diff options
Diffstat (limited to 'src/wavpack/words.c')
| -rw-r--r-- | src/wavpack/words.c | 560 |
1 files changed, 560 insertions, 0 deletions
diff --git a/src/wavpack/words.c b/src/wavpack/words.c new file mode 100644 index 00000000..0e5a3db7 --- /dev/null +++ b/src/wavpack/words.c @@ -0,0 +1,560 @@ +//////////////////////////////////////////////////////////////////////////// +// **** WAVPACK **** // +// Hybrid Lossless Wavefile Compressor // +// Copyright (c) 1998 - 2006 Conifer Software. // +// All Rights Reserved. // +// Distributed under the BSD Software License (see license.txt) // +//////////////////////////////////////////////////////////////////////////// + +// words.c + +// This module provides entropy word encoding and decoding functions using +// a variation on the Rice method. This was introduced in version 3.93 +// because it allows splitting the data into a "lossy" stream and a +// "correction" stream in a very efficient manner and is therefore ideal +// for the "hybrid" mode. For 4.0, the efficiency of this method was +// significantly improved by moving away from the normal Rice restriction of +// using powers of two for the modulus divisions and now the method can be +// used for both hybrid and pure lossless encoding. + +// Samples are divided by median probabilities at 5/7 (71.43%), 10/49 (20.41%), +// and 20/343 (5.83%). Each zone has 3.5 times fewer samples than the +// previous. Using standard Rice coding on this data would result in 1.4 +// bits per sample average (not counting sign bit). However, there is a +// very simple encoding that is over 99% efficient with this data and +// results in about 1.22 bits per sample. + +#include "wavpack.h" + +#include <string.h> + +//////////////////////////////// local macros ///////////////////////////////// + +#define LIMIT_ONES 16 // maximum consecutive 1s sent for "div" data + +// these control the time constant "slow_level" which is used for hybrid mode +// that controls bitrate as a function of residual level (HYBRID_BITRATE). +#define SLS 8 +#define SLO ((1 << (SLS - 1))) + +// these control the time constant of the 3 median level breakpoints +#define DIV0 128 // 5/7 of samples +#define DIV1 64 // 10/49 of samples +#define DIV2 32 // 20/343 of samples + +// this macro retrieves the specified median breakpoint (without frac; min = 1) +#define GET_MED(med) (((c->median [med]) >> 4) + 1) + +// These macros update the specified median breakpoints. Note that the median +// is incremented when the sample is higher than the median, else decremented. +// They are designed so that the median will never drop below 1 and the value +// is essentially stationary if there are 2 increments for every 5 decrements. + +#define INC_MED0() (c->median [0] += ((c->median [0] + DIV0) / DIV0) * 5) +#define DEC_MED0() (c->median [0] -= ((c->median [0] + (DIV0-2)) / DIV0) * 2) +#define INC_MED1() (c->median [1] += ((c->median [1] + DIV1) / DIV1) * 5) +#define DEC_MED1() (c->median [1] -= ((c->median [1] + (DIV1-2)) / DIV1) * 2) +#define INC_MED2() (c->median [2] += ((c->median [2] + DIV2) / DIV2) * 5) +#define DEC_MED2() (c->median [2] -= ((c->median [2] + (DIV2-2)) / DIV2) * 2) + +#define count_bits(av) ( \ + (av) < (1 << 8) ? nbits_table [av] : \ + ( \ + (av) < (1L << 16) ? nbits_table [(av) >> 8] + 8 : \ + ((av) < (1L << 24) ? nbits_table [(av) >> 16] + 16 : nbits_table [(av) >> 24] + 24) \ + ) \ +) + +///////////////////////////// local table storage //////////////////////////// + +const char nbits_table [] = { + 0, 1, 2, 2, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4, // 0 - 15 + 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, // 16 - 31 + 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, // 32 - 47 + 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, // 48 - 63 + 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // 64 - 79 + 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // 80 - 95 + 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // 96 - 111 + 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // 112 - 127 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, // 128 - 143 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, // 144 - 159 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, // 160 - 175 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, // 176 - 191 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, // 192 - 207 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, // 208 - 223 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, // 224 - 239 + 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8 // 240 - 255 +}; + +static const uchar log2_table [] = { + 0x00, 0x01, 0x03, 0x04, 0x06, 0x07, 0x09, 0x0a, 0x0b, 0x0d, 0x0e, 0x10, 0x11, 0x12, 0x14, 0x15, + 0x16, 0x18, 0x19, 0x1a, 0x1c, 0x1d, 0x1e, 0x20, 0x21, 0x22, 0x24, 0x25, 0x26, 0x28, 0x29, 0x2a, + 0x2c, 0x2d, 0x2e, 0x2f, 0x31, 0x32, 0x33, 0x34, 0x36, 0x37, 0x38, 0x39, 0x3b, 0x3c, 0x3d, 0x3e, + 0x3f, 0x41, 0x42, 0x43, 0x44, 0x45, 0x47, 0x48, 0x49, 0x4a, 0x4b, 0x4d, 0x4e, 0x4f, 0x50, 0x51, + 0x52, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, 0x5c, 0x5d, 0x5e, 0x5f, 0x60, 0x61, 0x62, 0x63, + 0x64, 0x66, 0x67, 0x68, 0x69, 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x74, 0x75, + 0x76, 0x77, 0x78, 0x79, 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x80, 0x81, 0x82, 0x83, 0x84, 0x85, + 0x86, 0x87, 0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93, 0x94, 0x95, + 0x96, 0x97, 0x98, 0x99, 0x9a, 0x9b, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f, 0xa0, 0xa1, 0xa2, 0xa3, 0xa4, + 0xa5, 0xa6, 0xa7, 0xa8, 0xa9, 0xa9, 0xaa, 0xab, 0xac, 0xad, 0xae, 0xaf, 0xb0, 0xb1, 0xb2, 0xb2, + 0xb3, 0xb4, 0xb5, 0xb6, 0xb7, 0xb8, 0xb9, 0xb9, 0xba, 0xbb, 0xbc, 0xbd, 0xbe, 0xbf, 0xc0, 0xc0, + 0xc1, 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc6, 0xc7, 0xc8, 0xc9, 0xca, 0xcb, 0xcb, 0xcc, 0xcd, 0xce, + 0xcf, 0xd0, 0xd0, 0xd1, 0xd2, 0xd3, 0xd4, 0xd4, 0xd5, 0xd6, 0xd7, 0xd8, 0xd8, 0xd9, 0xda, 0xdb, + 0xdc, 0xdc, 0xdd, 0xde, 0xdf, 0xe0, 0xe0, 0xe1, 0xe2, 0xe3, 0xe4, 0xe4, 0xe5, 0xe6, 0xe7, 0xe7, + 0xe8, 0xe9, 0xea, 0xea, 0xeb, 0xec, 0xed, 0xee, 0xee, 0xef, 0xf0, 0xf1, 0xf1, 0xf2, 0xf3, 0xf4, + 0xf4, 0xf5, 0xf6, 0xf7, 0xf7, 0xf8, 0xf9, 0xf9, 0xfa, 0xfb, 0xfc, 0xfc, 0xfd, 0xfe, 0xff, 0xff +}; + +static const uchar exp2_table [] = { + 0x00, 0x01, 0x01, 0x02, 0x03, 0x03, 0x04, 0x05, 0x06, 0x06, 0x07, 0x08, 0x08, 0x09, 0x0a, 0x0b, + 0x0b, 0x0c, 0x0d, 0x0e, 0x0e, 0x0f, 0x10, 0x10, 0x11, 0x12, 0x13, 0x13, 0x14, 0x15, 0x16, 0x16, + 0x17, 0x18, 0x19, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1d, 