lib/librte_acl/acl_run_altivec.h

   1 /*
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  * Copyright (C) IBM Corporation 2016.
   4  */
   5
   6 #include "acl_run.h"
   7 #include "acl_vect.h"
   8
   9 struct _altivec_acl_const {
  10         rte_xmm_t xmm_shuffle_input;
  11         rte_xmm_t xmm_index_mask;
  12         rte_xmm_t xmm_ones_16;
  13         rte_xmm_t range_base;
  14 } altivec_acl_const  __attribute__((aligned(RTE_CACHE_LINE_SIZE))) = {
  15         {
  16                 .u32 = {0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c}
  17         },
  18         {
  19                 .u32 = {RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX,
  20                 RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX}
  21         },
  22         {
  23                 .u16 = {1, 1, 1, 1, 1, 1, 1, 1}
  24         },
  25         {
  26                 .u32 = {0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c}
  27         },
  28 };
  29
  30 /*
  31  * Resolve priority for multiple results (altivec version).
  32  * This consists comparing the priority of the current traversal with the
  33  * running set of results for the packet.
  34  * For each result, keep a running array of the result (rule number) and
  35  * its priority for each category.
  36  */
  37 static inline void
  38 resolve_priority_altivec(uint64_t transition, int n,
  39         const struct rte_acl_ctx *ctx, struct parms *parms,
  40         const struct rte_acl_match_results *p, uint32_t categories)
  41 {
  42         uint32_t x;
  43         xmm_t results, priority, results1, priority1;
  44         vector bool int selector;
  45         xmm_t *saved_results, *saved_priority;
  46
  47         for (x = 0; x < categories; x += RTE_ACL_RESULTS_MULTIPLIER) {
  48
  49                 saved_results = (xmm_t *)(&parms[n].cmplt->results[x]);
  50                 saved_priority =
  51                         (xmm_t *)(&parms[n].cmplt->priority[x]);
  52
  53                 /* get results and priorities for completed trie */
  54                 results = *(const xmm_t *)&p[transition].results[x];
  55                 priority = *(const xmm_t *)&p[transition].priority[x];
  56
  57                 /* if this is not the first completed trie */
  58                 if (parms[n].cmplt->count != ctx->num_tries) {
  59
  60                         /* get running best results and their priorities */
  61                         results1 = *saved_results;
  62                         priority1 = *saved_priority;
  63
  64                         /* select results that are highest priority */
  65                         selector = vec_cmpgt(priority1, priority);
  66                         results = vec_sel(results, results1, selector);
  67                         priority = vec_sel(priority, priority1,
  68                                 selector);
  69                 }
  70
  71                 /* save running best results and their priorities */
  72                 *saved_results = results;
  73                 *saved_priority = priority;
  74         }
  75 }
  76
  77 /*
  78  * Check for any match in 4 transitions
  79  */
  80 static __rte_always_inline uint32_t
  81 check_any_match_x4(uint64_t val[])
  82 {
  83         return (val[0] | val[1] | val[2] | val[3]) & RTE_ACL_NODE_MATCH;
  84 }
  85
  86 static __rte_always_inline void
  87 acl_match_check_x4(int slot, const struct rte_acl_ctx *ctx, struct parms *parms,
  88         struct acl_flow_data *flows, uint64_t transitions[])
  89 {
  90         while (check_any_match_x4(transitions)) {
  91                 transitions[0] = acl_match_check(transitions[0], slot, ctx,
  92                         parms, flows, resolve_priority_altivec);
  93                 transitions[1] = acl_match_check(transitions[1], slot + 1, ctx,
  94                         parms, flows, resolve_priority_altivec);
  95                 transitions[2] = acl_match_check(transitions[2], slot + 2, ctx,
  96                         parms, flows, resolve_priority_altivec);
  97                 transitions[3] = acl_match_check(transitions[3], slot + 3, ctx,
  98                         parms, flows, resolve_priority_altivec);
  99         }
 100 }
 101
 102 /*
 103  * Process 4 transitions (in 2 XMM registers) in parallel
 104  */
 105 static inline __attribute__((optimize("O2"))) xmm_t
