lib/librte_acl/acl_run_altivec.h

   1 /*
   2  *   BSD LICENSE
   3  *
   4  *   Copyright (C) IBM Corporation 2016.
   5  *
   6  *   Redistribution and use in source and binary forms, with or without
   7  *   modification, are permitted provided that the following conditions
   8  *   are met:
   9  *
  10  *     * Redistributions of source code must retain the above copyright
  11  *       notice, this list of conditions and the following disclaimer.
  12  *     * Redistributions in binary form must reproduce the above copyright
  13  *       notice, this list of conditions and the following disclaimer in
  14  *       the documentation and/or other materials provided with the
  15  *       distribution.
  16  *     * Neither the name of IBM Corporation nor the names of its
  17  *       contributors may be used to endorse or promote products derived
  18  *       from this software without specific prior written permission.
  19  *
  20  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  24  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  25  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  26  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  27  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  28  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  29  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  30  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  31 */
  32
  33 #include "acl_run.h"
  34 #include "acl_vect.h"
  35
  36 struct _altivec_acl_const {
  37         rte_xmm_t xmm_shuffle_input;
  38         rte_xmm_t xmm_index_mask;
  39         rte_xmm_t xmm_ones_16;
  40         rte_xmm_t range_base;
  41 } altivec_acl_const  __attribute__((aligned(RTE_CACHE_LINE_SIZE))) = {
  42         {
  43                 .u32 = {0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c}
  44         },
  45         {
  46                 .u32 = {RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX,
  47                 RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX}
  48         },
  49         {
  50                 .u16 = {1, 1, 1, 1, 1, 1, 1, 1}
  51         },
  52         {
  53                 .u32 = {0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c}
  54         },
  55 };
  56
  57 /*
  58  * Resolve priority for multiple results (altivec version).
  59  * This consists comparing the priority of the current traversal with the
  60  * running set of results for the packet.
  61  * For each result, keep a running array of the result (rule number) and
  62  * its priority for each category.
  63  */
  64 static inline void
  65 resolve_priority_altivec(uint64_t transition, int n,
  66         const struct rte_acl_ctx *ctx, struct parms *parms,
  67         const struct rte_acl_match_results *p, uint32_t categories)
  68 {
  69         uint32_t x;
  70         xmm_t results, priority, results1, priority1;
  71         vector bool int selector;
  72         xmm_t *saved_results, *saved_priority;
  73
  74         for (x = 0; x < categories; x += RTE_ACL_RESULTS_MULTIPLIER) {
  75
  76                 saved_results = (xmm_t *)(&parms[n].cmplt->results[x]);
  77                 saved_priority =
  78                         (xmm_t *)(&parms[n].cmplt->priority[x]);
  79
  80                 /* get results and priorities for completed trie */
  81                 results = *(const xmm_t *)&p[transition].results[x];
  82                 priority = *(const xmm_t *)&p[transition].priority[x];
  83
  84                 /* if this is not the first completed trie */
  85                 if (parms[n].cmplt->count != ctx->num_tries) {
  86
  87                         /* get running best results and their priorities */
  88                         results1 = *saved_results;
  89                         priority1 = *saved_priority;
  90
  91                         /* select results that are highest priority */
  92                         selector = vec_cmpgt(priority1, priority);
  93                         results = vec_sel(results, results1, selector);
  94                         priority = vec_sel(priority, priority1,
  95                                 selector);
  96                 }
  97
  98                 /* save running best results and their priorities */
  99                 *saved_results = results;
 100                 *saved_priority = priority;
 101         }
 102 }
 103
 104 /*
 105  * Check for any match in 4 transitions
 106  */
 107 static inline __attribute__((always_inline)) uint32_t
 108 check_any_match_x4(uint64_t val[])
 109 {
 110         return (val[0] | val[1] | val[2] | val[3]) & RTE_ACL_NODE_MATCH;
 111 }
 112
 113 static inline __attribute__((always_inline)) void
 114 acl_match_check_x4(int slot, const struct rte_acl_ctx *ctx, struct parms *parms,