0x1e, 0x1f, 0x20, 0x20, 0x21, 0x22, 0x23, + 0x24, 0x24, 0x25, 0x26, 0x27, 0x28, 0x28, 0x29, 0x2a, 0x2b, 0x2c, 0x2c, 0x2d, 0x2e, 0x2f, 0x30, + 0x30, 0x31, 0x32, 0x33, 0x34, 0x35, 0x35, 0x36, 0x37, 0x38, 0x39, 0x3a, 0x3a, 0x3b, 0x3c, 0x3d, + 0x3e, 0x3f, 0x40, 0x41, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47, 0x48, 0x48, 0x49, 0x4a, 0x4b, + 0x4c, 0x4d, 0x4e, 0x4f, 0x50, 0x51, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57, 0x58, 0x59, 0x5a, + 0x5b, 0x5c, 0x5d, 0x5e, 0x5e, 0x5f, 0x60, 0x61, 0x62, 0x63, 0x64, 0x65, 0x66, 0x67, 0x68, 0x69, + 0x6a, 0x6b, 0x6c, 0x6d, 0x6e, 0x6f, 0x70, 0x71, 0x72, 0x73, 0x74, 0x75, 0x76, 0x77, 0x78, 0x79, + 0x7a, 0x7b, 0x7c, 0x7d, 0x7e, 0x7f, 0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x87, 0x88, 0x89, 0x8a, + 0x8b, 0x8c, 0x8d, 0x8e, 0x8f, 0x90, 0x91, 0x92, 0x93, 0x95, 0x96, 0x97, 0x98, 0x99, 0x9a, 0x9b, + 0x9c, 0x9d, 0x9f, 0xa0, 0xa1, 0xa2, 0xa3, 0xa4, 0xa5, 0xa6, 0xa8, 0xa9, 0xaa, 0xab, 0xac, 0xad, + 0xaf, 0xb0, 0xb1, 0xb2, 0xb3, 0xb4, 0xb6, 0xb7, 0xb8, 0xb9, 0xba, 0xbc, 0xbd, 0xbe, 0xbf, 0xc0, + 0xc2, 0xc3, 0xc4, 0xc5, 0xc6, 0xc8, 0xc9, 0xca, 0xcb, 0xcd, 0xce, 0xcf, 0xd0, 0xd2, 0xd3, 0xd4, + 0xd6, 0xd7, 0xd8, 0xd9, 0xdb, 0xdc, 0xdd, 0xde, 0xe0, 0xe1, 0xe2, 0xe4, 0xe5, 0xe6, 0xe8, 0xe9, + 0xea, 0xec, 0xed, 0xee, 0xf0, 0xf1, 0xf2, 0xf4, 0xf5, 0xf6, 0xf8, 0xf9, 0xfa, 0xfc, 0xfd, 0xff +}; + +static const char ones_count_table [] = { + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,5, + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,6, + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,5, + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,7, + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,5, + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,6, + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,5, + 0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,4,0,1,0,2,0,1,0,3,0,1,0,2,0,1,0,8 +}; + +///////////////////////////// executable code //////////////////////////////// + +void init_words (WavpackStream *wps) +{ + CLEAR (wps->w); +} + +static int mylog2 (uint32_t avalue); + +// Read the median log2 values from the specifed metadata structure, convert +// them back to 32-bit unsigned values and store them. If length is not +// exactly correct then we flag and return an error. + +int read_entropy_vars (WavpackStream *wps, WavpackMetadata *wpmd) +{ + uchar *byteptr = wpmd->data; + + if (wpmd->byte_length != ((wps->wphdr.flags & MONO_DATA) ? 6 : 12)) + return FALSE; + + wps->w.c [0].median [0] = exp2s (byteptr [0] + (byteptr [1] << 8)); + wps->w.c [0].median [1] = exp2s (byteptr [2] + (byteptr [3] << 8)); + wps->w.c [0].median [2] = exp2s (byteptr [4] + (byteptr [5] << 8)); + + if (!(wps->wphdr.flags & MONO_DATA)) { + wps->w.c [1].median [0] = exp2s (byteptr [6] + (byteptr [7] << 8)); + wps->w.c [1].median [1] = exp2s (byteptr [8] + (byteptr [9] << 8)); + wps->w.c [1].median [2] = exp2s (byteptr [10] + (byteptr [11] << 8)); + } + + return TRUE; +} + +// Read the hybrid related values from the specifed metadata structure, convert +// them back to their internal formats and store them. The extended profile +// stuff is not implemented yet, so return an error if we get more data than +// we know what to do with. + +int read_hybrid_profile (WavpackStream *wps, WavpackMetadata *wpmd) +{ + uchar *byteptr = wpmd->data; + uchar *endptr = byteptr + wpmd->byte_length; + + if (wps->wphdr.flags & HYBRID_BITRATE) { + wps->w.c [0].slow_level = exp2s (byteptr [0] + (byteptr [1] << 8)); + byteptr += 2; + + if (!