Skip to contentSkip to conversation

drm/amdgpu: cleanup coding style a bit

PATCH In review Christian König 4 participants 9 messages Jul 18, 2018 7 messages not in the archive

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

PATCH 1/3 drm/amdgpu: cleanup coding style a bit

No need to bitcast a boolean and even if we should use "!!" instead.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c | 20 ++++++++++----------
 1 file changed, 10 insertions(+), 10 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +10 -10
index 258b6f73cbdf..e451a3f25beb 100644
@@ -328,35 +328,35 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 		case AMDGPU_HW_IP_GFX:
 			type = AMD_IP_BLOCK_TYPE_GFX;
 			for (i = 0; i < adev->gfx.num_gfx_rings; i++)
-				ring_mask |= ((adev->gfx.gfx_ring[i].ready ? 1 : 0) << i);
+				ring_mask |= adev->gfx.gfx_ring[i].ready << i;
 			ib_start_alignment = 32;
 			ib_size_alignment = 32;
 			break;
 		case AMDGPU_HW_IP_COMPUTE:
 			type = AMD_IP_BLOCK_TYPE_GFX;
 			for (i = 0; i < adev->gfx.num_compute_rings; i++)
-				ring_mask |= ((adev->gfx.compute_ring[i].ready ? 1 : 0) << i);
+				ring_mask |= adev->gfx.compute_ring[i].ready << i;
 			ib_start_alignment = 32;
 			ib_size_alignment = 32;
 			break;
 		case AMDGPU_HW_IP_DMA:
 			type = AMD_IP_BLOCK_TYPE_SDMA;
 			for (i = 0; i < adev->sdma.num_instances; i++)
-				ring_mask |= ((adev->sdma.instance[i].ring.ready ? 1 : 0) << i);
+				ring_mask |= adev->sdma.instance[i].ring.ready << i;
 			ib_start_alignment = 256;
 			ib_size_alignment = 4;
 			break;
 		case AMDGPU_HW_IP_UVD:
 			type = AMD_IP_BLOCK_TYPE_UVD;
 			for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-				ring_mask |= ((adev->uvd.inst[i].ring.ready ? 1 : 0) << i);
+				ring_mask |= adev->uvd.inst[i].ring.ready << i;
 			ib_start_alignment = 64;
 			ib_size_alignment = 64;
 			break;
 		case AMDGPU_HW_IP_VCE:
 			type = AMD_IP_BLOCK_TYPE_VCE;
 			for (i = 0; i < adev->vce.num_rings; i++)
-				ring_mask |= ((adev->vce.ring[i].ready ? 1 : 0) << i);
+				ring_mask |= adev->vce.ring[i].ready << i;
 			ib_start_alignment = 4;
 			ib_size_alignment = 1;
 			break;
@@ -365,27 +365,27 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 			for (i = 0; i < adev->uvd.num_uvd_inst; i++)
 				for (j = 0; j < adev->uvd.num_enc_rings; j++)
 					ring_mask |=
-					((adev->uvd.inst[i].ring_enc[j].ready ? 1 : 0) <<
-					(j + i * adev->uvd.num_enc_rings));
+					adev->uvd.inst[i].ring_enc[j].ready <<
+					(j + i * adev->uvd.num_enc_rings);
 			ib_start_alignment = 64;
 			ib_size_alignment = 64;
 			break;
 		case AMDGPU_HW_IP_VCN_DEC:
 			type = AMD_IP_BLOCK_TYPE_VCN;
-			ring_mask = adev->vcn.ring_dec.ready ? 1 : 0;
+			ring_mask = adev->vcn.ring_dec.ready;
 			ib_start_alignment = 16;
 			ib_size_alignment = 16;
 			break;
 		case AMDGPU_HW_IP_VCN_ENC:
 			type = AMD_IP_BLOCK_TYPE_VCN;
 			for (i = 0; i < adev->vcn.num_enc_rings; i++)
-				ring_mask |= ((adev->vcn.ring_enc[i].ready ? 1 : 0) << i);
+				ring_mask |= adev->vcn.ring_enc[i].ready << i;
 			ib_start_alignment = 64;
 			ib_size_alignment = 1;
 			break;
 		case AMDGPU_HW_IP_VCN_JPEG:
 			type = AMD_IP_BLOCK_TYPE_VCN;
-			ring_mask = adev->vcn.ring_jpeg.ready ? 1 : 0;
+			ring_mask = adev->vcn.ring_jpeg.ready;
 			ib_start_alignment = 16;
 			ib_size_alignment = 16;
 			break;
Signature · 6 lines
2.14.1

_______________________________________________
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): 20180718125553.72259-1-christian.koenig-5C7GfCeVMHo@public.gmane.org