 106 transition4(xmm_t next_input, const uint64_t *trans,
 107         xmm_t *indices1, xmm_t *indices2)
 108 {
 109         xmm_t addr, tr_lo, tr_hi;
 110         xmm_t in, node_type, r, t;
 111         xmm_t dfa_ofs, quad_ofs;
 112         xmm_t *index_mask, *tp;
 113         vector bool int dfa_msk;
 114         vector signed char zeroes = {};
 115         union {
 116                 uint64_t d64[2];
 117                 uint32_t d32[4];
 118         } v;
 119
 120         /* Move low 32 into tr_lo and high 32 into tr_hi */
 121         tr_lo = (xmm_t){(*indices1)[0], (*indices1)[2],
 122                         (*indices2)[0], (*indices2)[2]};
 123         tr_hi = (xmm_t){(*indices1)[1], (*indices1)[3],
 124                         (*indices2)[1], (*indices2)[3]};
 125
 126          /* Calculate the address (array index) for all 4 transitions. */
 127         index_mask = (xmm_t *)&altivec_acl_const.xmm_index_mask.u32;
 128         t = vec_xor(*index_mask, *index_mask);
 129         in = vec_perm(next_input, (xmm_t){},
 130                 *(vector unsigned char *)&altivec_acl_const.xmm_shuffle_input);
 131
 132         /* Calc node type and node addr */
 133         node_type = vec_and(vec_nor(*index_mask, *index_mask), tr_lo);
 134         addr = vec_and(tr_lo, *index_mask);
 135
 136         /* mask for DFA type(0) nodes */
 137         dfa_msk = vec_cmpeq(node_type, t);
 138
 139         /* DFA calculations. */
 140         r = vec_sr(in, (vector unsigned int){30, 30, 30, 30});
 141         tp = (xmm_t *)&altivec_acl_const.range_base.u32;
 142         r = vec_add(r, *tp);
 143         t = vec_sr(in, (vector unsigned int){24, 24, 24, 24});
 144         r = vec_perm(tr_hi, (xmm_t){(uint16_t)0 << 16},
 145                 (vector unsigned char)r);
 146
 147         dfa_ofs = vec_sub(t, r);
 148
 149         /* QUAD/SINGLE caluclations. */
 150         t = (xmm_t)vec_cmpgt((vector signed char)in, (vector signed char)tr_hi);
 151         t = (xmm_t)vec_sel(
 152                 vec_sel(
 153                         (vector signed char)vec_sub(
 154                                 zeroes, (vector signed char)t),
 155                         (vector signed char)t,
 156                         vec_cmpgt((vector signed char)t, zeroes)),
 157                 zeroes,
 158                 vec_cmpeq((vector signed char)t, zeroes));
 159
 160         t = (xmm_t)vec_msum((vector signed char)t,
 161                 (vector unsigned char)t, (xmm_t){});
 162         quad_ofs = (xmm_t)vec_msum((vector signed short)t,
 163                 *(vector signed short *)&altivec_acl_const.xmm_ones_16.u16,
 164                 (xmm_t){});
 165
 166         /* blend DFA and QUAD/SINGLE. */
 167         t = vec_sel(quad_ofs, dfa_ofs, dfa_msk);
 168
 169         /* calculate address for next transitions. */
 170         addr = vec_add(addr, t);
 171
 172         v.d64[0] = (uint64_t)trans[addr[0]];
 173         v.d64[1] = (uint64_t)trans[addr[1]];
 174         *indices1 = (xmm_t){v.d32[0], v.d32[1], v.d32[2], v.d32[3]};
 175         v.d64[0] = (uint64_t)trans[addr[2]];
 176         v.d64[1] = (uint64_t)trans[addr[3]];
 177         *indices2 = (xmm_t){v.d32[0], v.d32[1], v.d32[2], v.d32[3]};
 178
 179         return vec_sr(next_input,
 180                 (vector unsigned int){CHAR_BIT, CHAR_BIT, CHAR_BIT, CHAR_BIT});
 181 }
 182
 183 /*
 184  * Execute trie traversal with 8 traversals in parallel
 185  */
 186 static inline int
 187 search_altivec_8(const struct rte_acl_ctx *ctx, const uint8_t **data,
 188         uint32_t *results, uint32_t total_packets, uint32_t categories)
 189 {
 190         int n;
 191         struct acl_flow_data flows;
 192         uint64_t index_array[MAX_SEARCHES_ALTIVEC8];
 193         struct completion cmplt[MAX_SEARCHES_ALTIVEC8];
 194         struct parms parms[MAX_SEARCHES_ALTIVEC8];
 195         xmm_t input0, input1;
 196
 197         acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
 198                 total_packets, categories, ctx->trans_table);
 199
 200         for (n = 0; n < MAX_SEARCHES_ALTIVEC8; n++) {
 201                 cmplt[n].count = 0;
 202                 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
 203         }
 204
 205          /* Check for any matches. */
 206         acl_match_check_x4(0, ctx, parms, &flows, (uint64_t *)&index_array[0]);
