Skip to contentSkip to conversation

drm/amdgpu/mes: Simplify hqd mask initialization

PATCH In review Lang Yu 3 participants 3 messages Jan 9

Subsystems (2)
lore
Share & tools
lore

Lang Yu

author
#0 lore raw

PATCH v2 1/1 drm/amdgpu/mes: Simplify hqd mask initialization

"adev->mes.compute_hqd_mask[i] = adev->gfx.disable_kq ? 0xF" is incorrect for MEC with 8 queues per pipe. Let's calculate hqd mask with number of queues per pipe and number of gfx/compute queues kernel used, and get rid of version check and hardcode.

Currently, only MEC1 is used for both kernel/user compute queue. To enable more MECs, we need to redistribute queue per pipe and adjust queue resource shared with kfd that needs a separate patch. Skip other MECs for now.

v2: Force reserved queues to 0 if kernel queue is explicitly disabled.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 59 +++++++++++++++----------
 1 file changed, 35 insertions(+), 24 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +35 -24
index dffa0f7276b7..bed37e50d45b 100644
@@ -31,7 +31,6 @@
 
 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
 #define AMDGPU_ONE_DOORBELL_SIZE 8
-#define AMDGPU_MES_RESERVED_QUEUES	2
 
 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
 {
@@ -89,12 +88,25 @@ static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
 	bitmap_free(adev->mes.doorbell_bitmap);
 }
 
+static inline u32 amdgpu_mes_get_hqd_mask(u32 num_pipe, u32 num_hqd_per_pipe, u32 num_reserved_hqd)
+{
+	u32 total_hqd_mask = (u32)((1ULL << num_hqd_per_pipe) - 1);
+	u32 reserved_hqd_mask = (u32)((1ULL << DIV_ROUND_UP(num_reserved_hqd, num_pipe)) - 1);
+
+	return (total_hqd_mask & ~reserved_hqd_mask);
+}
+
 int amdgpu_mes_init(struct amdgpu_device *adev)
 {
 	int i, r, num_pipes;
 	u32 total_vmid_mask, reserved_vmid_mask;
-	u32 queue_mask, reserved_queue_mask;
 	int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
+	u32 gfx_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.me.num_pipe_per_me,
+				adev->gfx.me.num_queue_per_pipe,
+				adev->gfx.disable_kq ? 0 : adev->gfx.num_gfx_rings);
+	u32 compute_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.mec.num_pipe_per_mec,
+				adev->gfx.mec.num_queue_per_pipe,
+				adev->gfx.disable_kq ? 0 : adev->gfx.num_compute_rings);
 
 	adev->mes.adev = adev;
 
@@ -115,9 +127,6 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 	adev->mes.vmid_mask_mmhub = 0xFF00;
 	adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;
 
-	queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
-	reserved_queue_mask = (u32)(1UL << AMDGPU_MES_RESERVED_QUEUES) - 1;
-
 	num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
 	if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
 		dev_warn(adev->dev, "more gfx pipes than supported by MES! (%d vs %d)\n",
@@ -126,22 +135,8 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 	for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++) {
 		if (i >= num_pipes)
 			break;
-		if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
-		    IP_VERSION(12, 0, 0))
-			/*
-			 * GFX V12 has only one GFX pipe, but 8 queues in it.
-			 * GFX pipe 0 queue 0 is being used by Kernel queue.
-			 * Set GFX pipe 0 queue 1-7 for MES scheduling
-			 * mask = 1111 1110b
-			 */
-			adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0xFF : 0xFE;
-		else
-			/*
-			 * GFX pipe 0 queue 0 is being used by Kernel queue.
-			 * Set GFX pipe 0 queue 1 for MES scheduling
-			 * mask = 10b
-			 */
-			adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0x3 : 0x2;
+
+		adev->mes.gfx_hqd_mask[i] = gfx_hqd_mask;
 	}
 
 	num_pipes = adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_mec;
@@ -150,10 +145,15 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 			 num_pipes, AMDGPU_MES_MAX_COMPUTE_PIPES);
 
 	for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
-		if (i >= num_pipes)
+		/*
+		 * Currently, only MEC1 is used for both kernel and user compute queue.
+		 * To enable more MECs, we need to redistribute queue per pipe and
+		 * adjust queue resource shared with kfd. Skip other MECs for now.
+		 */
+		if (i >= adev->gfx.mec.num_pipe_per_mec)
 			break;
-		adev->mes.compute_hqd_mask[i] =
-			adev->gfx.disable_kq ? 0xF : (queue_mask & ~reserved_queue_mask);
+
+		adev->mes.compute_hqd_mask[i] = compute_hqd_mask;
 	}
 
 	num_pipes = adev->sdma.num_instances;
@@ -167,6 +167,17 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 		adev->mes.sdma_hqd_mask[i] = 0xfc;
 	}
 
+	dev_info(adev->dev,
+			 "MES: vmid_mask_mmhub 0x%08x, vmid_mask_gfxhub 0x%08x\n",
+			 adev->mes.vmid_mask_mmhub,
+			 adev->mes.vmid_mask_gfxhub);
+
+	dev_info(adev->dev,
+			 "MES: gfx_hqd_mask 0x%08x, compute_hqd_mask 0x%08x, sdma_hqd_mask 0x%08x\n",
+			 adev->mes.gfx_hqd_mask[0],
+			 adev->mes.compute_hqd_mask[0],
+			 adev->mes.sdma_hqd_mask[0]);
+
 	for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
 		r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
 		if (r) {
Signature · 1 line
2.34.1
Reviewed-by on v2 1/1
#1 lore raw
Quoted 6 lines · Lang Yu

"adev->mes.compute_hqd_mask[i] = adev->gfx.disable_kq ? 0xF" is incorrect for MEC with 8 queues per pipe. Let's calculate hqd mask with number of queues per pipe and number of gfx/compute queues kernel used, and get rid of version check and hardcode.

Quoted 8 lines

Currently, only MEC1 is used for both kernel/user compute queue. To enable more MECs, we need to redistribute queue per pipe and adjust queue resource shared with kfd that needs a separate patch. Skip other MECs for now.

v2: Force reserved queues to 0 if kernel queue is explicitly disabled.

Quoted 118 lines
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 59 +++++++++++++++----------
 1 file changed, 35 insertions(+), 24 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +35 -25
index dffa0f7276b7..bed37e50d45b 100644
@@ -31,7 +31,6 @@
 
 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024
 #define AMDGPU_ONE_DOORBELL_SIZE 8
-#define AMDGPU_MES_RESERVED_QUEUES     2
 
 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)
 {
@@ -89,12 +88,25 @@ static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
        bitmap_free(adev->mes.doorbell_bitmap);
 }
 
+static inline u32 amdgpu_mes_get_hqd_mask(u32 num_pipe, u32 num_hqd_per_pipe, u32 num_reserved_hqd)
+{
+       u32 total_hqd_mask = (u32)((1ULL << num_hqd_per_pipe) - 1);
+       u32 reserved_hqd_mask = (u32)((1ULL << DIV_ROUND_UP(num_reserved_hqd, num_pipe)) - 1);
+
+       return (total_hqd_mask & ~reserved_hqd_mask);
+}
+
 int amdgpu_mes_init(struct amdgpu_device *adev)
 {
        int i, r, num_pipes;
        u32 total_vmid_mask, reserved_vmid_mask;
-       u32 queue_mask, reserved_queue_mask;
        int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;
+       u32 gfx_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.me.num_pipe_per_me,
+                               adev->gfx.me.num_queue_per_pipe,
+                               adev->gfx.disable_kq ? 0 : adev->gfx.num_gfx_rings);
+       u32 compute_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.mec.num_pipe_per_mec,
+                               adev->gfx.mec.num_queue_per_pipe,
+                               adev->gfx.disable_kq ? 0 : adev->gfx.num_compute_rings);
 
        adev->mes.adev = adev;
 
@@ -115,9 +127,6 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
        adev->mes.vmid_mask_mmhub = 0xFF00;
        adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;
 
-       queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
-       reserved_queue_mask = (u32)(1UL << AMDGPU_MES_RESERVED_QUEUES) - 1;
-
        num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
        if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
                dev_warn(adev->dev, "more gfx pipes than supported by MES! (%d vs %d)\n",
@@ -126,22 +135,8 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
        for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++) {
                if (i >= num_pipes)
                        break;
-               if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
-                   IP_VERSION(12, 0, 0))
-                       /*
-                        * GFX V12 has only one GFX pipe, but 8 queues in it.
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1-7 for MES scheduling
-                        * mask = 1111 1110b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0xFF : 0xFE;
-               else
-                       /*
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1 for MES scheduling
-                        * mask = 10b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0x3 : 0x2;
+
+               adev->mes.gfx_hqd_mask[i] = gfx_hqd_mask;
        }
 
        num_pipes = adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_mec;
@@ -150,10 +145,15 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
                         num_pipes, AMDGPU_MES_MAX_COMPUTE_PIPES);
 
        for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
-               if (i >= num_pipes)
+               /*
+                * Currently, only MEC1 is used for both kernel and user compute queue.
+                * To enable more MECs, we need to redistribute queue per pipe and
+                * adjust queue resource shared with kfd. Skip other MECs for now.
+                */
+               if (i >= adev->gfx.mec.num_pipe_per_mec)
                        break;
-               adev->mes.compute_hqd_mask[i] =
-                       adev->gfx.disable_kq ? 0xF : (queue_mask & ~reserved_queue_mask);
+
+               adev->mes.compute_hqd_mask[i] = compute_hqd_mask;
        }
 
        num_pipes = adev->sdma.num_instances;
@@ -167,6 +167,17 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
                adev->mes.sdma_hqd_mask[i] = 0xfc;
        }
 
+       dev_info(adev->dev,
+                        "MES: vmid_mask_mmhub 0x%08x, vmid_mask_gfxhub 0x%08x\n",
+                        adev->mes.vmid_mask_mmhub,
+                        adev->mes.vmid_mask_gfxhub);
+
+       dev_info(adev->dev,
+                        "MES: gfx_hqd_mask 0x%08x, compute_hqd_mask 0x%08x, sdma_hqd_mask 0x%08x\n",
+                        adev->mes.gfx_hqd_mask[0],
+                        adev->mes.compute_hqd_mask[0],
+                        adev->mes.sdma_hqd_mask[0]);
+
        for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
                r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
                if (r) {
--

2.34.1

#2 lore raw
In reply to Lang Yu

[AMD Official Use Only - AMD Internal Distribution Only]

I just found in mes_v12_1 , it still hardcode the pkt->gfx_hqd_mask[0] = 0xFE in function mes_v12_1_set_gfx_hqd_mask, I think this is wrong and unnecessary.

Regards
Shaoyun.liu

-----Original Message-----
From: amd-gfx <[email protected]> On Behalf Of Lang Yu
Sent: Friday, January 9, 2026 5:10 AM
To: [email protected]; Deucher, Alexander <[email protected]>
Cc: Zhang, Hawking <[email protected]>; Yu, Lang <[email protected]>
Subject: [PATCH v2] drm/amdgpu/mes: Simplify hqd mask initialization

"adev->mes.compute_hqd_mask[i] = adev->gfx.disable_kq ? 0xF" is incorrect for MEC with 8 queues per pipe. Let's calculate hqd mask with number of queues per pipe and number of gfx/compute queues kernel used, and get rid of version check and hardcode.

Currently, only MEC1 is used for both kernel/user compute queue.
To enable more MECs, we need to redistribute queue per pipe and adjust queue resource shared with kfd that needs a separate patch. Skip other MECs for now.

v2: Force reserved queues to 0 if kernel queue is explicitly disabled.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 59 +++++++++++++++----------
 1 file changed, 35 insertions(+), 24 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +0 -1
index dffa0f7276b7..bed37e50d45b 100644
@@ -31,7 +31,6 @@
 
 #define AMDGPU_MES_MAX_NUM_OF_QUEUES_PER_PROCESS 1024  #define AMDGPU_ONE_DOORBELL_SIZE 8
-#define AMDGPU_MES_RESERVED_QUEUES     2
 
 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev)  { @@ -89,12 +88,25 @@ static void amdgpu_mes_doorbell_free(struct amdgpu_device *adev)
        bitmap_free(adev->mes.doorbell_bitmap);
 }

+static inline u32 amdgpu_mes_get_hqd_mask(u32 num_pipe, u32 +num_hqd_per_pipe, u32 num_reserved_hqd) {
+ u32 total_hqd_mask = (u32)((1ULL << num_hqd_per_pipe) - 1);
+ u32 reserved_hqd_mask = (u32)((1ULL << DIV_ROUND_UP(num_reserved_hqd,
+num_pipe)) - 1);
+
+ return (total_hqd_mask & ~reserved_hqd_mask); }
+

 int amdgpu_mes_init(struct amdgpu_device *adev)  {
        int i, r, num_pipes;
        u32 total_vmid_mask, reserved_vmid_mask;
-       u32 queue_mask, reserved_queue_mask;
        int num_xcc = adev->gfx.xcc_mask ? NUM_XCC(adev->gfx.xcc_mask) : 1;

+ u32 gfx_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.me.num_pipe_per_me,
+ adev->gfx.me.num_queue_per_pipe,
+ adev->gfx.disable_kq ? 0 : adev->gfx.num_gfx_rings);
+ u32 compute_hqd_mask = amdgpu_mes_get_hqd_mask(adev->gfx.mec.num_pipe_per_mec,
+ adev->gfx.mec.num_queue_per_pipe,
+ adev->gfx.disable_kq ? 0 : adev->gfx.num_compute_rings);

        adev->mes.adev = adev;

@@ -115,9 +127,6 @@ int amdgpu_mes_init(struct amdgpu_device *adev)

        adev->mes.vmid_mask_mmhub = 0xFF00;
        adev->mes.vmid_mask_gfxhub = total_vmid_mask & ~reserved_vmid_mask;

- queue_mask = (u32)(1UL << adev->gfx.mec.num_queue_per_pipe) - 1;
- reserved_queue_mask = (u32)(1UL << AMDGPU_MES_RESERVED_QUEUES) - 1;
-

        num_pipes = adev->gfx.me.num_pipe_per_me * adev->gfx.me.num_me;
        if (num_pipes > AMDGPU_MES_MAX_GFX_PIPES)
                dev_warn(adev->dev, "more gfx pipes than supported by MES! (%d vs %d)\n", @@ -126,22 +135,8 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
        for (i = 0; i < AMDGPU_MES_MAX_GFX_PIPES; i++) {
                if (i >= num_pipes)
                        break;
-               if (amdgpu_ip_version(adev, GC_HWIP, 0) >=
-                   IP_VERSION(12, 0, 0))
-                       /*
-                        * GFX V12 has only one GFX pipe, but 8 queues in it.
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1-7 for MES scheduling
-                        * mask = 1111 1110b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0xFF : 0xFE;
-               else
-                       /*
-                        * GFX pipe 0 queue 0 is being used by Kernel queue.
-                        * Set GFX pipe 0 queue 1 for MES scheduling
-                        * mask = 10b
-                        */
-                       adev->mes.gfx_hqd_mask[i] = adev->gfx.disable_kq ? 0x3 : 0x2;
+
+               adev->mes.gfx_hqd_mask[i] = gfx_hqd_mask;
        }

        num_pipes = adev->gfx.mec.num_pipe_per_mec * adev->gfx.mec.num_mec; @@ -150,10 +145,15 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
                         num_pipes, AMDGPU_MES_MAX_COMPUTE_PIPES);

        for (i = 0; i < AMDGPU_MES_MAX_COMPUTE_PIPES; i++) {
-               if (i >= num_pipes)
+               /*
+                * Currently, only MEC1 is used for both kernel and user compute queue.
+                * To enable more MECs, we need to redistribute queue per pipe and
+                * adjust queue resource shared with kfd. Skip other MECs for now.
+                */
+               if (i >= adev->gfx.mec.num_pipe_per_mec)
                        break;
-               adev->mes.compute_hqd_mask[i] =
-                       adev->gfx.disable_kq ? 0xF : (queue_mask & ~reserved_queue_mask);
+
+               adev->mes.compute_hqd_mask[i] = compute_hqd_mask;
        }

        num_pipes = adev->sdma.num_instances;
@@ -167,6 +167,17 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
                adev->mes.sdma_hqd_mask[i] = 0xfc;
        }

+ dev_info(adev->dev,

+                        "MES: vmid_mask_mmhub 0x%08x, vmid_mask_gfxhub 0x%08x\n",
+                        adev->mes.vmid_mask_mmhub,
+                        adev->mes.vmid_mask_gfxhub);
+
+       dev_info(adev->dev,
+                        "MES: gfx_hqd_mask 0x%08x, compute_hqd_mask 0x%08x, sdma_hqd_mask 0x%08x\n",
+                        adev->mes.gfx_hqd_mask[0],
+                        adev->mes.compute_hqd_mask[0],
+                        adev->mes.sdma_hqd_mask[0]);
+
        for (i = 0; i < AMDGPU_MAX_MES_PIPES * num_xcc; i++) {
                r = amdgpu_device_wb_get(adev, &adev->mes.sch_ctx_offs[i]);
                if (r) {

--
2.34.1