#2 lore raw

PATCH 2/3 drm/amdgpu: expose only the first UVD instance for now

Going to completely rework the context to ring mapping with Nayan's GSoC work, but for now just stopping to expose the second UVD instance should do it.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 13 +++++--------
 drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c |  9 ++-------
 2 files changed, 7 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -8
index e451a3f25beb..75da3c41f3b3 100644
@@ -286,7 +286,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 	struct drm_crtc *crtc;
 	uint32_t ui32 = 0;
 	uint64_t ui64 = 0;
-	int i, j, found;
+	int i, found;
 	int ui32_size = sizeof(ui32);
 
 	if (!info->return_size || !info->return_pointer)
@@ -348,8 +348,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 			break;
 		case AMDGPU_HW_IP_UVD:
 			type = AMD_IP_BLOCK_TYPE_UVD;
-			for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-				ring_mask |= adev->uvd.inst[i].ring.ready << i;
+			ring_mask |= adev->uvd.inst[0].ring.ready;
 			ib_start_alignment = 64;
 			ib_size_alignment = 64;
 			break;
@@ -362,11 +361,9 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
 			break;
 		case AMDGPU_HW_IP_UVD_ENC:
 			type = AMD_IP_BLOCK_TYPE_UVD;
-			for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-				for (j = 0; j < adev->uvd.num_enc_rings; j++)
-					ring_mask |=
-					adev->uvd.inst[i].ring_enc[j].ready <<
-					(j + i * adev->uvd.num_enc_rings);
+			for (i = 0; i < adev->uvd.num_enc_rings; i++)
+				ring_mask |=
+					adev->uvd.inst[0].ring_enc[i].ready << i;
 			ib_start_alignment = 64;
 			ib_size_alignment = 64;
 			break;
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +2 -7
index ea9850c9224d..d8357290ad09 100644
@@ -66,8 +66,6 @@ static int amdgpu_identity_map(struct amdgpu_device *adev,
 			       u32 ring,
 			       struct amdgpu_ring **out_ring)
 {
-	u32 instance;
-
 	switch (mapper->hw_ip) {
 	case AMDGPU_HW_IP_GFX:
 		*out_ring = &adev->gfx.gfx_ring[ring];
@@ -79,16 +77,13 @@ static int amdgpu_identity_map(struct amdgpu_device *adev,
 		*out_ring = &adev->sdma.instance[ring].ring;
 		break;
 	case AMDGPU_HW_IP_UVD:
-		instance = ring;
-		*out_ring = &adev->uvd.inst[instance].ring;
+		*out_ring = &adev->uvd.inst[0].ring;
 		break;
 	case AMDGPU_HW_IP_VCE:
 		*out_ring = &adev->vce.ring[ring];
 		break;
 	case AMDGPU_HW_IP_UVD_ENC:
-		instance = ring / adev->uvd.num_enc_rings;
-		*out_ring =
-		&adev->uvd.inst[instance].ring_enc[ring%adev->uvd.num_enc_rings];
+		*out_ring = &adev->uvd.inst[0].ring_enc[ring];
 		break;
 	case AMDGPU_HW_IP_VCN_DEC:
 		*out_ring = &adev->vcn.ring_dec;
Signature · 6 lines
2.14.1

_______________________________________________
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): 20180718125553.72259-2-christian.koenig-5C7GfCeVMHo@public.gmane.org

#4 lore raw

Re: [PATCH 2/3] drm/amdgpu: expose only the first UVD instance for now

Quoted 67 lines · Christian König

Going to completely rework the context to ring mapping with Nayan's GSoC work, but for now just stopping to expose the second UVD instance should do it.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 13 +++++--------
  drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c |  9 ++-------
  2 files changed, 7 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +5 -8
index e451a3f25beb..75da3c41f3b3 100644
@@ -286,7 +286,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
  	struct drm_crtc *crtc;
  	uint32_t ui32 = 0;
  	uint64_t ui64 = 0;
-	int i, j, found;
+	int i, found;
  	int ui32_size = sizeof(ui32);
 
