[PATCH 2/2] fbdev: Improve performance of sys_imageblit()

[Thomas Zimmermann]

Hi Sam

Am 18.02.22 um 11:14 schrieb Sam Ravnborg:
> Hi Thomas,
> 
> On Thu, Feb 17, 2022 at 11:34:05AM +0100, Thomas Zimmermann wrote:
>> Improve the performance of sys_imageblit() by manually unrolling
>> the inner blitting loop and moving some invariants out. The compiler
>> failed to do this automatically. The resulting binary code was even
>> slower than the cfb_imageblit() helper, which uses the same algorithm,
>> but operates on I/O memory.
> 
> It would be super to have the same optimization done to cfb_imageblit(),
> to prevent that the two codebases diverge more than necessary.
> Also I think cfb_ version would also see a performance gain from this.

Yes, I can do that.

> 
> The actual implementation looks good.
> So with or without the extra un-rolling the patch is:
> Acked-by: Sam Ravnborg <sam at ravnborg.org>
> 
> One small nit belwo.
> 
> 	Sam
> 
>>
>> A microbenchmark measures the average number of CPU cycles
>> for sys_imageblit() after a stabilizing period of a few minutes
>> (i7-4790, FullHD, simpledrm, kernel with debugging). The value
>> for CFB is given as a reference.
>>
>>    sys_imageblit(), new: 25934 cycles
>>    sys_imageblit(), old: 35944 cycles
>>    cfb_imageblit():      30566 cycles
>>
>> In the optimized case, sys_imageblit() is now ~30% faster than before
>> and ~20% faster than cfb_imageblit().
>>
>> Signed-off-by: Thomas Zimmermann <tzimmermann at suse.de>
>> ---
>>   drivers/video/fbdev/core/sysimgblt.c | 51 +++++++++++++++++++++-------
>>   1 file changed, 39 insertions(+), 12 deletions(-)
>>
>> diff --git a/drivers/video/fbdev/core/sysimgblt.c b/drivers/video/fbdev/core/sysimgblt.c
>> index a4d05b1b17d7..d70d65af6fcb 100644
>> --- a/drivers/video/fbdev/core/sysimgblt.c
>> +++ b/drivers/video/fbdev/core/sysimgblt.c
>> @@ -188,23 +188,32 @@ static void fast_imageblit(const struct fb_image *image, struct fb_info *p,
>>   {
>>   	u32 fgx = fgcolor, bgx = bgcolor, bpp = p->var.bits_per_pixel;
>>   	u32 ppw = 32/bpp, spitch = (image->width + 7)/8;
>> -	u32 bit_mask, end_mask, eorx, shift;
>> +	u32 bit_mask, eorx;
>>   	const char *s = image->data, *src;
>>   	u32 *dst;
>> -	const u32 *tab = NULL;
>> -	int i, j, k;
>> +	const u32 *tab;
>> +	size_t tablen;
>> +	u32 colortab[16];
>> +	int i, j, k, jdecr;
>> +
>> +	if ((uintptr_t)dst1 % 8)
>> +		return;
> This check is new - and should not trigger ever. Maybe add an unlikely
> and a WARN_ON_ONCE()?

I think I can remove this test. It was supposed to tell the compiler 
that dst1 is 8-aligned, but I don't think it worked.

Best regards
Thomas

> 
> 
>>   
>>   	switch (bpp) {
>>   	case 8:
>>   		tab = fb_be_math(p) ? cfb_tab8_be : cfb_tab8_le;
>> +		tablen = 16;
>>   		break;
>>   	case 16:
>>   		tab = fb_be_math(p) ? cfb_tab16_be : cfb_tab16_le;
>> +		tablen = 4;
>>   		break;
>>   	case 32:
>> -	default:
>>   		tab = cfb_tab32;
>> +		tablen = 2;
>>   		break;
>> +	default:
>> +		return;
>>   	}
>>   
>>   	for (i = ppw-1; i--; ) {
>> @@ -217,19 +226,37 @@ static void fast_imageblit(const struct fb_image *image, struct fb_info *p,
>>   	bit_mask = (1 << ppw) - 1;
>>   	eorx = fgx ^ bgx;
>>   	k = image->width/ppw;
>> +	jdecr = 8 / ppw;
>> +
>> +	for (i = 0; i < tablen; ++i)
>> +		colortab[i] = (tab[i] & eorx) ^ bgx;
> This code could have been embedded with the switch (bpp) {
> That would have made some sense I think.
> But both ways works, so this was just a small observation.
> 
> 	Sam

-- 
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Maxfeldstr. 5, 90409 Nürnberg, Germany
(HRB 36809, AG Nürnberg)
Geschäftsführer: Ivo Totev
-------------- next part --------------
A non-text attachment was scrubbed...
Name: OpenPGP_signature
Type: application/pgp-signature
Size: 840 bytes
Desc: OpenPGP digital signature
URL: <https://lists.freedesktop.org/archives/dri-devel/attachments/20220218/ffed4824/attachment.sig>