Commit Graph

93 Commits

Author SHA1 Message Date
Konstantin Zhuravlyov c1ad82a6b7 Loader updates for code object v3
- Fix loading in some cases
  - Fix symbol kind

Change-Id: I721b4a35972b6d2a6d0ac733ab770b096cc74e17
2019-01-18 15:41:01 -05:00
Sean Keely 269be0be2e Disable forced explicit selection of public vs internal HSA interfaces.
Temporary to reenable OCL builds on TC.

Change-Id: Ia81f2f9a9dd10ae8ce9627313247a586a8711584
2018-11-16 15:26:26 -06:00
Sean Keely 4e8597681b Cache KFD Events used by user allocated InterruptSignals.
Change-Id: I7f102f880fea9c78febe28cd262f93ee77f03184
2018-11-12 22:37:42 -06:00
Sean Keely 8323b2e1d7 Add pooling for Signal ABI blocks (SharedSignal).
Makes better use of memory and greatly reduces mmap count.

Change-Id: Ib444cd1ccd144986adbcc7cec297a966e2c08bc7
2018-11-12 22:37:28 -06:00
Sean Keely 936ecd1885 Remove legacy SVM region concept.
Also rename blit_agent to region_gpu and add comments to clarify
its role in deprecated region API support rather than to do blits.

Change-Id: I80b1043db2e1c5d40a58fc801eef70a688ea9169
2018-11-09 06:27:53 -06:00
Sean Keely 9ec37b5103 Ensure runtime cleanup when hsa_init ref count reaches 0.
Delete the runtime object when the last hsa_shut_down occurs.

Change-Id: I2005d52d06702eaef166714fd5e471cc277924db
2018-10-22 19:32:00 -05:00
Sean Keely 757502ccd6 Report internal queue creation to tools.
Debug agent requires handles to internal queues for single step debugging.
Added tools only API hsa_amd_runtime_queue_create_register for reporting.

hsa_amd_runtime_queue_create_register sets a callback which is invoked
when internal queues are created.

Change-Id: Ia5190ae724fadba686c15f25b2cd085350eeff0e
2018-10-20 23:12:27 -04:00
Sean Keely 1e0d690948 Use ptrinfo rather than apertures in hsa_memory_copy
Apertures now overlap with the change to 48bit addressing which
precludes using aperture checks to discover buffer ownership.
Switches to ptrinfo to decide which device a buffer owned by.

This corrects faults in the legacy hsa_memory_copy api.

Change-Id: I5c7ce0216e1cdc96f836fc6fec9c3defdf4b9d90
2018-10-11 13:34:53 -04:00
Ramesh Errabolu 01eea21d6c Capture number of Numa Nodes present on system
Change-Id: Ic789a6b9da8e316cb483e50b0fe9faa03798f97c
2018-09-18 16:27:30 -05:00
Sean Keely 56ed5c8904 Refactor blocking sdma commands.
Remove fence pool and use two signals.  Two signals allows overlapped
submission and copy while reducing thread busy polling.

Change-Id: Idb5f8e4c7f482a596ffce9e7799191fdd785a216
2018-08-29 19:13:23 -04:00
Sean Keely e0839ab27e Implement SDMA copy rect for gfx9.
Fix pitch overflow due to small element detection.
Add wide pitch 2D copy handling.
Cleanup code duplication.

Change-Id: I93b1584aba8e5964957eb7ab3544df806ca3e2f9
2018-08-29 19:13:07 -04:00
Jay Cornwall e388a23344 Add hsa_amd_queue_set_priority extension function
Controls dispatch and wavefront scheduling arbitration across quees.

Change-Id: I498f4898b544f79b8fb8514bf7e789ca9da29462
2018-06-19 19:41:28 -05:00
Jay Cornwall 536823482b Handle llvm.trap only in gfx9 trap handler
llvm.debugtrap and other trap IDs are reserved and should not place
the queue into an error state.