  	if (!info->return_size || !info->return_pointer)
@@ -348,8 +348,7 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
  			break;
  		case AMDGPU_HW_IP_UVD:
  			type = AMD_IP_BLOCK_TYPE_UVD;
-			for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-				ring_mask |= adev->uvd.inst[i].ring.ready << i;
+			ring_mask |= adev->uvd.inst[0].ring.ready;
  			ib_start_alignment = 64;
  			ib_size_alignment = 64;
  			break;
@@ -362,11 +361,9 @@ static int amdgpu_info_ioctl(struct drm_device *dev, void *data, struct drm_file
  			break;
  		case AMDGPU_HW_IP_UVD_ENC:
  			type = AMD_IP_BLOCK_TYPE_UVD;
-			for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-				for (j = 0; j < adev->uvd.num_enc_rings; j++)
-					ring_mask |=
-					adev->uvd.inst[i].ring_enc[j].ready <<
-					(j + i * adev->uvd.num_enc_rings);
+			for (i = 0; i < adev->uvd.num_enc_rings; i++)
+				ring_mask |=
+					adev->uvd.inst[0].ring_enc[i].ready << i;
  			ib_start_alignment = 64;
  			ib_size_alignment = 64;
  			break;
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -4
index ea9850c9224d..d8357290ad09 100644
@@ -66,8 +66,6 @@ static int amdgpu_identity_map(struct amdgpu_device *adev,
  			       u32 ring,
  			       struct amdgpu_ring **out_ring)
  {
-	u32 instance;
-
  	switch (mapper->hw_ip) {
  	case AMDGPU_HW_IP_GFX:
  		*out_ring = &adev->gfx.gfx_ring[ring];
@@ -79,16 +77,13 @@ static int amdgpu_identity_map(struct amdgpu_device *adev,
  		*out_ring = &adev->sdma.instance[ring].ring;
  		break;
  	case AMDGPU_HW_IP_UVD:
-		instance = ring;
-		*out_ring = &adev->uvd.inst[instance].ring;

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c static int amdgpu_identity_map(struct amdgpu_device *adev, · line 86

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +2 -3
+		*out_ring = &adev->uvd.inst[0].ring;
  		break;
  	case AMDGPU_HW_IP_VCE:
  		*out_ring = &adev->vce.ring[ring];
  		break;
  	case AMDGPU_HW_IP_UVD_ENC:
-		instance = ring / adev->uvd.num_enc_rings;
-		*out_ring =
-		&adev->uvd.inst[instance].ring_enc[ring%adev->uvd.num_enc_rings];
+		*out_ring = &adev->uvd.inst[0].ring_enc[ring];

Hi Christian,

If here uses amdgpu_identity_map, I think only first instance's ring is in use.

In current Vega20 situation, should we change to use amdgpu_lru_map or something else to schedule which instance/ring for use ?

James

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c static int amdgpu_identity_map(struct amdgpu_device *adev, · line 89

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -0
  		break;
  	case AMDGPU_HW_IP_VCN_DEC:
  		*out_ring = &adev->vcn.ring_dec;

_______________________________________________
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): e0cd7fc0-d4b6-b903-eda3-e557f3181547-5C7GfCeVMHo@public.gmane.org

#6 lore raw

Re: [PATCH 2/3] drm/amdgpu: expose only the first UVD instance for now

Am 18.07.2018 um 15:40 schrieb James Zhu:

Quoted 86 lines · Christian König
Quoted 73 lines

Going to completely rework the context to ring mapping with Nayan's GSoC work, but for now just stopping to expose the second UVD instance should do it.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 13 +++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c |  9 ++-------

  2 files changed, 7 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -1
index e451a3f25beb..75da3c41f3b3 100644
@@ -286,7 +286,7 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
       struct drm_crtc *crtc;
       uint32_t ui32 = 0;
       uint64_t ui64 = 0;
-    int i, j, found;
+    int i, found;
       int ui32_size = sizeof(ui32);
         if (!info->return_size || !info->return_pointer)
@@ -348,8 +348,7 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
               break;

          case AMDGPU_HW_IP_UVD:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++) -                ring_mask |= adev->uvd.inst[i].ring.ready << i; +            ring_mask |= adev->uvd.inst[0].ring.ready; ib_start_alignment = 64; ib_size_alignment = 64;
              break;

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
@@ -362,11 +361,9 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
               break;

          case AMDGPU_HW_IP_UVD_ENC:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++) -                for (j = 0; j < adev->uvd.num_enc_rings; j++) -                    ring_mask |=
-                    adev->uvd.inst[i].ring_enc[j].ready << -                    (j + i * adev->uvd.num_enc_rings); +            for (i = 0; i < adev->uvd.num_enc_rings; i++) +                ring_mask |=
+                    adev->uvd.inst[0].ring_enc[i].ready << i; ib_start_alignment = 64; ib_size_alignment = 64;
              break;

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -2
index ea9850c9224d..d8357290ad09 100644
@@ -66,8 +66,6 @@ static int amdgpu_identity_map(struct amdgpu_device 
 *adev,
                      u32 ring,
                      struct amdgpu_ring **out_ring)
   {
-    u32 instance;
-
       switch (mapper->hw_ip) {

      case AMDGPU_HW_IP_GFX:
          *out_ring = &adev->gfx.gfx_ring[ring];

