1. Use s_mov_b32 to move 0xcafe to s18. s_movk_i32 is a sign extention move
instruction. Oxcafe will be extended to 0xffffcafe which is not desired
2. Add wait to s_load_dword instruction to make sure memory read finish before
the next store instruction.
Change-Id: I665d1d471019edfaba5693e07cdc567d4103573f
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
IPC was failing due to calling fork when HSA was open. The fix
was correcting incomplete cleanup in several other tests.
TestBase::Close (via CommonCleanUp) now checks that HSA is properly
closed between tests.
rocrtstPerf.Memory_Async_Copy uses hwloc which uses OpenCL which
has no shutdown routine. Consequently this test can not cleanup
properly. I added a hack to force HSA refcount to the value
it should have if OpenCL were cleaning up but this leaks resources
and potentially puts hwloc & OpenCL in a bad state.
OpenCL loads LLVM which installs some exit handlers. Those handlers
can't execute in a child process and can't be removed since OpenCL
doesn't cleanup. IPC hacks around this by aborting rather than exiting
in the child process.
Change-Id: I92326a73d7b11632208717d99728e6dafdc7d3ca
If TTM eviction and restore happens, it may takes very long time if
retry, the longest time is 5 minutes during my test. There is chance
packet is submited to queue while eviction, we have to increase the
Wait4PacketConsumption timeout.
The queue will continue to execute after eviction and restore. If we
upmap the memory from GPU while queue is evicted, this will cause VM
fault. Change to unmap memory after queue is destroyed.
Change-Id: I1b44e2274ea7b83398b2e3293578dad6947cb5af
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Because dma32 zone is on node 0, use all system memory on node 0 will
cause TTM eviction to free dma32 zone for other devices which only
work with 32bit physical address. The TTM eviction and restore may take
too long and cause queue timeout.
Running on other NUMA nodes, the NUMA default memory policy is
MPOL_PREFERRED, means TTM will get pages from local node first, and then
get remaining pages from other nodes. Check /proc/buddyinfo can confirm
this.
Reset NUMA bind to all after the test.
Change-Id: I39b373c07a2d5aa396f5c7602bffabab0481930f
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
CPUClockCounter is not NTP adjusted (CLOCK_MONOTONIC_RAW) so should be
better for measurements. However, it is implemented with syscall while
CLOCK_MONOTONIC is implemented via vDSO. The latency increase becomes
significant when language layers make corresponding clock measurements.
Reverting to CLOCK_MONOTONIC will reduce latency and allow small
duration events to be measured at the cost of incorporating NTP
frequency skew errors. NTP may adjust frequency by 500ppm so limits us
to ~3 decimals in elapsed time.
Change-Id: I920b9f707f47109d80d6c256c475638c03fb8d76
PSDB and other jenkins jobs are currently failing on several kfd tests.
This is blocking user throughput for screening patches by PSDB.
Blacklist multiple tests and submit JIRA's.
KFDIPCTest.BasicTest (ROCMOPS-459) .CMABasicTest (ROCMOPS-460) .CrossMemoryAttachTest (ROCMOPS-461)
KFDMemoryTest.BigBufferStressTest (ROCMOPS-462)
KFDQMTest.MultipleSdmaQueues (ROCMOPS-463) (ROCMOPS-416)
KFDEvictTest.BurstyTest (ROCMOPS-464)
Change-Id: I2c7cdeabc26654f39823201ce86d4113b3a98a0e
Signed-off-by: Cole Nelson <cole.nelson@amd.com>
This relates to the following commits:
1. commit aa7c13264a
2. commit 54807526b9
3. commit 6df62c78b8
Change-Id: I3d0d3214baba403b4709b358132b6756a15f42d7
Signed-off-by: Ori Messinger <Ori.Messinger@amd.com>
Description was inconsistent with itself and code. Existing behavior
returns HSA_AMD_MEMORY_POOL_INFO_ACCESSIBLE_BY_ALL == true for system
memory pools only and system memory pools do require hsa_amd_agents_allow_access.
