Skip to contentSkip to conversation

drm/i915: Reduce locking in command submission

RFC In review Tvrtko Ursulin 4 participants 9 messages Dec 11, 2014 – Jan 16, 2015

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

RFC 1/1 drm/i915: Reduce locking in command submission

From: Tvrtko Ursulin <[email protected]>

Eliminate six needless spin lock/unlock pairs when writing ELSP.

RFC for now with some #define copy and paste.

---
 drivers/gpu/drm/i915/intel_lrc.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/intel_lrc.c +9 -7
index a82020e..f2f4a28 100644
@@ -276,6 +276,10 @@ static uint64_t execlists_ctx_descriptor(struct drm_i915_gem_object *ctx_obj)
 	return desc;
 }
 
+#define __raw_i915_read32(dev_priv__, reg__) readl((dev_priv__)->regs + (reg__))
+#define __raw_i915_write32(dev_priv__, reg__, val__) writel(val__, (dev_priv__)->regs + (reg__))
+#define __raw_posting_read(dev_priv__, reg__) (void)__raw_i915_read32(dev_priv__, reg__)
+
 static void execlists_elsp_write(struct intel_engine_cs *ring,
 				 struct drm_i915_gem_object *ctx_obj0,
 				 struct drm_i915_gem_object *ctx_obj1)
@@ -323,19 +327,17 @@ static void execlists_elsp_write(struct intel_engine_cs *ring,
 			dev_priv->uncore.funcs.force_wake_get(dev_priv,
 							      FORCEWAKE_ALL);
 	}
-	spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
 
-	I915_WRITE(RING_ELSP(ring), desc[1]);
-	I915_WRITE(RING_ELSP(ring), desc[0]);
-	I915_WRITE(RING_ELSP(ring), desc[3]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[1]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[0]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[3]);
 	/* The context is automatically loaded after the following */
-	I915_WRITE(RING_ELSP(ring), desc[2]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[2]);
 
 	/* ELSP is a wo register, so use another nearby reg for posting instead */
-	POSTING_READ(RING_EXECLIST_STATUS(ring));
+	__raw_posting_read(dev_priv, RING_EXECLIST_STATUS(ring));
 
 	/* Release Force Wakeup (see the big comment above). */
-	spin_lock_irqsave(&dev_priv->uncore.lock, flags);
 	if (IS_CHERRYVIEW(dev) || INTEL_INFO(dev)->gen >= 9) {
 		if (--dev_priv->uncore.fw_rendercount == 0)
 			dev_priv->uncore.funcs.force_wake_put(dev_priv,
Signature · 6 lines
2.1.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#1 lore raw
Quoted 8 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

Eliminate six needless spin lock/unlock pairs when writing ELSP.

RFC for now with some #define copy and paste.

Yeah makes sense. I'm on the fence whether we should do an all-uppercase conversion of the raw mmio macros, would be a nothc more consistent. And some perf data for this patch would be good, too.
-Daniel

Quoted 51 lines
---
 drivers/gpu/drm/i915/intel_lrc.c | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)
drivers/gpu/drm/i915/intel_lrc.c +9 -7
index a82020e..f2f4a28 100644
@@ -276,6 +276,10 @@ static uint64_t execlists_ctx_descriptor(struct drm_i915_gem_object *ctx_obj)
 	return desc;
 }
 
+#define __raw_i915_read32(dev_priv__, reg__) readl((dev_priv__)->regs + (reg__))
+#define __raw_i915_write32(dev_priv__, reg__, val__) writel(val__, (dev_priv__)->regs + (reg__))
+#define __raw_posting_read(dev_priv__, reg__) (void)__raw_i915_read32(dev_priv__, reg__)
+
 static void execlists_elsp_write(struct intel_engine_cs *ring,
 				 struct drm_i915_gem_object *ctx_obj0,
 				 struct drm_i915_gem_object *ctx_obj1)
@@ -323,19 +327,17 @@ static void execlists_elsp_write(struct intel_engine_cs *ring,
 			dev_priv->uncore.funcs.force_wake_get(dev_priv,
 							      FORCEWAKE_ALL);
 	}
-	spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
 
-	I915_WRITE(RING_ELSP(ring), desc[1]);
-	I915_WRITE(RING_ELSP(ring), desc[0]);
-	I915_WRITE(RING_ELSP(ring), desc[3]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[1]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[0]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[3]);
 	/* The context is automatically loaded after the following */
-	I915_WRITE(RING_ELSP(ring), desc[2]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[2]);
 
 	/* ELSP is a wo register, so use another nearby reg for posting instead */
-	POSTING_READ(RING_EXECLIST_STATUS(ring));
+	__raw_posting_read(dev_priv, RING_EXECLIST_STATUS(ring));
 