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -0
@@ -79,16 +77,13 @@ static int amdgpu_identity_map(struct 
 amdgpu_device *adev,
           *out_ring = &adev->sdma.instance[ring].ring;
           break;

      case AMDGPU_HW_IP_UVD:
-        instance = ring;
-        *out_ring = &adev->uvd.inst[instance].ring; +        *out_ring = &adev->uvd.inst[0].ring; break;

Quoted 8 lines

      case AMDGPU_HW_IP_VCE:
          *out_ring = &adev->vce.ring[ring];
          break;
      case AMDGPU_HW_IP_UVD_ENC:
-        instance = ring / adev->uvd.num_enc_rings;
-        *out_ring =
- &adev->uvd.inst[instance].ring_enc[ring%adev->uvd.num_enc_rings]; +        *out_ring = &adev->uvd.inst[0].ring_enc[ring];

Hi Christian,

Quoted 3 lines

If here uses amdgpu_identity_map, I think only first instance's ring is in use.

In current Vega20 situation, should we change to use amdgpu_lru_map or something else to schedule which instance/ring for use ?

Yes and no. I'm going to completely remove the queue manager and replace it with Nayan's work on the scheduler.

So no more lru map, but instead a function which picks the best ring based on actual load during context creation.

Regards,
Christian.

Quoted 7 lines

James

Quoted 3 lines

          break;
      case AMDGPU_HW_IP_VCN_DEC:
          *out_ring = &adev->vcn.ring_dec;

_______________________________________________
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): dadd17b4-fe49-9ed2-34f4-84d08803364a-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#8 lore raw

Re: [PATCH 2/3] drm/amdgpu: expose only the first UVD instance for now

Quoted 89 lines · Christian König

Am 18.07.2018 um 15:40 schrieb James Zhu:

Quoted 88 lines · Christian König
Quoted 73 lines

Going to completely rework the context to ring mapping with Nayan's GSoC work, but for now just stopping to expose the second UVD instance should do it.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 13 +++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c |  9 ++-------

  2 files changed, 7 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -1
index e451a3f25beb..75da3c41f3b3 100644
@@ -286,7 +286,7 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
       struct drm_crtc *crtc;
       uint32_t ui32 = 0;
       uint64_t ui64 = 0;
-    int i, j, found;
+    int i, found;
       int ui32_size = sizeof(ui32);
         if (!info->return_size || !info->return_pointer)
@@ -348,8 +348,7 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
               break;

          case AMDGPU_HW_IP_UVD:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++) -                ring_mask |= adev->uvd.inst[i].ring.ready << i; +            ring_mask |= adev->uvd.inst[0].ring.ready; ib_start_alignment = 64; ib_size_alignment = 64;
              break;

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
@@ -362,11 +361,9 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
               break;

          case AMDGPU_HW_IP_UVD_ENC:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++) -                for (j = 0; j < adev->uvd.num_enc_rings; j++) -                    ring_mask |=
-                    adev->uvd.inst[i].ring_enc[j].ready << -                    (j + i * adev->uvd.num_enc_rings); +            for (i = 0; i < adev->uvd.num_enc_rings; i++) +                ring_mask |=
+                    adev->uvd.inst[0].ring_enc[i].ready << i; ib_start_alignment = 64; ib_size_alignment = 64;
              break;

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -2
index ea9850c9224d..d8357290ad09 100644
@@ -66,8 +66,6 @@ static int amdgpu_identity_map(struct 
 amdgpu_device *adev,
                      u32 ring,
                      struct amdgpu_ring **out_ring)
   {
-    u32 instance;
-
       switch (mapper->hw_ip) {

      case AMDGPU_HW_IP_GFX:
          *out_ring = &adev->gfx.gfx_ring[ring];

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -0
@@ -79,16 +77,13 @@ static int amdgpu_identity_map(struct 
 amdgpu_device *adev,
           *out_ring = &adev->sdma.instance[ring].ring;
           break;

      case AMDGPU_HW_IP_UVD:
-        instance = ring;
-        *out_ring = &adev->uvd.inst[instance].ring; +        *out_ring = &adev->uvd.inst[0].ring; break;

Quoted 8 lines

      case AMDGPU_HW_IP_VCE:
          *out_ring = &adev->vce.ring[ring];
          break;
      case AMDGPU_HW_IP_UVD_ENC:
-        instance = ring / adev->uvd.num_enc_rings;
-        *out_ring =
- &adev->uvd.inst[instance].ring_enc[ring%adev->uvd.num_enc_rings]; +        *out_ring = &adev->uvd.inst[0].ring_enc[ring];

Hi Christian,

If here uses amdgpu_identity_map, I think only first instance's ring is in use.

