Am 26.05.21 um 15:55 schrieb Felix Kuehling:
Am 2021-05-25 um 1:56 p.m. schrieb Eric Huang:
With XGMI connection flushing HDP on PCIe is unnecessary,
it is also to optimize memory allocation latency.
Signed-off-by: Eric Huang <jinhuieric.huang@xxxxxxx>
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
index ac45d9c7a4e9..aefb3d2927d5 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm_cpu.c
@@ -108,6 +108,8 @@ static int amdgpu_vm_cpu_update(struct
amdgpu_vm_update_params *p,
static int amdgpu_vm_cpu_commit(struct amdgpu_vm_update_params *p,
struct dma_fence **fence)
{
+ if (p->adev->gmc.xgmi.connected_to_cpu)
+ return 0;
I wonder if this check should be inside hdp_v4_0_flush_hdp instead so it
catches all unnecessary HDP flushing. On the other hand, that would
still leave the mb(). But that mb() is probably needed anyway to ensure
that the GPU will see any previous memory writes.
Yes, that was also my concern. But it looks like my reply never made it
to the mailing list.
I think we should have a function in amdgpu_device.c to decide if HDP
flush/Invalidate are necessary or not.
This should then be used at all the different places where we trigger
this (CS, GART and here).
Regards,
Christian.
Regards,
Felix
/* Flush HDP */
mb();
amdgpu_asic_flush_hdp(p->adev, NULL);
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
