| 1 | /* |
| 2 | * mixed_intersection.c |
| 3 | * |
| 4 | */ |
| 5 | |
| 6 | #include <roaring/array_util.h> |
| 7 | #include <roaring/bitset_util.h> |
| 8 | #include <roaring/containers/convert.h> |
| 9 | #include <roaring/containers/mixed_intersection.h> |
| 10 | |
| 11 | /* Compute the intersection of src_1 and src_2 and write the result to |
| 12 | * dst. */ |
| 13 | void array_bitset_container_intersection(const array_container_t *src_1, |
| 14 | const bitset_container_t *src_2, |
| 15 | array_container_t *dst) { |
| 16 | if (dst->capacity < src_1->cardinality) { |
| 17 | array_container_grow(dst, src_1->cardinality, false); |
| 18 | } |
| 19 | int32_t newcard = 0; // dst could be src_1 |
| 20 | const int32_t origcard = src_1->cardinality; |
| 21 | for (int i = 0; i < origcard; ++i) { |
| 22 | uint16_t key = src_1->array[i]; |
| 23 | // this branchless approach is much faster... |
| 24 | dst->array[newcard] = key; |
| 25 | newcard += bitset_container_contains(src_2, key); |
| 26 | /** |
| 27 | * we could do it this way instead... |
| 28 | * if (bitset_container_contains(src_2, key)) { |
| 29 | * dst->array[newcard++] = key; |
| 30 | * } |
| 31 | * but if the result is unpredictible, the processor generates |
| 32 | * many mispredicted branches. |
| 33 | * Difference can be huge (from 3 cycles when predictible all the way |
| 34 | * to 16 cycles when unpredictible. |
| 35 | * See |
| 36 | * https://github.com/lemire/Code-used-on-Daniel-Lemire-s-blog/blob/master/extra/bitset/c/arraybitsetintersection.c |
| 37 | */ |
| 38 | } |
| 39 | dst->cardinality = newcard; |
| 40 | } |
| 41 | |
| 42 | /* Compute the size of the intersection of src_1 and src_2. */ |
| 43 | int array_bitset_container_intersection_cardinality( |
| 44 | const array_container_t *src_1, const bitset_container_t *src_2) { |
| 45 | int32_t newcard = 0; |
| 46 | const int32_t origcard = src_1->cardinality; |
| 47 | for (int i = 0; i < origcard; ++i) { |
| 48 | uint16_t key = src_1->array[i]; |
| 49 | newcard += bitset_container_contains(src_2, key); |
| 50 | } |
| 51 | return newcard; |
| 52 | } |
| 53 | |
| 54 | |
| 55 | bool array_bitset_container_intersect(const array_container_t *src_1, |
| 56 | const bitset_container_t *src_2) { |
| 57 | const int32_t origcard = src_1->cardinality; |
| 58 | for (int i = 0; i < origcard; ++i) { |
| 59 | uint16_t key = src_1->array[i]; |
| 60 | if(bitset_container_contains(src_2, key)) return true; |
| 61 | } |
| 62 | return false; |
| 63 | } |
| 64 | |
| 65 | /* Compute the intersection of src_1 and src_2 and write the result to |
| 66 | * dst. It is allowed for dst to be equal to src_1. We assume that dst is a |
| 67 | * valid container. */ |
| 68 | void array_run_container_intersection(const array_container_t *src_1, |
| 69 | const run_container_t *src_2, |
| 70 | array_container_t *dst) { |
| 71 | if (run_container_is_full(src_2)) { |
| 72 | if (dst != src_1) array_container_copy(src_1, dst); |
| 73 | return; |
| 74 | } |
| 75 | if (dst->capacity < src_1->cardinality) { |
| 76 | array_container_grow(dst, src_1->cardinality, false); |
| 77 | } |
| 78 | if (src_2->n_runs == 0) { |
| 79 | return; |
| 80 | } |
| 81 | int32_t rlepos = 0; |
| 82 | int32_t arraypos = 0; |
| 83 | rle16_t rle = src_2->runs[rlepos]; |
| 84 | int32_t newcard = 0; |