 207         acl_match_check_x4(4, ctx, parms, &flows, (uint64_t *)&index_array[4]);
 208
 209         while (flows.started > 0) {
 210
 211                 /* Gather 4 bytes of input data for each stream. */
 212                 input0 = (xmm_t){GET_NEXT_4BYTES(parms, 0),
 213                                 GET_NEXT_4BYTES(parms, 1),
 214                                 GET_NEXT_4BYTES(parms, 2),
 215                                 GET_NEXT_4BYTES(parms, 3)};
 216
 217                 input1 = (xmm_t){GET_NEXT_4BYTES(parms, 4),
 218                                 GET_NEXT_4BYTES(parms, 5),
 219                                 GET_NEXT_4BYTES(parms, 6),
 220                                 GET_NEXT_4BYTES(parms, 7)};
 221
 222                  /* Process the 4 bytes of input on each stream. */
 223
 224                 input0 = transition4(input0, flows.trans,
 225                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 226                 input1 = transition4(input1, flows.trans,
 227                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 228
 229                 input0 = transition4(input0, flows.trans,
 230                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 231                 input1 = transition4(input1, flows.trans,
 232                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 233
 234                 input0 = transition4(input0, flows.trans,
 235                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 236                 input1 = transition4(input1, flows.trans,
 237                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 238
 239                 input0 = transition4(input0, flows.trans,
 240                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 241                 input1 = transition4(input1, flows.trans,
 242                         (xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
 243
 244                  /* Check for any matches. */
 245                 acl_match_check_x4(0, ctx, parms, &flows,
 246                         (uint64_t *)&index_array[0]);
 247                 acl_match_check_x4(4, ctx, parms, &flows,
 248                         (uint64_t *)&index_array[4]);
 249         }
 250
 251         return 0;
 252 }
 253
 254 /*
 255  * Execute trie traversal with 4 traversals in parallel
 256  */
 257 static inline int
 258 search_altivec_4(const struct rte_acl_ctx *ctx, const uint8_t **data,
 259          uint32_t *results, int total_packets, uint32_t categories)
 260 {
 261         int n;
 262         struct acl_flow_data flows;
 263         uint64_t index_array[MAX_SEARCHES_ALTIVEC4];
 264         struct completion cmplt[MAX_SEARCHES_ALTIVEC4];
 265         struct parms parms[MAX_SEARCHES_ALTIVEC4];
 266         xmm_t input;
 267
 268         acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
 269                 total_packets, categories, ctx->trans_table);
 270
 271         for (n = 0; n < MAX_SEARCHES_ALTIVEC4; n++) {
 272                 cmplt[n].count = 0;
 273                 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
 274         }
 275
 276         /* Check for any matches. */
 277         acl_match_check_x4(0, ctx, parms, &flows, index_array);
 278
 279         while (flows.started > 0) {
 280
 281                 /* Gather 4 bytes of input data for each stream. */
 282                 input = (xmm_t){GET_NEXT_4BYTES(parms, 0),
 283                                 GET_NEXT_4BYTES(parms, 1),
 284                                 GET_NEXT_4BYTES(parms, 2),
 285                                 GET_NEXT_4BYTES(parms, 3)};
 286
 287                 /* Process the 4 bytes of input on each stream. */
 288                 input = transition4(input, flows.trans,
 289                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 290                 input = transition4(input, flows.trans,
 291                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 292                 input = transition4(input, flows.trans,
 293                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 294                 input = transition4(input, flows.trans,
 295                         (xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
 296
 297                 /* Check for any matches. */
 298                 acl_match_check_x4(0, ctx, parms, &flows, index_array);
 299         }
 300
 301         return 0;
 302 }