Skip to contentSkip to conversation

drm/amdgpu/gfx: keep all compute queues on the same pipe

PATCH In review Alex Deucher 3 participants 4 messages Jul 11, 2017 – Jul 13, 2017 3 messages not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdgpu/gfx: keep all compute queues on the same pipe

Spreading them causes performance regressions using compute queues.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2 -1
index e26108a..4f6c68f 100644
@@ -125,7 +125,8 @@ void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
 		if (mec >= adev->gfx.mec.num_mec)
 			break;
 
-		if (adev->gfx.mec.num_mec > 1) {
+		/* FIXME: spreading the queues across pipes causes perf regressions */
+		if (0) {
 			/* policy: amdgpu owns the first two queues of the first MEC */
 			if (mec == 0 && queue < 2)
 				set_bit(i, adev->gfx.mec.queue_bitmap);
Signature · 6 lines
2.5.5

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

A message that is not in the archive (sent off-list or not yet received): 1499785992-14693-1-git-send-email-alexander.deucher-5C7GfCeVMHo@public.gmane.org

#2 lore raw

Hey Alex,

Which apps are having perf problems?

Also, is the issue present when the queue gets mapped to pipe0? (i.e. it the perf regression only happens on pipe1+)

Regards,
Andres

Quoted 15 lines · Alex Deucher

Spreading them causes performance regressions using compute queues.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +0 -0
index e26108a..4f6c68f 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev) · line 132

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +2 -1
@@ -125,7 +125,8 @@ void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
  		if (mec >= adev->gfx.mec.num_mec)
  			break;
 
-		if (adev->gfx.mec.num_mec > 1) {
+		/* FIXME: spreading the queues across pipes causes perf regressions */
+		if (0) {
  			/* policy: amdgpu owns the first two queues of the first MEC */
  			if (mec == 0 && queue < 2)
  				set_bit(i, adev->gfx.mec.queue_bitmap);

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

A message that is not in the archive (sent off-list or not yet received): 20625ee9-1944-63af-a2c9-22a82375042c-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#4 lore raw
Quoted 6 lines · Andres Rodriguez

Hey Alex,

Which apps are having perf problems?

Also, is the issue present when the queue gets mapped to pipe0? (i.e. it the perf regression only happens on pipe1+)

It's a couple of OCL mandlebrot tests. We are only seeing the regression on Polaris 11, but it's a 40% drop. Investigating further.

Alex

Quoted 25 lines

Regards,
Andres

Quoted 20 lines · Alex Deucher

Spreading them causes performance regressions using compute queues.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +0 -0
index e26108a..4f6c68f 100644
@@ -125,7 +125,8 @@ void amdgpu_gfx_compute_queue_acquire(struct
 amdgpu_device *adev)
                if (mec >= adev->gfx.mec.num_mec)
                        break;
  -             if (adev->gfx.mec.num_mec > 1) {
Quoted 9 lines
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +1 -0
+               /* FIXME: spreading the queues across pipes causes perf

regressions */

+               if (0) {
                        /* policy: amdgpu owns the first two queues of the
first MEC */
                        if (mec == 0 && queue < 2)
                                set_bit(i, adev->gfx.mec.queue_bitmap);

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

A message that is not in the archive (sent off-list or not yet received): CADnq5_Neb2bXjj=0ALr6_ZciSMFaqEGPRd1Wz6Sx=OENVq3Huw-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#6 lore raw

Patch Acked-by: Jim Qu <[email protected]>

Thanks
JimQu

________________________________________
发件人: Alex Deucher <[email protected]>
发送时间: 2017年7月12日 23:56
收件人: Andres Rodriguez
抄送: amd-gfx list; Deucher, Alexander; Qu, Jim
主题: Re: [PATCH] drm/amdgpu/gfx: keep all compute queues on the same pipe

Quoted 6 lines · Andres Rodriguez

Hey Alex,

Which apps are having perf problems?

Also, is the issue present when the queue gets mapped to pipe0? (i.e. it the perf regression only happens on pipe1+)

It's a couple of OCL mandlebrot tests. We are only seeing the regression on Polaris 11, but it's a 40% drop. Investigating further.

Alex

Quoted 25 lines

Regards,
Andres

Quoted 20 lines · Alex Deucher

Spreading them causes performance regressions using compute queues.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +0 -0
index e26108a..4f6c68f 100644
@@ -125,7 +125,8 @@ void amdgpu_gfx_compute_queue_acquire(struct
 amdgpu_device *adev)
                if (mec >= adev->gfx.mec.num_mec)
                        break;
  -             if (adev->gfx.mec.num_mec > 1) {
Quoted 9 lines
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c +1 -0
+               /* FIXME: spreading the queues across pipes causes perf

regressions */

+               if (0) {
                        /* policy: amdgpu owns the first two queues of the
first MEC */
                        if (mec == 0 && queue < 2)
                                set_bit(i, adev->gfx.mec.queue_bitmap);

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