| 85 | while (arraypos < src_1->cardinality) { |
| 86 | const uint16_t arrayval = src_1->array[arraypos]; |
| 87 | while (rle.value + rle.length < |
| 88 | arrayval) { // this will frequently be false |
| 89 | ++rlepos; |
| 90 | if (rlepos == src_2->n_runs) { |
| 91 | dst->cardinality = newcard; |
| 92 | return; // we are done |
| 93 | } |
| 94 | rle = src_2->runs[rlepos]; |
| 95 | } |
| 96 | if (rle.value > arrayval) { |
| 97 | arraypos = advanceUntil(src_1->array, arraypos, src_1->cardinality, |
| 98 | rle.value); |
| 99 | } else { |
| 100 | dst->array[newcard] = arrayval; |
| 101 | newcard++; |
| 102 | arraypos++; |
| 103 | } |
| 104 | } |
| 105 | dst->cardinality = newcard; |
| 106 | } |
| 107 | |
| 108 | /* Compute the intersection of src_1 and src_2 and write the result to |
| 109 | * *dst. If the result is true then the result is a bitset_container_t |
| 110 | * otherwise is a array_container_t. If *dst == src_2, an in-place processing |
| 111 | * is attempted.*/ |
| 112 | bool run_bitset_container_intersection(const run_container_t *src_1, |
| 113 | const bitset_container_t *src_2, |
| 114 | void **dst) { |
| 115 | if (run_container_is_full(src_1)) { |
| 116 | if (*dst != src_2) *dst = bitset_container_clone(src_2); |
| 117 | return true; |
| 118 | } |
| 119 | int32_t card = run_container_cardinality(src_1); |
| 120 | if (card <= DEFAULT_MAX_SIZE) { |
| 121 | // result can only be an array (assuming that we never make a |
| 122 | // RunContainer) |
| 123 | if (card > src_2->cardinality) { |
| 124 | card = src_2->cardinality; |
| 125 | } |
| 126 | array_container_t *answer = array_container_create_given_capacity(card); |
| 127 | *dst = answer; |
| 128 | if (*dst == NULL) { |
| 129 | return false; |
| 130 | } |
| 131 | for (int32_t rlepos = 0; rlepos < src_1->n_runs; ++rlepos) { |
| 132 | rle16_t rle = src_1->runs[rlepos]; |
| 133 | uint32_t endofrun = (uint32_t)rle.value + rle.length; |
| 134 | for (uint32_t runValue = rle.value; runValue <= endofrun; |
| 135 | ++runValue) { |
| 136 | answer->array[answer->cardinality] = (uint16_t)runValue; |
| 137 | answer->cardinality += |
| 138 | bitset_container_contains(src_2, runValue); |
| 139 | } |
| 140 | } |
| 141 | return false; |
| 142 | } |
| 143 | if (*dst == src_2) { // we attempt in-place |
| 144 | bitset_container_t *answer = (bitset_container_t *)*dst; |
| 145 | uint32_t start = 0; |
| 146 | for (int32_t rlepos = 0; rlepos < src_1->n_runs; ++rlepos) { |
| 147 | const rle16_t rle = src_1->runs[rlepos]; |
| 148 | uint32_t end = rle.value; |
| 149 | bitset_reset_range(src_2->array, start, end); |
| 150 | |
| 151 | start = end + rle.length + 1; |
| 152 | } |
| 153 | bitset_reset_range(src_2->array, start, UINT32_C(1) << 16); |
| 154 | answer->cardinality = bitset_container_compute_cardinality(answer); |
| 155 | if (src_2->cardinality > DEFAULT_MAX_SIZE) { |
| 156 | return true; |
| 157 | } else { |
| 158 | array_container_t *newanswer = array_container_from_bitset(src_2); |
| 159 | if (newanswer == NULL) { |
| 160 | *dst = NULL; |
| 161 | return false; |
| 162 | } |
| 163 | *dst = newanswer; |
| 164 | return false; |
| 165 | } |
| 166 | } else { // no inplace |
| 167 | // we expect the answer to be a bitmap (if we are lucky) |
| 168 | bitset_container_t *answer = bitset_container_clone(src_2); |
| 169 | |
| 170 | *dst = answer; |