Change-Id: I64b287bff9fdb21688aa169296e410edf1b209b5
Allocating these before the big memory allocations minimizes the chances
of spurious out of memory errors.
Change-Id: I94aff9ec7ea34d4dc98ae08ac4cf9dc335b3df7f
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
This reduces thrashing due to graphics submissions only and
significantly speeds up the BasicTest when keeping idle compute
processes evicted. In the BasicTest compute is always idle, so
only one compute eviction and no restore is triggered. Then
graphics submissions complete quickly without thrashing each other.
Change-Id: Iae6da98903b20424a5097f235e1d09cf13e4b41b
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
1. umc error injection only accepts parameter "0 0".
2. flush output to file in order to make writing happen
immediately.
Change-Id: I8d3bde287caee6b90b6eec56c760f5a228be7595
Signed-off-by: Eric Huang <JinhuiEric.Huang@amd.com>
The path was wrong based on assumption that GPU dri render
node starts from 0, because if there is a VGA device on
board, node 0 will be VGA and node 1 will be GPU. So the fix
will look at the name of GPU minor node and find the correct
primary node on which RAS debugfs entry exists.
Change-Id: Icc5e63ce48698d5d29105c0417e3bec8afa0a7c8
Signed-off-by: Eric Huang <JinhuiEric.Huang@amd.com>
Check if it is true or not. The string() call would define this to an
empty string, which would pass. This would then leave a trailing -
in the version string, which dpkg would error on during package
installation.
Change-Id: Ifb5fc15f5dde506e96bff7881a5d3f22d983406e
Search the local src directories first. If using a system
installed hsakmt, this would pick the installed hsa headers.
Change-Id: I9746d6e9db1749a130e4d93e024556754a537083
Remove the HSA_DEBUG environment variable that controlled the
creation of these mappings.
This should allow the debugger to attach to a running process and
access VRAM buffers through ptrace without having to do anything
special.
On processes that create many small VRAM mappings, this may cause
regressions due to the per-process mmap limit. However, the
sub-allocator in ROCr should consolidate most small allocations
into 2MB blocks nowadays, for good TLB efficiency. So this is
unlikely to cause problems.
Change-Id: I929da1be0f6cb51ec00a02f3f241d16083e4d95f
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Joined threads can not be joined more than once nor can they be detached.
Thread library wait and close allows multiple waits and separate close so
this fixes the pthread implementation.
Change-Id: I0019271a438f11ed4c6c11854011f5c4f6e16b65
The queue IDs passed over to the kernel via kfd_ioctl_dbg_trap_args->ptr
should be a list of uint32_t's. Need to convert from the passed in
64 bit HSA_QUEUEID to 32 bit uint32_t's.
Change-Id: I8718566d9f9ffc90ce0b2ecc129b10c49d73186a
Signed-off-by: Philip Cox <Philip.Cox@amd.com>
Small times may be given to time conversion if GPU clocks are used to
accumulate elapsed time. Because HSA APIs deal in absolute time this
leads to large conversion offsets of order system uptime. Variation
in relative clock ratio estimation may be amplified in this case,
destroying elapsed time measurements.
This patch fixes the relative clock ratio used for times which predate
the call to hsa_init. This correlates errors in such times allowing
the elapsed time to be correctly computed.
The effective maximum system uptime before elapsed time conversion becomes
inaccurate is ~3.5 months. GPU event timestamps are good for process uptime
of ~3.5 months. These are limited by double's mantissa precision.
Change-Id: I48752ff354920439d91016d6f2b0c8ddfa60b445
This can cause build failures on unknown of future compiler versions.
Only enable it if explicitly enabled by an environment variable. This
allows us to continue building with -Werror in internal builds with
known compiler versions.
Change-Id: Ic1cd9d223218cc4e4cddba49df93bb357c1cbd40
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
There was a mistake and RESUME was used when it should
have been suspend in two places in the suspend resume
code. This fixes that error.
