The problem -muniform-simt is trying to address is to make sure that a register produced outside an openmp simd region is available when used in a lane inside an simd region. The solution is to, outside an simd region, execute in all warp lanes, thus producing consistent values in result registers in each warp thread. [ Note that this solution is: as-produced, asap. Openacc has the same problem, but deals with it: as-needed, alap. ] This approach doesn't work when executing in all warp lanes multiplies the side effects from 1 to 32 separate side effects, which is the case for for instance atomic insns. So atomic insns are rewritten to execute only in the master lane, and if there are any results, propagate those to the other threads in the warp. [ And likewise for system calls malloc, free, vprintf. ] [ The corresponding reorg pass nvptx_reorg_uniform_simt potentially rewrites all statements, be those inside or outside an simd region. But care is taken that the rewrite only has effect outside the simd region. ] Now, take a non-atomic update: ld, add, store. The store has side effects, are those multiplied as well? Now, pre-sm_70 we have the guarantee that warps execute in lock step. So: - the load will load the same value into the result register across the warp, - the add will write the same value into the result register across the warp, - the store will write the same value to the same memory location, 32 times, at once, having the result of a single store. So, no side-effect multiplication (well, at least that's the observation). Starting sm_70, the threads in a warp are no longer guaranteed to execute in lockstep. Consequently, we can have the following execution trace: - some threads load a value into the result register - those threads do an add and write the result into the result register - that result is stored - the other threads arrive, and now load the now updated, thus different value into the result register - the other threads do an add and write a different result into their result register - the updated result is stored So, we both have now the side effect multiplied, and the registers are no longer in sync.
We could try the same solution as for atomic: predicate ld/st to only execute in lane 0, and propagate ld result. Another solution might be to wrap each ld/st in two bar.warp.sync.
Created attachment 52629 [details] Attempt, runs into driver internal error FTR, this is an attempt at a fix. It does the "predicate ld/st to only execute in lane 0, and propagate ld result", expect for modes where we run into a problem with producing a shuffle (HFmode, V2SImode), where we use the "wrap each ld/st in two bar.warp.sync" approach instead. For libgomp/testsuite/libgomp.c++/target_teams_distribute__parallel_for__simd.C (copied from OvO) I end up with: ... Linking Link complete: 0.000000ms Link log warning : Stack size for entry function '_Z48test_target_teams_distribute__parallel_for__simdv$_omp_fn$0' cannot be statically determined libgomp: Link error log fatal : Internal error: reference to deleted section ... This may be due to a problem with the patch or the driver, I'm not sure.
Anyway, having reread the volta architecture whitepaper again, I think it's ok to use the solution I already found that does work (see PR104783): add a warp sync at simt exit. The tricky bit is that we rely on the warps to stay uniform after that warp sync until the next simt region entry. Which AFAICT is not something enforced by specification, but rather by a assuming that the 'Convergence optimizer' has a sane implementation.
The master branch has been updated by Tom de Vries <vries@gcc.gnu.org>: https://gcc.gnu.org/g:a624388b9546b066250be8baa118b7d50c403c25 commit r12-7765-ga624388b9546b066250be8baa118b7d50c403c25 Author: Tom de Vries <tdevries@suse.de> Date: Wed Mar 9 10:35:14 2022 +0100 [nvptx] Add warp sync at simt exit Consider this code (with N defined to 1024): ... float v = 0.0; #pragma omp target map(tofrom: v) #pragma omp parallel for simd for (int i = 0 ; i < N; i++) { #pragma omp atomic update v = v + 1.0; } ... It hangs when executing on target board unix/-foffload=-misa=sm_75, using drivers 470.103.01 and 510.54 on a T400 board (sm_75). I'm tentatively identifying the problem as a bug in -muniform-simt for architectures that support Independent Thread Scheduling (sm_70 and later). The problem -muniform-simt is trying to address is to make sure that a register produced outside an openmp simd region is available when used in any lane inside an simd region. The solution is to, outside an simd region, execute in all warp lanes, thus producing consistent values in result registers in each warp thread. This approach doesn't work when executing in all warp lanes multiplies the side effects from 1 to 32 separate side effects, which is the case for atomic insns. So atomic insns are rewritten to execute only in lane 0, and if there are any results, those are propagated to the other threads in the warp. [ And likewise for system calls malloc, free, vprintf. ] Now, consider a non-atomic update: ld, add, store. The store has side effects, are those multiplied or not? Pre-sm_70 we can assume that at the end of an SIMT region, any divergent control flow has reconverged, and we have a uniform warp, executing in lock step. So: - the load will load the same value into the result register across the warp, - the add will write the same value into the result register across the warp, - the store will write the same value to the same memory location, 32 times, at once, having the result of a single store. So, no side-effect multiplication (well, at least that's the observation). Starting sm_70, the threads in a warp are no longer guaranteed to reconverge after divergence. There's a "Convergence Optimizer" that can can identify that it is safe for a warp to reconverge, but that works only as long as the code does not contain "synchronizing operations". Consequently, the ld, add, store sequence can be executed by a non-uniform warp, which means the side effects can have multiplied, and the registers are no longer guarantueed to be in sync. The atomic update in the example above is translated using an atom.cas loop, which means that we have divergence (because only one thread is allowed to succeed at a time) and the "Convergence Optimizer" doesn't reconverge probably because the atom.cas counts as a "synchronizing operation". So, it seems plausible that the root cause for the mentioned hang is the problem described above. Fix this by adding an explicit warp sync at simt exit. Note that we're assuming here that the warp will stay uniform until the next SIMT region entry. Tested on x86_64 with nvptx accelerator. gcc/ChangeLog: 2022-03-09 Tom de Vries <tdevries@suse.de> PR target/104916 PR target/104783 * config/nvptx/nvptx.md (define_expand "omp_simt_exit"): Emit warp sync (or uniform warp check for mptx < 6.0). libgomp/ChangeLog: 2022-03-15 Tom de Vries <tdevries@suse.de> PR target/104916 PR target/104783 * testsuite/libgomp.c/pr104783-2.c: New test.
Fixed.