Skip to contentSkip to conversation

drm/amdgpu: Config VCN instance for soc_v1_0"

PATCH New [email protected] 1 participant 13 messages Aug 10

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

PATCH 1/13 drm/amdgpu: Config VCN instance for soc_v1_0"

From: Sonny Jiang <[email protected]>

Set up the default VCN instance mask and related instance counts during soc_v1_0 SOC configuration, before IP discovery is enabled.

---
 drivers/gpu/drm/amd/amdgpu/soc_v1_0.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/gpu/drm/amd/amdgpu/soc_v1_0.c +5 -0
index 43e92e2a85e8..37e2068b6fda 100644
@@ -849,6 +849,11 @@ int soc_v1_0_init_soc_config(struct amdgpu_device *adev)
 	adev->sdma.sdma_mask = sdma_mask;
 	adev->sdma.num_instances = NUM_XCC(adev->sdma.sdma_mask);
 
+	adev->vcn.harvest_config = 0;
+	adev->vcn.num_inst_per_aid = 1;
+	adev->vcn.inst_mask = 1;
+	adev->vcn.num_vcn_inst = hweight32(adev->vcn.inst_mask);
+
 	ret = soc_v1_0_xcp_mgr_init(adev);
 	if (ret)
 		return ret;
Signature · 1 line
2.43.0
#1 lore raw

PATCH 2/13 drm/amdgpu: Config JPEG instance for soc_v1_0

From: Sonny Jiang <[email protected]>

Set up the default JPEG instance mask and related instance counts during soc_v1_0 SOC configuration, before IP discovery is enabled.

---
 drivers/gpu/drm/amd/amdgpu/soc_v1_0.c | 5 +++++
 1 file changed, 5 insertions(+)
drivers/gpu/drm/amd/amdgpu/soc_v1_0.c +5 -0
index 37e2068b6fda..04c49e23d79b 100644
@@ -854,6 +854,11 @@ int soc_v1_0_init_soc_config(struct amdgpu_device *adev)
 	adev->vcn.inst_mask = 1;
 	adev->vcn.num_vcn_inst = hweight32(adev->vcn.inst_mask);
 
+	adev->jpeg.harvest_config = 0;
+	adev->jpeg.num_inst_per_aid = 1;
+	adev->jpeg.inst_mask = 1;
+	adev->jpeg.num_jpeg_inst = hweight32(adev->jpeg.inst_mask);
+
 	ret = soc_v1_0_xcp_mgr_init(adev);
 	if (ret)
 		return ret;
Signature · 1 line
2.43.0
#2 lore raw

PATCH 3/13 drm/amdgpu: correct VCN V5_0_2 doorbell range_size

In reply to [email protected]

From: Sonny Jiang <[email protected]>

Each VCN 5.0.2 instance has 1 VCN ring and 10 JPEG rings, requiring 11 consecutive doorbells. Increase the NBIO doorbell range size from 8 to 11 to cover the full ring set.

---
 drivers/gpu/drm/amd/amdgpu/nbio_v6_3_2.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/nbio_v6_3_2.c +1 -1
index 5e8f466f23ad..d249011b2cf7 100644
@@ -266,7 +266,7 @@ static void nbio_v6_3_2_vcn_doorbell_range(struct amdgpu_device *adev,
 		doorbell_range = REG_SET_FIELD(doorbell_range,
 					       GDC_S2A0_S2A_DOORBELL_ENTRY_2_CTRL,
 					       S2A_DOORBELL_PORT2_RANGE_SIZE,
-					       8);
+					       11);
 		doorbell_range = REG_SET_FIELD(doorbell_range,
 					       GDC_S2A0_S2A_DOORBELL_ENTRY_2_CTRL,
 					       S2A_DOORBELL_PORT2_AWADDR_31_28_VALUE,
Signature · 1 line
2.43.0
#3 lore raw

PATCH 4/13 drm/amdgpu: Correct VCN V5_0_2 doorbell index setting

In reply to [email protected]

From: Sonny Jiang <[email protected]>

Each VCN 5.0.2 instance has 11 doorbells. Use an 11-doorbell when calculating the per-instance VCN ring doorbell index.

