lib/librte_acl/acl_run_sse.h

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include "acl_run.h"
  35 #include "acl_vect.h"
  36
  37 enum {
  38         SHUFFLE32_SLOT1 = 0xe5,
  39         SHUFFLE32_SLOT2 = 0xe6,
  40         SHUFFLE32_SLOT3 = 0xe7,
  41         SHUFFLE32_SWAP64 = 0x4e,
  42 };
  43
  44 static const rte_xmm_t xmm_shuffle_input = {
  45         .u32 = {0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c},
  46 };
  47
  48 static const rte_xmm_t xmm_ones_16 = {
  49         .u16 = {1, 1, 1, 1, 1, 1, 1, 1},
  50 };
  51
  52 static const rte_xmm_t xmm_match_mask = {
  53         .u32 = {
  54                 RTE_ACL_NODE_MATCH,
  55                 RTE_ACL_NODE_MATCH,
  56                 RTE_ACL_NODE_MATCH,
  57                 RTE_ACL_NODE_MATCH,
  58         },
  59 };
  60
  61 static const rte_xmm_t xmm_index_mask = {
  62         .u32 = {
  63                 RTE_ACL_NODE_INDEX,
  64                 RTE_ACL_NODE_INDEX,
  65                 RTE_ACL_NODE_INDEX,
  66                 RTE_ACL_NODE_INDEX,
  67         },
  68 };
  69
  70 static const rte_xmm_t xmm_range_base = {
  71         .u32 = {
  72                 0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c,
  73         },
  74 };
  75
  76 /*
  77  * Resolve priority for multiple results (sse version).
  78  * This consists comparing the priority of the current traversal with the
  79  * running set of results for the packet.
  80  * For each result, keep a running array of the result (rule number) and
  81  * its priority for each category.
  82  */
  83 static inline void
  84 resolve_priority_sse(uint64_t transition, int n, const struct rte_acl_ctx *ctx,
  85         struct parms *parms, const struct rte_acl_match_results *p,
  86         uint32_t categories)
  87 {
  88         uint32_t x;
  89         xmm_t results, priority, results1, priority1, selector;
  90         xmm_t *saved_results, *saved_priority;
  91
  92         for (x = 0; x < categories; x += RTE_ACL_RESULTS_MULTIPLIER) {
  93
  94                 saved_results = (xmm_t *)(&parms[n].cmplt->results[x]);
  95                 saved_priority =
  96                         (xmm_t *)(&parms[n].cmplt->priority[x]);
  97
  98                 /* get results and priorities for completed trie */
  99                 results = _mm_loadu_si128(
 100                         (const xmm_t *)&p[transition].results[x]);
 101                 priority = _mm_loadu_si128(
 102                         (const xmm_t *)&p[transition].priority[x]);
 103
 104                 /* if this is not the first completed trie */
 105                 if (parms[n].cmplt->count != ctx->num_tries) {
 106
 107                         /* get running best results and their priorities */
 108                         results1 = _mm_loadu_si128(saved_results);
 109                         priority1 = _mm_loadu_si128(saved_priority);
 110
 111                         /* select results that are highest priority */
 112                         selector = _mm_cmpgt_epi32(priority1, priority);
 113                         results = _mm_blendv_epi8(results, results1, selector);
 114                         priority = _mm_blendv_epi8(priority, priority1,
 115                                 selector);
 116                 }
 117
 118                 /* save running best results and their priorities */
 119                 _mm_storeu_si128(saved_results, results);
 120                 _mm_storeu_si128(saved_priority, priority);
 121         }
 122 }
 123
 124 /*
 125  * Extract transitions from an XMM register and check for any matches
 126  */
 127 static void
 128 acl_process_matches(xmm_t *indices, int slot, const struct rte_acl_ctx *ctx,
 129         struct parms *parms, struct acl_flow_data *flows)
 130 {
 131         uint64_t transition1, transition2;
 132
 133         /* extract transition from low 64 bits. */
 134         transition1 = _mm_cvtsi128_si64(*indices);
 135
 136         /* extract transition from high 64 bits. */
 137         *indices = _mm_shuffle_epi32(*indices, SHUFFLE32_SWAP64);
 138         transition2 = _mm_cvtsi128_si64(*indices);
 139
 140         transition1 = acl_match_check(transition1, slot, ctx,
 141                 parms, flows, resolve_priority_sse);
 142         transition2 = acl_match_check(transition2, slot + 1, ctx,
 143                 parms, flows, resolve_priority_sse);
 144
 145         /* update indices with new transitions. */
 146         *indices = _mm_set_epi64x(transition2, transition1);