Change-Id: I98193a35ac7da94c4a42ee75d87754ee552ebea0
2018-05-04 13:15:23 -05:00
Sean Keely 5f25619bb7 Enable large scratch on GFX8.
Ensure system release fence is set on GFX8 large scratch using packets.

Change-Id: I13cfdcd35969482ea6e95e0b352f5cb3a0454b86
2018-04-30 07:24:53 -04:00
Sean Keely 7cd6e366ed Workaround SDMA poll packet preemption.
Use async. signal handler to satisfy dependencies for SDMA blits.

Change-Id: Ifa8d3ee6810509f400a568ca2387ac6ab3ab7c36
2018-04-26 02:00:33 -05:00
Qingchuan Shi 49d2175c74 debug suport for queue error.
1/ Revised debug event handler to handle different events.
2/ Added queue error handler using the callback in queue create, which will print out wave info when queue in error state.
3/ Preempt queue instead of destory queue when queue error state.

Change-Id: Ib727d208de9caf1c72c76d42268483b24aaebde8
2018-04-20 14:25:16 -04:00
Konstantin Zhuravlyov 7ef70f7eaa Bring naming on par with the spec (hsa-runtime)
Change-Id: Ie1903c90a195cf95b186eb5552131a20af408adf
2018-04-10 09:15:02 -04:00
Sean Keely 7caf9633f6 Support large scratch allocations and reclaim.
Also improve small_heap used for scratch region allocation.

Change-Id: Ib7311b663b38968d88ebc355b81e12c0863dc541
2018-04-05 21:51:56 -04:00
Jay Cornwall df964343a3 Support new first-level trap handler ABI
- Ignore exceptions passed to the second-level handler
- Restore SQ_WAVE_IB_STS and SQ_WAVE_STATUS before exiting trap

Change-Id: I872c111c030d94eae644ae073df3c2e508f42f45
2018-04-04 11:01:14 -05:00
Sean Keely b6f0248f53 Respect new memory model requirements at queue destroy.
Spec requires GPU release fences and CPU acquire fences at queue destroy.
Also update the recognized status codes.

Change-Id: If9166f5149f65417c7057ff7c0f69f6ac094d6ab
2018-04-04 08:13:00 -04:00
Sean Keely 6df9ba97ce Sequence queue error callbacks with queue destroy.
HSA v1.2 update.

Change-Id: I13975e71b2c1ea5b7738236f5d02df84312ad00c
2018-04-04 08:12:58 -04:00
Ramesh Errabolu 987f3f97aa Enable sDMA packet HDP Flush on Gfx9 and later devices
Change-Id: I85922e5266883ef7e9eed3565e2c3b209009d294
2018-04-02 11:47:59 -04:00
Sean Keely cd46954cc4 Cleanup in blit kernel management code.
Remove unused function (FenceRelease), add comments to barrier packet settings,
correct profiling controls to work with queue wrappers.

Change-Id: I45bb26227bcc2b78edb8ad5dc497603c33234e18
2018-03-20 22:19:54 -04:00
Sean Keely f4521ce782 Revert "Reduce to only one internal compute queue."
This reverts commit 0eb534e3cf.

Change-Id: Ifcc5e148457243a6cf9ef277da7ab7c4e10f6fc9
2018-03-20 22:19:44 -04:00
Wilkin 8e3d26c617 ROCm Runtime Support for respecting target xnack setting
This includes the changes provided by Konstantin, "Add xnack from elf header" (Change 136389).

Change-Id: I95e51141caa0d7c21903b09212c02e4906ec54a3
2018-03-20 16:57:15 -04:00
Sean Keely ac5ccb45b7 Use atomic variable for Runtime ref_count_.
Change-Id: Ic4d0ad9ff93d0cc52cfe2df006ee3436d5960b07
2018-03-06 03:45:14 -06:00
Sean Keely 0eb534e3cf Reduce to only one internal compute queue.
Change-Id: Ie42ecb3b242077624d74caeabfcd418dbbd9ff3e
2018-02-09 14:02:15 -05:00
Sean Keely bd5dd47ca1 Defer creation of internal queues and blits until first needed.
Change-Id: I2e61d7e102f38389d806d9eb24beda910573157b
2018-02-09 14:02:07 -05:00
Sean Keely d2e70bb999 Cleanup Signal interfaces for doorbells.
Create an interface for doorbell signals to reduce code duplication.
No functional changes.

