mirror of https://lore.kernel.org/lkml/
 help / color / mirror / Atom feed
* [PATCH] drm/fixed: use fls64() in drm_fixp_msbset()
@ 2026-11-17 11:32 Oarora Etimis
  2026-11-17 15:25 ` Louis Chauvet
  0 siblings, 1 reply; 2+ messages in thread
From: Oarora Etimis @ 2026-11-17 11:32 UTC (permalink / raw)
  To: Maarten Lankhorst, Maxime Ripard, Thomas Zimmermann,
	David Airlie, Simona Vetter
  Cc: dri-devel, linux-kernel, Oarora Etimis

Replace the loop with fls64() on the value with the sign folded away.

Output is unchanged. The loop runs once per bit it has to skip, so a
channel value takes it about 47 iterations and 347 executed
instructions. The replacement is 6, and the work is one bit scan: bsr on
x86-64, two bsrl on i386, clz on arm and riscv.

Signed-off-by: Oarora Etimis <oaroraetimis@gmail.com>
---
drm_fixp_msbset() walks from bit 62 down, one bit at a time, to find the
highest bit that differs from the sign bit. fls64() returns the same
value in one step.

In the other users it is only called a few times, so the loop does not
show. In vkms it runs for every channel of every pixel of every frame,
and a channel value is 16 bits wide, so each run starts 47 compares from
its answer.

  Link: https://lore.kernel.org/20261108-lut-v1-0-f762d70ae4cd@gmail.com

Fold the sign away and let the CPU find the highest set bit. The range
drm_fixp_div() shifts by is unchanged, and the output is bit-identical.

Counted with callgrind, gcc 16.2.1 at -O2 on x86-64, function body only.
---
 include/drm/drm_fixed.h | 9 +++------
 1 file changed, 3 insertions(+), 6 deletions(-)

diff --git a/include/drm/drm_fixed.h b/include/drm/drm_fixed.h
index 21d822aeed55..7dbe89ca30c6 100644
--- a/include/drm/drm_fixed.h
+++ b/include/drm/drm_fixed.h
@@ -28,6 +28,7 @@
 #include <linux/math64.h>
 #include <linux/types.h>
 #include <linux/wordpart.h>
+#include <linux/bitops.h>
 
 typedef union dfixed {
 	u32 full;
@@ -121,13 +122,9 @@ static inline int drm_fixp2int_ceil(s64 a)
 
 static inline unsigned drm_fixp_msbset(s64 a)
 {
-	unsigned shift, sign = (a >> 63) & 1;
+	u64 v = (a < 0) ? ~(u64)a : (u64)a;
 
-	for (shift = 62; shift > 0; --shift)
-		if (((a >> shift) & 1) != sign)
-			return shift;
-
-	return 0;
+	return fls64(v | 1) - 1;
 }
 
 static inline s64 drm_fixp_mul(s64 a, s64 b)

---
base-commit: fc72cd52fb284f3707fa36a68ea5a22543a51d71
change-id: 20261113-fixp-msbset-5965c14bcb9b

Best regards,
--  
Oarora Etimis <oaroraetimis@gmail.com>


^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [PATCH] drm/fixed: use fls64() in drm_fixp_msbset()
  2026-11-17 11:32 [PATCH] drm/fixed: use fls64() in drm_fixp_msbset() Oarora Etimis
@ 2026-11-17 15:25 ` Louis Chauvet
  0 siblings, 0 replies; 2+ messages in thread
From: Louis Chauvet @ 2026-11-17 15:25 UTC (permalink / raw)
  To: Oarora Etimis, Maarten Lankhorst, Maxime Ripard,
	Thomas Zimmermann, David Airlie, Simona Vetter
  Cc: dri-devel, linux-kernel



On 11/17/26 12:32, Oarora Etimis wrote:
> Replace the loop with fls64() on the value with the sign folded away.
> 
> Output is unchanged. The loop runs once per bit it has to skip, so a
> channel value takes it about 47 iterations and 347 executed
> instructions. The replacement is 6, and the work is one bit scan: bsr on
> x86-64, two bsrl on i386, clz on arm and riscv.
> 
> Signed-off-by: Oarora Etimis <oaroraetimis@gmail.com>

I did some quick testing, it seems to really improve the VKMS 
performance on some workload (up to 3x faster)

Reviewed-by: Louis Chauvet <louis.chauvet@bootlin.com>

> ---
> drm_fixp_msbset() walks from bit 62 down, one bit at a time, to find the
> highest bit that differs from the sign bit. fls64() returns the same
> value in one step.
> 
> In the other users it is only called a few times, so the loop does not
> show. In vkms it runs for every channel of every pixel of every frame,
> and a channel value is 16 bits wide, so each run starts 47 compares from
> its answer.
> 
>    Link: https://lore.kernel.org/20261108-lut-v1-0-f762d70ae4cd@gmail.com
> 
> Fold the sign away and let the CPU find the highest set bit. The range
> drm_fixp_div() shifts by is unchanged, and the output is bit-identical.
> 
> Counted with callgrind, gcc 16.2.1 at -O2 on x86-64, function body only.
> ---
>   include/drm/drm_fixed.h | 9 +++------
>   1 file changed, 3 insertions(+), 6 deletions(-)
> 
> diff --git a/include/drm/drm_fixed.h b/include/drm/drm_fixed.h
> index 21d822aeed55..7dbe89ca30c6 100644
> --- a/include/drm/drm_fixed.h
> +++ b/include/drm/drm_fixed.h
> @@ -28,6 +28,7 @@
>   #include <linux/math64.h>
>   #include <linux/types.h>
>   #include <linux/wordpart.h>
> +#include <linux/bitops.h>
>   
>   typedef union dfixed {
>   	u32 full;
> @@ -121,13 +122,9 @@ static inline int drm_fixp2int_ceil(s64 a)
>   
>   static inline unsigned drm_fixp_msbset(s64 a)
>   {
> -	unsigned shift, sign = (a >> 63) & 1;
> +	u64 v = (a < 0) ? ~(u64)a : (u64)a;
>   
> -	for (shift = 62; shift > 0; --shift)
> -		if (((a >> shift) & 1) != sign)
> -			return shift;
> -
> -	return 0;
> +	return fls64(v | 1) - 1;
>   }
>   
>   static inline s64 drm_fixp_mul(s64 a, s64 b)
> 
> ---
> base-commit: fc72cd52fb284f3707fa36a68ea5a22543a51d71
> change-id: 20261113-fixp-msbset-5965c14bcb9b
> 
> Best regards,
> --
> Oarora Etimis <oaroraetimis@gmail.com>
> 


^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-17 15:25 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-17 11:32 [PATCH] drm/fixed: use fls64() in drm_fixp_msbset() Oarora Etimis
2026-11-17 15:25 ` Louis Chauvet

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox

all inboxes | Powered by JetHome®