[Libpgf-devel] Improvement to bitplane decoder

[Jan W. <was...@fo...>]

Dear developers,

thank you for providing libPGF in open-source form!

I am endeavouring to speed up loading of large (200 MB) images by
decompressing them on the fly during IO.
While PGF is quite fast compared to JPEG-LS, its decompression =
throughput
turns out to be slower than reading from the hard-drive. Unfortunately =
that
makes it unsuitable for the current application, but I'd like to pass on
several minor changes that were observed to improve performance on this
Athlon x2 system:

- Decoder.h SetSign: OR-in the sign bit directly (avoids a conditional
branch)

	void  SetSign(UINT32 pos, bool sign)
	{
		pos |=3D ((UINT32)sign) << 31;
//		if (sign && m_value[pos] >=3D 0)
//			m_value[pos] =3D -m_value[pos];
	}


- BitStream.h SeekBitRange: use pointer instead of base/displacement
addressing, restructure loop to reduce branches, make use of integer
overflow when left-shifting

inline UINT32 SeekBitRange(UINT32* stream, UINT32 pos, UINT32 len) {
	UINT32 count =3D 0;
	UINT32 testMask =3D 1 << (pos%WordWidth);
	UINT32* word =3D stream + (pos >> WordWidthLog);

	for(;;)
	{
		if(*word & testMask)
			return count;
		count++;
		if(count >=3D len)
			return len;
		testMask *=3D 2;
		if(!testMask)	// overflow
		{
			word++; testMask =3D 1;

			// fast steps if all bits in a word are zero
			while ((count + WordWidth <=3D len) && (*word =3D=3D 0)) {
				word++;=20
				count +=3D WordWidth;
			}
		}
	}
}

Hopefully this is helpful.

Do you know of any efforts to sacrifice compression ratios for =
decompression
throughput that is competitive with hard drives? My layman's impression =
is
that simple LZ compression of the wavelet coefficients (analogous to =
PNG's
zlib compression of Paeth-filtered pixels) may be successful. Has this =
been
tried during development?

Best Regards
Jan Wassenberg