 147 }
 148
 149 /*
 150  * Check for any match in 4 transitions (contained in 2 SSE registers)
 151  */
 152 static inline __attribute__((always_inline)) void
 153 acl_match_check_x4(int slot, const struct rte_acl_ctx *ctx, struct parms *parms,
 154         struct acl_flow_data *flows, xmm_t *indices1, xmm_t *indices2,
 155         xmm_t match_mask)
 156 {
 157         xmm_t temp;
 158
 159         /* put low 32 bits of each transition into one register */
 160         temp = (xmm_t)_mm_shuffle_ps((__m128)*indices1, (__m128)*indices2,
 161                 0x88);
 162         /* test for match node */
 163         temp = _mm_and_si128(match_mask, temp);
 164
 165         while (!_mm_testz_si128(temp, temp)) {
 166                 acl_process_matches(indices1, slot, ctx, parms, flows);
 167                 acl_process_matches(indices2, slot + 2, ctx, parms, flows);
 168
 169                 temp = (xmm_t)_mm_shuffle_ps((__m128)*indices1,
 170                                         (__m128)*indices2,
 171                                         0x88);
 172                 temp = _mm_and_si128(match_mask, temp);
 173         }
 174 }
 175
 176 /*
 177  * Process 4 transitions (in 2 XMM registers) in parallel
 178  */
 179 static inline __attribute__((always_inline)) xmm_t
 180 transition4(xmm_t next_input, const uint64_t *trans,
 181         xmm_t *indices1, xmm_t *indices2)
 182 {
 183         xmm_t addr, tr_lo, tr_hi;
 184         uint64_t trans0, trans2;
 185
 186         /* Shuffle low 32 into tr_lo and high 32 into tr_hi */
 187         ACL_TR_HILO(mm, __m128, *indices1, *indices2, tr_lo, tr_hi);
 188
 189          /* Calculate the address (array index) for all 4 transitions. */
 190         ACL_TR_CALC_ADDR(mm, 128, addr, xmm_index_mask.x, next_input,
 191                 xmm_shuffle_input.x, xmm_ones_16.x, xmm_range_base.x,
 192                 tr_lo, tr_hi);
 193
 194          /* Gather 64 bit transitions and pack back into 2 registers. */
 195
 196         trans0 = trans[_mm_cvtsi128_si32(addr)];
 197
 198         /* get slot 2 */
 199
 200         /* {x0, x1, x2, x3} -> {x2, x1, x2, x3} */
 201         addr = _mm_shuffle_epi32(addr, SHUFFLE32_SLOT2);
 202         trans2 = trans[_mm_cvtsi128_si32(addr)];
 203
 204         /* get slot 1 */
 205
 206         /* {x2, x1, x2, x3} -> {x1, x1, x2, x3} */
 207         addr = _mm_shuffle_epi32(addr, SHUFFLE32_SLOT1);
 208         *indices1 = _mm_set_epi64x(trans[_mm_cvtsi128_si32(addr)], trans0);
 209
 210         /* get slot 3 */
 211
 212         /* {x1, x1, x2, x3} -> {x3, x1, x2, x3} */
 213         addr = _mm_shuffle_epi32(addr, SHUFFLE32_SLOT3);
 214         *indices2 = _mm_set_epi64x(trans[_mm_cvtsi128_si32(addr)], trans2);
 215
 216         return _mm_srli_epi32(next_input, CHAR_BIT);
 217 }
 218
 219 /*
 220  * Execute trie traversal with 8 traversals in parallel
 221  */
 222 static inline int
 223 search_sse_8(const struct rte_acl_ctx *ctx, const uint8_t **data,
 224         uint32_t *results, uint32_t total_packets, uint32_t categories)
 225 {
 226         int n;
 227         struct acl_flow_data flows;
 228         uint64_t index_array[MAX_SEARCHES_SSE8];
 229         struct completion cmplt[MAX_SEARCHES_SSE8];
 230         struct parms parms[MAX_SEARCHES_SSE8];
 231         xmm_t input0, input1;
 232         xmm_t indices1, indices2, indices3, indices4;
 233
 234         acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
 235                 total_packets, categories, ctx->trans_table);
 236
 237         for (n = 0; n < MAX_SEARCHES_SSE8; n++) {
 238                 cmplt[n].count = 0;
 239                 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
 240         }
 241
 242         /*
 243          * indices1 contains index_array[0,1]
 244          * indices2 contains index_array[2,3]
 245          * indices3 contains index_array[4,5]
 246          * indices4 contains index_array[6,7]
 247          */
 248
 249         indices1 = _mm_loadu_si128((xmm_t *) &index_array[0]);
 250         indices2 = _mm_loadu_si128((xmm_t *) &index_array[2]);
 251
 252         indices3 = _mm_loadu_si128((xmm_t *) &index_array[4]);
 253         indices4 = _mm_loadu_si128((xmm_t *) &index_array[6]);
 254
 255          /* Check for any matches. */
 256         acl_match_check_x4(0, ctx, parms, &flows,
 257                 &indices1, &indices2, xmm_match_mask.x);
 258         acl_match_check_x4(4, ctx, parms, &flows,
