This change adds the initial classes for the AIE agent and AIE AQL
queue.
An AIE agent list is added to the core runtime object.
Change-Id: I84b02f52171b80726dfb2c8431582a3ea2986eb3
Recommended SDMA engines for DMA copies are now exposed for better
GPU-GPU performance. ROCr can now select those DMA engines.
Also lock-in host-device copies to SDMA0 and device-host copies to
SDMA1 for better stability and performance.
Change-Id: Ideff2e13daf537104efecb8b837bd49ee5096cb5
This fixes an issue for missing HW events when out of HW events.
We cannot determine whether a HW event has occurred unless we call the
underlying drivers with hsaKmtWaitOnMultipleEvents_Ext. Previous logic
in Signal::WaitAny would switch to ACTIVE_WAIT state if we run out of
hardware events (signal->EopEvent() == NULL) and this would cause the
hsaKmtWaitOnMultipleEvents_Ext call to be skipped. But also, when we
have some signals without hardware events, calling
hsaKmtWaitOnMultipleEvents_Ext with a timeout of 0 so that we can poll
for remaining signals adds overhead with an IOCTL call and may cause
extra delay. Separating AsyncEventLoop into two separate threads so
that:
1. We can have a new Signal::WaitAnyExceptions to wait for HW events
This function can be simpler as it does not have to perform all the
timer calculations because it is expected to be always waiting on
hsaKmtWaitOnMultipleEvents_Ext through the lifetime of a process.
2. Signal::WaitAny does not need to have extra code to check for HW
exceptions as it only needs to handle HSA_EVENTTYPE_SIGNAL events. It
can also skip the calls to hsaKmtWaitOnMultipleEvents_Ext if needed.
Change-Id: I52ba99fd6e483e0cb477b7931a0dcc03520aa523
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
If user application tries to free memory that is currently being used by
the underlying HW device, the hsaKmtFreeMemory function call will fail.
This would be caused by an incorrect call by the user application. A
system memory error is raised and the user application is expected to
abort when this happens.
Note: This leaves the allocation_map_ table in an inconsistent state as
this address entry is removed from it while the pointer is not actually
free'd. But re-organising the FreeMemory() function would require the
memory_lock_ to be held for much longer and may affect performance.
Since this is a very unlikely and invalid use case, we prefer to leave
the FreeMemory() function as is.
Change-Id: I24279eb98620c32d34f4c5ad1b7a0a30cb65835d
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Force mem_flags to be explicit passed in then calling Queue constructor
to avoid ambiguity with calls to Queue constructor trying to only pass
the agent_node_id.
Change-Id: Ib6fedcb9e52d6c9f35f9051dfa989343456ca368
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
New API to accept a file stream for logging
Co-authored-by: David Yat Sin <David.YatSin@amd.com>
Change-Id: Ie09c35ae14ca86a97eb25f61251be287c55d7169
Signed-off-by: Chris Freehill <cfreehil@amd.com>
New API to support alignment parameter when reserving virtual addresses.
If the alignment is 0, then the default size is used. Otherwise the
alignment needs to be a power of 2 and greater than or equal to page
size.
Existing hsa_amd_vmem_address_reserve marked for future deprecation.
Change-Id: I17cee75420183dea5842fc1ecc2514cdcd760bac
Signed-off-by: Chris Freehill <cfreehil@amd.com>
Packet for GFX12 is incompatible with pre-GFX12 as some fields changed
location. Implement code path and packet specific to GFX12.
This fixes some issues with SDMA blits and 3D images.
Signed-off-by: David Belanger <david.belanger@amd.com>
Change-Id: I56c204aaa12160e563ec960bd3b226cfa94e142d
Signed-off-by: Chris Freehill <cfreehil@amd.com>
Added GFX12 implementation for InitScratchSRD and for compute_tmpring.
Implementation for compute_tmpring could be combined with GFX11 with some
refactoring as a possible future improvement.
Signed-off-by: David Belanger <david.belanger@amd.com>
Change-Id: I8013cbe4438786bf41bbfd03f6a5d3b9ef51e7bf
Signed-off-by: Chris Freehill <cfreehil@amd.com>
Updated struct definitions, field size changes and new fields in
registers.h.
Added resource_gfx12.h and updated fields in BUF/IMG/SAMP descriptor
structs based on documentation.
Signed-off-by: David Belanger <david.belanger@amd.com>
Change-Id: I08f05ba30f54c40e7b823a6a105829a1e8590b3d
Signed-off-by: Chris Freehill <cfreehil@amd.com>
Add a new driver interface as a core ROCr component.