(wps->wphdr.flags & MONO_DATA)) { + wps->w.c [1].slow_level = exp2s (byteptr [0] + (byteptr [1] << 8)); + byteptr += 2; + } + } + + wps->w.bitrate_acc [0] = (int32_t)(byteptr [0] + (byteptr [1] << 8)) << 16; + byteptr += 2; + + if (!(wps->wphdr.flags & MONO_DATA)) { + wps->w.bitrate_acc [1] = (int32_t)(byteptr [0] + (byteptr [1] << 8)) << 16; + byteptr += 2; + } + + if (byteptr < endptr) { + wps->w.bitrate_delta [0] = exp2s ((short)(byteptr [0] + (byteptr [1] << 8))); + byteptr += 2; + + if (!(wps->wphdr.flags & MONO_DATA)) { + wps->w.bitrate_delta [1] = exp2s ((short)(byteptr [0] + (byteptr [1] << 8))); + byteptr += 2; + } + + if (byteptr < endptr) + return FALSE; + } + else + wps->w.bitrate_delta [0] = wps->w.bitrate_delta [1] = 0; + + return TRUE; +} + +// This function is called during both encoding and decoding of hybrid data to +// update the "error_limit" variable which determines the maximum sample error +// allowed in the main bitstream. In the HYBRID_BITRATE mode (which is the only +// currently implemented) this is calculated from the slow_level values and the +// bitrate accumulators. Note that the bitrate accumulators can be changing. + +void update_error_limit (struct words_data *w, uint32_t flags) +{ + int bitrate_0 = (w->bitrate_acc [0] += w->bitrate_delta [0]) >> 16; + + if (flags & MONO_DATA) { + if (flags & HYBRID_BITRATE) { + int slow_log_0 = (w->c [0].slow_level + SLO) >> SLS; + + if (slow_log_0 - bitrate_0 > -0x100) + w->c [0].error_limit = exp2s (slow_log_0 - bitrate_0 + 0x100); + else + w->c [0].error_limit = 0; + } + else + w->c [0].error_limit = exp2s (bitrate_0); + } + else { + int bitrate_1 = (w->bitrate_acc [1] += w->bitrate_delta [1]) >> 16; + + if (flags & HYBRID_BITRATE) { + int slow_log_0 = (w->c [0].slow_level + SLO) >> SLS; + int slow_log_1 = (w->c [1].slow_level + SLO) >> SLS; + + if (flags & HYBRID_BALANCE) { + int balance = (slow_log_1 - slow_log_0 + bitrate_1 + 1) >> 1; + + if (balance > bitrate_0) { + bitrate_1 = bitrate_0 * 2; + bitrate_0 = 0; + } + else if (-balance > bitrate_0) { + bitrate_0 = bitrate_0 * 2; + bitrate_1 = 0; + } + else { + bitrate_1 = bitrate_0 + balance; + bitrate_0 = bitrate_0 - balance; + } + } + + if (slow_log_0 - bitrate_0 > -0x100) + w->c [0].error_limit = exp2s (slow_log_0 - bitrate_0 + 0x100); + else + w->c [0].error_limit = 0; + + if (slow_log_1 - bitrate_1 > -0x100) + w->c [1].error_limit = exp2s (slow_log_1 - bitrate_1 + 0x100); + else + w->c [1].error_limit = 0; + } + else { + w->c [0].error_limit = exp2s (bitrate_0); + w->c [1].error_limit = exp2s (bitrate_1); + } + } +} + +static uint32_t read_code (Bitstream *bs, uint32_t maxcode); + +// Read the next word from the bitstream "wvbits" and return the value. This +// function can be used for hybrid or lossless streams, but since an +// optimized version is available for lossless this function would normally +// be used for hybrid only. If a hybrid lossless stream is being read then +// the "correction" offset is written at the specified pointer. A return value +// of WORD_EOF indicates that the end of the bitstream was reached (all 1s) or +// some other error occurred. + +int32_t get_words (int32_t *buffer, int nsamples, uint32_t flags, + struct words_data *w, Bitstream *bs) +{ + register struct entropy_data *c = w->c; + int csamples; + + if (!