 	/* Release Force Wakeup (see the big comment above). */
-	spin_lock_irqsave(&dev_priv->uncore.lock, flags);
 	if (IS_CHERRYVIEW(dev) || INTEL_INFO(dev)->gen >= 9) {
 		if (--dev_priv->uncore.fw_rendercount == 0)
 			dev_priv->uncore.funcs.force_wake_put(dev_priv,
Signature · 6 lines
2.1.1

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
Signature · 7 lines
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#2 lore raw
Quoted 5 lines · Daniel Vetter
Quoted 4 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

Eliminate six needless spin lock/unlock pairs when writing ELSP.

Quoted 8 lines
Quoted 4 lines

RFC for now with some #define copy and paste.

Yeah makes sense. I'm on the fence whether we should do an all-uppercase conversion of the raw mmio macros, would be a nothc more consistent. And some perf data for this patch would be good, too.

I know perf data would be good but I had no time to set up a suitable platform for testing. This was more like a drive-by since it's not pretty so it annoyed me. Will see if I can get back to this in the near future.

Regards,

Tvrtko
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#3 lore raw
In reply to Daniel Vetter
Quoted 5 lines · Daniel Vetter
Quoted 4 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

Eliminate six needless spin lock/unlock pairs when writing ELSP.

Quoted 8 lines
Quoted 4 lines

RFC for now with some #define copy and paste.

Yeah makes sense. I'm on the fence whether we should do an all-uppercase conversion of the raw mmio macros, would be a nothc more consistent. And some perf data for this patch would be good, too.

With regards to perf data, Ben Widawsky was kind enough to give this patch a spin on his perf test bed (CHV), on a range of OGL benchmarks.

Apparently only two results have "confidence t-score" > 95% (statistics is not my area), bench_OglBatch4 and bench_OglDeferred which show 0.51% and 0.73% gains respectively.

Looking just on the basis of those two, I'd say the patch is worth cleaning up since it is a good gain for such a simple change.

Other results show anything from 4.29% slowdown (!*) (bench_OglTexFilterAniso) to a 7.08% gain (bench_OglMultithreaded).

Average across all benchmarks is a 0.38% gain.

Thoughts?

Regards,

Tvrtko

* I can't really understand regressions for some tests?!
_______________________________________________
Intel-gfx mailing list [email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#4 lore raw

PATCH 1/1 drm/i915: Reduce locking in command submission

In reply to Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

This eliminates six needless spin lock/unlock pairs when writing out ELSP. Apart from tidier code main benefit is between 0.51% and 0.73% speedup on some OGL tests under CHV (bench_OglBatch4 bench_OglDeferred respectively).

Kindly benchmarked by Ben Widawsky.

Cc Dave Gordon <[email protected]>
Cc Daniel Vetter <[email protected]>
Cc Ben Widawsky <[email protected]>
---
 drivers/gpu/drm/i915/i915_drv.h     | 15 +++++++++++++++
 drivers/gpu/drm/i915/intel_lrc.c    | 13 ++++++-------
 drivers/gpu/drm/i915/intel_uncore.c | 14 --------------
 3 files changed, 21 insertions(+), 21 deletions(-)
drivers/gpu/drm/i915/i915_drv.h +15 -0
index 66f0c60..33d577a 100644
@@ -3197,6 +3197,21 @@ int vlv_freq_opcode(struct drm_i915_private *dev_priv, int val);
 #define POSTING_READ(reg)	(void)I915_READ_NOTRACE(reg)
 #define POSTING_READ16(reg)	(void)I915_READ16_NOTRACE(reg)
 
