From bbdba9b8d71f25af5bc0caaad48c998963f7a111 Mon Sep 17 00:00:00 2001 From: Dale Weiler Date: Tue, 2 Oct 2012 12:06:25 -0400 Subject: [PATCH] some awesome documentation about CRC for future viewers --- util.c | 22 ++++++++++++++++++++++ 1 file changed, 22 insertions(+) diff --git a/util.c b/util.c index bdc3f0f..006a814 100644 --- a/util.c +++ b/util.c @@ -233,6 +233,28 @@ void util_endianswap(void *m, int s, int l) { } } +/* + * CRC algorithms vary in the width of the polynomial, the value of said polynomial, + * the initial value used for the register, weather the bits of each byte are reflected + * before being processed, weather the algorithm itself feeds input bytes through the + * register or XORs them with a byte from one end and then straight into the table, as + * well as (but not limited to the idea of reflected versions) where the final register + * value becomes reversed, and finally weather the value itself is used to XOR the final + * register value. AS such you can already imagine how painfully annoying CRCs are, + * of course we stand to target Quake, which expects it's certian set of rules for proper + * calculation of a CRC. + * + * In most traditional CRC algorithms on uses a reflected table driven method where a value + * or register is reflected if it's bits are swapped around it's center. For example: + * take the bits 0101 is the 4-bit reflection of 1010, and respectfully 0011 would be the + * reflection of 1100. Quakle however expects a NON-Reflected CRC on the output, but still + * requires a final XOR on the values (0xFFFF and 0x0000) this is a standard CCITT CRC-16 + * which I respectfully as a programmer don't agree with. + * + * So now you know what we target, and why we target it, despite how unsettling it may seem + * but those are what Quake seems to request. + */ + /* * This is an implementation of CRC32 & CRC16. The polynomials have been * offline computed for faster generation at the cost of larger code size. -- 2.39.2