On 10/28/25 23:06, Timur Kristóf wrote:
> Add an extra 16M (4096 pages) to the GART before GTT.
> This space is going to be used for the VCE VCPU BO.
>
> Split this into a separate patch to make it easier to bisect,
> in case there are any errors in the future.
>
> Signed-off-by: Timur Kristóf <[email protected]>
> ---
> drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
> b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
> index 499dfd78092d..bfeb60cfbf62 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v6_0.c
> @@ -214,6 +214,9 @@ static void gmc_v6_0_vram_gtt_location(struct
> amdgpu_device *adev,
> amdgpu_gmc_set_agp_default(adev, mc);
> amdgpu_gmc_vram_location(adev, mc, base);
> amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_LOW);
> +
> + /* Add space for VCE's VCPU BO so that VCE1 can access it. */
> + mc->num_gart_pages_before_gtt += 4096;
4096*4KiB=16MiB. Do we really need so much?
> }
>
> static void gmc_v6_0_mc_program(struct amdgpu_device *adev)
> @@ -338,7 +341,7 @@ static int gmc_v6_0_mc_init(struct amdgpu_device *adev)
> case CHIP_TAHITI: /* UVD, VCE do not support GPUVM */
> case CHIP_PITCAIRN: /* UVD, VCE do not support GPUVM */
> case CHIP_OLAND: /* UVD, VCE do not support GPUVM */
> - adev->gmc.gart_size = 1024ULL << 20;
> + adev->gmc.gart_size = 1040ULL << 20;
Ideally that should be a power of two.
We can in theory increase it in units of 2MiB without wasting memory, but I'm
not 100% sure if that is actually tested everywhere.
Regards,
Christian.
> break;
> }
> } else {