---
 drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c +1 -1
index b9f6ae75ea72..5291ebbc5c3e 100644
@@ -129,7 +129,7 @@ static int vcn_v5_0_2_sw_init(struct amdgpu_ip_block *ip_block)
 		ring->use_doorbell = true;
 
 		ring->doorbell_index =
-			(adev->doorbell_index.vcn.vcn_ring0_1 << 1) + 32 * vcn_inst;
+			(adev->doorbell_index.vcn.vcn_ring0_1 << 1) + 11 * vcn_inst;
 
 		ring->vm_hub = AMDGPU_MMHUB0(adev->vcn.inst[i].aid_id);
 		sprintf(ring->name, "vcn_unified_%d", adev->vcn.inst[i].aid_id);
Signature · 1 line
2.43.0
#4 lore raw

PATCH 5/13 drm/amdgpu: VCN 5_0_2 add support for multi-instance

In reply to [email protected]

From: Sonny Jiang <[email protected]>

VCN 5.0.2 has two VCN instances per MID. Update the SOC configuration and VCN 5.0.2 initialization to handle multiple instances

---
 drivers/gpu/drm/amd/amdgpu/soc_v1_0.c   |  2 +-
 drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c | 54 ++++++++++++++++++++-----
 2 files changed, 45 insertions(+), 11 deletions(-)
drivers/gpu/drm/amd/amdgpu/soc_v1_0.c +1 -1
index 04c49e23d79b..db48bf2bc213 100644
@@ -850,7 +850,7 @@ int soc_v1_0_init_soc_config(struct amdgpu_device *adev)
 	adev->sdma.num_instances = NUM_XCC(adev->sdma.sdma_mask);
 
 	adev->vcn.harvest_config = 0;
-	adev->vcn.num_inst_per_aid = 1;
+	adev->vcn.num_inst_per_aid = 2;
 	adev->vcn.inst_mask = 1;
 	adev->vcn.num_vcn_inst = hweight32(adev->vcn.inst_mask);
 
drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c +44 -10

95 lines not on this page · raw message

Signature · 1 line
2.43.0
#5 lore raw

PATCH 6/13 drm/amdgpu: VCN 5_0_2 disable VCN_RRMT

In reply to [email protected]

From: Sonny Jiang <[email protected]>

When RRMT is enabled, VCN firmware's register access is remapped so that VCN instances on different dies (VCN0/1 on MID0, VCN2/3 on MID1) require different target addresses for the same logical register.

Remove the RRMT capability detection so the driver always assumes RRMT is disabled, using local-die SMN addresses for all instances.

---
 drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c | 2 --
 1 file changed, 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c +0 -2
index 6ffb34974a0d..2842d8a9b42b 100644
@@ -208,8 +208,6 @@ static int vcn_v5_0_2_hw_init(struct amdgpu_ip_block *ip_block)
 	int i, r, vcn_inst;
 	uint32_t tmp;
 
-	if (RREG32_SOC15(VCN, GET_INST(VCN, 0), regVCN_RRMT_CNTL) & 0x200)
-		adev->vcn.caps |= AMDGPU_VCN_CAPS(RRMT_ENABLED);
 	for (i = 0; i < adev->vcn.num_vcn_inst; ++i) {
 		vcn_inst = GET_INST(VCN, i);
 		ring = &adev->vcn.inst[i].ring_enc[0];
Signature · 1 line
2.43.0
#6 lore raw

PATCH 7/13 drm/amdgpu: VCN 5_0_2 enable multi-instance

In reply to [email protected]

From: Sonny Jiang <[email protected]>

Remove forcing the VCN instance mask to a single instance during soc_v1_0 SOC setup. Use the configured instance mask so multi-instance VCN 5.0.2 topologies can be enabled.

