Re: [Linux-fbdev-devel] [PATCH 1/1 2.6.13] framebuffer: bit_putcs() optimization for 8x* fonts

[Knut Petersen <Knut_Petersen@t-online.de>]

 Hi Roman!

> > +static inline void __fb_pad_aligned_buffer(u8 *dst, u32 d_pitch, u8 *src, +
> > u32 s_pitch, u32 height)
> > +{
> > +	int i, j;
> > +
> > +	if (likely(s_pitch==1))
> > +		for(i=0; i < height; i++)
> > +			dst[d_pitch*i] = src[i];
> >    

I added the multiply back because gcc (v. 3.3.4) does generate the 
fastest code
if I write it this way. I compiled, inspected the generated assembly and 
benchmarked
about a  dozend variations of the code (benchmark as previously described).

The special case for s_pitch == 1 saves about 10 ms system time (770 ms 
-> 760 ms)

The special case for s_pitch == 2 saves about 270 ms system time (2120 
-> 1850ms)
with a 16x30 font.

The third case is for even bigger fonts ... I believe that it will not 
often be used but
something like that must be present.

> You have now 3 slightly different variants of the same, which isn't really 
> an improvement. In my example I showed you how to generate the first and 
> last version from the same source.
The first version will only be generated when gcc can be sure that 
s_pitch is 1.
Therefore you had to explicitly call __fb_pad_aligned_buffer with that 
value:

	if (likely(idx == 1))
		__fb_pad_aligned_buffer(dst, pitch, src, 1, image.height);
	else
		fb_pad_aligned_buffer(dst, pitch, src, idx, image.height);

With the version I propose it´s enough to write

		__fb_pad_aligned_buffer(dst, pitch, src, idx, image.height);

instead, and you will get good performance for all cases. If the value of
idx/s_pitch is know at compile time, the compiler can and will ignore the
other cases.

> If you also want to optimize for other sizes, you might want to always 
> inline the function, if the function call overhead is the largest part 
> anyway, the special case for 2 bytes might not be needed anymore.
>  
fb_pad_aligned_buffer() is  usefull to save some space in cases like 
softcursor.
It´s also used by some  drivers (nvidia and riva), but the authors of 
those drivers
have to decide if they prefer the inlined version or the version fixed 
in fbmem.

> BTW this version saves another condition:
>
> static inline void __fb_pad_aligned_buffer(u8 *dst, u32 d_pitch, u8 *src, u32 s_pitch, u32 height)
> {
> 	int i, j;
>
> 	d_pitch -= s_pitch;
> 	i = height;
> 	do {
> 		/* s_pitch is a few bytes at the most, memcpy is suboptimal */
> 		j = s_pitch;
> 		do
> 			*dst++ = *src++;
> 		while (--j > 0);
> 		dst += d_pitch;
> 	} while (--i > 0);
> }
I tested that code, together with the followig code in but_putcs():

            if(idx==1)
                   __fb_pad_aligned_buffer(dst, pitch, src, 1, 
image.height);
            else if (idx==2)
                   __fb_pad_aligned_buffer(dst, pitch, src, 2, 
image.height);
            else
                   fb_pad_aligned_buffer(dst, pitch, src, idx, 
image.height);
            dst += width;

It´s as fast/slow as your previous version, the measurements are almost 
identical.

Let´s summarize:

Your version of __fb_pad_aligned_buffer looks much better, but it needs 
not so nice
conditionals when used. My version looks bad, but it is easier to use 
and it is
_faster_.

cu,
knut



-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/