The driver component provides an interface for ROCr to interact with
agent kernel-model drivers in a generic way. This interface will be used
to interact with the XDNA NPU driver. Eventually, the ROCt library's
functionality should be implemented behind this interface.
For now the interface provides basic queue and memory allocation
for supporting HSA queues and signals and matches the thunk API
closely.
Change-Id: I37ac9f2dcbadc86ce45999f76b0e9ce753fd0c06
Signed-off-by: Chris Freehill <cfreehil@amd.com>
Support contiguous physical memory allocation flag. Allocations with
this flag will have contiguous physical memory. This is dependent on KFD
support for this flag and the AllocateKfdMemory(..) function call will
fail when it is not supported.
Change-Id: I6c51c8b061f7b026fdcc2aa2c37c74ecc13d95b6
Fix Musl libc NULL errors and unsupported pthread funcs for compatibility.
Also ensures cleanup and error handling irrespective of CPU affinity override.
Fix submitted by github dev - AngryLoki
https://github.com/ROCm/ROCR-Runtime/issues/181
Change-Id: Ia487315e504112be5d3370756f23f6e23b9ae4be
New hsa_amd_queue_get_info API to support:
- HSA_AMD_QUEUE_INFO_AGENT: Agent that owns the underlying HW queue
- HSA_AMD_QUEUE_INFO_DOORBELL_ID: KFD doorbell ID of the queue
completion signal.
Change-Id: I98842131bcbdd08552649791a5d43e578a615808
Flush is used by the client to retrieve data that are currently stored
in the buffers. This is used by the client to retrieve current data when
the buffers are not full.
Change-Id: Ib8304dcdfb2797cb060ec72df4970d95cf6be348
Retrieve data from the buffers previously set in the 2nd level trap
handler TMA. We use a double buffering mechanism to allow the 2nd level
trap handler to write to one buffer while we are copying data from the
other.
Co-authored by: Joseph Greathouse <Joseph.Greathouse@amd.com>
Co-authored by: James Zhu <James.Zhu@amd.com>
Change-Id: I252c381ea06b8cf927c4f9af6ea59dedc3717fbb
Code is valid for gfx9 GPUs excluding gfx94x.
1st level trap handler will use TTMP13[22] to indicate host trap and
TTMP13[21] to indicate stochastic trap.
For each PC sampling method (hosttrap and stochastic), we use a double
buffering mechanism to transfer data between GPU and host.
The GPU will dump data into one buffer while CPU may be reading data
from the other buffer. There are 2 separate signals, one for each
buffer.
When signal != 0, the buffer belongs to the GPU and the GPU can write
to it. Once the buffer has reached the high watermark, the GPU will
set the signal to 0 to wake up the host and so that the host can try
to switch the buffers and read the data.
Co-authored-by: David Yat Sin <David.YatSin@amd.com>
Change-Id: If3eb0913e52fb4788059a71e5feca334612f3d5d
Create dedicated CP queue with highest priority for PC Sampling. Reduce
the highest priority that LRT's can set for existing API so that PC
Sampling queue will always have highest priority over any other CP
queues
Change-Id: Ia70d74415edc83b4862a3e18dbdbd7cebe73ab47
Implement PC Sampling session create and destroy APIs.
Change-Id: I93370d3d01b74ee15e71b8b0e20feb8f0066a3dc
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Signed-off-by: Vladimir Indic <Vladimir.Indic@amd.com>
Change-Id: Ib0c64356a1a4616b12d5dbeebe16273fe2a84abe
Add new PC Sampling API to list the supported PC Sampling methods and
options on a specific agent. If there is already a PC Sampling session
active on this agent, the list of methods returned will be reduced to
methods that can be run simultaneously with the current active session.
Change-Id: I42ac2b8f30d5c368faf8ed4cf37ca4134db22985
ExecutePM4() function can optionally accept extra arguments for
acquire fence scope, release fence scope andcompletion signal. When
a completion signal is provided, ExecutePM4() does not wait for the
commands to complete.
Change-Id: Ib2a433b7bce1cb6260be8b76fe902335bd5dfada
Hard limit for scratch is 4GB per XCC and checks in case user specifies
values exceeding this value
Change-Id: Ib3cade762ff66c7e7d6a2d311e482cacbcf2b0de
- Per-executable contexts should be used from now on
- Global contexts are left as is for now for backwards
compatibility and will be phased out in follow up
patches.
Change-Id: I6291abf865c7ed24ee71f5065e539afc23f5ce64
This reverts commit 5c520f4544c654e5f18e05cabd1c63d64473cfab.
