github-code/lvgl
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(decoder): use shift and mask to deal with rounding error (#5042)

Browse Source 
Signed-off-by: Xu Xingliang <xuxingliang@xiaomi.com>
This commit is contained in:
Neo Xu
2023-12-19 03:16:33 +08:00
committed by GitHub
parent 8af0cfa7aa
commit 347b5eb23d
1 changed files with 23 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
src/libs/bin_decoder/lv_bin_decoder.c
View File

@@ -672,6 +672,23 @@ static lv_result_t decode_rgb(lv_image_decoder_t * decoder, lv_image_decoder_dsc
}
#endif
/**
* Extend A1/2/4 to A8 with interpolation to reduce rounding error.
*/
static inline uint8_t bit_extend(uint8_t value, uint8_t bpp)
{
if(value == 0) return 0;
uint8_t res = value;
uint8_t bpp_now = bpp;
while(bpp_now < 8) {
res |= value << (8 - bpp_now);
bpp_now += bpp;
};
return res;
}
static lv_result_t decode_alpha_only(lv_image_decoder_t * decoder, lv_image_decoder_dsc_t * dsc)
{
LV_UNUSED(decoder);
@@ -720,19 +737,14 @@ static lv_result_t decode_alpha_only(lv_image_decoder_t * decoder, lv_image_deco
/**
* Rounding error:
* Take bpp = 4 as example, alpha value of 0x0 to 0x0F should be
* mapped to 0x00 to 0xFF, thus the equation should be below Equation 3.
* mapped to 0x00 to 0xFF. Using below equation will give us 0x00 to 0xF0
* thus causes error. We can simply interpolate the value to fix it.
*
* But it involves division and multiplication, which is slow. So, if
* we ignore the rounding errors, Equation1, 2 could be faster. But it
* will either has error when alpha is 0xff or 0x00.
*
* We use Equation 3 here for maximum accuracy.
*
* Equation 1: *out = ((*in >> shift) & mask) << (8 - bpp);
* Equation 2: *out = ((((*in >> shift) & mask) + 1) << (8 - bpp)) - 1;
* Equation 3: *out = ((*in >> shift) & mask) * 255 / ((1L << bpp) - 1) ;
* Equation: *out = ((*in >> shift) & mask) << (8 - bpp);
* Ideal: *out = ((*in >> shift) & mask) * 255 / ((1L << bpp) - 1)
*/
*out = ((*in >> shift) & mask) * 255L / ((1L << bpp) - 1) ;
uint8_t value = ((*in >> shift) & mask);
*out = bit_extend(value, bpp);
shift += bpp;
if(shift >= 8) {
shift = 0;