https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102544
Bug ID: 102544
Summary: GCN offloading not working for
'amdgcn-amd-amdhsa--gfx906:sramecc+:xnack-'
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Keywords: openacc, openmp
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, caplanr at predsci dot com,
jules at gcc dot gnu.org
Target Milestone: ---
Target: GCN
Ron Caplan (in CC) is always eager to try his computational codes with GCC (in
addition to other/proprietary/vendor-provided compilers), and I very much
appreciate him providing feedback of any kind, which I find is always
constructive.
He's recently reported that "using Ubuntu 20.04 with the latest rocm and a
Radeon VII", they got GCN offloading "to compile but when we run it we get:
'libgomp: target function wasn't mapped'".
He sent me 'rocm-smi' and 'rocminfo' output, and comparing his vs. that of our
amd_ryzen3 system, I find the following differences:
-/+ Agent [...]
*******
Name: gfx906
-/+ Uuid: [...]
- Marketing Name: Vega 20
+ Marketing Name: Vega 20 [Radeon VII]
Vendor Name: AMD
[...]
- Node: 1
+ Node: 2
Device Type: GPU
Cache Info:
L1: 16(0x10) KB
+ L2: 8192(0x2000) KB
Chip ID: 26287(0x66af)
Cacheline Size: 64(0x40)
-/+ Max Clock Freq. (MHz): [...]
- BDFID: 2816
+ BDFID: 33536
- Internal Node ID: 1
+ Internal Node ID: 2
Compute Unit: 60
SIMDs per CU: 4
Shader Engines: 4
@@ -128,7 +188,7 @@ Agent 2
Accessible by all: FALSE
ISA Info:
ISA 1
- Name: amdgcn-amd-amdhsa--gfx906:sramecc-:xnack-
+ Name: amdgcn-amd-amdhsa--gfx906:sramecc+:xnack-
The very last item, I suppose, is the crucial one: 'sramecc-' vs. 'sramecc+' --
Andrew?
(a) How do we improve upon the run-time 'libgomp: target function wasn't
mapped' error reporting mode?
(b) What needs to be done to make this work?
