On Ubuntu 20.04, in Release mode, gcc fails with this error:
In file included from /usr/include/string.h:495,
from /opt/rocm/include/hsa/hsa_api_trace.h:57,
from ../rocprofiler/src/util/hsa_rsrc_factory.h:29,
from ../rocprofiler/src/util/hsa_rsrc_factory.cpp:25:
In function ‘char* strncpy(char*, const char*, size_t)’,
inlined from ‘const util::AgentInfo* util::HsaRsrcFactory::AddAgentInfo(hsa_agent_t)’ at ../rocprofiler/src/util/hsa_rsrc_factory.cpp:323:12:
/usr/include/x86_64-linux-gnu/bits/string_fortified.h:106:34: error: ‘char* __builtin___strncpy_chk(char*, const char*, long unsigned int, long unsigned int)’ specified bound depends on the length of the source argument [-Werror=stringop-overflow=]
106 | return __builtin___strncpy_chk (__dest, __src, __len, __bos (__dest));
| ~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
../rocprofiler/src/util/hsa_rsrc_factory.cpp: In member function ‘const util::AgentInfo* util::HsaRsrcFactory::AddAgentInfo(hsa_agent_t)’:
../rocprofiler/src/util/hsa_rsrc_factory.cpp:322:39: note: length computed here
322 | const int gfxip_label_len = strlen(agent_info->name) - 2;
| ~~~~~~^~~~~~~~~~~~~~~~~~
The error is caused by the following 2 lines:
const int gfxip_label_len = strlen(agent_info->name) - 2;
strncpy(agent_info->gfxip, agent_info->name, gfxip_label_len);
The size argument to strncpy should not depend on the input string.
Since the terminating character is not considered (the copy is at
most len - 2 bytes), using memcpy is preferable. Also, make sure
the destination does not overflow by clamping the size.
Change-Id: I0c5cf7e0daf4cd6fcf7092efb1d9fd4c02a6c639
Concurrent profiling relies on the aqlprofile read_api
and tracker. This patch set those options to enable
the concurrent profiling.
Change-Id: Ib97d4d8facfbc11f2684d83109397cd13f117d5e
This patch adds barrier packets, together with extra signals,
to enforce the completion order of read packets w.r.t dispatch.
And, PmcStopper is added to stop the profiling finally.
Change-Id: I8e8d3a41d86e42be1d9e5afd44c247be876cf1a5
The profiling was only enabled in serial mode, i.e., kernels
are serialized in execution, and counters are reset at each
kernel start and read at kernel completion. This patch adds
the concurrent mode, by issuing the process-level start
packet to reset counters, and then reading twice at kernel
start and end time to obtain the counter value difference.
The new concurrent profiling usage needs the integration
with the corresponding augment at aqlprofile side.
Change-Id: I94b4442eadc8c64b8fba51b1e4916fc8b895ad21
Rachida Kebichi
Laurent Morichetti
Evgeny
Xianwei Zhang