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
Add handler to handle HW exception events reported by underlying
drivers. These events are generally caused by GPU resets and need the
application to abort.
As an improvement, in the future, we can provide additional information
about the exception (e.g mode-reset level)
Change-Id: If3fb5f19f9fce181a9d3b5e34a5506725856e7b0
An AQL packet header field is stored using an atomic release, and needs
to be read using atomic acquire if it may be written by another thread.
Change-Id: I1d75587fd93f9c6216deebffc9a627b404a7e749
Define AMD_AQL_FORMAT_INTERCEPT_MARKER AMD vendor AQL packet. Add
support to intercept queue to invoke a callback for these packets.
Change-Id: Ia58d5fe2171f563632b4edd6343e02585f49d149
It is possible that packet rewriting an initial packet for the intercept
queue produces more packets that the size of the wrapped queue. The code
would never submit the such a set of packets as it attempted to submit
all or none. This can result in an infinite loop.
This is corrected to submit what will fit if the rewrite is larger than
the wrapped queue.
Change-Id: I8f03228c2e15151287e25de46eaee998f829c62a
Describe the assumption being made when checking if there is a retry
barrier packet on the queue. Also enforce the consequential requirement
of the minimum queue size.
Change-Id: I0efaffc5a79b9e2fdab3655b8b74270118a5c2ff
Using new ExtendedCoherent KFD HSA memory flag to achieve system
scope coherence on atomic instructions. Non-compliant systems may
have the need to perform explicit HDP flushes to achieve system
scope coherence using this flag.
Change-Id: Ic6b47c0e97285086fa1f52bbfa4597b81cadafeb
Silence warnings on more stringent compile checks for lack of override
declaration.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Change-Id: Iaa54dfc3dd74f5ee55763cafbbcf2db73493bb21
On busy systems, the memory allocation can take long duration and
increase calls to hsa_signal_create/hsa_amd_signal_create. This
mitigates this issue.
Change-Id: Ib7640273262ebc3dbf1f07049ce5da10b1d6b158
Some GFX9 devices will drop commands if ring buffer submission is less
than 64 DWORDs. Pad submission with a NOP head an trailing null
DWORDs in this case.
Change-Id: I850af490fb699f7efe8aef96d97c600a8e76516b
When oversubscribing SDMA gangs, a circular deadlock can occur since
gang enqueue is staggered with respect to SDMA engine leader based
on source to destination.
As a result, an enqueued leader may be waiting on a gang item that is
waiting on another enqueued leader or gang item and so on.
To prevent this, first lock the submission to ensure dma status query
and submissions are atomic. Once this is in place, be more stringent
with ganging in that all SDMA engines must be available in order to gang.
Finally, re-enable SDMA ganging by default.
Change-Id: I4511e3487db9d26475b5aece4897f10168cc5322
Use all available SDMA engines capped by xGMI bandwith for
all D2D copies within a hive.
By default, set the latency boundary copy size as 4KB and below.
Any copy size in within this boundary will not gang.
Avoid oversubscribing engines by not ganging on engines with
pending non-ganged work.
An enviroment variable HSA_ENABLE_SDMA_GANG has been provided
to override default ganging behaviour.
Change-Id: Iccde76aa1af1d47ea2a151789432c9db4f0ffa8d
Add agent info query to return nearest CPU agent. This can be used to
determine which CPU agent is in the same NUMA region as the GPU agent.
Change-Id: I5400b4347ffbf4d2a836df31c4de443a38b0ecd1
Aqlpacket:IsValid() function: Replaced bitwise AND operator (&) with the logical
AND operator (&&) when evaluating AQL packet type
Change-Id: I59980bc206cc7eff424023fff0bb92b618aa8c70
Support function to retain allocation handle for memory mappings.
The get allocation properties function will return the current
allocation properties for existing memory mappings.
This is part of patch series for Virtual Memory API.
Change-Id: I0a53a11b6efc2b5bf9d463512a489a2abd812551
Support exporting and importing dmabuf file descriptors for memory
mappings. The exported dmabuf file descriptors are shareable posix
file descriptors that can be used for cross-vendor, cross-device
and cross-process memory sharing.
