https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105421
Bug ID: 105421
Summary: GCN offloading, raised '-mgang-private-size':
'HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION'
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Keywords: openacc
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: tschwinge at gcc dot gnu.org
CC: ams at gcc dot gnu.org, jules at gcc dot gnu.org
Target Milestone: ---
Target: GCN
As discussed in
<http://mid.mail-archive.com/87fslzspgo.fsf@dem-tschwing-1.ger.mentorg.com>,
after commit r12-8252-gb2202431910e30d8505c94d1cb9341cac7080d10 "fortran: Fix
up gfc_trans_oacc_construct [PR104717]", we need commit
r12-8311-g2a570f11a2fecf23998d7fe1d5cabad62cfe5cec "Fix up
'libgomp.oacc-fortran/print-1.f90' GCN offloading compilation [PR104717]",
however:
> That only works with the default GCN multilib '-march=fiji', testing
> on gfx803 amdfury2 system. For all of '-march=gfx900' (amdnano2),
> '-march=gfx906' (amd_ryzen3), '-march=gfx908' (amd-instinct1), I get:
>
> libgomp: GCN fatal error: Asynchronous queue error
> Runtime message: HSA_STATUS_ERROR_MEMORY_APERTURE_VIOLATION: The agent
> attempted to access memory beyond the largest legal address.
>
> ..., and I still get that if lowering the allocation to the minimum,
> '-foffload-options=amdgcn-amdhsa=-mgang-private-size=560'.
>
> This is a really simple OpenACC 'parallel' construct:
>
> !$acc parallel
> write (0, '("The answer is ", I2)') var
> !$acc end parallel
>
> ..., which ought to launch a 1-gang x 1-worker x 1-vector GPU kernel, so
> I'd assume '-mgang-private-size=560' (or '-mgang-private-size=13579' in
> fact) is not a problem?