Quoted 8 lines
Quoted 1 line

In current Vega20 situation, should we change to use amdgpu_lru_map or something else to schedule which instance/ring for use ?

Yes and no. I'm going to completely remove the queue manager and replace it with Nayan's work on the scheduler.

So no more lru map, but instead a function which picks the best ring based on actual load during context creation.

Regards,
Christian.

Got it. So do we totally remove instance and ring concept out of user space scope? This design is good for use, but it lost flexibility on fine resource management. It may be requested by cgroup's use case in the future.

Regards!
James

Quoted 9 lines
Quoted 8 lines

James

Quoted 3 lines

          break;
      case AMDGPU_HW_IP_VCN_DEC:
          *out_ring = &adev->vcn.ring_dec;

_______________________________________________
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): c05c6514-ae4c-095b-acdd-f5da3c15a255-5C7GfCeVMHo@public.gmane.org

#10 lore raw

Re: [PATCH 2/3] drm/amdgpu: expose only the first UVD instance for now

Am 18.07.2018 um 16:22 schrieb James Zhu:

Quoted 94 lines · Christian König
Quoted 91 lines

Am 18.07.2018 um 15:40 schrieb James Zhu:

Quoted 89 lines · Christian König
Quoted 73 lines

Going to completely rework the context to ring mapping with Nayan's GSoC work, but for now just stopping to expose the second UVD instance should do it.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 13 +++++-------- drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c |  9 ++-------

  2 files changed, 7 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +1 -1
index e451a3f25beb..75da3c41f3b3 100644
@@ -286,7 +286,7 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
       struct drm_crtc *crtc;
       uint32_t ui32 = 0;
       uint64_t ui64 = 0;
-    int i, j, found;
+    int i, found;
       int ui32_size = sizeof(ui32);
         if (!info->return_size || !info->return_pointer)
@@ -348,8 +348,7 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
               break;

          case AMDGPU_HW_IP_UVD:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++) -                ring_mask |= adev->uvd.inst[i].ring.ready << i; +            ring_mask |= adev->uvd.inst[0].ring.ready; ib_start_alignment = 64; ib_size_alignment = 64;
              break;

drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +0 -0
@@ -362,11 +361,9 @@ static int amdgpu_info_ioctl(struct drm_device 
 *dev, void *data, struct drm_file
               break;

          case AMDGPU_HW_IP_UVD_ENC:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++) -                for (j = 0; j < adev->uvd.num_enc_rings; j++) -                    ring_mask |=
-                    adev->uvd.inst[i].ring_enc[j].ready << -                    (j + i * adev->uvd.num_enc_rings); +            for (i = 0; i < adev->uvd.num_enc_rings; i++) +                ring_mask |=
+                    adev->uvd.inst[0].ring_enc[i].ready << i; ib_start_alignment = 64; ib_size_alignment = 64;
              break;

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -2
index ea9850c9224d..d8357290ad09 100644
@@ -66,8 +66,6 @@ static int amdgpu_identity_map(struct 
 amdgpu_device *adev,
                      u32 ring,
                      struct amdgpu_ring **out_ring)
   {
-    u32 instance;
-
       switch (mapper->hw_ip) {

      case AMDGPU_HW_IP_GFX:
          *out_ring = &adev->gfx.gfx_ring[ring];

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -0
@@ -79,16 +77,13 @@ static int amdgpu_identity_map(struct 
 amdgpu_device *adev,
           *out_ring = &adev->sdma.instance[ring].ring;
           break;

      case AMDGPU_HW_IP_UVD:
-        instance = ring;
-        *out_ring = &adev->uvd.inst[instance].ring; +        *out_ring = &adev->uvd.inst[0].ring; break;

Quoted 8 lines

      case AMDGPU_HW_IP_VCE:
          *out_ring = &adev->vce.ring[ring];
          break;
      case AMDGPU_HW_IP_UVD_ENC:
-        instance = ring / adev->uvd.num_enc_rings;
-        *out_ring =
- &adev->uvd.inst[instance].ring_enc[ring%adev->uvd.num_enc_rings]; +        *out_ring = &adev->uvd.inst[0].ring_enc[ring];

Hi Christian,

If here uses amdgpu_identity_map, I think only first instance's ring is in use.

In current Vega20 situation, should we change to use amdgpu_lru_map or something else to schedule which instance/ring for use ?

Quoted 7 lines
Quoted 6 lines

Yes and no. I'm going to completely remove the queue manager and replace it with Nayan's work on the scheduler.

So no more lru map, but instead a function which picks the best ring based on actual load during context creation.