+/* Raw MMIO access with no forcewake handling, use with care. */
+#define __raw_i915_read8(dev_priv__, reg__) readb((dev_priv__)->regs + (reg__))
+#define __raw_i915_write8(dev_priv__, reg__, val__) writeb(val__, (dev_priv__)->regs + (reg__))
+
+#define __raw_i915_read16(dev_priv__, reg__) readw((dev_priv__)->regs + (reg__))
+#define __raw_i915_write16(dev_priv__, reg__, val__) writew(val__, (dev_priv__)->regs + (reg__))
+
+#define __raw_i915_read32(dev_priv__, reg__) readl((dev_priv__)->regs + (reg__))
+#define __raw_i915_write32(dev_priv__, reg__, val__) writel(val__, (dev_priv__)->regs + (reg__))
+
+#define __raw_i915_read64(dev_priv__, reg__) readq((dev_priv__)->regs + (reg__))
+#define __raw_i915_write64(dev_priv__, reg__, val__) writeq(val__, (dev_priv__)->regs + (reg__))
+
+#define __raw_posting_read(dev_priv__, reg__) (void)__raw_i915_read32(dev_priv__, reg__)
+
 /* "Broadcast RGB" property */
 #define INTEL_BROADCAST_RGB_AUTO 0
 #define INTEL_BROADCAST_RGB_FULL 1
drivers/gpu/drm/i915/intel_lrc.c +6 -7
index e405b61..e22b866 100644
@@ -305,6 +305,7 @@ static void execlists_elsp_write(struct intel_engine_cs *ring,
 	 * Instead, we do the runtime_pm_get/put when creating/destroying requests.
 	 */
 	spin_lock_irqsave(&dev_priv->uncore.lock, flags);
+
 	if (IS_CHERRYVIEW(dev) || INTEL_INFO(dev)->gen >= 9) {
 		if (dev_priv->uncore.fw_rendercount++ == 0)
 			dev_priv->uncore.funcs.force_wake_get(dev_priv,
@@ -322,19 +323,17 @@ static void execlists_elsp_write(struct intel_engine_cs *ring,
 			dev_priv->uncore.funcs.force_wake_get(dev_priv,
 							      FORCEWAKE_ALL);
 	}
-	spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
 
-	I915_WRITE(RING_ELSP(ring), desc[1]);
-	I915_WRITE(RING_ELSP(ring), desc[0]);
-	I915_WRITE(RING_ELSP(ring), desc[3]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[1]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[0]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[3]);
 	/* The context is automatically loaded after the following */
-	I915_WRITE(RING_ELSP(ring), desc[2]);
+	__raw_i915_write32(dev_priv, RING_ELSP(ring), desc[2]);
 
 	/* ELSP is a wo register, so use another nearby reg for posting instead */
-	POSTING_READ(RING_EXECLIST_STATUS(ring));
+	__raw_posting_read(dev_priv, RING_EXECLIST_STATUS(ring));
 
 	/* Release Force Wakeup (see the big comment above). */
-	spin_lock_irqsave(&dev_priv->uncore.lock, flags);
 	if (IS_CHERRYVIEW(dev) || INTEL_INFO(dev)->gen >= 9) {
 		if (--dev_priv->uncore.fw_rendercount == 0)
 			dev_priv->uncore.funcs.force_wake_put(dev_priv,
drivers/gpu/drm/i915/intel_uncore.c +0 -14
index e9561de..9a31932 100644
@@ -26,20 +26,6 @@
 
 #define FORCEWAKE_ACK_TIMEOUT_MS 2
 
-#define __raw_i915_read8(dev_priv__, reg__) readb((dev_priv__)->regs + (reg__))
-#define __raw_i915_write8(dev_priv__, reg__, val__) writeb(val__, (dev_priv__)->regs + (reg__))
-
-#define __raw_i915_read16(dev_priv__, reg__) readw((dev_priv__)->regs + (reg__))
-#define __raw_i915_write16(dev_priv__, reg__, val__) writew(val__, (dev_priv__)->regs + (reg__))
-
-#define __raw_i915_read32(dev_priv__, reg__) readl((dev_priv__)->regs + (reg__))
-#define __raw_i915_write32(dev_priv__, reg__, val__) writel(val__, (dev_priv__)->regs + (reg__))
-
-#define __raw_i915_read64(dev_priv__, reg__) readq((dev_priv__)->regs + (reg__))
-#define __raw_i915_write64(dev_priv__, reg__, val__) writeq(val__, (dev_priv__)->regs + (reg__))
-
-#define __raw_posting_read(dev_priv__, reg__) (void)__raw_i915_read32(dev_priv__, reg__)
-
 static void
 assert_device_not_suspended(struct drm_i915_private *dev_priv)
 {
Signature · 6 lines
2.2.0

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
#5 lore raw
Quoted 5 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

This eliminates six needless spin lock/unlock pairs when writing out ELSP. Apart from tidier code main benefit is between 0.51% and 0.73% speedup on some OGL tests under CHV (bench_OglBatch4 bench_OglDeferred respectively).

With 95% confidence t-test on n=5

Quoted 2 lines

Kindly benchmarked by Ben Widawsky.

FWIW, as I mentioned on IRC, I think the reduction of the unnecessary forcewake (someone should fix the shadow register list) is probably more beneficial than removing the spin on an uncontested lock. I was tempted to try that myself, but I didn't have time or much interest since your patch accomplishes the same thing.

The sucky thing, which I actually care about since I've been doing a lot of profiling, is the raw MMIO doesn't show up with our i915 trace functions. It's obtainable still, but then I get a mess of other stuff I don't want.