Reason for revert: This patch is introducing a synchronization related bug in Unit_hipGetSetDevice_MultiThreaded testcase.
Change-Id: I367e4d4f1d75b21658ac1127c58982894a97cedb
Instead of caching shared memory fds for export on the exporter side,
only export the FD in the async handler when requested.
The importer should request export fd closure once import is done.
Change-Id: I469e0cd1749beeb9c506c8a6461745fb039d9c3b
Optimizations include:
- Greedy gang by placing gang leaders on first D2D sdma blit context
to avoid dead locking with other gang leaders and items. Note that
this is fine since we can't avoid an oversubscription problem when
there is only 1 xGMI link anyways, so treat all xGMI links as a single
pipe for ganging.
- Non-leader gang items don't have to poll on dependency signals so this
opens up more non-blocking SDMA channels.
- unlock gang lock when gangs are not needed.
- Change gang factor lookup from vector pair to map and register all
gpus in gang factor lookup regardless of link type so that we can take
advantage of the O(logN) direct key/value lookup time.
Fixes include:
- HSA_PAGE_SIZE_4KB was an incorrect macro to use for gang size limit.
As a result, small copies ended up ganging and hitting latency limit.
Use hardcoded 4096 bytes instead.
- Cap auxillary gang factor to the number of non-XGMI SDMA engines.
Change-Id: Ic23fde131502906a807134a04599aa6d012e8cbb
As the KFD IPC IOCTLs will not be upstreamed, change runtime
implementation to use DMA bufs.
DMA buf fds will be passed over abstract unix domain sockets.
The exporter spins a thread that creates a socket server.
The importer connects to the server to fetch the fd.
libDRM will be required to do a manual import and GPU map for
memory that is not already imported and mapped.
For now, use the legacy IPC implementation by default as a
follow on patch will disable the HSA_ENABLE_IPC_MODE_LEGACY
environment variable.
Change-Id: Ifd8469e9adfc81f8a1ea78d6010fb10b515ba1b4
Implement HybridMutex to improve latencies compared to KernelMutex when
there is contention between several threads calling hsa_signal_create
and hsa_amd_signal_async_handler.
Change-Id: If53377033e749b0050727964c9303f09b02527cc
Fixes issue where t1_ counters may not be updated when doing dispatch
profiling, causing a divide by 0.
Change-Id: I91060ac3f9fd2183d277e6e7cd810398a453a87f
Add query to return flags for GPU agent memory properties and AQL
extensions.
Implement flag to determine that GPU agent is an APU
Change-Id: Ic04c51290b2b9763e14989c117f35a2e22297453
Member added to KFDVersion to report if KFD supports core dump
mechanism. This is done through hsaKmtRuntimeEnable API call while
the topology is being built. It also dictates if core dump will be
generated by either KFD or hsa-runtime.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Change-Id: I2e9d4166563402f78613d728446feb692c52d9d1
This API consists in one function to be called from a fault event at the
hsa-runtime to generate a core dump.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Change-Id: Ib1b90d5beb13f93c4e8ebd21fd61705ebb12ca5d
The alternate scratch memory is used for dispatches that have a low
number of waves but relatively large wave size.
This allows us to keep the tmpring_size.bits.WAVES field of the main
scratch to full occupancy.
Change-Id: I32d240fac4b7d38200d1eebc1b0fdc8a823920d3
For devices where the CP FW supports asynchronous scratch reclaim, ROCr
is able to claw-back scratch memory that was assigned to an AQL queue.
With that ability, ROCr does not have to rely on using USO
(use-scratch-once) when assigning large amounts of memory to a queue.
If we reach a situation where we are running low on device memory, ROCr
will attempt to claw-back the scratch memory.
Change-Id: Iddf8ec84e37ab8b9fdc58bafbe2b61fe2acb6eb7
SDMA4.4 and SDMA5.2+ has increased it's available copy size to 2^30 bytes
represented by exponent as bits set in the COUNT field of the
linear copy.
Also note that the full 2^22 byte limit is available from SDMA4 onwards
as it has corrected the 0x3fffe0 HW limitation from SDMA3.
As copy limit has increase, this can change system performance
so provide env var HSA_ENABLE_SDMA_COPY_SIZE_OVERRIDE=0 to fall
back to the original 0x3fffe0 limit for debugging purposes.
Change-Id: I0fb6e5378f68e5b8a00ff559271691a943ee06ee
Tony Gutierrez
Jonathan Kim
David Yat Sin
Shweta.Khatri
Saleel Kudchadker
David Belanger
Joseph Greathouse
Konstantin Zhuravlyov
pvanhout
Mythreya
Alex Sierra