(). This likely makes threading smoother.
This is likely more efficient for a number of algorithms than a singular barrier.parallel_memset(data, 0); // The 1024 threads go off and, in parallel + some efficient manner, initialize the datastructure arrive(); do_spare_work(); // some extra work before checking for a wait wait(); // We ran out of main work + spare work, so now we must fully stop and wait do_stuff(); // This probably relies on all default data being set before starting---------
As far as the L2 cache thing specifically: I don't know.