This is part of patch series for Virtual Memory API.
Change-Id: I3673fc009f7e73bc26be8349e19f66e20d0607c5
Mapping memory handles to virtual memory addresses do not make them
accessible. The set access function is needed to make the memory
mappings accessible to specific agents. The get access function
returns current access properties for individual agents.
This is part of patch series for Virtual Memory API.
Change-Id: I152ba0557fd2a802eb9d840568b68cdd1911b72c
Add support for mapping and unmapping memory handles to virtual
address ranges.
This is part of patch series for Virtual Memory API.
Change-Id: If512d49ff4211e68f2064249add607a3200e458a
Add support for creating and releasing memory handles. Memory
handles are memory allocations on device memory without a virtual
address.
This is part of patch series for Virtual Memory API.
Change-Id: I5dfb162eb1661621cce171b2870a3c93b24d840e
Add support for reserving virtual address ranges. Virtual address
ranges are addresses without any memory backing. These address ranges
need to be mapped to memory handles later.
This is part of patch series for Virtual Memory API.
Change-Id: I5d066e7421d6896f933f524312afc230a13d594e
Change initialize libdrm device and file descriptor initialization
to use new APIs from Thunk. Libdrm recommends that we re-use the same
file descriptor thoughout the life of a process instead of re-creating
new one each time.
This is part of patch series for Virtual Memory API.
Change-Id: I1c0b8d1bd660cd25478b5f94c84071b90d93fc6c
Checks whether version of libdrm library installed on current
system supports the amdgpu_device_get_fd API. This API is
required to support the virtual memory API functions. The
amdgpu_device_get_fd function was introduced in libdrm-2.4.109.
Using a runtime check test instead of static dependency to be
able to support previous APIs on older versions of libdrm.
Add query for virtual memory API support.
This is part of patch series for Virtual Memory API.
Change-Id: Iec831eb24b5d1689c392e50ae86f4d52d4870ac4
Unless SDMA blits have actually been used for copies, prevent the DMA
copy status from querying the blit's pending byte status to avoid
creating an unnecessary HW queue.
Change-Id: Ied1fbed73c08f0408f0e3583f9b56f2768c71708
Querying pending bytes on a blit kernel is unnecessary when runtime
runs out of SDMA resource since we are returning an SDMA availabilty
mask.
Change-Id: I347efba0c85b70ea3ba8749d76a499afc23909e8
Added HSA_AMD_MEMORY_POOL_GLOBAL_FLAG_EXT_SCOPE_FINE_GRAINED flag to enable extended scope memory region
where the device-scope atomics act as system-scope atomics
Change-Id: I79fc3207cb630dfc68bed2f8aabd75f35fe80b12
KFD kernel version 1.13 starts to support event age
tracking which help elimating unncessary busy wait.
Change-Id: Ib447ed6e0350f3110a4d6b9b80a0388000dd0e72
Signed-off-by: James Zhu <James.Zhu@amd.com>
Earlier, hsa-runtime was unable to find symbols from a stripped ELF-image becasue
no support to find symbols from ".dynsym" section.
Looking for symbols in .dynsym is enabled by LOADER_USE_DYNSYM=1
environment variable
Change-Id: I4f0e8dd0eb053a6066d4d49b670c52e51149531a
Adding support for gfx941 and gfx942 ISAs.
gfx940 ISA will use sc0:1 sc1:1 on load/store operations
gfx942 ISA will use default load/store operations
Change-Id: If1efbef86f59e2cf2d48fe359cd4166405a0a579
When compiling in ASAN mode, remap the first page of device allocations
to system memory. ASAN's memory allocator uses a small amount of extra
memory to store data for housekeeping purpose. But because this memory
is from the GPU memory pool, it might have uncommon memory type for host
to access. Mapping this section of memory to the host makes this memory
accessible to ASAN.
Change-Id: I36f659d616a4d15558372592439a8723c5c84a69
Signed-off-by: Bing Ma <Bing.Ma@amd.com>
David Yat Sin
Jonathan Kim
Mythreya
Alex Sierra
Tony Tye
Shweta Khatri
James Zhu
raghavmedicherla