This patch adds the support for gfx90c apu. So far we treat it as "dgpu" and
gfx900. Will update hsa gfxip table while the isa/llvm is implemented on gfx90c.
Change-Id: I6ef164bf3e751fe6dd6287cac212a500dce84b1a
Signed-off-by: Huang Rui <ray.huang@amd.com>
On NUMA system, allocate queue ctx save restore area on the closest NUMA
node to the GPU which the queue is going to run. This will improve
performance on NUMA system generally by reducing schedule latency and
fix the multi-node rccl-tests unstable performance issue.
If the closest NUMA node has no memory available, set flags NoNUMABind=1
to bypass mbind, to use default NUMA memory policy to allocate system
memory.
Change-Id: Ic62bfa5bb2efbf4f6ae79ff403e9610ddf18d45c
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
on NUMA system, node 0 may have no memory, application pass node id
0 to hsaKmtAllocMemory will fail because mbind to specify the allocation
from node 0 return EINVAL.
Add new flag NoNUMABind for application to pass it to hsaKmtAllocMemory
to skip mbind.
hsaKmtCreateEvent and hsaKmtCreateQueue specify the new flag NoNUMABind
to allocate system memory for event page and CWSR area, don't bind the
system memory to a specific NUMA node.
Change-Id: I854e5a57502c7807c4c5ff2e441d499ae515c309
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Originally reserved 100 SGPRs per wave. Pre-gfx10 needs 102 SGPRs
and gfx10 needs 128 SGPRs. Reserve 128 SGPRs per wave for all ASICs
to simplify calculation.
Also double VGPR register size for gfx908 family
Change-Id: I98b741cbfa051f49ed37ff25d99f851f124be7b6
Signed-off-by: Jay Cornwall <Jay.Cornwall@amd.com>
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
The NodeId parameter is redundant and can be retrieved
from QueueId parameter.
Change-Id: I12853849b868b304bd27633fa7653ba644d69026
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
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>
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>
While this may not be supported in the runtime, the kernel/firmware
support it
Change-Id: I7fe4536a6b3055f39e25f453060e899938645d91
Signed-off-by: Kent Russell <kent.russell@amd.com>
The previous name doorbell_offset is used too extensively throughout
the code and did not reflect the true usage.
Change-Id: I50d33f5c00e82c46cdf4264a78b8f925705bed6a
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Hit queue create failure when do kfdtest with --gtest_repet=-1
fix: 4bb90d04("Remove the use of IS_DGPU()")
Change-Id: I04fa73f90cef13a5517dbaceb89c41dc0f821a79
Signed-off-by: xinhui pan <xinhui.pan@amd.com>
The information can be obtained directly from node id. Also improve the
whole logic for future compatibility.
Change-Id: I130733be4e7930d5953d5e81409905e60c2ec35e
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
There were two doorbells, one embedded in another, which are very confusing.
Change the member variable name to mapping to differentiate them. Also,
rename doorbells_mutex to just mutext for brevity.
Change-Id: Iaa14a1a3ee09449a9089fc1fb39c916fdf32fb44
Signed-off-by: Yong Zhao <yong.zhao@amd.com>
Upstream KFD doesn't support mapping doorbells to GPUVM yet. Fall
back to the old method.
Change-Id: I452a6fc59b88329b833844e3914c480c2f13c82d
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
The main point is to move update_ctx_save_restore_size() out of if()
condition.
Change-Id: I58a1a4f3edca2d1c510fdd0e31e59b5c41e92a14
Signed-off-by: Yong Zhao <yong.zhao@amd.com>
On discrete GPUs place the EOP queue in VRAM. The reader/writer of this
queue is the CP and the size is small. Dispatch latency improves
through lower read latency in AQL completion phase.
Change-Id: Id8351dcddbd21fd7c7d699803c96434c9132db71
Signed-off-by: Jay Cornwall <Jay.Cornwall@amd.com>
Marking heap memory as executable using mprotect() is not allowed
by SELinux. mprotect() calls that try to do this will fail on systems
with SELinux enabled. This is also a security risk, so it should be
fixed even on systems that allow this.
Any memory we want to mark as executable must be allocated using mmap().
See https://www.akkadia.org/drepper/selinux-mem.html
The two places where we try to mark heap memory as executable both use
posix_memalign() to allocate the heap memory. In both cases, the
alignment value passed into this function is always equal to PAGE_SIZE,
which means that they are safe to replace with mmap(), which guarantees
alignment to PAGE_SIZE. In this case PAGE_SIZE has been set to
sysconf(_SC_PAGESIZE);
v2:
- Use MAP_PRIVATE instead of MAP_SHARED. This matches the behavior
of memory allocated by posix_memalign()
- Ignore alignment hints instead of returning error when we can't
accommodate them.
