Commit Graph

731 Commits

Author SHA1 Message Date
Brzak, Branislav 494c0e3201 SWDEV-531996 - Add legacy fallback Mesa loading via dlsym (#512) 2025-06-11 08:02:51 +05:30
Brzak, Branislav 5fabfedbd1 SWDEV-531996 SWDEV-530154 - Use *getprocaddress to obtain Mesa functions (#504) 2025-06-10 13:31:32 +05:30
Andryeyev, German 5cc172c99c SWDEV-533074 - Expose hipDeviceAttributeNumberOfXccs attribute (#336) 2025-05-21 22:35:42 +05:30
Jayaprakash, Karthik 12131de4a9 SWDEV-529929 - hipMemGetHandleForAddressRange implementation. (#245) 2025-05-20 15:56:04 -04:00
Kudchadker, Saleel 1b0ea080e4 SWDEV-523279 - Use preferred engine mask for SDMA (#317)
- ROCr now reports preferred engine for copy status. We can leverage
this for max bandwidth for inter-GPU copies
- Cleanup logging
2025-05-19 16:04:51 -07:00
Kudchadker, Saleel 5712944c7c SWDEV-531518 - Fix offset accumulation (#333)
srcAddress/dstAddress accumulation was cumulative, which shouldnt be
done if we increment offset.
2025-05-19 18:03:06 +05:30
Jayaprakash, Karthik bb7750a946 Revert "SWDEV-522707 - Set phys_mem_handle type to sizeof(size_t) to avoid blocking address range. (#105)" (#348)
This reverts commit 6811fd90b8.
2025-05-19 15:19:16 +05:30
Jayaprakash, Karthik f5b8db33f1 SWDEV-531711 - Report correct error code based on device failure. (#286) 2025-05-17 06:33:13 -04:00
Andryeyev, German da198ac5b2 SWDEV-531678 - Remove split path from the dispatch (#283)
The split path for blit kernels are no longer necessary, since the new blit kernels
don't use the copy size as the global workload
2025-05-12 12:50:32 -04:00
Xie, Jiabao(Jimbo) a320a3f214 SWDEV-528913 - support gfx950 in rocsetting (#217)
* SWDEV-528913 - support gfx950 in rocsetting

---------

Co-authored-by: Jimbo Xie <jiabaxie@amd.com>
2025-05-07 15:44:49 -04:00
Andryeyev, German 65a0181a7c SWDEV-528808 - Release all HW queues even if only one is idle (#240)
Pytorch may not explicitly idle each queue. Thus, some queues can be considered as busy,
but have idle state in reality
2025-05-05 19:09:01 -04:00
Andryeyev, German 9b018165ce SWDEV-528808 - Disable dynamic queue by default (#256)
Dynamic queue management will be disabled by default and
the original sort logic is restored
2025-05-05 10:56:35 -04:00
Assiouras, Ioannis d3fb8eda8b SWDEV-525593, SWDEV-527293 - Acquire active queue after xferQueue is created (#165)
For xferQueue VirtualGPU::create is called after ProfilingBegin
so the active queue needs to be acquired.
2025-04-30 09:21:11 +01:00
Jiang, Julia eb62fe9f62 SWDEV-522634 - Fix device properties in hipInfo (#203) 2025-04-29 11:29:47 -04:00
Sang, Tao 96cadbc9e9 SWDEV-520352 - Remove HostThread and legacy monitor (#230)
* SWDEV-520352 - Remove HostThread and legacy monitor

Remove HostThread, semaphore and  legacy monitor.
Make original logics of thread and command queue stricker.
Add more comments to make logics clearer.
Some other minor improvement.

Also part of SWDEV-458943.
2025-04-29 09:55:24 -04:00
Jayaprakash, Karthik 6811fd90b8 SWDEV-522707 - Set phys_mem_handle type to sizeof(size_t) to avoid blocking address range. (#105) 2025-04-29 11:19:16 +05:30
Jayaprakash, Karthik b2388dfb88 SWDEV-506467 - Skip Abort in case of crash from the device. (#60)
Change-Id: I964b2f2647d068202e9c38fcddb1337da754df8d
2025-04-29 11:19:02 +05:30
Kudchadker, Saleel 072fb0804e SWDEV-521647 - Fix tracking of hw_event (#206)
- When a command may possibly have two packets(like device heap
  initializer), and if there is no signal on the main kernel packet the
tracking was broken as it marked HW event of the command as the first
packet signal.
- Make sure if no completion signal is attached to the second packet
  then clear the HW event for the command.
2025-04-25 08:46:44 -07:00
Kudchadker, Saleel ce24936970 SWDEV-510186 - Improve logging (#220)
- Print all arguments for logs, this is useful for debug
2025-04-25 08:40:31 -07:00
Sang, Tao 1113eff3f9 SWDEV-493275 - Support scratch limit (#20)
Support programmatic query and change of scratch limit on
AMD devices.

