GC 9.4.3 to set gfx target version to 9.4.x dependent on revision and
capabilities. Due to this, where applicable, mask off the gfx target
stepping version and only check major/minor version (9.4). There are no
collisions due to this change since GC 9.4.3 is the only ASIC that uses
gfx target version 9.4.x.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I72803e594c421f054d18ccfa7e92c507128fa5be
Use q->total_mem_alloc_size for munmap in SVM codepath of free_queue.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I2fecaa1ddb337b1fe71f9cbba45a0c9467eff0c0
Currently, on queue destroy, context save restore memory is freed
only for a single XCC. Instead, we need to free the entire context
save restore memory, which was allocated for all XCCs.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Change-Id: I51ebb12fa8d5ebed41979d68e74f7c5392dca062
Do not allocate the EOP buffer when not required.
Signed-off-by: David Belanger <david.belanger@amd.com>
Change-Id: I1664a3f0a882219a72278174006cdb8d46fd4f5e
Allocate debug area big enough for all XCCs in the partition. Also, fix
the cu_num calculations as driver now reports cu_num as the total number
of CUs in the partition.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Change-Id: I6e80d57196b770bb3c2506bc58cb366c0046084b
Add gfx version for VGPR size per CU calc, add FAMILY_AV to KfdFamilyId,
add blacklist filter to kfdtest.exclude.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I9b8072e45f4d497e0a8fd3f8f97f1425238e8b42
Fixed VGPR memory size, size was too small for some GPU, causing a memory overflow.
Refactored macro code into a function.
Thanks to Jay Cornwall for locating the problem and proposing the fix.
Change-Id: Iffedea1c4f341967f02c56d810ff048225b02c16
Signed-off-by: David Belanger <david.belanger@amd.com>
This is a temporary work around for GPU hang issues observed on GFX11.
Change-Id: I98fbedbbd1c51fe402c2116b35ca548931a390c9
Signed-off-by: David Belanger <david.belanger@amd.com>
When is hsaKmtCreateQueue called first time for node
doorbells[NodeId].size is initialized to zero in init_process_doorbells
but used to calculate the doorbell offset. It works just by accident
because doorbells[NodeId].size is uint32_t so -1 will be 0xFFFFFFFF which
is zero extended into 0x00000000FFFFFFFF and it will work as long as mmap
offset bits are not within lower 32 bits.
Bug: https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/issues/78
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: Ia791adfc51363d4704cb50fa4f01137b7dd48a75
fork process copy-on-write MMU nitifier on CWSR range will evict user
queues, and then update GPU mapping and resume queues, use MADV_DONTFORK
to avoid COW MMU notifier callback on CWSR SVM range.
Use mmap to alloc SVM range for CWSR because posix_memalign don't alloc
new range in child process, this fails to register svm range as range is
invalid address in forked child process.
Change-Id: Ibaea56a691dd6f577ed2e1f2d43f4a3500b8316f
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
GFX1036(ISA version) is not included in the previous range.
This patch can really include all gfx10 series ASICs.
Signed-off-by: Aaron Liu <aaron.liu@amd.com>
Change-Id: I0e28dbfc031c216166b306b9fb39f644f75a330f
Remove potential double free condition when free_queue() is called
after hsaKmtDestroyQueue() if mapping doorbell fails during queue
creation.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Change-Id: If2aa19c455b30d2940b232dbafb9cc1eaad721a5
Queue ctx_save_restore memory is allocated with size
ctx_save_restore_size + debug_memory_size, use the same size
in free_queue to free ctx_save_restore memory.
Change-Id: I4902ff15fb82ddea64b8342b89776a1bf5c38d13
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
The debugger depends on the CWSR area being executable. Set the right
flag when registering SVM memory.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Tested-by: Laurent Morichetti <Laurent.Morichetti@amd.com>
Change-Id: I7441e214d1a4da8324d775e777976fabd1c81a6f
To improve performance on queue preemption, allocate ctx s/r
area in VRAM instead of system memory, and migrate it back
to system memory when VRAM is full.
Signed-off-by: Eric Huang <jinhuieric.huang@amd.com>
Change-Id: If775782027188dbe84b6868260e429373675434c
Non-paged allocation for queue memory necessary for binding wptr to
GART. Required to support usermode queue oversubscription with MES on
GFX11.