Change-Id: I101a8997dd582ff99e1537758c804b21fe3bb6af
2017-11-28 22:12:19 -06:00
Sean Keely b93ffafdc7 Pull from github (tstellar):
Prefer using memfd_create() for the ring buffer.

We were using /dev/shm, but this won't work on systems that
either don't have /dev/shm or have mounted it with noexec, because
for everything other than gfx700 we map the ring buffer with PROT_EXEC.

memfd_create() is Linux specific and was added in Linux 3.17, so we
will fallback to using /dev/shm on systems where memfd_create() is
not available.

Change-Id: I58fb533eebc362f6d29dc3e316a80801014d50e8
2017-11-28 20:47:12 -05:00
Evgeny 6e1b9288f6 aqlprofil API: removing from HSA hsa_api_trace/hsa_ext_interface
Change-Id: I12fac55ea9ccfdb119899bf9d000e3c8b0bf4bbb
2017-11-11 10:01:12 -06:00
Sean Keely 6455a69b03 Fix bad casts in tools.
Also virtualize queue profiling enable.

Change-Id: I761b41269be3df7eb64a5914ee9951ed6b51bb04
2017-11-08 15:50:02 -05:00
Sean Keely a6d8a48cbc Add callback exception forwarding.
Modified callbacks for intercept queue, queue error, iterate agent and
iterate region.

Change-Id: I8bdd67f2312510ea7eb9caec93babca244938b40
2017-11-08 15:50:02 -05:00
Sean Keely f312a7386e Exception support for Queue.
Remove "zombie" queue state and report queue creation failure via
exceptions.  Make Shared object a final container and support array
objects with Shared.  Add message printing to hsa_exception in
debug builds.

Change-Id: I459f38c80846018acbf45538874e95f91dd6b195
2017-11-08 15:50:02 -05:00
Sean Keely 0c7dde2d1f Add queue intercept support to the runtime.
Queue intercept is exposed as two tools-only APIs via the API
intercept table.

Change-Id: Iac9602ed3143974d85c3569e9092295ad18037f8
2017-11-08 15:50:01 -05:00
Sean Keely d93f92f42d Make HostQueue::queue_count_ a portable atomic type.
Also make lint happy.

Change-Id: I0f965df6a76fd959df9eb411d1f1b11847159790
2017-10-31 02:38:25 -05:00
Qingchuan Shi ce6aee01ed Add APIs to support debugging vm fault
1. Add hsa ext api hsa_amd_register_vmfault_handler for debugger to register callback in case of VM fault.
2. Extend hsa_ven_amd_loader API to:
   (1) iterate loaded code objects in executable:
       hsa_ven_amd_loader_executable_iterate_loaded_code_objects
   (2) get loaded code object info:
       hsa_ven_amd_loader_loaded_code_object_get_info
3. Make the id of hsa_queue the same as the one used in communication with thunk (for amd_aql_queue)

Change-Id: I68910809e59e24297350d262606f00e96c14bcbd
2017-10-28 21:48:26 -04:00
Sean Keely e9a6f2c3e6 Support hsa_amd_agents_allow_access on page fragments.
Since access may only be manipulated on whole pages, suballocator fragments must cooperate to set the page's access.
Since the KFD does not migrate memory on access changes this implementation makes agent access sticky across the requests in a fragmented page.

Change-Id: I88479ed45fb40e9782b704526a7b8ffb22e7bd76
2017-09-27 19:04:04 -05:00
Sean Keely 117be0b55a Add suballocator for ordinary VRAM allocations smaller than 2MB.
Track pointer info for sub 2MB fragment allocations in allocation_map_.

