Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2020 Arm Limited. All rights reserved. |
| 3 | * |
| 4 | * SPDX-License-Identifier: Apache-2.0 |
| 5 | * |
| 6 | * Licensed under the Apache License, Version 2.0 (the License); you may |
| 7 | * not use this file except in compliance with the License. |
| 8 | * You may obtain a copy of the License at |
| 9 | * |
| 10 | * www.apache.org/licenses/LICENSE-2.0 |
| 11 | * |
| 12 | * Unless required by applicable law or agreed to in writing, software |
| 13 | * distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| 14 | * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 15 | * See the License for the specific language governing permissions and |
| 16 | * limitations under the License. |
| 17 | */ |
| 18 | |
| 19 | #include <stdio.h> |
| 20 | #include <stdlib.h> |
| 21 | #include <stdint.h> |
| 22 | #include <stdbool.h> |
| 23 | #include <string.h> |
| 24 | #include <assert.h> |
| 25 | #include <math.h> |
| 26 | #include <stdarg.h> |
| 27 | #include <math.h> |
| 28 | #include "mlw_common.h" |
| 29 | #include "mlw_decode.h" |
| 30 | |
| 31 | |
| 32 | /////////////////////////////// Read from bitstream |
| 33 | |
| 34 | typedef struct bitbuf { |
| 35 | uint8_t *buf; |
| 36 | int buf_size; // in bytes |
| 37 | int pos; // bit pos of next bit |
| 38 | int log_symbols; |
| 39 | } bitbuf_t; |
| 40 | |
| 41 | |
| 42 | // size in byte |
| 43 | static void bitbuf_init( bitbuf_t *bb, uint8_t *buf, int size, int log_symbols) { |
| 44 | bb->buf = buf; |
| 45 | bb->pos = 0; |
| 46 | bb->buf_size = size; |
| 47 | bb->log_symbols = log_symbols; |
| 48 | } |
| 49 | |
| 50 | static int bitbuf_getbit( bitbuf_t *bb) { |
| 51 | int byte_pos = bb->pos>>3; |
| 52 | int bit_pos = bb->pos&7; |
| 53 | if ( byte_pos < 0 || byte_pos >= bb->buf_size ) { |
| 54 | printf("bitbuf_getbit: underrun, bit_pos %3d byte_pos %3d buf_size %3d\n", bit_pos, byte_pos, bb->buf_size); |
| 55 | exit(1); |
| 56 | } |
| 57 | int bit = bb->buf[ byte_pos ] & (1<<bit_pos) ? 1 : 0; |
| 58 | bb->pos++; |
| 59 | return bit; |
| 60 | } |
| 61 | |
| 62 | static int bitbuf_get( bitbuf_t *bb, const char *name, int len) { |
| 63 | int i, data=0, save_pos=bb->pos; |
| 64 | if (len>0) { |
| 65 | for(i=0; i<len; i++) { |
| 66 | data |= bitbuf_getbit(bb)<<i; |
| 67 | } |
| 68 | if (bb->log_symbols) |
| 69 | printf("bitbuf: pos %3d %7s len %d data %x\n", save_pos, name, len, data); |
| 70 | } |
| 71 | return data; |
| 72 | } |
| 73 | |
| 74 | // Decode the given weight stream |
| 75 | // inbuf compressed bitstream |
| 76 | // inbuf_size size of compressed bitstream in bytes |
| 77 | // outbuf uncompressed 9bit signed weights, buffer malloced |
| 78 | // verbose if non-zero, printf log |
| 79 | // Return value is the number of uncompressed weights |
| 80 | int mlw_decode( uint8_t *inbuf, int inbuf_size, int16_t **outbuf, int verbose) { |
| 81 | int nvalues; |
| 82 | int w_grc_div; |
| 83 | int w_grc_trunc; |
| 84 | int w_uncompressed; |
| 85 | int z_grc_div, z_prev_grc_div=0; |
| 86 | int new_palette; |
| 87 | int palsize=0, palbits=0; |
| 88 | int direct_offset=0; |
| 89 | int16_t palette[512]; |
| 90 | int first=1; |
| 91 | int use_zero_run, i, j; |
| 92 | int outbuf_size=0; |
| 93 | int nchunks=0; |
| 94 | |
| 95 | *outbuf=0; |
| 96 | |
| 97 | bitbuf_t bitbuf_s, *bb=&bitbuf_s; |
| 98 | bitbuf_init( bb, inbuf, inbuf_size, (verbose&2)?1:0 ); |
| 99 | |
| 100 | // Loop over all slices |
| 101 | while(1) { |
| 102 | // Decode slice header |
| 103 | z_grc_div = bitbuf_get( bb, "ZDIV", 3 ); |
