Skip to contentSkip to conversation

Revert "drm/amdgpu: disable gpu-sched load balance for uvd"

PATCH In review Nirmoy Das 3 participants 4 messages Sep 1, 2020 – Sep 2, 2020

Subsystems (2)
lore
Share & tools
lore

Nirmoy Das

author
#0 lore raw

PATCH 1/2 Revert "drm/amdgpu: disable gpu-sched load balance for uvd"

This reverts commit e0300ed8820d19fe108006cf1b69fa26f0b4e3fc.

We should also disable load balance for AMDGPU_HW_IP_UVD_ENC jobs.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +1 -3
index 7cd398d25498..59032c26fc82 100644
@@ -114,9 +114,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
 	scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
 	num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
 
-	if (hw_ip == AMDGPU_HW_IP_VCN_ENC ||
-	    hw_ip == AMDGPU_HW_IP_VCN_DEC ||
-	    hw_ip == AMDGPU_HW_IP_UVD) {
+	if (hw_ip == AMDGPU_HW_IP_VCN_ENC || hw_ip == AMDGPU_HW_IP_VCN_DEC) {
 		sched = drm_sched_pick_best(scheds, num_scheds);
 		scheds = &sched;
 		num_scheds = 1;
Signature · 6 lines
2.28.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
author
#1 lore raw

PATCH 2/2 drm/amdgpu: disable gpu-sched load balance for uvd

On hardware with multiple uvd instances, dependent uvd jobs may get scheduled to different uvd instances. Because uvd jobs retain hw context, dependent jobs should always run on the same uvd instance. This patch disables gpu scheduler's load balancer for a context that binds jobs from same the context to a uvd instance.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +4 -1
index 59032c26fc82..fc392dfd1789 100644
@@ -114,7 +114,10 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
 	scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
 	num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
 
-	if (hw_ip == AMDGPU_HW_IP_VCN_ENC || hw_ip == AMDGPU_HW_IP_VCN_DEC) {
+	if (hw_ip == AMDGPU_HW_IP_UVD ||
+	    hw_ip == AMDGPU_HW_IP_UVD_ENC ||
+	    hw_ip == AMDGPU_HW_IP_VCN_ENC ||
+	    hw_ip == AMDGPU_HW_IP_VCN_DEC) {
 		sched = drm_sched_pick_best(scheds, num_scheds);
 		scheds = &sched;
 		num_scheds = 1;

--
2.28.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

maintainer
#2 lore raw
In reply to Nirmoy Das

Am 01.09.20 um 21:49 schrieb Nirmoy Das:

Quoted 3 lines

This reverts commit e0300ed8820d19fe108006cf1b69fa26f0b4e3fc.

We should also disable load balance for AMDGPU_HW_IP_UVD_ENC jobs.

Well revert and re-apply is usually not the best option. Just provide a delta patch and Alex might decide to squash it into the original one during upstreaming.

Christian.

Quoted 11 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 4 +---
  1 file changed, 1 insertion(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +0 -0
index 7cd398d25498..59032c26fc82 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip, · line 120

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +1 -3
@@ -114,9 +114,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip,
  	scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
  	num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
 
-	if (hw_ip == AMDGPU_HW_IP_VCN_ENC ||
-	    hw_ip == AMDGPU_HW_IP_VCN_DEC ||
-	    hw_ip == AMDGPU_HW_IP_UVD) {
+	if (hw_ip == AMDGPU_HW_IP_VCN_ENC || hw_ip == AMDGPU_HW_IP_VCN_DEC) {
  		sched = drm_sched_pick_best(scheds, num_scheds);
  		scheds = &sched;
  		num_scheds = 1;

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#3 lore raw
Quoted 6 lines · Christian König

Am 01.09.20 um 21:49 schrieb Nirmoy Das:

Quoted 3 lines

This reverts commit e0300ed8820d19fe108006cf1b69fa26f0b4e3fc.

We should also disable load balance for AMDGPU_HW_IP_UVD_ENC jobs.

Well revert and re-apply is usually not the best option. Just provide a delta patch and Alex might decide to squash it into the original one during upstreaming.

I wasn't sure how to handle that. Thanks,  I will send a delta patch.

Regards,

Nirmoy

Quoted 9 lines

Christian.

Quoted 6 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 4 +---

  1 file changed, 1 insertion(+), 3 deletions(-)
Quoted 17 lines
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +0 -0
index 7cd398d25498..59032c26fc82 100644
@@ -114,9 +114,7 @@ static int amdgpu_ctx_init_entity(struct 
 amdgpu_ctx *ctx, u32 hw_ip,
       scheds = adev->gpu_sched[hw_ip][hw_prio].sched;
       num_scheds = adev->gpu_sched[hw_ip][hw_prio].num_scheds;
Quoted 8 lines

  -    if (hw_ip == AMDGPU_HW_IP_VCN_ENC ||
-        hw_ip == AMDGPU_HW_IP_VCN_DEC ||
-        hw_ip == AMDGPU_HW_IP_UVD) {
+    if (hw_ip == AMDGPU_HW_IP_VCN_ENC || hw_ip == AMDGPU_HW_IP_VCN_DEC) {
          sched = drm_sched_pick_best(scheds, num_scheds); scheds = &sched;
          num_scheds = 1;

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx