Skip to contentSkip to conversation

drm/amdgpu: Read clock counter via MMIO to reduce delay

PATCH In review YuBiao Wang 4 participants 4 messages Jun 29, 2021 – Jun 30, 2021

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

PATCH 1/1 drm/amdgpu: Read clock counter via MMIO to reduce delay

[Why]
GPU timing counters are read via KIQ under sriov, which will introduce a delay.

[How]
It could be directly read by MMIO.

v2: Add additional check to prevent carryover issue.
v3: Only check for carryover for once to prevent performance issue.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +10 -2
index ff7e9f49040e..82a5b7ab8dc5 100644
@@ -7610,6 +7610,7 @@ static int gfx_v10_0_soft_reset(void *handle)
 static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)
 {
 	uint64_t clock;
+	uint64_t clock_lo, clock_hi, hi_check;
 
 	amdgpu_gfx_off_ctrl(adev, false);
 	mutex_lock(&adev->gfx.gpu_clock_mutex);
@@ -7620,8 +7621,15 @@ static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)
 			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER_Vangogh) << 32ULL);
 		break;
 	default:
-		clock = (uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER) |
-			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER) << 32ULL);
+		clock_hi = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+		hi_check = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		// If carryover happens, update lower count again.
+		if (hi_check != clock_hi) {
+			clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+			clock_hi = hi_check;
+		}
+		clock = (uint64_t)clock_lo | ((uint64_t)clock_hi << 32ULL);
 		break;
 	}
 	mutex_unlock(&adev->gfx.gpu_clock_mutex);
Signature · 6 lines
2.25.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Acked-by Acked-by on v1 1/1
#1 lore raw

[AMD Official Use Only]

-----Original Message-----
From: YuBiao Wang <[email protected]>
Sent: Tuesday, June 29, 2021 6:01 PM
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Quan, Evan <[email protected]>; Chen, Horace <[email protected]>; Tuikov, Luben <[email protected]>; Koenig, Christian <[email protected]>; Deucher, Alexander <[email protected]>; Xiao, Jack <[email protected]>; Zhang, Hawking <[email protected]>; Liu, Monk <[email protected]>; Xu, Feifei <[email protected]>; Wang, Kevin(Yang) <[email protected]>; Wang, YuBiao <[email protected]>
Subject: [PATCH 1/1] drm/amdgpu: Read clock counter via MMIO to reduce delay

[Why]
GPU timing counters are read via KIQ under sriov, which will introduce a delay.

[How]
It could be directly read by MMIO.

v2: Add additional check to prevent carryover issue.
v3: Only check for carryover for once to prevent performance issue.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +10 -2
index ff7e9f49040e..82a5b7ab8dc5 100644
@@ -7610,6 +7610,7 @@ static int gfx_v10_0_soft_reset(void *handle)  static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)  {
 	uint64_t clock;
+	uint64_t clock_lo, clock_hi, hi_check;
 
 	amdgpu_gfx_off_ctrl(adev, false);
 	mutex_lock(&adev->gfx.gpu_clock_mutex);
@@ -7620,8 +7621,15 @@ static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)
 			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER_Vangogh) << 32ULL);
 		break;
 	default:
-		clock = (uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER) |
-			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER) << 32ULL);
+		clock_hi = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+		hi_check = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		// If carryover happens, update lower count again.
+		if (hi_check != clock_hi) {
+			clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+			clock_hi = hi_check;
+		}
+		clock = (uint64_t)clock_lo | ((uint64_t)clock_hi << 32ULL);
 		break;
 	}
 	mutex_unlock(&adev->gfx.gpu_clock_mutex);

--
2.25.1
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Acked-by Horace Chen on v1 1/1
#2 lore raw
In reply to YuBiao Wang

[AMD Official Use Only]

Please do not use "//" in linux kernel patch , use "/* */" instead

After this part fixed the patch is : Reviewed-by: Monk Liu <[email protected]>

Thanks

------------------------------------------
Monk Liu | Cloud-GPU Core team
------------------------------------------

-----Original Message-----
From: YuBiao Wang <[email protected]>
Sent: Tuesday, June 29, 2021 6:01 PM
To: [email protected]
Cc: Grodzovsky, Andrey <[email protected]>; Quan, Evan <[email protected]>; Chen, Horace <[email protected]>; Tuikov, Luben <[email protected]>; Koenig, Christian <[email protected]>; Deucher, Alexander <[email protected]>; Xiao, Jack <[email protected]>; Zhang, Hawking <[email protected]>; Liu, Monk <[email protected]>; Xu, Feifei <[email protected]>; Wang, Kevin(Yang) <[email protected]>; Wang, YuBiao <[email protected]>
Subject: [PATCH 1/1] drm/amdgpu: Read clock counter via MMIO to reduce delay

[Why]
GPU timing counters are read via KIQ under sriov, which will introduce a delay.

[How]
It could be directly read by MMIO.

v2: Add additional check to prevent carryover issue.
v3: Only check for carryover for once to prevent performance issue.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +10 -2
index ff7e9f49040e..82a5b7ab8dc5 100644
@@ -7610,6 +7610,7 @@ static int gfx_v10_0_soft_reset(void *handle)  static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)  {
 	uint64_t clock;
+	uint64_t clock_lo, clock_hi, hi_check;
 
 	amdgpu_gfx_off_ctrl(adev, false);
 	mutex_lock(&adev->gfx.gpu_clock_mutex);
@@ -7620,8 +7621,15 @@ static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)
 			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER_Vangogh) << 32ULL);
 		break;
 	default:
-		clock = (uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER) |
-			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER) << 32ULL);
+		clock_hi = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+		hi_check = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		// If carryover happens, update lower count again.
+		if (hi_check != clock_hi) {
+			clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+			clock_hi = hi_check;
+		}
+		clock = (uint64_t)clock_lo | ((uint64_t)clock_hi << 32ULL);
 		break;
 	}
 	mutex_unlock(&adev->gfx.gpu_clock_mutex);

--
2.25.1
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#3 lore raw
In reply to YuBiao Wang

Am 29.06.21 um 12:01 schrieb YuBiao Wang:

Quoted 21 lines

[Why]
GPU timing counters are read via KIQ under sriov, which will introduce a delay.

[How]
It could be directly read by MMIO.

v2: Add additional check to prevent carryover issue.
v3: Only check for carryover for once to prevent performance issue.

---
  drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 12 ++++++++++--
  1 file changed, 10 insertions(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +0 -0
index ff7e9f49040e..82a5b7ab8dc5 100644

drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c static int gfx_v10_0_soft_reset(void *handle) · line 7613

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +1 -0
@@ -7610,6 +7610,7 @@ static int gfx_v10_0_soft_reset(void *handle)
  static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)
  {
  	uint64_t clock;
+	uint64_t clock_lo, clock_hi, hi_check;

You could put that on one line.

drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev) · line 7627

Quoted 13 lines
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +4 -2
  	amdgpu_gfx_off_ctrl(adev, false);
  	mutex_lock(&adev->gfx.gpu_clock_mutex);
@@ -7620,8 +7621,15 @@ static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev)
  			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER_Vangogh) << 32ULL);
  		break;
  	default:
-		clock = (uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER) |
-			((uint64_t)RREG32_SOC15(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER) << 32ULL);
+		clock_hi = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+		hi_check = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_UPPER);
+		// If carryover happens, update lower count again.

That is obvious and doesn't need a comment, but what you should comment is how unlikely a carry over is.

E.g. something like carry over only happens every x clock cycles which are roughly ~y days/weeks/month etc...

And as Monk noted as well please use kernel style comments.

Regards,
Christian.

drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c static uint64_t gfx_v10_0_get_gpu_clock_counter(struct amdgpu_device *adev) · line 7635

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +5 -0
+		if (hi_check != clock_hi) {
+			clock_lo = RREG32_SOC15_NO_KIQ(SMUIO, 0, mmGOLDEN_TSC_COUNT_LOWER);
+			clock_hi = hi_check;
+		}
+		clock = (uint64_t)clock_lo | ((uint64_t)clock_hi << 32ULL);
  		break;
  	}
  	mutex_unlock(&adev->gfx.gpu_clock_mutex);

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