- Drop alignment parameter of allocate_exec_aligned_memory() since
the only alignment supported is sysconf(_SC_PAGESIZE).
- Remove extra parameter from fmm_release().
- Add error path to fmm_allocate_host_cpu() for when mmap fails.
v3:
- Avoid use after free.
Change-Id: I7d51279790d9700bc3fa761c44bfde1c1936019b
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Hardware block testing is done with the workgroup state offset
initialized to the control stack size on all ASICs. MEC microcode
assumes this space is available when the workgroup state offset is
reset after a context restore event.
Fixes context save area overrun when the full save area is used.
Change-Id: I8eeb62f97140c6fe409fe78b4497d833584feea8
Signed-off-by: Jay Cornwall <Jay.Cornwall@amd.com>
ctl_stack_copy is allocated from malloc. It should be freed by free.
Change-Id: Ib924da20200d91f52f106fe173464d47862759a8
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
For items in HsaQueueInfo, control stack information comes from KFD, CU
mask information is maintained in Thunk, and others (queue detail error
and queue type extended) are ignored (value = 0) at this point.
Change-Id: Ib21370b0f52b2bb4ebe6a9b4b6ec6139cccb25ca
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
Use checkpatch.pl to fix the majority of errors. Some that remain and
will be excluded:
Use of typedefs/externs/volatile/sscanf
Lines over 80 characters
Remaining errors are due to misunderstanding the * symbol with typedefs
Also use this opportunity to spell manageable properly
Change-Id: I0b335e9cb3e1eea38bee27eaa1f582b2c9b09b38
Integrate the supported device ID list distributed in topology, queue, and
pmc into one place: topology.
Change-Id: If035cf8e4a6fc6caff6c94ec627647cfb11c3d79
The control stack memory for CWSR is allocate in kernel together with MQD
allocation.
Change-Id: Ib1c0ab9402df3431e9555649394320380d6c6dd8
Signed-off-by: shaoyun.liu <shaoyun.liu@amd.com>
On SOC15 chips, the ABI for the create_queue ioctl is changed to
allow doorbell allocation independent of the queue ID. This is
necessary to accommodate doorbell routing to specific engines in
the BIF.
Change-Id: Ie98d0a758758149dd5fc09ae088afccc29904124
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
On gfx900 we need 64-bit for all doorbells and SDMA WPTRs.
Change-Id: I9b922e16442e967599ae3c928308451d5cc470b3
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
This is in preparation for gfx900, which uses 64-bit doorbells. We
maintain the same number of doorbells per process by making the
doorbell page size bigger.
KFD will need to implement the same rule.
Change-Id: I3c4110869b191b83943b5a390a48edfc94d941d8
events.c and queues.c were accidently changed to 755 by change
fc70f0c30976f4021f7d763bfc10d76a76029553. Change them back.
Change-Id: If51c0b91139afc23e9051cf94c83d61fc20297e6
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Add gfx803 10/11 device IDs that were recently added to KFD.
Change-Id: Id40b117ae47bacedefa6e333fdfdf58dea92cd2d
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
If fork() is called, clear all duplicated data that is invalid in the
child process.
Change-Id: I4e27198060db593c630c6337b7071dfbd0d80b83
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
CWSR buffers can be large on dGPUs (~21MB on gfx803). Allocating them
in VRAM limits the number of queues that can be created unnecessarily.
Also make freeing of per-queue buffers symmetric with allocation. All
buffers are now allocated with allocate_exec_aligned_memory on dGPUs
and APUs, so use free_exec_aligned_memory to free them.
Change-Id: I45e8cb1801857d0268750202cdd422426611e457
gfx802 requires a workaround for a VM TLB bug in which lookups use
the ACTIVE bit of the 8th PTE within any aligned group of 8 PTEs.
Until this is fixed in amdgpu the GPUVM doorbell logic will fail.
Change-Id: I5ec7b1fcd8b7677011a141d27cfc486c45d9a415
Signed-off-by: Jay Cornwall <Jay.Cornwall@amd.com>
Allocate doorbells for dGPUs in the SVM aperture and map them for
GPU access. This is necessary to allow GPU-initiated submissions to
user mode queues.
Depends on new doorbell BO allocation flag in KFD.
Change-Id: I0737bef4a4764bb4a66c43846707ead2108f6601
Huang Rui
Philip Yang
shaoyunl
Yong Zhao
Jay Cornwall
Oak Zeng
Philip Cox
Kent Russell
Gang Ba
xinhui pan
Felix Kuehling
Laurent Morichetti
Tom Stellard
Amber Lin
Harish Kasiviswanathan