 115         struct acl_flow_data *flows, uint64_t transitions[])
 116 {
 117         while (check_any_match_x4(transitions)) {
 118                 transitions[0] = acl_match_check(transitions[0], slot, ctx,
 119                         parms, flows, resolve_priority_altivec);
 120                 transitions[1] = acl_match_check(transitions[1], slot + 1, ctx,
 121                         parms, flows, resolve_priority_altivec);
 122                 transitions[2] = acl_match_check(transitions[2], slot + 2, ctx,
 123                         parms, flows, resolve_priority_altivec);
 124                 transitions[3] = acl_match_check(transitions[3], slot + 3, ctx,
 125                         parms, flows, resolve_priority_altivec);
 126         }
 127 }
 128
 129 /*
 130  * Process 4 transitions (in 2 XMM registers) in parallel
 131  */
 132 static inline __attribute__((optimize("O2"))) xmm_t
 133 transition4(xmm_t next_input, const uint64_t *trans,
 134         xmm_t *indices1, xmm_t *indices2)
 135 {
 136         xmm_t addr, tr_lo, tr_hi;
 137         xmm_t in, node_type, r, t;
 138         xmm_t dfa_ofs, quad_ofs;
 139         xmm_t *index_mask, *tp;
 140         vector bool int dfa_msk;
 141         vector signed char zeroes = {};
 142         union {
 143                 uint64_t d64[2];
 144                 uint32_t d32[4];
 145         } v;
 146
 147         /* Move low 32 into tr_lo and high 32 into tr_hi */
 148         tr_lo = (xmm_t){(*indices1)[0], (*indices1)[2],
 149                         (*indices2)[0], (*indices2)[2]};
 150         tr_hi = (xmm_t){(*indices1)[1], (*indices1)[3],
 151                         (*indices2)[1], (*indices2)[3]};
 152
 153          /* Calculate the address (array index) for all 4 transitions. */
 154         index_mask = (xmm_t *)&altivec_acl_const.xmm_index_mask.u32;
 155         t = vec_xor(*index_mask, *index_mask);
 156         in = vec_perm(next_input, (xmm_t){},
 157                 *(vector unsigned char *)&altivec_acl_const.xmm_shuffle_input);
 158
 159         /* Calc node type and node addr */
 160         node_type = vec_and(vec_nor(*index_mask, *index_mask), tr_lo);
 161         addr = vec_and(tr_lo, *index_mask);
 162
 163         /* mask for DFA type(0) nodes */
 164         dfa_msk = vec_cmpeq(node_type, t);
 165
 166         /* DFA calculations. */
 167         r = vec_sr(in, (vector unsigned int){30, 30, 30, 30});
 168         tp = (xmm_t *)&altivec_acl_const.range_base.u32;
 169         r = vec_add(r, *tp);
 170         t = vec_sr(in, (vector unsigned int){24, 24, 24, 24});
 171         r = vec_perm(tr_hi, (xmm_t){(uint16_t)0 << 16},
 172                 (vector unsigned char)r);
 173
 174         dfa_ofs = vec_sub(t, r);
 175
 176         /* QUAD/SINGLE caluclations. */
 177         t = (xmm_t)vec_cmpgt((vector signed char)in, (vector signed char)tr_hi);
 178         t = (xmm_t)vec_sel(
 179                 vec_sel(
 180                         (vector signed char)vec_sub(
 181                                 zeroes, (vector signed char)t),
 182                         (vector signed char)t,
 183                         vec_cmpgt((vector signed char)t, zeroes)),
 184                 zeroes,
 185                 vec_cmpeq((vector signed char)t, zeroes));
 186
 187         t = (xmm_t)vec_msum((vector signed char)t,
 188                 (vector unsigned char)t, (xmm_t){});
 189         quad_ofs = (xmm_t)vec_msum((vector signed short)t,
 190                 *(vector signed short *)&altivec_acl_const.xmm_ones_16.u16,
 191                 (xmm_t){});
 192
 193         /* blend DFA and QUAD/SINGLE. */
 194         t = vec_sel(quad_ofs, dfa_ofs, dfa_msk);
 195
 196         /* calculate address for next transitions. */
 197         addr = vec_add(addr, t);
 198
 199         v.d64[0] = (uint64_t)trans[addr[0]];
 200         v.d64[1] = (uint64_t)trans[addr[1]];
 201         *indices1 = (xmm_t){v.d32[0], v.d32[1], v.d32[2], v.d32[3]};
 202         v.d64[0] = (uint64_t)trans[addr[2]];
 203         v.d64[1] = (uint64_t)trans[addr[3]];
 204         *indices2 = (xmm_t){v.d32[0], v.d32[1], v.d32[2], v.d32[3]};
 205
 206         return vec_sr(next_input,
 207                 (vector unsigned int){CHAR_BIT, CHAR_BIT, CHAR_BIT, CHAR_BIT});
 208 }
 209
 210 /*
 211  * Execute trie traversal with 8 traversals in parallel
 212  */
 213 static inline int
 214 search_altivec_8(const struct rte_acl_ctx *ctx, const uint8_t **data,
 215         uint32_t *results, uint32_t total_packets, uint32_t categories)
 216 {
 217         int n;
 218         struct acl_flow_data flows;
 219         uint64_t index_array[MAX_SEARCHES_ALTIVEC8];
 220         struct completion cmplt[MAX_SEARCHES_ALTIVEC8];
 221         struct parms parms[MAX_SEARCHES_ALTIVEC8];
 222         xmm_t input0, input1;
 223
 224         acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