Change ensures queue memory does not specify ATS.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I10b23b0205c90dad902c711a88cfb5e9b4979617
Currently, context save area size passed to KFD includes the
size of the debug area. Change this to report the actual size
of the context save area to KFD.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Change-Id: I5d440ae802255a97ade046775f6a000bae79d5d5
to faciliate debugging, print errno when queue destroy fail
current log give very little information when fail:
[ RUN ] KFDQMTest.AllSdmaQueues
[ ] Regular SDMA engines number: 1 SDMA queues per engine: 2
[ OK ] KFDQMTest.AllSdmaQueues (11 ms)
[ RUN ] KFDQMTest.AllXgmiSdmaQueues
[ ] XGMI SDMA engines number: 0 SDMA queues per engine: 2
[ OK ] KFDQMTest.AllXgmiSdmaQueues (6 ms)
[ RUN ] KFDQMTest.AllQueues
/home/foreman/build/hsakmt-roct-amdgpu-1.0.9.40500/sources/libhsakmt/tests/kfdtest/src/KFDQMTest.cpp:381: Failure
Value of: (cpQueues[i].Destroy())
Actual: 1
Expected: HSAKMT_STATUS_SUCCESS
Which is: 0
Signed-off-by: Yifan Zhang <yifan1.zhang@amd.com>
Change-Id: I5b1b5b616a5fd7ff198360c893a7aeed685022bd
All conditional logic previously dependent on asic_family now dependent
on gfx version. As such, HSA_FORCE_ASIC_TYPE is now deprecated and
HSA_OVERRIDE_GFX_VERSION should be used instead.
device_info structs containing asic_family, eop_buffer_size, and
doorbell_size removed (alongside dev_lookup_table[]). eop_buffer_size
and doorbell_size now determined via macro, with eop_buffer_size and
full gfx version field being added to queue struct.
Certain funcs previously dependent on DID or GPU ID now dependent on
Node ID (due to use of get_gfxv_by_node_id()).
Fixes sleeper bug where ctl_stack_size and debug_memory_size were being
incorrectly programmed on gfx90a.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: Iaabdc5c73920ad83b4b379cd6086992357b8ff10
Expand check control stack size for Beige Goby, Van Gogh and Yellow Carp.
Signed-off-by: Aaron Liu <aaron.liu@amd.com>
Change-Id: I6c8345bbb751eb9179fc740fe1674504b4295a5f
Update hsaKmtCreateQueue to initialize the new save area header with the
exception payload and event ID.
Signed-by-off: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Sean Keely <sean.keely@amd.com>
Change-Id: Icd38062dc982cb29b30644699014eeb0b3e26d00
This change is for the A+A bring-up branch as it needs to made more
generic to handle all ASICs.
For A+A all the system buffers are mapped as NC (non coherent) unless
explicitly marked as UC (uncached). The coherency is then expected to be
handled by shader by explicitly using acquire/release instructions.
However, CP doesn't have same feature. The buffers used by CP thus have
to UC. For now queue buffer and Signal handler memory is marked as UC.
This change shouldn't affect other ASICs since Uncached flag is not used
in those. However, this change still need to be made more generic.
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Change-Id: I56c37a809913f7f08c94d01b0572d0f4864939aa
On gfx9, the maximum number of wavefronts per queue is the minimum of
40 waves per compute units, or 512 waves per shader engine. On gfx10,
there can only be 32 waves per compute units.
Signed-off-by: Laurent Morichetti <laurent.morichetti@amd.com>
Change-Id: I148d1a4fe6c07cdbfaa1f77939eb29311c81c008
Reserve some space in the context save area for the debugger's
use. There should be 32 bytes per wave for a given queue.
Change-Id: I65ddb6123d0f6afd3149844617ad19023009101d
On gfx9, the maximum number of wavefronts per queue is the minimum of
40 waves per compute units, or 512 waves per shader engine. On gfx10,
there can only be 32 waves per compute units.
Signed-off-by: Laurent Morichetti <laurent.morichetti@amd.com>
Change-Id: I148d1a4fe6c07cdbfaa1f77939eb29311c81c008
Reserve some space in the context save area for the debugger's
use. There should be 32 bytes per wave for a given queue.
Change-Id: I65ddb6123d0f6afd3149844617ad19023009101d
The queue control stack size cannot exceed 0x7000 on ASICs
gfx1010 through gfx1031. The lower limit is not achievable
with AQL so this should have no practical effect.
Fixes control stack size overflow on large ASICs.
Change-Id: Ib78cf6e4c5f096044bf8de24debe211689891caa
Signed-off-by: Jay Cornwall <jay.cornwall@amd.com>
KFD already implemented the fallback path for APU. Thunk will use flag
which exposed by kfd to configure is_dgpu instead of hardcode before.
Signed-off-by: Huang Rui <ray.huang@amd.com>
Change-Id: I445f6cf668f9484dd06cd9ae1bb3cfe7428ec7eb
Gfx10 need 12bytes/wave control stack
Change-Id: I6c6f2819572e6b43aa3140d4dbe79d930e4c1c9c
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Jay Cornwall <Jay.Cornwall@amd.com>
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>
Graham Sider
Mukul Joshi
David Belanger
Felix Kuehling
Philip Yang
Aaron Liu
Eric Huang
Sean Keely
Yifan Zhang
chen gong
Jonathan Kim
Chengming Gui
Mike Li
Kent Russell
Harish Kasiviswanathan
Laurent Morichetti
Jay Cornwall
Huang Rui
Oak Zeng
Yong Zhao
shaoyunl