| 104 | while(z_grc_div==ZDIV_EOS) { // TODO: change to ZDIV_PAD |
| 105 | // End of stream |
| 106 | // Byte align |
| 107 | bitbuf_get( bb, "BYTEALIGN", (8-(bb->pos&7))&7 ); |
| 108 | first=1; |
| 109 | if ( (bb->pos/8) == inbuf_size) { |
| 110 | // Quit if we actually reached end of input stream |
| 111 | break; |
| 112 | } |
| 113 | z_grc_div = bitbuf_get( bb, "ZDIV", 3 ); |
| 114 | } |
| 115 | if ( (bb->pos/8) == inbuf_size) { |
| 116 | break; // reached end of input stream |
| 117 | } |
| 118 | assert(z_grc_div<4 || z_grc_div==ZDIV_DISABLE); |
| 119 | use_zero_run = z_grc_div!=ZDIV_DISABLE; // alternating grc |
| 120 | nvalues = bitbuf_get( bb, "SLICELEN", 15 )+1; |
| 121 | w_grc_div = bitbuf_get( bb, "WDIV", 3 ); |
| 122 | w_grc_trunc = bitbuf_get( bb, "WTRUNC", 1 ); |
| 123 | new_palette = bitbuf_get( bb, "NEWPAL", 1 ); |
| 124 | if (first) { |
| 125 | // the first slice must have a palette/direct mode setup |
| 126 | assert(new_palette); |
| 127 | first=0; |
| 128 | } |
| 129 | if (!new_palette) { |
| 130 | // At the moment it is not supported to change between alternating |
| 131 | // and non-alternating without redefining the palette (this is because |
| 132 | // the zero is not included in the palette in case of alternating) |
| 133 | int prev_use_zero_run = z_prev_grc_div!=ZDIV_DISABLE; |
| 134 | (void)(prev_use_zero_run); |
| 135 | assert( use_zero_run == prev_use_zero_run); |
| 136 | } |
| 137 | z_prev_grc_div = z_grc_div; |
| 138 | if (new_palette) { |
| 139 | direct_offset = bitbuf_get( bb, "DIROFS", 5 ); |
| 140 | palsize = bitbuf_get( bb, "PALSIZE", 5 ); |
| 141 | if (palsize>0) |
| 142 | palsize++; |
| 143 | palbits = bitbuf_get( bb, "PALBITS", 3 )+2; |
| 144 | for(i=0; i<palsize; i++) { |
| 145 | palette[i] = bitbuf_get( bb, "PALETTE", palbits ); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | if (w_grc_div==WDIV_UNCOMPRESSED) { |
| 150 | // Uncompressed mode |
| 151 | w_uncompressed = 1; |
| 152 | int uncompressed_bits; |
| 153 | if (palsize>0) { |
| 154 | // Uncompressed bits is given by palette size. |
| 155 | uncompressed_bits=0; |
| 156 | while( (1<<uncompressed_bits) < palsize ) |
| 157 | uncompressed_bits++; |
| 158 | } else { |
| 159 | // No palette. PALBITS is used to specify uncompressed bits. |
| 160 | uncompressed_bits=palbits; |
| 161 | } |
| 162 | // In uncompressed mode there's only a remainder part (no unary) |
| 163 | // This is achieved by setting w_grc_div to index bit width |
| 164 | w_grc_div = uncompressed_bits; |
| 165 | } else { |
| 166 | w_uncompressed = 0; |
| 167 | assert(w_grc_div<6); |
| 168 | } |
| 169 | |
| 170 | // Decode the slice |
| 171 | int z_nvalues = nvalues + (new_palette?1:0); |
| 172 | int *w_value = malloc( nvalues*sizeof(int) ); |
| 173 | int *z_value = malloc( z_nvalues*sizeof(int) ); |
| 174 | int w_pos=0, z_pos=0; |
| 175 | int w_prev_pos=0, z_prev_pos=0; |
| 176 | int w_unary0=0, w_unary1=0, w_unary1_len=0, w_q[12]={0}, w_carry=0; |
| 177 | int z_unary=0, z_q[12]={0}, z_carry=0; |
| 178 | int w_nsymbols=0; |
| 179 | int w_prev_enable=0, w_prev_nsymbols=0, w_prev_q[12]={0}; |
| 180 | int z_nsymbols=0; |
| 181 | int z_prev_enable=0, z_prev_nsymbols=0, z_prev_q[12]={0}; |
| 182 | int total_zcnt=0; |
| 183 | int z_unary_len = z_grc_div<3 ? 12 : 8; |
| 184 | |
| 185 | // Loop over all chunks in the slice |
| 186 | do { |
| 187 | // Flow control to possibly throttle either the weights or zero-runs |
| 188 | int balance = use_zero_run ? w_pos - z_pos : 0; |
| 189 | int w_enable = (balance<8 || !use_zero_run) && w_pos<nvalues; |
