Exactly that, yes. The idea is that we want to create at least some
pressure on VRAM when VRAM|GTT is specified.
Essentially we would like to have that for CS as well, but it's just
hard to find the right balance between no pressure and the full pressure
of VRAM placement for VRAM|GTT BOs.
It's on my todo list to tackle that and I'm currently working on an
userspace tool to better simulate those setups, but it's just a side
project.
Christian.
Am 17.09.21 um 09:35 schrieb Pan, Xinhui:
[AMD Official Use Only]
Why? just to evict some inactive vram BOs?
------------------------------------------------------------------------
*From:* Koenig, Christian <[email protected]>
*Sent:* Friday, September 17, 2021 3:06:16 PM
*To:* Pan, Xinhui <[email protected]>; [email protected]
<[email protected]>
*Cc:* Deucher, Alexander <[email protected]>
*Subject:* Re: [PATCH] drm/amdgpu: Let BO created in its allowed_domain
Am 17.09.21 um 09:04 schrieb xinhui pan:
> For any user BO, it is allowed to alloc memory in both GTT and VRAM
> domain. That means one VRAM BO could be eviced to GTT.
>
> Now if device is already lack of VRAM, this BO could alloc memory from
> GTT directly without evicting other VRAM BO. In this way we can reduce
> memory evictions.
NAK, that is intentional pressure on VRAM during allocation of new memory.
Christian.
>
> Signed-off-by: xinhui pan <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index 66bb8a53bb20..9a547bb38cda 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -579,7 +579,7 @@ int amdgpu_bo_create(struct amdgpu_device *adev,
> AMDGPU_GEM_DOMAIN_GDS))
> amdgpu_bo_placement_from_domain(bo,
AMDGPU_GEM_DOMAIN_CPU);
> else
> - amdgpu_bo_placement_from_domain(bo, bp->domain);
> + amdgpu_bo_placement_from_domain(bo, bo->allowed_domains);
> if (bp->type == ttm_bo_type_kernel)
> bo->tbo.priority = 1;
>