| 171 | if (answer == NULL) { |
| 172 | return true; |
| 173 | } |
| 174 | uint32_t start = 0; |
| 175 | for (int32_t rlepos = 0; rlepos < src_1->n_runs; ++rlepos) { |
| 176 | const rle16_t rle = src_1->runs[rlepos]; |
| 177 | uint32_t end = rle.value; |
| 178 | bitset_reset_range(answer->array, start, end); |
| 179 | start = end + rle.length + 1; |
| 180 | } |
| 181 | bitset_reset_range(answer->array, start, UINT32_C(1) << 16); |
| 182 | answer->cardinality = bitset_container_compute_cardinality(answer); |
| 183 | |
| 184 | if (answer->cardinality > DEFAULT_MAX_SIZE) { |
| 185 | return true; |
| 186 | } else { |
| 187 | array_container_t *newanswer = array_container_from_bitset(answer); |
| 188 | bitset_container_free((bitset_container_t *)*dst); |
| 189 | if (newanswer == NULL) { |
| 190 | *dst = NULL; |
| 191 | return false; |
| 192 | } |
| 193 | *dst = newanswer; |
| 194 | return false; |
| 195 | } |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | /* Compute the size of the intersection between src_1 and src_2 . */ |
| 200 | int array_run_container_intersection_cardinality(const array_container_t *src_1, |
| 201 | const run_container_t *src_2) { |
| 202 | if (run_container_is_full(src_2)) { |
| 203 | return src_1->cardinality; |
| 204 | } |
| 205 | if (src_2->n_runs == 0) { |
| 206 | return 0; |
| 207 | } |
| 208 | int32_t rlepos = 0; |
| 209 | int32_t arraypos = 0; |
| 210 | rle16_t rle = src_2->runs[rlepos]; |
| 211 | int32_t newcard = 0; |
| 212 | while (arraypos < src_1->cardinality) { |
| 213 | const uint16_t arrayval = src_1->array[arraypos]; |
| 214 | while (rle.value + rle.length < |
| 215 | arrayval) { // this will frequently be false |
| 216 | ++rlepos; |
| 217 | if (rlepos == src_2->n_runs) { |
| 218 | return newcard; // we are done |
| 219 | } |
| 220 | rle = src_2->runs[rlepos]; |
| 221 | } |
| 222 | if (rle.value > arrayval) { |
| 223 | arraypos = advanceUntil(src_1->array, arraypos, src_1->cardinality, |
| 224 | rle.value); |
| 225 | } else { |
| 226 | newcard++; |
| 227 | arraypos++; |
| 228 | } |
| 229 | } |
| 230 | return newcard; |
| 231 | } |
| 232 | |
| 233 | /* Compute the intersection between src_1 and src_2 |
| 234 | **/ |
| 235 | int run_bitset_container_intersection_cardinality( |
| 236 | const run_container_t *src_1, const bitset_container_t *src_2) { |
| 237 | if (run_container_is_full(src_1)) { |
| 238 | return bitset_container_cardinality(src_2); |
| 239 | } |
| 240 | int answer = 0; |
| 241 | for (int32_t rlepos = 0; rlepos < src_1->n_runs; ++rlepos) { |
| 242 | rle16_t rle = src_1->runs[rlepos]; |
| 243 | answer += |
| 244 | bitset_lenrange_cardinality(src_2->array, rle.value, rle.length); |
| 245 | } |
| 246 | return answer; |
| 247 | } |
| 248 | |
| 249 | |
| 250 | bool array_run_container_intersect(const array_container_t *src_1, |
| 251 | const run_container_t *src_2) { |
| 252 | if( run_container_is_full(src_2) ) { |
| 253 | return !array_container_empty(src_1); |
| 254 | } |
| 255 | if (src_2->n_runs == 0) { |
| 256 | return false; |
| 257 | } |
| 258 | int32_t rlepos = 0; |
| 259 | int32_t arraypos = 0; |
| 260 | rle16_t rle = src_2->runs[rlepos]; |
| 261 | while (arraypos < src_1->cardinality) { |
| 262 | const uint16_t arrayval = src_1->array[arraypos]; |
| 263 | while (rle.value + rle.length < |