 225                 total_packets, categories, ctx->trans_table);
 226
 227         for (n = 0; n < MAX_SEARCHES_ALTIVEC8; n++) {
 228                 cmplt[n].count = 0;
 229                 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
 230         }
 231
 232          /* Check for any matches. */
 233         acl_match_check_x4(0, ctx, parms, &flows, (uint64_t *)&index_array[0]);
 234         acl_match_check_x4(4, ctx, parms, &flows, (uint64_t *)&index_array[4]);
 235
 236         while (flows.started > 0) {
 237
 238                 /* Gather 4 bytes of input data for each stream. */
 239                 input0 = (xmm_t){GET_NEXT_4BYTES(parms, 0),
 240                                 GET_NEXT_4BYTES(parms, 1),
 241                                 GET_NEXT_4BYTES(parms, 2),
 242                                 GET_NEXT_4BYTES(parms, 3)};
 243
 244                 input1 = (xmm_t){GET_NEXT_4BYTES(parms, 4),
 245                                 GET_NEXT_4BYTES(parms, 5),
 246                                 GET_NEXT_4BYTES(parms, 6),
 247                                 GET_NEXT_4BYTES(parms, 7)};
 248
 249                  /* Process the 4 bytes of input on each stream. */
 250
 251                 input0 = transition4(input0, flows.trans,
 252                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 253                 input1 = transition4(input1, flows.trans,
 254                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 255
 256                 input0 = transition4(input0, flows.trans,
 257                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 258                 input1 = transition4(input1, flows.trans,
 259                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 260
 261                 input0 = transition4(input0, flows.trans,
 262                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 263                 input1 = transition4(input1, flows.trans,
 264                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 265
 266                 input0 = transition4(input0, flows.trans,
 267                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 268                 input1 = transition4(input1, flows.trans,
 269                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 270
 271                  /* Check for any matches. */
 272                 acl_match_check_x4(0, ctx, parms, &flows,
 273                         (uint64_t *)&index_array[0]);
 274                 acl_match_check_x4(4, ctx, parms, &flows,
 275                         (uint64_t *)&index_array[4]);
 276         }
 277
 278         return 0;
 279 }
 280
 281 /*
 282  * Execute trie traversal with 4 traversals in parallel
 283  */
 284 static inline int
 285 search_altivec_4(const struct rte_acl_ctx *ctx, const uint8_t **data,
 286          uint32_t *results, int total_packets, uint32_t categories)
 287 {
 288         int n;
 289         struct acl_flow_data flows;
 290         uint64_t index_array[MAX_SEARCHES_ALTIVEC4];
 291         struct completion cmplt[MAX_SEARCHES_ALTIVEC4];
 292         struct parms parms[MAX_SEARCHES_ALTIVEC4];
 293         xmm_t input;
 294
 295         acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
 296                 total_packets, categories, ctx->trans_table);
 297
 298         for (n = 0; n < MAX_SEARCHES_ALTIVEC4; n++) {
 299                 cmplt[n].count = 0;
 300                 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
 301         }
 302
 303         /* Check for any matches. */
 304         acl_match_check_x4(0, ctx, parms, &flows, index_array);
 305
 306         while (flows.started > 0) {
 307
 308                 /* Gather 4 bytes of input data for each stream. */
 309                 input = (xmm_t){GET_NEXT_4BYTES(parms, 0),
 310                                 GET_NEXT_4BYTES(parms, 1),
 311                                 GET_NEXT_4BYTES(parms, 2),
 312                                 GET_NEXT_4BYTES(parms, 3)};
 313
 314                 /* Process the 4 bytes of input on each stream. */
 315                 input = transition4(input, flows.trans,
 316                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 317                 input = transition4(input, flows.trans,
 318                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 319                 input = transition4(input, flows.trans,
 320                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 321                 input = transition4(input, flows.trans,
 322                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 323
 324                 /* Check for any matches. */
 325                 acl_match_check_x4(0, ctx, parms, &flows, index_array);
 326         }
 327
 328         return 0;
 329 }