Regards,
Christian.

Got it. So do we totally remove instance and ring concept out of user space scope?

Yes, that's at least the idea.

Quoted 2 lines

This design is good for use, but it lost flexibility on fine resource management. It may be requested by cgroup's use case in the future.

I actually do this to allow simpler cgroup implementation on top of this.

In other words cgroup can still be used to limit certain applications to certain hardware engines and/or sw priorities, etc...

Regards,
Christian.

Quoted 13 lines

Regards!
James

Quoted 9 lines
Quoted 8 lines

James

Quoted 3 lines

          break;
      case AMDGPU_HW_IP_VCN_DEC:
          *out_ring = &adev->vcn.ring_dec;

_______________________________________________
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): 35334103-7b93-592c-90e8-0933b83e9fac-5C7GfCeVMHo@public.gmane.org

#12 lore raw

Re: [PATCH 2/3] drm/amdgpu: expose only the first UVD instance for now

Quoted 113 lines · Christian König

Am 18.07.2018 um 16:22 schrieb James Zhu:

Quoted 112 lines · Christian König
Quoted 108 lines

Am 18.07.2018 um 15:40 schrieb James Zhu:

Quoted 93 lines · Christian König
Quoted 76 lines

Going to completely rework the context to ring mapping with Nayan's GSoC
work, but for now just stopping to expose the second UVD instance should
do it.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 13 +++++--------
  drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c |  9 ++-------
  2 files changed, 7 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +2 -3
index e451a3f25beb..75da3c41f3b3 100644
@@ -286,7 +286,7 @@ static int amdgpu_info_ioctl(struct drm_device
 *dev, void *data, struct drm_file
      struct drm_crtc *crtc;
      uint32_t ui32 = 0;
      uint64_t ui64 = 0;
-    int i, j, found;
+    int i, found;
      int ui32_size = sizeof(ui32);
        if (!info->return_size || !info->return_pointer)
@@ -348,8 +348,7 @@ static int amdgpu_info_ioctl(struct drm_device
 *dev, void *data, struct drm_file
              break;
          case AMDGPU_HW_IP_UVD:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-                ring_mask |= adev->uvd.inst[i].ring.ready << i;
+            ring_mask |= adev->uvd.inst[0].ring.ready;
              ib_start_alignment = 64;
              ib_size_alignment = 64;
              break;
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +3 -5
@@ -362,11 +361,9 @@ static int amdgpu_info_ioctl(struct drm_device
 *dev, void *data, struct drm_file
              break;
          case AMDGPU_HW_IP_UVD_ENC:
              type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-                for (j = 0; j < adev->uvd.num_enc_rings; j++)
-                    ring_mask |=
-                    adev->uvd.inst[i].ring_enc[j].ready <<
-                    (j + i * adev->uvd.num_enc_rings);
+            for (i = 0; i < adev->uvd.num_enc_rings; i++)
+                ring_mask |=
+                    adev->uvd.inst[0].ring_enc[i].ready << i;
              ib_start_alignment = 64;
              ib_size_alignment = 64;
              break;
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -2
index ea9850c9224d..d8357290ad09 100644
@@ -66,8 +66,6 @@ static int amdgpu_identity_map(struct amdgpu_device
 *adev,
                     u32 ring,
                     struct amdgpu_ring **out_ring)
  {
-    u32 instance;
-
      switch (mapper->hw_ip) {
      case AMDGPU_HW_IP_GFX:
          *out_ring = &adev->gfx.gfx_ring[ring];
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -2
@@ -79,16 +77,13 @@ static int amdgpu_identity_map(struct amdgpu_device
 *adev,
          *out_ring = &adev->sdma.instance[ring].ring;
          break;
      case AMDGPU_HW_IP_UVD:
-        instance = ring;
-        *out_ring = &adev->uvd.inst[instance].ring;

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c static int amdgpu_identity_map(struct amdgpu_device · line 87

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +2 -3
+        *out_ring = &adev->uvd.inst[0].ring;
          break;
      case AMDGPU_HW_IP_VCE:
          *out_ring = &adev->vce.ring[ring];
          break;
      case AMDGPU_HW_IP_UVD_ENC:
-        instance = ring / adev->uvd.num_enc_rings;
-        *out_ring =
- &adev->uvd.inst[instance].ring_enc[ring%adev->uvd.num_enc_rings];
+        *out_ring = &adev->uvd.inst[0].ring_enc[ring];

Hi Christian,

Quoted 5 lines

If here uses amdgpu_identity_map, I think only first instance's ring is in use.

In current Vega20 situation, should we change to use amdgpu_lru_map or something else to schedule which instance/ring for use ?