| 264 | arrayval) { // this will frequently be false |
| 265 | ++rlepos; |
| 266 | if (rlepos == src_2->n_runs) { |
| 267 | return false; // we are done |
| 268 | } |
| 269 | rle = src_2->runs[rlepos]; |
| 270 | } |
| 271 | if (rle.value > arrayval) { |
| 272 | arraypos = advanceUntil(src_1->array, arraypos, src_1->cardinality, |
| 273 | rle.value); |
| 274 | } else { |
| 275 | return true; |
| 276 | } |
| 277 | } |
| 278 | return false; |
| 279 | } |
| 280 | |
| 281 | /* Compute the intersection between src_1 and src_2 |
| 282 | **/ |
| 283 | bool run_bitset_container_intersect(const run_container_t *src_1, |
| 284 | const bitset_container_t *src_2) { |
| 285 | if( run_container_is_full(src_1) ) { |
| 286 | return !bitset_container_empty(src_2); |
| 287 | } |
| 288 | for (int32_t rlepos = 0; rlepos < src_1->n_runs; ++rlepos) { |
| 289 | rle16_t rle = src_1->runs[rlepos]; |
| 290 | if(!bitset_lenrange_empty(src_2->array, rle.value,rle.length)) return true; |
| 291 | } |
| 292 | return false; |
| 293 | } |
| 294 | |
| 295 | /* |
| 296 | * Compute the intersection between src_1 and src_2 and write the result |
| 297 | * to *dst. If the return function is true, the result is a bitset_container_t |
| 298 | * otherwise is a array_container_t. |
| 299 | */ |
| 300 | bool bitset_bitset_container_intersection(const bitset_container_t *src_1, |
| 301 | const bitset_container_t *src_2, |
| 302 | void **dst) { |
| 303 | const int newCardinality = bitset_container_and_justcard(src_1, src_2); |
| 304 | if (newCardinality > DEFAULT_MAX_SIZE) { |
| 305 | *dst = bitset_container_create(); |
| 306 | if (*dst != NULL) { |
| 307 | bitset_container_and_nocard(src_1, src_2, |
| 308 | (bitset_container_t *)*dst); |
| 309 | ((bitset_container_t *)*dst)->cardinality = newCardinality; |
| 310 | } |
| 311 | return true; // it is a bitset |
| 312 | } |
| 313 | *dst = array_container_create_given_capacity(newCardinality); |
| 314 | if (*dst != NULL) { |
| 315 | ((array_container_t *)*dst)->cardinality = newCardinality; |
| 316 | bitset_extract_intersection_setbits_uint16( |
| 317 | ((const bitset_container_t *)src_1)->array, |
| 318 | ((const bitset_container_t *)src_2)->array, |
| 319 | BITSET_CONTAINER_SIZE_IN_WORDS, ((array_container_t *)*dst)->array, |
| 320 | 0); |
| 321 | } |
| 322 | return false; // not a bitset |
| 323 | } |
| 324 | |
| 325 | bool bitset_bitset_container_intersection_inplace( |
| 326 | bitset_container_t *src_1, const bitset_container_t *src_2, void **dst) { |
| 327 | const int newCardinality = bitset_container_and_justcard(src_1, src_2); |
| 328 | if (newCardinality > DEFAULT_MAX_SIZE) { |
| 329 | *dst = src_1; |
| 330 | bitset_container_and_nocard(src_1, src_2, src_1); |
| 331 | ((bitset_container_t *)*dst)->cardinality = newCardinality; |
| 332 | return true; // it is a bitset |
| 333 | } |
| 334 | *dst = array_container_create_given_capacity(newCardinality); |
| 335 | if (*dst != NULL) { |
| 336 | ((array_container_t *)*dst)->cardinality = newCardinality; |
| 337 | bitset_extract_intersection_setbits_uint16( |
| 338 | ((const bitset_container_t *)src_1)->array, |
| 339 | ((const bitset_container_t *)src_2)->array, |
| 340 | BITSET_CONTAINER_SIZE_IN_WORDS, ((array_container_t *)*dst)->array, |
| 341 | 0); |
| 342 | } |
| 343 | return false; // not a bitset |
| 344 | } |
| 345 | |