---
 drivers/gpu/drm/amd/amdgpu/soc_v1_0.c | 1 -
 1 file changed, 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/soc_v1_0.c +0 -1
index db48bf2bc213..39224d4c9c10 100644
@@ -851,7 +851,6 @@ int soc_v1_0_init_soc_config(struct amdgpu_device *adev)
 
 	adev->vcn.harvest_config = 0;
 	adev->vcn.num_inst_per_aid = 2;
-	adev->vcn.inst_mask = 1;
 	adev->vcn.num_vcn_inst = hweight32(adev->vcn.inst_mask);
 
 	adev->jpeg.harvest_config = 0;
Signature · 1 line
2.43.0
#7 lore raw

PATCH 8/13 drm/amdgpu: JPEG 5_0_2 Add multi-instance interrupt support

In reply to [email protected]

From: Sonny Jiang <[email protected]>

JPEG 5.0.2 has two JPEG instances per MID. Register JPEG interrupt sources for both VCN IH clients and add a JPEG 5.0.2-specific interrupt handler.

---
 drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.c |   2 +-
 drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.h |   8 --
 drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_2.c | 100 ++++++++++++++++++++++-
 3 files changed, 97 insertions(+), 13 deletions(-)
drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.c +1 -1
index db3fbe4eabb9..8846cb3ed12b 100644
@@ -772,7 +772,7 @@ static int jpeg_v5_0_1_set_ras_interrupt_state(struct amdgpu_device *adev,
 
 
 
-int jpeg_v5_0_1_process_interrupt(struct amdgpu_device *adev,
+static int jpeg_v5_0_1_process_interrupt(struct amdgpu_device *adev,
 					 struct amdgpu_irq_src *source,
 					 struct amdgpu_iv_entry *entry)
 {
drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.h +0 -8
index 67346faecb47..a7e58d5fb246 100644
@@ -108,12 +108,4 @@ enum amdgpu_jpeg_v5_0_1_sub_block {
 	AMDGPU_JPEG_V5_0_1_MAX_SUB_BLOCK,
 };
 
-struct amdgpu_irq_src;
-struct amdgpu_iv_entry;
-struct amdgpu_device;
-
-int jpeg_v5_0_1_process_interrupt(struct amdgpu_device *adev,
-					struct amdgpu_irq_src *source,
-					struct amdgpu_iv_entry *entry);
-
 #endif /* __JPEG_V5_0_1_H__ */
drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_2.c +96 -4

124 lines not on this page · raw message

Signature · 1 line
2.43.0
#8 lore raw

PATCH 9/13 drm/amdgpu: VCN 5_0_2 correct process interrupt warn message

In reply to [email protected]

From: Sonny Jiang <[email protected]>

Fix the incorrect interrupt process warning message.

---
 drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/vcn_v5_0_2.c +2 -2
index 2842d8a9b42b..74aae5b5eec4 100644
@@ -1172,7 +1172,7 @@ static int vcn_v5_0_2_process_interrupt(struct amdgpu_device *adev, struct amdgp
 		break;
 	default:
 		dev_WARN_ONCE(adev->dev, 1,
-				"Interrupt received for unknown VCN instance %d",
+				"Interrupt received for unknown VCN node_id %d",
 				entry->node_id);
 		return 0;
 	}
@@ -1186,7 +1186,7 @@ static int vcn_v5_0_2_process_interrupt(struct amdgpu_device *adev, struct amdgp
 	if (i >= adev->vcn.num_vcn_inst) {
 		dev_WARN_ONCE(adev->dev, 1,
 				"Interrupt received for unknown VCN instance %d",
-				entry->node_id);
+				inst);
 		return 0;
 	}
 
Signature · 1 line
2.43.0
#9 lore raw

PATCH 10/13 drm/amdgpu: JPEG 5_0_2 set inst_per_aid to 2

In reply to [email protected]

