<div dir="ltr"><div>Excellent idea to put the error diffusion into the division. Just a bit of cleanup and changes for some suspected bugs (that are probably invisible but might as well get them):</div><div class="gmail_extra"><br><div class="gmail_quote">On Mon, Apr 11, 2016 at 7:36 PM, Søren Sandmann Pedersen <span dir="ltr"><<a href="mailto:soren.sandmann@gmail.com" target="_blank">soren.sandmann@gmail.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">There are a few bugs in the current normalization code<br>
<br>
(1) The normalization is based on the sum of the *floating point*<br>
values generated by integral(). But in order to get the sum to be<br>
close to pixman_fixed_1, the sum of the rounded fixed point values<br>
should be used.<br>
<br>
(2) The multiplications in the normalization loops often round the<br>
same way, so the residual error can fairly large.<br>
<br>
(3) The residual error is added to the sample located at index<br>
(width - width / 2), which is not the midpoint for odd widths (and<br>
for width 1 is in fact outside the array).<br>
<br>
This patch fixes these issues by (1) using the sum of the fixed point<br>
values as the total to divide by, (2) doing error diffusion in the<br>
normalization loop, and (3) putting any residual error (which is now<br>
guaranteed to be less than pixman_fixed_e) at the first sample, which<br>
is the only one that didn't get any error diffused into it.<br>
<br>
Signed-off-by: Søren Sandmann <<a href="mailto:soren.sandmann@gmail.com">soren.sandmann@gmail.com</a>><br>
---<br>
pixman/pixman-filter.c | 23 +++++++++++++++--------<br>
1 file changed, 15 insertions(+), 8 deletions(-)<br>
<br>
diff --git a/pixman/pixman-filter.c b/pixman/pixman-filter.c<br>
index 32aaa9a..4abd05f 100644<br>
--- a/pixman/pixman-filter.c<br>
+++ b/pixman/pixman-filter.c<br>
@@ -247,7 +247,7 @@ create_1d_filter (int width,<br>
double frac = step / 2.0 + i * step;<br>
pixman_fixed_t new_total;<br>
int x, x1, x2;<br>
- double total;<br>
+ double total, e;<br></blockquote><div><br></div><div>I think total should be a pixman_fixed_t as that is what it is summing.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
/* Sample convolution of reconstruction and sampling<br>
* filter. See rounding.txt regarding the rounding<br>
@@ -278,24 +278,31 @@ create_1d_filter (int width,<br>
ihigh - ilow);<br>
}<br>
<br>
- total += c;<br>
- *p++ = (pixman_fixed_t)(c * 65536.0 + 0.5);<br>
+ *p = (pixman_fixed_t)floor (c * 65536.0 + 0.5);<br></blockquote><div><br></div><div>floor probably is a good idea to make negative filter entries work. Is there a macro that does this conversion?</div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">+ total += *p;<br>
+ p++;<br>
}<br></blockquote><div><br></div><div>Might want to skip the normalize if total==pixman_fixed_1, though perhaps a test to see how often that happens would be informative.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br>
- /* Normalize */<br>
+ /* Normalize, with error diffusion */<br>
p -= width;<br>
- total = 1 / total;<br>
+ total = 65536.0 / total;<br></blockquote><div><br></div><div>Remove the division so total can be a pixman_fixed_t.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
new_total = 0;<br>
+ e = 0.0;<br></blockquote><div><br></div><div>Change this to 0.5</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
for (x = x1; x < x2; ++x)<br>
{<br>
- pixman_fixed_t t = (*p) * total + 0.5;<br>
+ double v = (*p) * total + e; </blockquote><div><br></div><div>Change to this so total can be a pixman_fixed_t:</div><div> double v = (*p) * (double)(pixman_fixed_1) / total + e;</div><div><br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+ pixman_fixed_t t = floor (v + 0.5);<br></blockquote><div><br></div><div>Change this to just floor(v). The 0.5 factor is incorporated into e. This version is in effect adding .5 to all the samples, though that is compensated for somewhat by floor in effect subtracting .5, but imagine what happens if e > 0.5. This bug was in Nuke for many years btw.</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
+ e = v - t;<br>
new_total += t;<br>
*p++ = t;<br>
}<br>
<br>
- if (new_total != pixman_fixed_1)<br>
- *(p - width / 2) += (pixman_fixed_1 - new_total);<br>
+ /* pixman_fixed_e's worth of error may remain; put it<br>
+ * at the first sample, since that is the only one that<br>
+ * hasn't had any error diffused into it.<br>
+ */<br>
+ *(p - width) += pixman_fixed_1 - new_total;<br></blockquote><div><br></div><div>I'm not absolutely convinced that the first sample is best. Dumping this on the central pixel may be better because that value is larger so the relative change is smaller:</div><div><br></div><div> *(p - (width+1)/2) += pixman_fixed_1 - new_total;</div><div> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
}<br>
}<br>
<span class="HOEnZb"><font color="#888888"><br>
--<br>
1.7.11.7<br>
<br>
</font></span></blockquote></div><br></div></div>