Add fragment support to IPC.

Change-Id: I00cfc2e2fa289aac90a4718c392f9bb056a61a87
2017-09-19 06:08:36 -04:00
Sean Keely c9642cf7af Initial IPC signal support.
Added an API for creating signals with attributes.
Added two APIs for IPC operations on signals.
Initial use of exceptions for error handling.

Add ref counting to signals.
Removed spin loops from signal destructors.
Signals are no longer to be destroyed with delete, use DeleteSignal instead.
Added delete safety to doorbells.
Added secondary hsa_signal_t -> Signal* translation path for IPC enabled signals.

Change-Id: Id59065d002f0c2566b0a9425694da2ed27cb7d7f
2017-08-11 18:41:34 -05:00
Sean Keely 2732b18092 Initial exception support for signals.
Also separate signal ABI block allocations from the runtime interface object.

Change-Id: If16763338db664f29163a1348f8f4c38cf0597b2
2017-08-11 18:41:34 -05:00
Sean Keely bc0bd00746 Fix queue interception in tools.
1. Correct amd::AqlQueue::ExecutePM4 to support interception.
2. Minor fixes to AqlPacket and SoftCP.
3. Minimal change to disable interception of runtime internal queues.

Change-Id: I103fece2ebf9a188d27f01e61221c737405d7253
2017-07-12 16:39:43 -04:00
Evgeny 4174f07fd1 hsa-runtime integration
Change-Id: I48968966ffe164218ebff88d0e3a1268e96bf1dd
2017-07-05 10:55:30 -04:00
Kenny Ho 5b4df54b10 Revert "Implement memory fault analysis through context save area"
This reverts commit 75c9506f9d.

Change-Id: Ibf11b764b383b9be291f3009a30550e1a1e2d115
2017-06-14 14:21:53 -04:00
Jay Cornwall 75c9506f9d Implement memory fault analysis through context save area
When a fatal memory fault occurs the scheduler context-saves all queues
in the process and notifies the runtime through the memory event. The
saved state contains all GPR/LDS data at the moment of the fault.

Retrieve this state and present it to the user if HSA_DEBUG_FAULT is set
to "analyze" and the wavefront caused the fault. If amdgcn-capable objdump
is in the PATH invoke this to disassemble code around the PC.

Queue lifetime is now managed by the runtime to allow querying the
context save state for all active queues.

Change-Id: I6fee662fad1c4f9aa125bf5c53d7d0ea1ab32f95
2017-06-13 23:12:28 -04:00
Sean Keely e38ff18990 Unmap GPUs when allow_access removes them from system pools.
Change-Id: Ib9eb88622fded43ebd9eddbf78ad6771a5b91e77
2017-05-17 20:58:05 -04:00
Jay Cornwall f0a1c7c4c6 Fix gfx9 trap handler to retrieve correct return address
The trap protocol changed between gfx8 and gfx9. The return address
is in trap temporaries [0,1] on gfx9 rather than [4,5] on gfx8.
Unfortunately SP3 changes the meaning of the ttmp register aliases
in gfx9, further confusing the issue.

Clean up later when LLVM assembly build is introduced to the runtime.

Change-Id: I84ea9bf3736f060dd95d0361f9d5a0f9a3576178
2017-04-05 17:33:49 -05:00
Christophe Paquot 05d587ef79 Add inc/ to some include
Change-Id: Id027b015c8785a132835a422d97a23b0bbce208a
2017-03-09 19:45:01 -08:00
Konstantin Zhuravlyov 9887c26113 Bring loader in sync with stg/sc
Change-Id: Iccce07b8fa03d37c4267a2a9bd343e6614dc43e7
2017-02-10 11:21:15 -05:00
Jay Cornwall 4d62b9482a Implement code cache invalidation for Gfx9
When a new enough microcode build is running use a vendor AQL packet
to submit the PM4 IB.

Change-Id: Icd3e2b322c418477420ba4a29f4455ce340ef0d2
2017-02-09 14:15:21 -05:00