Yes, you need to do this for the SDMA as well but in general that looks
like the idea I had in mind as well.
I would do it like this:
1. Change the special case when you only get one scheduler for an entity
to drop the pointer to the scheduler list.
This way we always use the same scheduler for the entity and can
pass in the array on the stack.
2. Change all callers which use more than one scheduler in the list to
pass in pointers which are not allocated on the stack.
This obviously also means that we build the list of schedulers for
each type only once during device init and not for each context init.
3. Make the scheduler list const and drop the kcalloc()/kfree() from the
entity code.
Regards,
Christian.
Am 08.12.19 um 20:57 schrieb Nirmoy:
On 12/6/19 8:41 PM, Christian König wrote:
Am 06.12.19 um 18:33 schrieb Nirmoy Das:
entity should not keep copy and maintain sched list for
itself.
That is a good step, but we need to take this further.
How about something like ?
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
index 0ae0a2715b0d..a71ee084b47a 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
@@ -269,8 +269,10 @@ struct amdgpu_gfx {
bool me_fw_write_wait;
bool cp_fw_write_wait;
struct amdgpu_ring gfx_ring[AMDGPU_MAX_GFX_RINGS];
+ struct drm_gpu_scheduler *gfx_sched_list[AMDGPU_MAX_GFX_RINGS];
unsigned num_gfx_rings;
struct amdgpu_ring compute_ring[AMDGPU_MAX_COMPUTE_RINGS];
+ struct drm_gpu_scheduler
*compute_sched_list[AMDGPU_MAX_COMPUTE_RINGS];
unsigned num_compute_rings;
struct amdgpu_irq_src eop_irq;
struct amdgpu_irq_src priv_reg_irq;
Regards,
Nirmoy
_______________________________________________
amd-gfx mailing list
amd-gfx@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
