Am 2022-05-17 um 20:32 schrieb [email protected]:
Hi,
There is an interesting feature divergence between amdgpu and amdkfd concerning
the shared/private apertures. I restricted my analysis to gfx9, it may differ
with other gfx.
In amdgpu, the pair of aperture start addresses is defined the following way
(in gmc_v9_0.c):
* shared_aperture_start = 0x2000000000000000ULL
* private_aperture_start = 0x1000000000000000ULL
In amdkfd, the pair is reversed (cf kfd_flat_memory.c and
kfd_device_queue_manager_v9.c):
* lds_base (ie shared_aperture_start) = (0x1UL) << 48
* scratch_base (ie private_aperture_start) = (0x2UL) << 48
It's more than just reversed. If I counted 0s correctly, the aperture
start addresses in gmc_v9_0.c expressed the same way would be
shared_aperture_start = 0x2ULL << 60;
private_aperture_start = 0x1ULL << 60;
Is there any reason why those two definitions diverge ?
I don't think there is any deliberate reason for them to diverge. But
there is also no reason for them to be the same. The sh_mem_bases
register is per-VMID, which means these settings can be different for
every address space. The address spaces for graphics and compute are
defined independently of each other, and there was no reason to
coordinate these settings.
Regards,
Felix
Thanks,
RD