Change-Id: I69be733d7ae7c14ce5ee8af57a307976e4212d62
This is updating to the new suspend and resume API for the
KFD and the thunk. We now support passing in a list of queues
to suspend, and not just all of the queues for the process.
The kfdtest testcase was also updated so it still compiles.
Change-Id: I71d1b178476bd9df0c311bdedaa6a891528cebcf
Signed-off-by: Philip Cox <Philip.Cox@amd.com>
hsaKmtGetQueueInfo needs to return the control stack size, and the
wave state size for the debugger. These changes are needed to support
returning the new values.
Change-Id: Ib4c60e0ea34446c06aef4a86996250989f348a69
Signed-off-by: Philip Cox <Philip.Cox@amd.com>
- Run more graphics command submissions with shorter delay between
them
- Synchronize after every graphics command submission
- Include the big VRAM BO in the BOList of the command submission
to trigger more evictions
- In QueueTest, run AMDGPU command submissions concurrently with
compute shader on the user mode queue
- Submit AMDGPU commands to GFX queue instead of compute queue to
avoid deadlocks between user-mode and kernel-mode queues on the
same pipe
- Allocate slightly less memory from KFD to avoid allocation errors
due to fragmentation or memory leaks in previous tests
- Running only two processes maximizes the number of KFD evictions
(probably because of lower chances of evicting non-KFD BOs)
Change-Id: If05d53f5fcf690b6488998a3f933f120ddaa71ee
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Add a MMIO_REMAP heap type and expose mmio virtual address
through HsaKmtGetNodeMemoryProperties
Change-Id: I1e585e6dfbec8fa7c85f1dda7b89b763a8e2c439
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
HDP conherence registers are remapped at driver level
to an empty page in mmio space (the remapped mmio page).
This change allocate and map the remapped mmio page to
process space.
Change-Id: I89c405c41870a79c5b58eea0d8e564aa35f55182
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
At the moment it is not possible to build ROCr with Clang. This is
a spurious limitation. The present PR addresses it by guarding GCC
only flags and by fixing some additional warnings that Clang triggers;
one of said warnings did outline a rather interesting issue with math
being done on void*s. - AlexVlx
Void ptr arithmetic had already been fixed in amd-master branch.
Change-Id: I5ee97e20b5c40b10dd73facecabe75f02ba46462
Non-paged memory can be IPC-shared even when HSA_USERPTR_FOR_PAGED_MEM
is enabled.
Change-Id: I8b1fa6d7a4a9327c78a77b3679697fbf55397093
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Those two tests cover the basic queue creation and destruction
without submitting packets to CP and SDMA user queues. During bringup,
they bring values in term of untangling the issues arising in queue
creation and packet execution, which are two very different kinds.
Because of those two tests, we also rename some existing tests as
follows:
CreateCpQueue -> SubmitPacketCpQueue
CreateSdmaQueue -> SubmitPacketSdmaQueue
CreateMultipleCpQueues -> MultipleCpQueues
CreateMultipleSdmaQueues -> MultipleSdmaQueues
Lastly, move MultipleCpQueues test closer to the CP queue section
rather than leaving it behind the SDMA queue section.
Change-Id: I110fb3f3fb21878339045dd1d1c8c9d61b8988b7
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
If a Render Node can't be found, we should finish off all child
processes immediately, then return.
Trying to do the check before forking the children results in the test
failing as well, regardless of the status of finding the render node,
which is likely why the forking occurred first in the test's initial
creation. This way we ensure that things finish cleanly before moving to
the next test
Change-Id: I2e1b62fed25c30ff1f179612127c23960da4ee5e
Oak Zeng
Evgeny
Sean Keely
Philip Yang
Cole Nelson
Chris Freehill
Ori Messinger
Felix Kuehling
Eric Huang
Matt Arsenault
Philip Cox
Kent Russell
Yong Zhao
Alex Voicu
Andreas Schneider