Support more scaling factors for JPEG decompression
Since libjpeg 7, decompression scaling factors of [1,16]/8 have been supported. Upscaling is slower, but using the entire range of downscaling factors is significantly faster for sizes between the power-of-two factors. Time to decompress a 5184x3456 image and scale to 1900x1200, slightly less than 3/8ths, changes from 251ms to 203ms. libjpeg versions prior to 7 will round up to the next largest factor they support, and continue to work as before. Change-Id: I00a0655df2ef057e739927a643bfe0b0cabd5602 Reviewed-by: Robin Burchell <robin+qt@viroteck.net> Reviewed-by: aavit <qt_aavit@ovi.com>
This commit is contained in:
parent
011b88a7b3
commit
56ff31f0c1
@ -45,6 +45,7 @@
|
||||
#include <qvariant.h>
|
||||
#include <qvector.h>
|
||||
#include <qbuffer.h>
|
||||
#include <qmath.h>
|
||||
#include <private/qsimd_p.h>
|
||||
|
||||
#include <stdio.h> // jpeglib needs this to be pre-included
|
||||
@ -321,27 +322,31 @@ static bool read_jpeg_image(QImage *outImage,
|
||||
}
|
||||
|
||||
// Determine the scale factor to pass to libjpeg for quick downscaling.
|
||||
if (!scaledSize.isEmpty()) {
|
||||
if (!scaledSize.isEmpty() && info->image_width && info->image_height) {
|
||||
if (clipRect.isEmpty()) {
|
||||
info->scale_denom =
|
||||
qMin(info->image_width / scaledSize.width(),
|
||||
info->image_height / scaledSize.height());
|
||||
} else {
|
||||
info->scale_denom =
|
||||
qMin(clipRect.width() / scaledSize.width(),
|
||||
clipRect.height() / scaledSize.height());
|
||||
}
|
||||
if (info->scale_denom < 2) {
|
||||
info->scale_denom = 1;
|
||||
} else if (info->scale_denom < 4) {
|
||||
info->scale_denom = 2;
|
||||
} else if (info->scale_denom < 8) {
|
||||
info->scale_denom = 4;
|
||||
} else {
|
||||
double f = qMin(double(info->image_width) / scaledSize.width(),
|
||||
double(info->image_height) / scaledSize.height());
|
||||
|
||||
// libjpeg supports M/8 scaling with M=[1,16]. All downscaling factors
|
||||
// are a speed improvement, but upscaling during decode is slower.
|
||||
info->scale_num = qBound(1, qCeil(8/f), 8);
|
||||
info->scale_denom = 8;
|
||||
}
|
||||
info->scale_num = 1;
|
||||
if (!clipRect.isEmpty()) {
|
||||
} else {
|
||||
info->scale_denom = qMin(clipRect.width() / scaledSize.width(),
|
||||
clipRect.height() / scaledSize.height());
|
||||
|
||||
// Only scale by powers of two when clipping so we can
|
||||
// keep the exact pixel boundaries
|
||||
if (info->scale_denom < 2)
|
||||
info->scale_denom = 1;
|
||||
else if (info->scale_denom < 4)
|
||||
info->scale_denom = 2;
|
||||
else if (info->scale_denom < 8)
|
||||
info->scale_denom = 4;
|
||||
else
|
||||
info->scale_denom = 8;
|
||||
info->scale_num = 1;
|
||||
|
||||
// Correct the scale factor so that we clip accurately.
|
||||
// It is recommended that the clip rectangle be aligned
|
||||
// on an 8-pixel boundary for best performance.
|
||||
|
Loading…
Reference in New Issue
Block a user