Change-Id: Id5da355a77366f97868e462847f3916e87fd2af6
2025-04-24 17:15:25 -04:00
Andryeyev, German 3fd7650fe3 SWDEV-459758 - Pass workgroup size explicitly (#185)
It's easier for compiler to move explicit kernel arguments into user SGPRs
2025-04-15 15:22:15 -04:00
Yao, Longlong 0de73eeaf8 SWDEV-518966 - Avoid creating Arena Memobj for VMM pointer (#39)
Change-Id: I69c6c0a1464d01e674ac929de34ab10047012f1a

Signed-off-by: Longlong Yao <Longlong.Yao@amd.com>
2025-04-11 16:55:53 +05:30
Andryeyev, German 4c363df3bf SWDEV-517481 - Add more restrictions to the queue management (#168) 2025-04-10 21:51:45 +05:30
Patel, Jaydeepkumar 9e7248aa36 SWDEV-521011 - Allow max stack size as per ISA. (#73) 2025-04-08 10:15:38 +05:30
Patel, Jaydeepkumar b5c9cbc236 SWDEV-508632 - Align address to 2 MBs for hidden heap allocation. (#29) 2025-04-02 16:33:29 +05:30
Arandjelovic, Marko e7ada4effe Revert SWDEV-512344 - Unmap all subbuffers (#26)
This reverts commit 0b69120cfcb5b4689d9f2037b1a01e274d85c20f.
2025-03-19 21:17:36 +05:30
Andryeyev, German 28967982b2 SWDEV-517481 - Add dynamic queue management (#37)
Enabled by defaulty. DEBUG_HIP_DYNAMIC_QUEUES controls the feature
2025-03-19 11:22:50 -04:00
Saleel Kudchadker 78d0ff2dbc SWDEV-519596 - Avoid passing dep signal to SDMA
- For D2H cases avoid passing dependent signals to SDMA, the signals
  take a while to resolve on SDMA engine

Change-Id: I569635228af977847f201c82ca897002f8f2f4a8
2025-03-07 17:37:21 -05:00
Branislav Brzak c2d1776ebd SWDEV-516564 - SWDEV-512817 - Remove mentions of gfx940 and gfx941
Change-Id: Ia069fcb9c6948c3fc9a00961593c9dcc59609375
2025-03-05 04:26:07 -05:00
Saleel Kudchadker e03e4f3b5d SWDEV-502365 - Track last used command
- This change tries to save extra synchronization packets we may insert
  as we didnt track the completion signals for every command. We track
the current enqueued command until it exits the enqueue stage. We also
record the exit scope to know if we flushed the caches
- Handle correct release scopes and store completion signal as HW events
- Use a new finishCommand implementation to only wait for the command
  passed as the argument

Change-Id: Ie4350c5dd24f5d48dfa6ccbabd892f0544caadcc
2025-03-04 16:05:02 -05:00
Marko Arandjelovic 3ec1d2d2f1 SWDEV-512344 - Unmap all subbuffers
Since hipMemMap can be called for multiple device handles on the same virtual memory, the same is true for hipMemUnmap, meaning that virtual memory can be "partially unmapped".

This means that the unmap function can be called for a specific part of the reserved address, meaning that only the designated subbuffer should be released. If unmap is called on the entire reserved memory, then all subbuffers should be released.

The main point is that for every hsa_amd_vmem_map, there should be a corresponding hsa_amd_vmem_unmap. Otherwise, if entire memory is unmapped by a single unmap call, then HSA will report the memory as "in use" if an attempt is made to delete it.

Change-Id: I039308eafb820decfb1c09f60347f26cdad1a362
2025-03-02 13:41:48 -05:00
Saleel Kudchadker 37d606d193 SWDEV-513197 - Unify getBuffer implementation
- Use getBuffer/releaseBuffer in BlitManager
- Cleanup XferBuffer as we use ManagedBuffer for both reads/writes

Change-Id: I2661b85dd012763b17a38a743fec1b1d79125f67
2025-02-28 12:47:51 -05:00
Saleel Kudchadker ca530c660b SWDEV-513197 - Improve launch perf for Device Heap kernels
- If any kernel uses device heap, the launch needs to be preceeded by an
  init kernel, Save on the extra barrier packet launch/flush between the
init heap kernel and user kernel

Change-Id: I8ebc6246188200e5f673dc464bc76a53bcb8b7c6
2025-02-27 19:17:51 -05:00
Rahul Manocha 4bf634dfca SWDEV-510849 - Restore pinned memory copy path
1) Create getBuffer method to return pinned host memory or staging buffer
2) for D2H path use managed buffer instead of static buffer
3) use staging buffer copy for 16KB < size < 1MB
4) use pinned memory copy for size > 1MB