Yes and no. I'm going to completely remove the queue manager and replace it with Nayan's work on the scheduler.

So no more lru map, but instead a function which picks the best ring based on actual load during context creation.

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

Regards,
Christian.

Got it. So do we totally remove instance and ring concept out of user space scope?

Yes, that's at least the idea.

I think we still may want to expose multiple queues/rings on an instance to allow userspace to make better decisions about parallelism. E.g., if the UMD knows there are two encode rings and it wants to run two encode tasks in parallel, it should be possible. Then the scheduler can deal with scheduling across instances optimally.

Alex

Quoted 30 lines
Quoted 3 lines

This design is good for use, but it lost flexibility on fine resource management. It may be requested by cgroup's use case in the future.

I actually do this to allow simpler cgroup implementation on top of this.

In other words cgroup can still be used to limit certain applications to certain hardware engines and/or sw priorities, etc...

Regards,
Christian.

Quoted 15 lines

Regards!
James

Quoted 11 lines
Quoted 10 lines

James

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -0
          break;
      case AMDGPU_HW_IP_VCN_DEC:
          *out_ring = &adev->vcn.ring_dec;
Quoted 4 lines

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

_______________________________________________
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_P7K+mr-VTT27oneuNKF6jL8vP=X0aZgwecMX+B_qMwbw-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#14 lore raw

Re: [PATCH 2/3] drm/amdgpu: expose only the first UVD instance for now

Am 18.07.2018 um 17:58 schrieb Alex Deucher:

Quoted 114 lines
Quoted 112 lines · Christian König

Am 18.07.2018 um 16:22 schrieb James Zhu:

Quoted 110 lines · Christian König
Quoted 97 lines

Am 18.07.2018 um 15:40 schrieb James Zhu:

Quoted 90 lines · Christian König
Quoted 75 lines

Going to completely rework the context to ring mapping with Nayan's GSoC
work, but for now just stopping to expose the second UVD instance should
do it.

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c       | 13 +++++--------
   drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c |  9 ++-------
   2 files changed, 7 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +2 -3
index e451a3f25beb..75da3c41f3b3 100644
@@ -286,7 +286,7 @@ static int amdgpu_info_ioctl(struct drm_device
 *dev, void *data, struct drm_file
       struct drm_crtc *crtc;
       uint32_t ui32 = 0;
       uint64_t ui64 = 0;
-    int i, j, found;
+    int i, found;
       int ui32_size = sizeof(ui32);
         if (!info->return_size || !info->return_pointer)
@@ -348,8 +348,7 @@ static int amdgpu_info_ioctl(struct drm_device
 *dev, void *data, struct drm_file
               break;
           case AMDGPU_HW_IP_UVD:
               type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-                ring_mask |= adev->uvd.inst[i].ring.ready << i;
+            ring_mask |= adev->uvd.inst[0].ring.ready;
               ib_start_alignment = 64;
               ib_size_alignment = 64;
               break;
drivers/gpu/drm/amd/amdgpu/amdgpu_kms.c +3 -5
@@ -362,11 +361,9 @@ static int amdgpu_info_ioctl(struct drm_device
 *dev, void *data, struct drm_file
               break;
           case AMDGPU_HW_IP_UVD_ENC:
               type = AMD_IP_BLOCK_TYPE_UVD;
-            for (i = 0; i < adev->uvd.num_uvd_inst; i++)
-                for (j = 0; j < adev->uvd.num_enc_rings; j++)
-                    ring_mask |=
-                    adev->uvd.inst[i].ring_enc[j].ready <<
-                    (j + i * adev->uvd.num_enc_rings);
+            for (i = 0; i < adev->uvd.num_enc_rings; i++)
+                ring_mask |=
+                    adev->uvd.inst[0].ring_enc[i].ready << i;
               ib_start_alignment = 64;
               ib_size_alignment = 64;
               break;
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -2
index ea9850c9224d..d8357290ad09 100644
@@ -66,8 +66,6 @@ static int amdgpu_identity_map(struct amdgpu_device
 *adev,
                      u32 ring,
                      struct amdgpu_ring **out_ring)
   {
-    u32 instance;
-
       switch (mapper->hw_ip) {
       case AMDGPU_HW_IP_GFX:
           *out_ring = &adev->gfx.gfx_ring[ring];
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -2
@@ -79,16 +77,13 @@ static int amdgpu_identity_map(struct amdgpu_device
 *adev,
           *out_ring = &adev->sdma.instance[ring].ring;
           break;
       case AMDGPU_HW_IP_UVD:
-        instance = ring;
-        *out_ring = &adev->uvd.inst[instance].ring;

drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c static int amdgpu_identity_map(struct amdgpu_device · line 87

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +2 -3
+        *out_ring = &adev->uvd.inst[0].ring;
           break;
       case AMDGPU_HW_IP_VCE:
           *out_ring = &adev->vce.ring[ring];
           break;
       case AMDGPU_HW_IP_UVD_ENC:
-        instance = ring / adev->uvd.num_enc_rings;
-        *out_ring =
- &adev->uvd.inst[instance].ring_enc[ring%adev->uvd.num_enc_rings];
+        *out_ring = &adev->uvd.inst[0].ring_enc[ring];

Hi Christian,

Quoted 5 lines

If here uses amdgpu_identity_map, I think only first instance's ring is in use.

