Backups/TwelveMonkeys
2
0
Fork 0
mirror of https://github.com/haraldk/TwelveMonkeys.git synced 2025-10-04 11:26:44 -04:00
Code Issues Projects Releases Wiki Activity

#417: Half precision support (clean-up)

Browse Source
This commit is contained in:
Harald Kuhr
2021-04-10 14:17:15 +02:00
parent 01a4e55185
commit 9adf0f4da3
3 changed files with 343 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
imageio/imageio-tiff/src/main/java/com/twelvemonkeys/imageio/plugins/tiff/TIFFImageReader.java
View File

@@ -42,6 +42,7 @@ import com.twelvemonkeys.imageio.metadata.iptc.IPTCReader;
import com.twelvemonkeys.imageio.metadata.jpeg.JPEG;
import com.twelvemonkeys.imageio.metadata.psd.PSD;
import com.twelvemonkeys.imageio.metadata.psd.PSDReader;
import com.twelvemonkeys.imageio.metadata.tiff.Half;
import com.twelvemonkeys.imageio.metadata.tiff.Rational;
import com.twelvemonkeys.imageio.metadata.tiff.TIFF;
import com.twelvemonkeys.imageio.metadata.tiff.TIFFReader;
@@ -2019,49 +2020,10 @@ public final class TIFFImageReader extends ImageReaderBase {
private void toFloat(final float[] rowDataFloat, final short[] rowDataShort) {
for (int i = 0; i < rowDataFloat.length; i++) {
rowDataFloat[i] = toFloat(rowDataShort[i]);
rowDataFloat[i] = Half.shortBitsToFloat(rowDataShort[i]);
}
}
/**
* Converts an IEEE 754 half-precision data type to single-precision.
*
* @param shortValue a 16 bit half precision value
* @return an IEE 754 single precision float
*
* @see <a href="https://stackoverflow.com/a/6162687/259991">Stack Overflow answer by x4u</a>
* @see <a href="https://en.wikipedia.org/wiki/Half-precision_floating-point_format>Wikipedia</a>
*/
private float toFloat(final short shortValue) {
int mantissa = shortValue & 0x03ff; // 10 bits mantissa
int exponent = shortValue & 0x7c00; // 5 bits exponent
if (exponent == 0x7c00) { // NaN/Inf
exponent = 0x3fc00; // -> NaN/Inf
}
else if (exponent != 0) { // Normalized value
exponent += 0x1c000; // exp - 15 + 127
// Smooth transition
if (mantissa == 0 && exponent > 0x1c400) {
return Float.intBitsToFloat((shortValue & 0x8000) << 16 | exponent << 13 | 0x3ff);
}
}
else if (mantissa != 0) { // && exp == 0 -> subnormal
exponent = 0x1c400; // Make it normal
do {
mantissa <<= 1; // mantissa * 2
exponent -= 0x400; // Decrease exp by 1
} while ((mantissa & 0x400) == 0); // while not normal
mantissa &= 0x3ff; // Discard subnormal bit
} // else +/-0 -> +/-0
// Combine all parts, sign << (31 - 15), value << (23 - 10)
return Float.intBitsToFloat((shortValue & 0x8000) << 16 | (exponent | mantissa) << 13);
}
private void clamp(final float[] rowDataFloat) {
for (int i = 0; i < rowDataFloat.length; i++) {
if (rowDataFloat[i] > 1f) {