(flags & MONO_DATA)) + nsamples *= 2; + + for (csamples = 0; csamples < nsamples; ++csamples) { + uint32_t ones_count, low, mid, high; + + if (!(flags & MONO_DATA)) + c = w->c + (csamples & 1); + + if (!(w->c [0].median [0] & ~1) && !w->holding_zero && !w->holding_one && !(w->c [1].median [0] & ~1)) { + uint32_t mask; + int cbits; + + if (w->zeros_acc) { + if (--w->zeros_acc) { + c->slow_level -= (c->slow_level + SLO) >> SLS; + *buffer++ = 0; + continue; + } + } + else { + for (cbits = 0; cbits < 33 && getbit (bs); ++cbits); + + if (cbits == 33) + break; + + if (cbits < 2) + w->zeros_acc = cbits; + else { + for (mask = 1, w->zeros_acc = 0; --cbits; mask <<= 1) + if (getbit (bs)) + w->zeros_acc |= mask; + + w->zeros_acc |= mask; + } + + if (w->zeros_acc) { + c->slow_level -= (c->slow_level + SLO) >> SLS; + CLEAR (w->c [0].median); + CLEAR (w->c [1].median); + *buffer++ = 0; + continue; + } + } + } + + if (w->holding_zero) + ones_count = w->holding_zero = 0; + else { + int next8; + + if (bs->bc < 8) { + if (++(bs->ptr) == bs->end) + bs->wrap (bs); + + next8 = (bs->sr |= *(bs->ptr) << bs->bc) & 0xff; + bs->bc += 8; + } + else + next8 = bs->sr & 0xff; + + if (next8 == 0xff) { + bs->bc -= 8; + bs->sr >>= 8; + + for (ones_count = 8; ones_count < (LIMIT_ONES + 1) && getbit (bs); ++ones_count); + + if (ones_count == (LIMIT_ONES + 1)) + break; + + if (ones_count == LIMIT_ONES) { + uint32_t mask; + int cbits; + + for (cbits = 0; cbits < 33 && getbit (bs); ++cbits); + + if (cbits == 33) + break; + + if (cbits < 2) + ones_count = cbits; + else { + for (mask = 1, ones_count = 0; --cbits; mask <<= 1) + if (getbit (bs)) + ones_count |= mask; + + ones_count |= mask; + } + + ones_count += LIMIT_ONES; + } + } + else { + bs->bc -= (ones_count = ones_count_table [next8]) + 1; + bs->sr >>= ones_count + 1; + } + + if (w->holding_one) { + w->holding_one = ones_count & 1; + ones_count = (ones_count >> 1) + 1; + } + else { + w->holding_one = ones_count & 1; + ones_count >>= 1; + } + + w->holding_zero = ~w->holding_one & 1; + } + + if ((flags & HYBRID_FLAG) && ((flags & MONO_DATA) || !(csamples & 1))) + update_error_limit (w, flags); + + if (ones_count == 0) { + low = 0; + high = GET_MED (0) - 1; + DEC_MED0 (); + } + else { + low = GET_MED (0); + INC_MED0 (); + + if (ones_count == 1) { + high = low + GET_MED (1) - 1; + DEC_MED1 (); + } + else { + low += GET_MED (1); + INC_MED1 (); + + if (ones_count == 2) { + high = low + GET_MED (2) - 1; + DEC_MED2 (); + } + else { + low += (ones_count - 2) * GET_MED (2); + high = low + GET_MED (2) - 1; + INC_MED2 (); + } + } + } + + mid = (high + low + 1) >> 1; + + if (!c->error_limit) + mid = read_code (bs, high - low) + low; + else while (high - low > c->error_limit) { + if (getbit (bs)) + mid = (high + (low = mid) + 1) >> 1; + else + mid = ((high = mid - 1) + low + 1) >> 1; + } + + *buffer++ = getbit (bs) ? ~mid : mid; + + if (flags & HYBRID_BITRATE) + c->slow_level = c->slow_level - ((c->slow_level + SLO) >> SLS) + mylog2 (mid); + } + + return (flags & MONO_DATA) ? csamples : (csamples / 2); +} + +// Read a single unsigned value from the specified bitstream with a value +// from 0 to maxcode. If there are exactly a power of two number of possible +// codes then this will read a fixed number of bits; otherwise it reads the +// minimum number of bits and then determines whether another bit is needed +// to define the code. + +static uint32_t read_code (Bitstream *bs, uint32_t maxcode) +{ + int bitcount = count_bits (maxcode); + uint32_t extras = (1L << bitcount) - maxcode - 1, code; + + if (!bitcount) + return 0; + + getbits (&code, bitcount - 1, bs); + code &= (1L << (bitcount - 1)) - 1; + + if (code >= extras) { + code = (code << 1) - extras; + + if (getbit (bs)) + ++code; + } + + return code; +} + +// The concept of a base 2 logarithm is used in many parts of WavPack. It is +// a way of sufficiently accurately representing 32-bit signed and unsigned +// values storing only 16 bits (actually fewer). It is also used in the hybrid +// mode for quickly comparing the relative magnitude of large values (i.e. +// division) and providing smooth exponentials using only addition. + +// These are not strict logarithms in that they become linear around zero and +// can therefore represent both zero and negative values. They have 8 bits +// of precision and in "roundtrip" conversions the total error never exceeds 1 +// part in 225 except for the cases of +/-115 and +/-195 (which error by 1). + + +// This function returns the log2 for the specified 32-bit unsigned value. +// The maximum value allowed is about 0xff800000 and returns 8447. + +static int mylog2 (uint32_t avalue) +{ + int dbits; + + if ((avalue += avalue >> 9) < (1 << 8)) { + dbits = nbits_table [avalue]; + return (dbits << 8) + log2_table [(avalue << (9 - dbits)) & 0xff]; + } + else { + if (avalue < (1L << 16)) + dbits = nbits_table [avalue >> 8] + 8; + else if (avalue < (1L << 24)) + dbits = nbits_table [avalue >> 16] + 16; + else + dbits = nbits_table [avalue >> 24] + 24; + + return (dbits << 8) + log2_table [(avalue >> (dbits - 9)) & 0xff]; + } +} + +// This function returns the log2 for the specified 32-bit signed value. +// All input values are valid and the return values are in the range of +// +/- 8192. + +int log2s (int32_t value) +{ + return (value < 0) ? -mylog2 (-value) : mylog2 (value); +} + +// This function returns the original integer represented by the supplied +// logarithm (at least within the provided accuracy). The log is signed, +// but since a full 32-bit value is returned this can be used for unsigned +// conversions as well (i.e. the input range is -8192 to +8447). + +int32_t exp2s (int log) +{ + uint32_t value; + + if (log < 0) + return -exp2s (-log); + + value = exp2_table [log & 0xff] | 0x100; + + if ((log >>= 8) <= 9) + return value >> (9 - log); + else + return value << (log - 9); +} + +// These two functions convert internal weights (which are normally +/-1024) +// to and from an 8-bit signed character version for storage in metadata. The +// weights are clipped here in the case that they are outside that range. + +int restore_weight (signed char weight) +{ + int result; + + if ((result = (int) weight << 3) > 0) + result += (result + 64) >> 7; + + return result; +} |