cheng/libpng
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[libpng16] Improved DIV65535 algorithm in simpleover.c. By experiment,

Browse Source 
(x + (x>>16) + 32769) works as well as the previous form and avoids one shift.
This commit is contained in:
John Bowler 2015-05-06 11:41:17 -05:00 committed by Glenn Randers-Pehrson
parent d6b1375bf5
commit 74428df2fd
1 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
contrib/examples/simpleover.c
View File

@ -60,12 +60,12 @@ struct sprite {
int main(void) {
double err = 0;
unsigned int xerr = 0;
unsigned int r = 32768;
unsigned int r = 32769;
{
unsigned int x = 0;
do {
unsigned int t = x + (x >> 16) + (x >> 31) + r;
unsigned int t = x + (x >> 16) + /*(x >> 31)*/ + r;
double v = x, errtest;
if (t < x) {
@ -149,7 +149,7 @@ sprite_op(const struct sprite *sprite, int x_offset, int y_offset,
* of the input alpha (1-alpha). A division is required but
* it is by the constant 65535. Approximate this as:
*
* (x + (x >> 16) + (x >> 31) + 32768) >> 16;
* (x + (x >> 16) + 32769) >> 16;
*
* This is exact (and does not overflow) for all values of
* x in the range 0..65535*65535. (Note that the calculation
@ -159,7 +159,7 @@ sprite_op(const struct sprite *sprite, int x_offset, int y_offset,
# define compose(c)\
tmp = out_pixel[c] * in_alpha;\
tmp = (tmp + (tmp >> 16) + (tmp >> 31) + 32768) >> 16;\
tmp = (tmp + (tmp >> 16) + 32769) >> 16;\
out_pixel[c] = tmp + in_pixel[c]
/* The following is very vectorizable... */
@ -488,7 +488,7 @@ simpleover_process(png_imagep output, png_bytep out_buf, int argc,
if (strcmp(sprites[isprite].name, name) == 0)
{
if (!add_sprite(output, out_buf, sprites+isprite, &argc, &argv))
goto out; /* error in addsprite */
goto out; /* error in add_sprite */
break;
}