In current Vega20 situation, should we change to use amdgpu_lru_map or something else to schedule which instance/ring for use ?

Quoted 8 lines

Yes and no. I'm going to completely remove the queue manager and replace it with Nayan's work on the scheduler.

So no more lru map, but instead a function which picks the best ring based on actual load during context creation.

Regards,
Christian.

Got it. So do we totally remove instance and ring concept out of user space scope?

Quoted 7 lines
Quoted 1 line

Yes, that's at least the idea.

I think we still may want to expose multiple queues/rings on an instance to allow userspace to make better decisions about parallelism. E.g., if the UMD knows there are two encode rings and it wants to run two encode tasks in parallel, it should be possible. Then the scheduler can deal with scheduling across instances optimally.

Well it depends.

I currently plan to expose the following engines to userspace independent of available instances for each block.

* 1 GFX ring (unchanged)
* 4 Compute rings (8 previously)
* 1 UVD decode ring (unchanged)
* 1 VCE encode ring (2/3 previously, but that doesn't make sense cause they have different priorities).

When userspace wants to handle multiple streams at the same time it should create multiple context anyway, so even with just 1 ring for each engine we should be on the save side for parallelism.

Christian.

Quoted 29 lines

Alex

Quoted 25 lines
Quoted 2 lines

This design is good for use, but it lost flexibility on fine resource management. It may be requested by cgroup's use case in the future.

I actually do this to allow simpler cgroup implementation on top of this.

In other words cgroup can still be used to limit certain applications to certain hardware engines and/or sw priorities, etc...

Regards,
Christian.

Quoted 9 lines

Regards!
James

Quoted 7 lines
Quoted 7 lines

James

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_queue_mgr.c +0 -0
           break;
       case AMDGPU_HW_IP_VCN_DEC:
           *out_ring = &adev->vcn.ring_dec;

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

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

#15 lore raw

PATCH 3/3 drm/amdgpu: patch the IBs for the second UVD instance

In reply to a message not in the archive

Patch the IBs for the second UVD instance so that userspace don't need to care about the instance they submit to.

---
 drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)
drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c +27 -0
index 831bb995b0ed..89fe910e5c9a 100644
@@ -1219,6 +1219,32 @@ static int uvd_v7_0_ring_test_ring(struct amdgpu_ring *ring)
 	return r;
 }
 
+/**
+ * uvd_v7_0_enc_ring_test_ring - test if UVD ENC ring is working
+ *
+ * @ring: the engine to test on
+ *
+ */
+static int uvd_v7_0_ring_parse_cs(struct amdgpu_cs_parser *p, uint32_t ib_idx)
+{
+	struct amdgpu_ib *ib = &p->job->ibs[ib_idx];
+	unsigned i;
+
+	/* No patching necessary for the first instance */
+	if (!p->ring->me)
+		return 0;
+
+	for (i = 0; i < ib->length_dw; i += 2) {
+		uint32_t reg = amdgpu_get_ib_value(p, ib_idx, i);
+
+		reg -= p->adev->reg_offset[UVD_HWIP][0][1];
+		reg += p->adev->reg_offset[UVD_HWIP][1][1];
+
+		amdgpu_set_ib_value(p, ib_idx, i, reg);
+	}
+	return 0;
+}
+
 /**
  * uvd_v7_0_ring_emit_ib - execute indirect buffer
  *
@@ -1711,6 +1737,7 @@ static const struct amdgpu_ring_funcs uvd_v7_0_ring_vm_funcs = {
 	.get_rptr = uvd_v7_0_ring_get_rptr,
 	.get_wptr = uvd_v7_0_ring_get_wptr,
 	.set_wptr = uvd_v7_0_ring_set_wptr,
+	.parse_cs = uvd_v7_0_ring_parse_cs,
 	.emit_frame_size =
 		6 + /* hdp invalidate */
 		SOC15_FLUSH_GPU_TLB_NUM_WREG * 6 +
Signature · 6 lines
2.14.1

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