Change-Id: I13d4d6ab60691bc6c7724239db1e11e23f0f3dc2
2025-02-26 11:25:02 -05:00
kjayapra-amd 1f583a6870 SWDEV-516303 - Remove SDMA retainer logic to select the engine.
Change-Id: I818129444131825cdb87e06cb495afa3e5cdb683
2025-02-20 11:34:38 -05:00
German Andryeyev ba8e740be4 SWDEV-515356 - Make the round-robin queue selection
- Add custom compare to the map of queues, which will help with
 the round-robin selection

Change-Id: Ie67a820bfb1a5b484a1b3edced967eed94228bb8
2025-02-20 11:09:54 -05:00
Vladana Stojiljkovic 061c5d877f SWDEV-510059 - Format CU mask properly
Change-Id: I80e94b4f3ea25f6988fc06d83aeb398e81ccddd1
2025-02-13 11:02:56 -05:00
Saleel Kudchadker 4c95ee5e1e SWDEV-504494 - Further copy improvements
- Fix regression for D2H pinned copies which adds systemscope release.
- Skip cpu wait for D2H unpinned copies as we can pass the signal of the
  barrier to rocr copy.
- Fix an old bug in sdmaEngineRetainCount_ logic
- Improve logging

Change-Id: If074bddb05564b15949b0d5f9bf12acd3692174e
2025-02-11 00:55:52 -05:00
Ioannis Assiouras d05ecea253 SWDEV-508435 - Added a fix for double free of hsaImageObject
Change-Id: I9397f7c9dbbad7c249b359155df312cb920eba6c
2025-02-05 22:21:24 +00:00
Ioannis Assiouras 3e01da3dac SWDEV-513323 - Fix for BatchMemOp on devices with no image support
BatchMemop should be positioned before the image support kernels
because the total number of kernels is determined by BlitLinearTotal,
when there is no image support on the device.

Change-Id: I8e53caf744ba54259ac04bad1762eef21806f3f2
2025-02-05 04:45:22 -05:00
kjayapra-amd cc62a82347 SWDEV-488290 - Remove Stream to Engine logic and rely on engine query status HSA API.
Change-Id: I469ab6679360c8ee8d4ee515678a8aa8d4578ebf
2025-02-04 13:00:16 -05:00
Jimbo Xie 4872b420c9 SWDEV-504383 - Cleaned up kForcedTimeout10us and removed IsHwEventReadyForcedWait
Also removed active_wait_timeout

Change-Id: I7a429f003c09a4df267b5c0983050704260094c6
2025-01-31 14:40:18 -05:00
taosang2 32daa8f384 SWDEV-501963 - Add missing codes for gfx950
Cherry-pick https://gerrit-git.amd.com/c/compute/ec/clr/+/1162997

Change-Id: I6b3c6bf55c61cffd43cd6f17b75998f751b75723
2025-01-31 14:34:49 -05:00
kjayapra-amd 0324014710 SWDEV-509280 - Combine multiple definitions of callbackQueue into a single function.
Change-Id: Ibbb56136bec2beed71c202d75e8aec9e82640a4e
2025-01-30 15:58:11 -05:00
Saleel Kudchadker 2d450e8b06 SWDEV-504494 - Resolve signal dependencies
- Resolve signal dependencies for barrier value packet if there are > 1
  depenent signals. Barrier Value packet accounts for only 1 dep signal
- Better log

Change-Id: Ia506ad5d80b91d598f92e7b539f41756e9b4b64b
2025-01-29 19:49:02 +00:00
German Andryeyev ea0b092af8 SWDEV-459826 - Add a crash dump for a failed queue
The logic can analyze the AQL queue state and
find a failed AQL packet with the kernel's name

Change-Id: I1a478fa2c25462cd07a194784958bdf22454b897
2025-01-28 14:27:46 -05:00
Saleel Kudchadker d208e8052f SWDEV-504494 - Set active engine for SDMA
Change-Id: I4cec84e71903c5813a7063e8b9ff1ea4473f4720
2025-01-27 17:54:36 -05:00
Ioannis Assiouras 21c223f8df SWDEV-510319 - Fixed random segfaults in graph tests
This change fixes random segfaults in graph tests that
are seen after the change make internal callbacks non-blocking.
The callback thread that decreases the GraphExec ref count
may now run after the runtime shutdown. This can cause a segfault
because the hip::device that is accessed in GraphExec destructor
is already destroyed during runtime shutdown. This patch ensures
that the hip::device object  stays alive until after the
callback thread completes.

Change-Id: I75a6ac01f27a0b2250bbd10ed389ebfb322927af
2025-01-25 09:54:15 -05:00
Saleel Kudchadker 9b7e0ad48a SWDEV-510186 - Improve logging of kernel names
- Demangle kernel names in logs

Change-Id: I9aa58e8c109becb45ef7fc747d991bd657c4190a
2025-01-24 11:43:02 -05:00
taosang2 799e54aa0d SWDEV-507969 - Fix wrong VGPRs for some devices
Change-Id: Ia8fc19564272e2c7171d991376bf896a99085a97
2025-01-22 10:11:47 -05:00