From: Sonny Jiang <[email protected]>

JPEG 5.0.2 has 2 JPEG instances per MID. Report 2 JPEG instances per AID

---
 drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_2.c | 4 ++--
 drivers/gpu/drm/amd/amdgpu/soc_v1_0.c    | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_2.c +2 -2
index 0ea341c238f5..059b6b927251 100644
@@ -167,7 +167,7 @@ static int jpeg_v5_0_2_sw_init(struct amdgpu_ip_block *ip_block)
 			ring->doorbell_index =
 				(adev->doorbell_index.vcn.vcn_ring0_1 << 1) +
 				1 + j + 11 * jpeg_inst;
-			sprintf(ring->name, "jpeg_dec_%d.%d", adev->jpeg.inst[i].aid_id, j);
+			sprintf(ring->name, "jpeg_dec_%d.%d", i, j);
 			r = amdgpu_ring_init(adev, ring, 512, &adev->jpeg.inst->irq, 0,
 					     AMDGPU_RING_PRIO_DEFAULT, NULL);
 			if (r)
@@ -243,7 +243,7 @@ static int jpeg_v5_0_2_hw_init(struct amdgpu_ip_block *ip_block)
 		if (ring->use_doorbell)
 			adev->nbio.funcs->vcn_doorbell_range(adev, ring->use_doorbell,
 				 (adev->doorbell_index.vcn.vcn_ring0_1 << 1) + 11 * jpeg_inst,
-				 adev->jpeg.inst[i].aid_id);
+				 jpeg_inst);
 
 		for (j = 0; j < adev->jpeg.num_jpeg_rings; ++j) {
 			ring = &adev->jpeg.inst[i].ring_dec[j];
drivers/gpu/drm/amd/amdgpu/soc_v1_0.c +1 -1
index 39224d4c9c10..7531890ec005 100644
@@ -854,7 +854,7 @@ int soc_v1_0_init_soc_config(struct amdgpu_device *adev)
 	adev->vcn.num_vcn_inst = hweight32(adev->vcn.inst_mask);
 
 	adev->jpeg.harvest_config = 0;
-	adev->jpeg.num_inst_per_aid = 1;
+	adev->jpeg.num_inst_per_aid = 2;
 	adev->jpeg.inst_mask = 1;
 	adev->jpeg.num_jpeg_inst = hweight32(adev->jpeg.inst_mask);
 
Signature · 1 line
2.43.0
#10 lore raw

PATCH 11/13 drm/amdgpu: Jpeg 5_0_2 enable doorbell

In reply to [email protected]

From: Sonny Jiang <[email protected]>

Enable JPEG 5_0_2 doorbell

---
 drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_2.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_2.c +1 -1
index 059b6b927251..341577db7d6b 100644
@@ -162,7 +162,7 @@ static int jpeg_v5_0_2_sw_init(struct amdgpu_ip_block *ip_block)
 
 		for (j = 0; j < adev->jpeg.num_jpeg_rings; ++j) {
 			ring = &adev->jpeg.inst[i].ring_dec[j];
-			ring->use_doorbell = false;
+			ring->use_doorbell = true;
 			ring->vm_hub = AMDGPU_MMHUB0(adev->jpeg.inst[i].aid_id);
 			ring->doorbell_index =
 				(adev->doorbell_index.vcn.vcn_ring0_1 << 1) +
Signature · 1 line
2.43.0
#11 lore raw

PATCH 12/13 drm/amdgpu: Jpeg uses shard VM invalidation engine

In reply to [email protected]

From: Sonny Jiang <[email protected]>

JPEG 5_0_2 has 2 instances per die. Each instance has 10 rings. Each ring consuming a unique VM invalidation engine exhausts the available engines per MMHUB.

All rings within a JPEG instance can safely share one engine since MMHUB serializes concurrent invalidation requests via hardware semaphore (see "drm/amdgpu: add invalidate semaphore support for gmc v12.1").

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c  |  9 +++++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c | 38 ++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.h |  2 ++
 3 files changed, 48 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c +8 -1
index 3dc8faa091d7..2f6d20c00ce2 100644
@@ -672,7 +672,8 @@ int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
 			continue;
 