Quoted 5 lines
Cc Dave Gordon <[email protected]>
Cc Daniel Vetter <[email protected]>
Cc Ben Widawsky <[email protected]>

[snip]

_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#6 lore raw
Quoted 10 lines · Ben Widawsky
Quoted 5 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

This eliminates six needless spin lock/unlock pairs when writing out ELSP. Apart from tidier code main benefit is between 0.51% and 0.73% speedup on some OGL tests under CHV (bench_OglBatch4 bench_OglDeferred respectively).

With 95% confidence t-test on n=5

Quoted 1 line
Quoted 7 lines
Quoted 1 line

Kindly benchmarked by Ben Widawsky.

FWIW, as I mentioned on IRC, I think the reduction of the unnecessary forcewake (someone should fix the shadow register list) is probably more beneficial than removing the spin on an uncontested lock. I was tempted to try that myself, but I didn't have time or much interest since your patch accomplishes the same thing.

I missed that IRC discussion, but I don't think it was doing forcewakes since the outer block in execlists_elsp_write bumps the counters which made I915_WRITE & co skip them.

Regards,

Tvrtko
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

#7 lore raw
Quoted 20 lines · Tvrtko Ursulin
Quoted 10 lines · Ben Widawsky
Quoted 5 lines · Tvrtko Ursulin

From: Tvrtko Ursulin <[email protected]>

This eliminates six needless spin lock/unlock pairs when writing out ELSP. Apart from tidier code main benefit is between 0.51% and 0.73% speedup on some OGL tests under CHV (bench_OglBatch4 bench_OglDeferred respectively).

With 95% confidence t-test on n=5

Quoted 1 line
Quoted 7 lines
Quoted 1 line

Kindly benchmarked by Ben Widawsky.

FWIW, as I mentioned on IRC, I think the reduction of the unnecessary forcewake (someone should fix the shadow register list) is probably more beneficial than removing the spin on an uncontested lock. I was tempted to try that myself, but I didn't have time or much interest since your patch accomplishes the same thing.

Quoted 7 lines

I missed that IRC discussion, but I don't think it was doing forcewakes since the outer block in execlists_elsp_write bumps the counters which made I915_WRITE & co skip them.

Regards,

Tvrtko

I didn't check the locking but it looks like it could actually get decremented once the spinlock is released. Probably never happens, but I think it's possible.

I completely missed that block somehow. I think my eyes skipped over it because how could getting forcewake take like 10+ lines :D
_______________________________________________
Intel-gfx mailing list
[email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Tested-by on v1 1/1
#8 lore raw
In reply to Tvrtko Ursulin

Task id: 5585 -------------------------------------Summary-------------------------------------
Platform Delta drm-intel-nightly Series Applied
PNV 353/353 353/353
ILK -1 200/200 199/200

SNB                                  400/422              400/422
IVB                                  487/487              487/487
BYT                                  296/296              296/296
HSW              +22-1              486/508              507/508
BDW                 -1              402/402              401/402
-------------------------------------Detailed-------------------------------------
Platform  Test                                drm-intel-nightly          Series Applied
*ILK  igt_gem_concurrent_blit_gtt-bcs-overwrite-source      PASS(2, M37)      NO_RESULT(1, M37)
 HSW  igt_kms_cursor_crc_cursor-size-change      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_kms_fence_pin_leak      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_kms_flip_event_leak      NSPT(2, M40)PASS(3, M20)      PASS(1, M20)
 HSW  igt_kms_flip_flip-vs-dpms-off-vs-modeset      DMESG_WARN(2, M20M40)PASS(1, M40)      DMESG_WARN(1, M20)
 HSW  igt_kms_mmio_vs_cs_flip_setcrtc_vs_cs_flip      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_kms_mmio_vs_cs_flip_setplane_vs_cs_flip      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_lpsp_non-edp      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_cursor      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_cursor-dpms      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_dpms-mode-unset-non-lpsp      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_dpms-non-lpsp      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_drm-resources-equal      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_fences      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_fences-dpms      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_gem-execbuf      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_gem-mmap-cpu      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_gem-mmap-gtt      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_gem-pread      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_i2c      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_modeset-non-lpsp      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_modeset-non-lpsp-stress-no-wait      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_pci-d3-state      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
 HSW  igt_pm_rpm_rte      NSPT(1, M40)PASS(4, M20M40)      PASS(1, M20)
*BDW  igt_gem_concurrent_blit_gtt-rcs-early-read-interruptible      PASS(7, M30M28)      DMESG_WARN(1, M30)

Note: You need to pay more attention to line start with '*'
_______________________________________________
Intel-gfx mailing list [email protected]
http://lists.freedesktop.org/mailman/listinfo/intel-gfx