OffTarget

CCE 20.0.2

94 observed issues; 2 regressions from the previous version, marked below. Most serious first: wrong results, then issues whose legality is settled, then regressions, then how widely an issue fails.

See them in the issue list

Wrong answers (35)

ftn (OpenACC): some of 200 integer(8) sum reductions on one parallel loop are wrong Open · regression
A declare target variable listed in shared on target teams loop is read with its host value Open
A firstprivate scalar allocated with omp_const_mem_alloc has the wrong value on the device Open
A sequential DO loop executed by one thread of a target parallel region runs only some of its iterations (-O2) Open
An array private to target teams distribute and shared by a nested parallel do loses the parallel region's writes Open
Lastprivate on distribute in a target region loses the final value, even with an explicit map Open
Lastprivate on target teams distribute parallel do simd collapse(2) with num_threads gives a varying wrong value Open
Lastprivate on target teams distribute parallel for does not imply map(tofrom), value lost Open
Schedule(static, 4) and schedule(static) on a parallel do in a target region deal iterations out one at a time Open
craycc: dist_schedule(static, 4) on target teams distribute parallel for puts every iteration on team 0 Open
craycc: map(delete:) on target exit data acts as release for an array entered twice (still present) Open
craycc: parallel for schedule(static) in a target region runs 1 of 64 iterations Open
craycc: sections in a target parallel region is accepted without a message and gives a wrong result Open
ftn -h acc: a loop seq construct inside a routine seq procedure gives wrong sums Open
ftn -h acc: a module scalar in declare create is treated as firstprivate in a serial region (write lost) Open
ftn: lastprivate on a parallel do in a target region gives a wrong final value when the variable is firstprivate in the target region (not mapped) Open
A lastprivate item smaller than 8 bytes on target teams distribute parallel for comes back as 0 or a garbage value Open
Lastprivate on target teams distribute parallel for accepted silently, value lost Open
SELECT TYPE in a device procedure takes the CLASS DEFAULT branch for mapped objects Open
Two host threads mapping disjoint halves of an allocatable array lose one half Open
a component of an array element mapped tofrom does not come back with separate memory Open · regression
ftn at -O2: a target region's write to a VALUE dummy, firstprivate there, becomes visible to the host Open
ftn: lastprivate(conditional:) on target simd does not imply map(tofrom), value lost Open
ftn: target update from(b(lo:hi)) inside a host parallel region moves no data in copy mode (HSA_XNACK=0) Open
target update of a strided array section (C) copies the contiguous range instead of every second element Open
A Fortran character scalar referenced in a target region is treated as a firstprivate scalar (writes lost) Fixed in 21.0.3
A reduction on target teams distribute parallel do with num_teams(65536) gives a wrong sum (8192 teams is correct) Fixed in 21.0.3
map(to: x) with map(tofrom: x%a) loses the component's copy back Open
the second of two member sections of a structure is not copied back Open
use_device_ptr does not convert a C_PTR list item: the region sees the host address Open
A minimized program with defaultmap and teams distribute computes checksum 0 on the device (fuzz seed 85) Mixed evidence
ftn: a target construct with in_reduction in an internal subroutine loses the writes to an implicitly mapped dummy array Open
With a defaultmap clause on target, an explicitly mapped map(tofrom) scalar is not copied back, and with defaultmap(tofrom) an array in map(alloc) or map(to) is copied back Open · regression
amdclang on MI300A: free() in a kernel appears not to release the stores to the block before it can be reused, so records written through malloc'd temporaries come out wrong (HSA_XNACK=1) Fixed in 21.0.3
ftn in copy mode: the map list item after a map clause with a mapper modifier is not copied back Fixed in 21.0.3

Crashes, hangs and errors at run time (18)

Compiler crashes and hangs (20)

Legal code refused at compile or link time (21)

Fixed or changed in this version

An array mapped always, from next to defaultmap(firstprivate: aggregate) is not copied back (fuzz seed 1978) Fixed in 20.0.2
An array mapped tofrom next to defaultmap(firstprivate: aggregate) is not copied back (fuzz seed 98) Fixed in 20.0.2
craycc 19: with 63 or more int64 sum reductions on target teams distribute parallel for, one is wrong Fixed in 20.0.2
ftn 19 (OpenACC): a two-dimensional automatic array in copyout crashes the run in copy mode Fixed in 20.0.2
ftn 19 does not finish compiling target regions with map(always, tofrom:) of a structure component and a depend on a declare target array Fixed in 20.0.2
craycc rejects the allocator modifier of the allocate clause Fixed in 20.0.2
ftn 19: internal compiler error (invalid addrspacecast) on reduction(task) Changed in 20.0.2