 		/* Skip if the ring is a shared ring */
-		if (amdgpu_sdma_is_shared_inv_eng(adev, ring))
+		if (amdgpu_sdma_is_shared_inv_eng(adev, ring) ||
+		    amdgpu_jpeg_is_shared_inv_eng(adev, ring))
 			continue;
 
 		inv_eng = ffs(vm_inv_engs[vmhub]);
@@ -702,6 +703,12 @@ int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
 					ring->name, ring->vm_inv_eng, shared_ring->name, ring->vm_hub);
 			continue;
 		}
+
+		/* All decode rings within one JPEG instance share one VM
+		 * invalidation engine (no-op unless ring is a JPEG instance's
+		 * first ring).
+		 */
+		amdgpu_jpeg_set_shared_inv_eng(adev, ring);
 	}
 
 	return 0;
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.c +38 -0
index 363b99168478..208566ffe898 100644
@@ -607,3 +607,41 @@ int amdgpu_jpeg_dec_parse_cs(struct amdgpu_cs_parser *parser,
 
 	return 0;
 }
+
+/**
+ * amdgpu_jpeg_is_shared_inv_eng - Check if a ring is a JPEG ring that shares a VM invalidation
+ * engine
+ * @adev: Pointer to the AMDGPU device structure
+ * @ring: Pointer to the ring structure to check
+ *
+ * All decode rings within one JPEG instance share a single VM invalidation
+ * engine (see amdgpu_jpeg_set_shared_inv_eng()). This returns true for every
+ * ring in the instance except the first (ring->pipe == 0), which owns the
+ * engine allocated by amdgpu_gmc_allocate_vm_inv_eng().
+ */
+bool amdgpu_jpeg_is_shared_inv_eng(struct amdgpu_device *adev, struct amdgpu_ring *ring)
+{
+	return ring->funcs->type == AMDGPU_RING_TYPE_VCN_JPEG && ring->pipe != 0;
+}
+
+/**
+ * amdgpu_jpeg_set_shared_inv_eng - Propagate a VM invalidation engine to the rest of a JPEG
+ * instance
+ * @adev: Pointer to the AMDGPU device structure
+ * @ring: Pointer to the ring that just had a VM invalidation engine allocated
+ *
+ * No-op unless @ring is the owner (ring->pipe == 0) of a JPEG instance.
+ * When it is, mirrors its freshly allocated vm_inv_eng onto the rest of the
+ * rings in that instance, so the whole instance shares one engine instead of
+ * consuming one per ring.
+ */
+void amdgpu_jpeg_set_shared_inv_eng(struct amdgpu_device *adev, struct amdgpu_ring *ring)
+{
+	int j;
+
+	if (ring->funcs->type != AMDGPU_RING_TYPE_VCN_JPEG || ring->pipe != 0)
+		return;
+
+	for (j = 1; j < adev->jpeg.num_jpeg_rings; j++)
+		adev->jpeg.inst[ring->me].ring_dec[j].vm_inv_eng = ring->vm_inv_eng;
+}
drivers/gpu/drm/amd/amdgpu/amdgpu_jpeg.h +2 -0
index fe95d9188713..05d58e6bbfab 100644
@@ -183,5 +183,7 @@ void amdgpu_jpeg_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_pri
 int amdgpu_jpeg_dec_parse_cs(struct amdgpu_cs_parser *parser,
 			     struct amdgpu_job *job,
 			     struct amdgpu_ib *ib);
+bool amdgpu_jpeg_is_shared_inv_eng(struct amdgpu_device *adev, struct amdgpu_ring *ring);
+void amdgpu_jpeg_set_shared_inv_eng(struct amdgpu_device *adev, struct amdgpu_ring *ring);
 
 #endif /*__AMDGPU_JPEG_H__*/
Signature · 1 line
2.43.0
#12 lore raw

PATCH 13/13 drm/amdgpu: JPEG 5_0_2 enable multi-instance

In reply to [email protected]

From: Sonny Jiang <[email protected]>

Enable JPEG 5_0_2 multi-instance.

---
 drivers/gpu/drm/amd/amdgpu/soc_v1_0.c | 1 -
 1 file changed, 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/soc_v1_0.c +0 -1
index 7531890ec005..7322526a09ba 100644
@@ -855,7 +855,6 @@ int soc_v1_0_init_soc_config(struct amdgpu_device *adev)
 
 	adev->jpeg.harvest_config = 0;
 	adev->jpeg.num_inst_per_aid = 2;
-	adev->jpeg.inst_mask = 1;
 	adev->jpeg.num_jpeg_inst = hweight32(adev->jpeg.inst_mask);
 
 	ret = soc_v1_0_xcp_mgr_init(adev);
Signature · 1 line
2.43.0