In Vulkan we have extensions to assist with capture in replay in a
world where addresses are returned to the application. This involves
creating buffers at the same VA during replay as they were during
capture.
By itself libdrm_amdgpu already has support for this, but there is
the obvious failure mode that if another buffer is already allocated
at that VA things fail spectacularly. This is an actual issue as
internal buffers, like winsys images or shader binaries also
participate in the same VA allocation.
To avoid this problem applications have to create buffers which
are going to be captured with a flag, and the implementation is to
separate VA allocation for those buffers to reduce the collision risk:
"Implementations are expected to separate such buffers in the GPU address
space so normal allocations will avoid using these addresses. Apps/tools
should avoid mixing app-provided and implementation-provided addresses for
buffers created with VK_BUFFER_CREATE_DEVICE_ADDRESS_CAPTURE_REPLAY_BIT,
to avoid address space allocation conflicts."
This patch implements that by adding a flag for these buffers and allocating
address space from the top of the address range instead of the bottom.
Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Christian König <christian.koenig@amd.com>
This was done with:
nm --dynamic --defined-only build/amdgpu/libdrm_amdgpu.so | \
grep amdgpu_ | \
cut -d' ' -f3 > /tmp/a.txt
while read sym; do
read f func line _ <<<$(cscope -d -L -1 $sym)
if [ ! -z "$f" ]; then
line=$((line-1))
sed -i "${line}s/^/drm_public /" $f
fi
done < /tmp/a.txt
Then the alignment of function arguments were manually fixed all over.
The idea here will be to switch the default visibility to hidden so we
don't export symbols we shouldn't.
Signed-off-by: Lucas De Marchi <lucas.demarchi@intel.com>
Acked-by: Eric Engestrom <eric.engestrom@intel.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
This will prevent any more missing `#include "config.h"` bug, at the
cost of having to recompile some files that didn't need to be when
changing build options.
Signed-off-by: Eric Engestrom <eric.engestrom@imgtec.com>
Reviewed-by: Emil Velikov <emil.velikov@collabora.com>
Return high addresses if requested and available.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
issue: UMD allocates top 4GB, but don't do anything, just reserve top 4GB space,
but the performance of VP13 drops from 162fps to 99fps.
root cause:
our va hole list of vamgr is too long by time going.
fix:
reusing old hole as much as possible can make the list shortest.
result:
performance recovers as non-list path, next patch will remove non-list code path.
Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Adding the extra reservation of the 32bit space to the 64bit manager is
complete nonsense and just a waste of memory and CPU cycles.
Signed-off-by: Christian König <christian.koenig@amd.com>
vamgr is an integral part of amdgpu_device. We don't need to calloc and free it.
This can save CPU time, reduce heap fragmentation.
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
Signed-off-by: Alex Xie <AlexBin.Xie@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
[Grazvydas Ignotas: rebase, correct a typo in commit message]
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
vamgr_32 is an integral part of amdgpu_device. We don't need to calloc and free it.
This can save CPU time, reduce heap fragmentation.
Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Emil Velikov <emil.l.velikov@gmail.com>
Signed-off-by: Alex Xie <AlexBin.Xie@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
[Grazvydas Ignotas: rebase, correct a typo in commit message]
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This patch fixes a use-after-free bug in the vamgr_deinit function.
Signed-off-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
In the function amdgpu_vamgr_find_va() the function would return
without unlocking the mutex if the base_required offset was below
the va managers base offset.
Signed-off-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Each device can have its own vamgr, so make it per device now.
This can fix the failure with multiple GPUs used in one single
process.
v2: rebase
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
The AMDGPU_VA_RANGE_32_BIT flag is added to request VA range in the
32bit address space for amdgpu_va_range_alloc.
The 32bit address space is reserved at initialization time, and managed
with a separate VAMGR as part of the global VAMGR. And if no enough VA
space available in range above 4GB, this reserved range can be used as
fallback.
v2: add comment for AMDGPU_VA_RANGE_32_BIT, and add vamgr to va_range
v3: rebase to Emil's drm_private series
v4: fix one warning
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Make it a generic function independent of the device info.
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-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: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-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: Emil Velikov <emil.l.velikov@gmail.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
The flags is added for extensibility to cover some special requirements
in the future, i.e, request VA range in the first 4GB of address space
Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
amdgpu_va_range_query interface is added so that client can query va
range supported by specific device.
Signed-off-by: Sabre Shao <Sabre.Shao@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Two new interfaces are added to support client request for
allocate virtual address without physical memory committed to.
The virtual address space can be managed by client itself.
Signed-off-by: Sabre Shao <Sabre.Shao@amd.com>
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
base_preferred parameter is added to amdgpu_vamgr_find_va
so UMD can specify preferred va address when allocating.
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
This is the first sub-patch of va interface task, the va task is
about adding more va management interfaces for UMD, by design, the
vamgr should be per-process rather than per-device.
Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
va+size will overflow if va equals to AMDGPU_INVALID_VA_ADDRESS, just
return when hit that condition
Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>
This is the new ioctl wrapper used by the new admgpu driver.
It's primarily used by xf86-video-amdgpu and mesa.
v2: fix amdgpu_drm.h install
v3: Integrate some of the sugestions from Emil:
clean up Makefile.am, configure.ac
capitalize header guards
fix _FILE_OFFSET_BITS with config.h
use drm_mmap/drm_munmap
Remove unused ARRAY_SIZE macro
use shared list implementation
use shared math implementation
use drmGetNodeTypeFromFd helper
v4: remove unused tiling defines
v5: include amdgpu.h in Makefile.am
v6: update amdgpu_drm.h
v7: libdrm.h -> libdrm_macros.h
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Bas Nieuwenhuizen
Lucas De Marchi
Eric Engestrom
Christian König
Andrey Grodzovsky
Chunming Zhou
Michel Dänzer
Alex Xie
Tom St Denis
Jammy Zhou
Emil Velikov
Sabre Shao
Ken Wang
monk.liu
Alex Deucher