The CONFIG_ prefix is reserved for Kconfig options in Make and CPP
syntax. Various static analysis tools rely on this naming convention
and check if CONFIG_ prefixed symbols are defined Kconfig. Hence add
yet another prefix AMD_ to CONFIG_REG_{BASE,END,SISE} to apply to this
convention and make static analysis tools happy.
Signed-off-by: Valentin Rothberg <valentinrothberg@gmail.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
amdgpu submits both kernel and user fences, but just need one interrupt,
disable user fence interrupt and don't effect user fence.
v2: fix merge error
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
04 is the minor version. API version is ver1.
So let's follow same scheme used on published version at 01.org.
If really needed the minor version a follow-up updated will be
done. But for now we need to move fwd and unblock end users.
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
This is the big pull request for amdgpu, the new driver for VI+ AMD
asics. I currently supports Tonga, Iceland, and Carrizo and also
contains a Kconfig option to build support for CI parts for testing.
All major functionality is supported (displays, gfx, compute, dma,
video decode/encode, etc.). Power management is working on Carrizo,
but is still being worked on for Tonga and Iceland.
* 'drm-next-4.2-amdgpu' of git://people.freedesktop.org/~agd5f/linux: (106 commits)
drm/amdgpu: only support IBs in the buffer list (v2)
drm/amdgpu: add vram_type and vram_bit_width for interface query (v2)
drm/amdgpu: add ib_size/start_alignment interface query
drm/amdgpu: add me/ce/pfp_feature_version interface query
drm/amdgpu add ce_ram_size for interface query
drm/amdgpu add max_memory_clock for interface query (v2)
drm/amdgpu: add hdp flush for gfx8 compute ring
drm/amdgpu: fix no hdp flush for compute ring
drm/amdgpu: add HEVC/H.265 UVD support
drm/amdgpu: stop loading firmware with pm.mutex locked
drm/amdgpu: remove mclk_lock
drm/amdgpu: fix description of vm_size module parameter (v2)
drm/amdgpu: remove all sh mem register modification in vm flush
drm/amdgpu: rename GEM_OP_SET_INITIAL_DOMAIN -> GEM_OP_SET_PLACEMENT
drm/amdgpu: fence should be added to shared slot
drm/amdgpu: sync fence of clear_invalids (v2)
drm/amdgpu: max_pde_used usage should be under protect
drm/amdgpu: fix bug of vm_bo_map (v2)
drm/amdgpu: implement the allocation range (v3)
drm/amdgpu: rename amdgpu_ip_funcs to amd_ip_funcs (v2)
...
This function could return a NULL pointer in case of handle not
present and in case of out of memory conditions however caller
function always returned EINVAL error hiding a possible ENOMEM.
This patch change the function to return the error instead to
be able to propagate the error instead of assuming EINVAL.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
In qxlhw_handle_to_bo we incremented counters twice, one time for release object
and one for reloc_info.
In the main function however reloc_info references was drop much earlier than
release so keeping the pointer only on release is safe and make cleaning
process easier.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
qxl_bo structure has two reference counters, one in the GEM object and
another in the TTM object. The GEM object keep a counter to the TTM object
so when GEM counter reached zero the TTM counter (using qxl_bo_unref) was
decremented. The qxl object is fully freed (both GEM and TTM part are cleaned)
when the TTM counter reach zero.
One issue was that surface idr structure has no owning on qxl_bo objects however
it contains a pointer to qxl_bo object. This caused some nasty race condition
for instance qxl_bo object was reaped even after counter was already zero.
This patch fix these races moving main counter (the one used by qxl_bo_(un)ref)
to GEM object which cleanup routine (qxl_gem_object_free) remove the idr pointer
(using qxl_surface_evict) when the counters are still valid.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Enable format string checks for qxl_io_log and remove resulting warnings
which could lead to memory errors on different platform or just printing
wrong information.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Free resources correctly if function fails
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
This function return handle to allocated release object which is an int.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Only EBUSY error was handled. This could cause code to believe
reserve was successful while it failed.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Is we are not able to get source bo object from handle we free
destination bo object and call cleanup code however destination
object was already inserted in reloc_info array (num_relocs was
already incremented) so on cleanup we free destination again.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
reloc_info[i] is not still initialized in the print statement.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
If the function fails reference counter to the object is not decremented
causing leaks.
This is hard to spot as it happens only on very low memory situations.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Cc: stable@vger.kernel.org
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
If objects are moved back from system memory to VRAM (and spice id
created again) memory is already initialized so we need to set flag
to not clear memory.
If you don't do it after a while using desktop many images turns to
black or transparents.
Signed-off-by: Frediano Ziglio <fziglio@redhat.com>
Cc: stable@vger.kernel.org
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
amdgpu_cs_find_mapping doesn't work without all buffers being validated,
so the TTM validation must be done first.
v2: only use amdgpu_cs_find_mapping for UVD/VCE VM emulation
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Track the type of vram on the board and provide a query for it.
User mode drivers and tools want this information for determining
bandwidth information and form informational purposes.
v2: fix build when CI support is not enabled
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Query the IB alignment requirements from the kernel rather
than hardcoding them in the user mode drivers.
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Provide this information to usermode drivers. We were previously
missing this info.
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Add a query for the CE ram size. User mode drivers
will want to use this to determine how much size
of the cache on the CE.
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewd-by: Jammy Zhou <Jammy.Zhou@amd.com>
Add a query for the max memory clock.
v2: handle the dpm enabled case properly
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewd-by: Jammy Zhou <Jammy.Zhou@amd.com>
We had forgotten to register the callback.
Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: jammy zhou <jammy.zhou@amd.com>
Loading firmware is a rather complicated process, in the end we add a
dependency between the pm mutex and the mm semaphore which results in
a harmless but annoying error message.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Leave that at the values set during init. No need to update
them repeatedly.
Signed-off-by: monk.liu <monk.liu@amd.com>
Signed-off-by: David Zhang <david1.zhang@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
bo_va may un-initialized, fix it.
Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>
Need to take the lock when accessing this.
Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>
call reservation_object_reserve_shared before amdgpu_bo_fence
Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>
Pass a ttm_placement pointer to amdgpu_bo_create_restricted
add min_offset to amdgpu_bo_pin_restricted. This makes it
easier to allocate memory with address restrictions. With
this patch we can also enable 2-ended allocation again.
v2: fix rebase conflicts
v3: memset placements before using
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
The structure is renamed and moved to amd_shared.h to make
the component independent. This makes it easier to add
new components in the future.
v2: fix include path
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: yanyang1 <young.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
It's causing issues with VMID handling and comparing the
fence value two times actually doesn't make handling faster.
Port of radeon commit "d6d5c5b8364bcc4d52cddc68bcb0a330d2af20f3".
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Enable UVD dpm (dynamic power management) and powergating. UVD dpm dynamically scales the UVD
clocks on demand. Powergating turns off the power to the block when it's not in use.
Signed-off-by: Sonny Jiang <sonny.jiang@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
VCE 3.0 has two indentical instances in the engine, they share
the same registers name in differrent memory block distinguished
by the grbm_gfx_index, we set to master instance after init, it
will dispatch task to slave instance. These two instances will
share the same firmware, but have their own stacks and heaps.
v2: add mutex for using grbm_gfx_index
Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Firmware required BO size changes in terms of ASIC family
Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Deny user and kernel mapping if we said we never want to do so.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Retry the dpcd fetch several times. Some eDP panels
fail several times before the fetch is successful.
bug:
https://bugs.freedesktop.org/show_bug.cgi?id=73530
Ported from radeon.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Use %*ph rather than walking the array.
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Properly protect the state and also handle submission failures.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Otherwise a process can access the GDS data of another process.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
The CE preamble IB can be dropped for the same context
v2: use the flags directly
v3: remove 'CE' for potential preamble usage by other rings
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
check the filp is not robust, and sometimes different contexts may
have same filp value.
v2: check both filp and ctx_id
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
It is required to support fence per context.
v2: add amdgpu_ctx_get/put
v3: improve get/put
v4: squash hlock fix
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Set up the CP and SDMA for proper unaligned memory access.
Required for OpenCL 2.x
v2: udpate commit message
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
The existing locks were protecting the list, but not the elements.
v2: rename hlock to lock
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
If ctx was released between put and get, then "get" would crash.
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
sg could point to array of contigiouse page*, only free page could lead
to memory leak.
v2: use iterator
Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This flag isn't used by user mode drivers, remove it to avoid
confusion. And rename GTT_WC to GTT_USWC to make it clear.
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Avoids possibility that info may leak via the uninitialized
_pad element.
Noticed-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Need to unlock the mutex on error.
Noticed-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
this could prevent mis-understanding, because libdrm side will consider
no bo_list created if handleis zero
Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Still need to handle ibs BO and validate them even bo_list is NULL
Signed-off-by: Monk.Liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_bo_create() calls amdgpu_ttm_placement_from_domain()
before ttm_bo_init() is called. amdgpu_ttm_placement_from_domain()
uses the ttm bo size to determine when to select top down
allocation but since the ttm bo is not initialized yet the
check is always false. It only took affect when buffers
were validated later. It also seemed to regress suspend
and resume on some systems possibly due to it not
taking affect in amdgpu_bo_create().
amdgpu_bo_create() and amdgpu_ttm_placement_from_domain()
need to be reworked substantially for this to be optimally
effective. Re-enable it at that point.
Ported from radeon commit:
a239118a24
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
This adds initial support for VI asics. This
includes Iceland, Tonga, and Carrizo. Our inital
focus as been Carrizo, so there are still gaps in
support for Tonga and Iceland, notably power
management.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch adds support for CIK parts. These parts
are also supported by radeon which is the preferred
option, so there is a config option to enable support
for CIK parts in amdgpu for testing.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Use readb() and memcpy_fromio() accessors instead.
Ported from radeon commit:
f2c9e560b4
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This adds the non-asic specific core driver code.
v2: remove extra kconfig option
v3: implement minor fixes from Fengguang Wu
v4: fix cast in amdgpu_ucode.c
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This is the main header file for amdgpu.
v2: remove stable comments
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This header defines asic families and attributes.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This header provides the smc message interface for the driver.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This header provides for format for the GCA blocks
clear state (i.e., default state). Each GCA version
has a specific clear state.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
These headers define the atombios table structure and
driver interface.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
These are register headers for the VCE (Video Codec Engine)
block on the GPU.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
These are register headers for the VCE (Video Codec Engine)
block on the GPU.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
These are register headers for the UVD (Universal Video Decoder)
block on the GPU.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
These are register headers for the UVD (Universal Video Decoder)
block on the GPU.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
These are register headers for the UVD (Universal Video Decoder)
block on the GPU.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
These are register headers for the SMU (System Management Unit)
block on the GPU.
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Oded Gabbay