@@ -711,7 +711,6 @@ struct AMDGPUDeviceImageTy : public DeviceImageTy {
711
711
private:
712
712
// / The executable loaded on the agent.
713
713
hsa_executable_t Executable;
714
- hsa_code_object_t CodeObject;
715
714
#if SANITIZER_AMDGPU
716
715
hsa_code_object_reader_t CodeObjectReader;
717
716
#endif
@@ -1709,8 +1708,8 @@ struct AMDGPUStreamTy {
1709
1708
1710
1709
// / Create an empty slot.
1711
1710
StreamSlotTy ()
1712
- : Signal(nullptr ), Callbacks({}), ActionArgs({} ),
1713
- OmptActionFunction ( nullptr ) {}
1711
+ : Signal(nullptr ), Callbacks({}), OmptActionFunction( nullptr ),
1712
+ ActionArgs ({} ) {}
1714
1713
1715
1714
// / Schedule a host memory copy action on the slot.
1716
1715
Error schedHostMemoryCopy (void *Dst, const void *Src, size_t Size) {
@@ -2025,8 +2024,8 @@ struct AMDGPUStreamTy {
2025
2024
assert (Args->Signal &&
2026
2025
" Invalid AMDGPUSignal Pointer in post kernel run processing" );
2027
2026
hsa_amd_profiling_dispatch_time_t TimeRec;
2028
- hsa_status_t Status = hsa_amd_profiling_get_dispatch_time (
2029
- Args-> Agent , Args-> Signal -> get (), &TimeRec);
2027
+ hsa_amd_profiling_get_dispatch_time (Args-> Agent , Args-> Signal -> get (),
2028
+ &TimeRec);
2030
2029
2031
2030
uint64_t StartTime = TimeRec.start * Args->TicksToTime ;
2032
2031
uint64_t EndTime = TimeRec.end * Args->TicksToTime ;
@@ -2913,7 +2912,7 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
2913
2912
" OMPX_ENABLE_GFX90A_COARSE_GRAIN_SHARED_ALLOC" , false ),
2914
2913
OMPX_StrictSanityChecks (" OMPX_STRICT_SANITY_CHECKS" , false ),
2915
2914
OMPX_SyncCopyBack (" LIBOMPTARGET_SYNC_COPY_BACK" , true ),
2916
- OMPX_APUPrefaultMemcopy (" LIBOMPTARGET_APU_PREFAULT_MEMCOPY" , " true" ),
2915
+ OMPX_APUPrefaultMemcopy (" LIBOMPTARGET_APU_PREFAULT_MEMCOPY" , true ),
2917
2916
OMPX_APUPrefaultMemcopySize (" LIBOMPTARGET_APU_PREFAULT_MEMCOPY_SIZE" ,
2918
2917
1 * 1024 * 1024 ), // 1MB
2919
2918
OMPX_DGPUMaps (" OMPX_DGPU_MAPS" , false ),
@@ -3886,6 +3885,9 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
3886
3885
case HSA_DEVICE_TYPE_DSP:
3887
3886
TmpCharPtr = " DSP" ;
3888
3887
break ;
3888
+ case HSA_DEVICE_TYPE_AIE:
3889
+ TmpCharPtr = " AIE" ;
3890
+ break ;
3889
3891
}
3890
3892
Info.add (" Device Type" , TmpCharPtr);
3891
3893
}
@@ -5060,7 +5062,6 @@ Error AMDGPUKernelTy::launchImpl(GenericDeviceTy &GenericDevice,
5060
5062
if (LaunchParams.Size )
5061
5063
std::memcpy (AllArgs, LaunchParams.Data , LaunchParams.Size );
5062
5064
5063
- uint64_t Buffer = 0 ;
5064
5065
AMDGPUDeviceTy &AMDGPUDevice = static_cast <AMDGPUDeviceTy &>(GenericDevice);
5065
5066
AMDGPUStreamTy *Stream = nullptr ;
5066
5067
if (auto Err = AMDGPUDevice.getStream (AsyncInfoWrapper, Stream))
@@ -5111,7 +5112,7 @@ void AMDGPUKernelTy::printAMDOneLineKernelTrace(GenericDeviceTy &GenericDevice,
5111
5112
// This line should print exactly as the one in the old plugin.
5112
5113
fprintf (
5113
5114
stderr,
5114
- " DEVID: %2d SGN:%d ConstWGSize:%-4d args:%2d teamsXthrds:(%4luX %4d) "
5115
+ " DEVID: %2d SGN:%d ConstWGSize:%-4d args:%2d teamsXthrds:(%4uX %4d) "
5115
5116
" reqd:(%4dX%4d) lds_usage:%uB sgpr_count:%u vgpr_count:%u agpr_count:%u "
5116
5117
" sgpr_spill_count:%u vgpr_spill_count:%u tripcount:%lu rpc:%d "
5117
5118
" md:%d md_LB:%ld md_UB:%ld Max Occupancy: %u Achieved Occupancy: "
@@ -5304,8 +5305,8 @@ static std::pair<uint64_t, uint64_t>
5304
5305
getKernelStartAndEndTime (const OmptKernelTimingArgsAsyncTy *Args) {
5305
5306
assert (Args->Signal && " Invalid AMDGPUSignal Pointer in OMPT profiling" );
5306
5307
hsa_amd_profiling_dispatch_time_t TimeRec;
5307
- hsa_status_t Status = hsa_amd_profiling_get_dispatch_time (
5308
- Args-> Agent , Args-> Signal -> get (), &TimeRec);
5308
+ hsa_amd_profiling_get_dispatch_time (Args-> Agent , Args-> Signal -> get (),
5309
+ &TimeRec);
5309
5310
5310
5311
uint64_t StartTime = TimeRec.start * Args->TicksToTime ;
5311
5312
uint64_t EndTime = TimeRec.end * Args->TicksToTime ;
@@ -5317,8 +5318,7 @@ static std::pair<uint64_t, uint64_t>
5317
5318
getCopyStartAndEndTime (const OmptKernelTimingArgsAsyncTy *Args) {
5318
5319
assert (Args->Signal && " Invalid AMDGPUSignal Pointer in OMPT profiling" );
5319
5320
hsa_amd_profiling_async_copy_time_t TimeRec;
5320
- hsa_status_t Status =
5321
- hsa_amd_profiling_get_async_copy_time (Args->Signal ->get (), &TimeRec);
5321
+ hsa_amd_profiling_get_async_copy_time (Args->Signal ->get (), &TimeRec);
5322
5322
uint64_t StartTime = TimeRec.start * Args->TicksToTime ;
5323
5323
uint64_t EndTime = TimeRec.end * Args->TicksToTime ;
5324
5324
0 commit comments