| 190 | int z_enable = balance>=0 && use_zero_run && z_pos<z_nvalues; |
| 191 | if (w_enable) { |
| 192 | if (!w_uncompressed) |
| 193 | w_unary0 = bitbuf_get( bb, "WUNARY0", 12 ); |
| 194 | else |
| 195 | w_unary0 = 0; |
| 196 | } |
| 197 | if (z_enable) { |
| 198 | z_unary = bitbuf_get( bb, "ZUNARY", z_unary_len ); |
| 199 | z_nsymbols=0; |
| 200 | int cnt = z_carry; |
| 201 | for(i=0; i<z_unary_len; i++) { |
| 202 | if (z_unary & (1<<i)) { |
| 203 | cnt++; |
| 204 | } else { |
| 205 | z_q[z_nsymbols++] = cnt; |
| 206 | cnt=0; |
| 207 | } |
| 208 | } |
| 209 | z_carry = cnt; |
| 210 | z_pos += z_nsymbols; |
| 211 | } |
| 212 | if (w_enable) { |
| 213 | w_unary1_len=0; |
| 214 | int max_symbols = w_uncompressed && w_grc_div>5 ? 8 : 12; |
| 215 | for(i=0; i<max_symbols; i++) { |
| 216 | if (w_unary0&(1<<i)) |
| 217 | w_unary1_len++; |
| 218 | } |
| 219 | w_unary1 = bitbuf_get( bb, "WUNARY1", w_unary1_len ); |
| 220 | w_nsymbols=0; |
| 221 | int cnt = w_carry; |
| 222 | for(i=0; i<max_symbols; i++) { |
| 223 | int code=0; |
| 224 | if (w_unary0 & (1<<i)) { |
| 225 | code++; |
| 226 | if (w_unary1&1) { |
| 227 | code++; |
| 228 | } |
| 229 | w_unary1 = w_unary1>>1; |
| 230 | } |
| 231 | cnt += code; |
| 232 | if (code<2 || w_grc_trunc) { |
| 233 | w_q[w_nsymbols++] = cnt; |
| 234 | cnt=0; |
| 235 | } |
| 236 | } |
| 237 | w_carry = cnt; |
| 238 | w_pos += w_nsymbols; |
| 239 | } |
| 240 | if (w_prev_enable) { |
| 241 | for(i=0; i<w_prev_nsymbols && w_prev_pos<nvalues; i++, w_prev_pos++) { |
| 242 | int remain = bitbuf_get( bb, "WREMAIN", w_grc_div ); |
| 243 | w_value[w_prev_pos] = (w_prev_q[i]<<w_grc_div) + remain; |
| 244 | } |
| 245 | } |
| 246 | if (z_prev_enable) { |
| 247 | for(i=0; i<z_prev_nsymbols && z_prev_pos<z_nvalues; i++, z_prev_pos++) { |
| 248 | int remain = bitbuf_get( bb, "ZREMAIN", z_grc_div ); |
| 249 | z_value[z_prev_pos] = (z_prev_q[i]<<z_grc_div) + remain; |
| 250 | total_zcnt += z_value[z_prev_pos]; |
| 251 | } |
| 252 | } |
| 253 | w_prev_enable = w_enable; |
| 254 | w_prev_nsymbols = w_nsymbols; |
| 255 | memcpy( w_prev_q, w_q, sizeof(w_prev_q)); |
| 256 | z_prev_enable = z_enable; |
| 257 | z_prev_nsymbols = z_nsymbols; |
| 258 | memcpy( z_prev_q, z_q, sizeof(z_prev_q)); |
| 259 | nchunks++; |
| 260 | } while( w_prev_enable || z_prev_enable ); |
| 261 | |
| 262 | // Interleave non-zero and zeros into the outbut buffer |
| 263 | // Increase the outbuffer to fit the new slice |
| 264 | *outbuf = realloc( *outbuf, (outbuf_size + nvalues + total_zcnt)*sizeof(int16_t)); |
| 265 | |
| 266 | int k=outbuf_size; |
| 267 | |
| 268 | // Insert initial zeros |
| 269 | // (slices redefining the palette may start with zeros) |
| 270 | if (new_palette && use_zero_run) { |
| 271 | for(j=0; j<z_value[0]; j++) { |
| 272 | (*outbuf)[k++] = 0; |
| 273 | } |
| 274 | } |
| 275 | |
| 276 | // Loop over all weights and insert zeros in-between |
| 277 | for(i=0; i<nvalues; i++) { |
| 278 | int val; |
| 279 | assert(w_value[i]<512); // HW supports 9bit |
| 280 | if (w_value[i]<palsize) { |
| 281 | val = palette[w_value[i]]; |
| 282 | } else { |
| 283 | val = w_value[i]-palsize+direct_offset; |
| 284 | } |
| 285 | int sign = val&1; |
| 286 | int mag = val>>1; |
| 287 | (*outbuf)[k++] = sign ? -mag : mag; |
| 288 | if (use_zero_run) { |
| 289 | for(j=0; j<z_value[i+(new_palette?1:0)]; j++) { |
| 290 | (*outbuf)[k++] = 0; |
| 291 | } |
| 292 | } |
| 293 | } |
| 294 | |
| 295 | outbuf_size = k; |
| 296 | free(w_value); |
| 297 | free(z_value); |
| 298 | } |
| 299 | return outbuf_size; |
| 300 | } |