 259                 &indices3, &indices4, xmm_match_mask.x);
 260
 261         while (flows.started > 0) {
 262
 263                 /* Gather 4 bytes of input data for each stream. */
 264                 input0 = _mm_cvtsi32_si128(GET_NEXT_4BYTES(parms, 0));
 265                 input1 = _mm_cvtsi32_si128(GET_NEXT_4BYTES(parms, 4));
 266
 267                 input0 = _mm_insert_epi32(input0, GET_NEXT_4BYTES(parms, 1), 1);
 268                 input1 = _mm_insert_epi32(input1, GET_NEXT_4BYTES(parms, 5), 1);
 269
 270                 input0 = _mm_insert_epi32(input0, GET_NEXT_4BYTES(parms, 2), 2);
 271                 input1 = _mm_insert_epi32(input1, GET_NEXT_4BYTES(parms, 6), 2);
 272
 273                 input0 = _mm_insert_epi32(input0, GET_NEXT_4BYTES(parms, 3), 3);
 274                 input1 = _mm_insert_epi32(input1, GET_NEXT_4BYTES(parms, 7), 3);
 275
 276                  /* Process the 4 bytes of input on each stream. */
 277
 278                 input0 = transition4(input0, flows.trans,
 279                         &indices1, &indices2);
 280                 input1 = transition4(input1, flows.trans,
 281                         &indices3, &indices4);
 282
 283                 input0 = transition4(input0, flows.trans,
 284                         &indices1, &indices2);
 285                 input1 = transition4(input1, flows.trans,
 286                         &indices3, &indices4);
 287
 288                 input0 = transition4(input0, flows.trans,
 289                         &indices1, &indices2);
 290                 input1 = transition4(input1, flows.trans,
 291                         &indices3, &indices4);
 292
 293                 input0 = transition4(input0, flows.trans,
 294                         &indices1, &indices2);
 295                 input1 = transition4(input1, flows.trans,
 296                         &indices3, &indices4);
 297
 298                  /* Check for any matches. */
 299                 acl_match_check_x4(0, ctx, parms, &flows,
 300                         &indices1, &indices2, xmm_match_mask.x);
 301                 acl_match_check_x4(4, ctx, parms, &flows,
 302                         &indices3, &indices4, xmm_match_mask.x);
 303         }
 304
 305         return 0;
 306 }
 307
 308 /*
 309  * Execute trie traversal with 4 traversals in parallel
 310  */
 311 static inline int
 312 search_sse_4(const struct rte_acl_ctx *ctx, const uint8_t **data,
 313          uint32_t *results, int total_packets, uint32_t categories)
 314 {
 315         int n;
 316         struct acl_flow_data flows;
 317         uint64_t index_array[MAX_SEARCHES_SSE4];
 318         struct completion cmplt[MAX_SEARCHES_SSE4];
 319         struct parms parms[MAX_SEARCHES_SSE4];
 320         xmm_t input, indices1, indices2;
 321
 322         acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
 323                 total_packets, categories, ctx->trans_table);
 324
 325         for (n = 0; n < MAX_SEARCHES_SSE4; n++) {
 326                 cmplt[n].count = 0;
 327                 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
 328         }
 329
 330         indices1 = _mm_loadu_si128((xmm_t *) &index_array[0]);
 331         indices2 = _mm_loadu_si128((xmm_t *) &index_array[2]);
 332
 333         /* Check for any matches. */
 334         acl_match_check_x4(0, ctx, parms, &flows,
 335                 &indices1, &indices2, xmm_match_mask.x);
 336
 337         while (flows.started > 0) {
 338
 339                 /* Gather 4 bytes of input data for each stream. */
 340                 input = _mm_cvtsi32_si128(GET_NEXT_4BYTES(parms, 0));
 341                 input = _mm_insert_epi32(input, GET_NEXT_4BYTES(parms, 1), 1);
 342                 input = _mm_insert_epi32(input, GET_NEXT_4BYTES(parms, 2), 2);
 343                 input = _mm_insert_epi32(input, GET_NEXT_4BYTES(parms, 3), 3);
 344
 345                 /* Process the 4 bytes of input on each stream. */
 346                 input = transition4(input, flows.trans, &indices1, &indices2);
 347                 input = transition4(input, flows.trans, &indices1, &indices2);
 348                 input = transition4(input, flows.trans, &indices1, &indices2);
 349                 input = transition4(input, flows.trans, &indices1, &indices2);
 350
 351                 /* Check for any matches. */
 352                 acl_match_check_x4(0, ctx, parms, &flows,
 353                         &indices1, &indices2, xmm_match_mask.x);
 354         }
 355
 356         return 0;
 357 }