4 * Copyright(c) 2010-2014 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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.
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.
38 SHUFFLE32_SLOT1 = 0xe5,
39 SHUFFLE32_SLOT2 = 0xe6,
40 SHUFFLE32_SLOT3 = 0xe7,
41 SHUFFLE32_SWAP64 = 0x4e,
44 static const rte_xmm_t xmm_shuffle_input = {
45 .u32 = {0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c},
48 static const rte_xmm_t xmm_shuffle_input64 = {
49 .u32 = {0x00000000, 0x04040404, 0x80808080, 0x80808080},
52 static const rte_xmm_t xmm_ones_16 = {
53 .u16 = {1, 1, 1, 1, 1, 1, 1, 1},
56 static const rte_xmm_t xmm_match_mask = {
65 static const rte_xmm_t xmm_match_mask64 = {
74 static const rte_xmm_t xmm_index_mask = {
83 static const rte_xmm_t xmm_index_mask64 = {
94 * Resolve priority for multiple results (sse version).
95 * This consists comparing the priority of the current traversal with the
96 * running set of results for the packet.
97 * For each result, keep a running array of the result (rule number) and
98 * its priority for each category.
101 resolve_priority_sse(uint64_t transition, int n, const struct rte_acl_ctx *ctx,
102 struct parms *parms, const struct rte_acl_match_results *p,
106 xmm_t results, priority, results1, priority1, selector;
107 xmm_t *saved_results, *saved_priority;
109 for (x = 0; x < categories; x += RTE_ACL_RESULTS_MULTIPLIER) {
111 saved_results = (xmm_t *)(&parms[n].cmplt->results[x]);
113 (xmm_t *)(&parms[n].cmplt->priority[x]);
115 /* get results and priorities for completed trie */
116 results = MM_LOADU((const xmm_t *)&p[transition].results[x]);
117 priority = MM_LOADU((const xmm_t *)&p[transition].priority[x]);
119 /* if this is not the first completed trie */
120 if (parms[n].cmplt->count != ctx->num_tries) {
122 /* get running best results and their priorities */
123 results1 = MM_LOADU(saved_results);
124 priority1 = MM_LOADU(saved_priority);
126 /* select results that are highest priority */
127 selector = MM_CMPGT32(priority1, priority);
128 results = MM_BLENDV8(results, results1, selector);
129 priority = MM_BLENDV8(priority, priority1, selector);
132 /* save running best results and their priorities */
133 MM_STOREU(saved_results, results);
134 MM_STOREU(saved_priority, priority);
139 * Extract transitions from an XMM register and check for any matches
142 acl_process_matches(xmm_t *indices, int slot, const struct rte_acl_ctx *ctx,
143 struct parms *parms, struct acl_flow_data *flows)
145 uint64_t transition1, transition2;
147 /* extract transition from low 64 bits. */
148 transition1 = MM_CVT64(*indices);
150 /* extract transition from high 64 bits. */
151 *indices = MM_SHUFFLE32(*indices, SHUFFLE32_SWAP64);
152 transition2 = MM_CVT64(*indices);
154 transition1 = acl_match_check(transition1, slot, ctx,
155 parms, flows, resolve_priority_sse);
156 transition2 = acl_match_check(transition2, slot + 1, ctx,
157 parms, flows, resolve_priority_sse);
159 /* update indices with new transitions. */
160 *indices = MM_SET64(transition2, transition1);
164 * Check for a match in 2 transitions (contained in SSE register)
166 static inline __attribute__((always_inline)) void
167 acl_match_check_x2(int slot, const struct rte_acl_ctx *ctx, struct parms *parms,
168 struct acl_flow_data *flows, xmm_t *indices, xmm_t match_mask)
172 temp = MM_AND(match_mask, *indices);
173 while (!MM_TESTZ(temp, temp)) {
174 acl_process_matches(indices, slot, ctx, parms, flows);
175 temp = MM_AND(match_mask, *indices);
180 * Check for any match in 4 transitions (contained in 2 SSE registers)
182 static inline __attribute__((always_inline)) void
183 acl_match_check_x4(int slot, const struct rte_acl_ctx *ctx, struct parms *parms,
184 struct acl_flow_data *flows, xmm_t *indices1, xmm_t *indices2,
189 /* put low 32 bits of each transition into one register */
190 temp = (xmm_t)MM_SHUFFLEPS((__m128)*indices1, (__m128)*indices2,
192 /* test for match node */
193 temp = MM_AND(match_mask, temp);
195 while (!MM_TESTZ(temp, temp)) {
196 acl_process_matches(indices1, slot, ctx, parms, flows);
197 acl_process_matches(indices2, slot + 2, ctx, parms, flows);
199 temp = (xmm_t)MM_SHUFFLEPS((__m128)*indices1,
202 temp = MM_AND(match_mask, temp);
207 * Calculate the address of the next transition for
208 * all types of nodes. Note that only DFA nodes and range
209 * nodes actually transition to another node. Match
212 static inline __attribute__((always_inline)) xmm_t
213 calc_addr_sse(xmm_t index_mask, xmm_t next_input, xmm_t shuffle_input,
214 xmm_t ones_16, xmm_t indices1, xmm_t indices2)
216 xmm_t addr, node_types, range, temp;
217 xmm_t dfa_msk, dfa_ofs, quad_ofs;
220 const xmm_t range_base = _mm_set_epi32(0xffffff0c, 0xffffff08,
221 0xffffff04, 0xffffff00);
224 * Note that no transition is done for a match
225 * node and therefore a stream freezes when
226 * it reaches a match.
229 /* Shuffle low 32 into temp and high 32 into indices2 */
230 temp = (xmm_t)MM_SHUFFLEPS((__m128)indices1, (__m128)indices2, 0x88);
231 range = (xmm_t)MM_SHUFFLEPS((__m128)indices1, (__m128)indices2, 0xdd);
233 t = MM_XOR(index_mask, index_mask);
235 /* shuffle input byte to all 4 positions of 32 bit value */
236 in = MM_SHUFFLE8(next_input, shuffle_input);
238 /* Calc node type and node addr */
239 node_types = MM_ANDNOT(index_mask, temp);
240 addr = MM_AND(index_mask, temp);
243 * Calc addr for DFAs - addr = dfa_index + input_byte
246 /* mask for DFA type (0) nodes */
247 dfa_msk = MM_CMPEQ32(node_types, t);
249 r = _mm_srli_epi32(in, 30);
250 r = _mm_add_epi8(r, range_base);
252 t = _mm_srli_epi32(in, 24);
253 r = _mm_shuffle_epi8(range, r);
255 dfa_ofs = _mm_sub_epi32(t, r);
258 * Calculate number of range boundaries that are less than the
259 * input value. Range boundaries for each node are in signed 8 bit,
260 * ordered from -128 to 127 in the indices2 register.
261 * This is effectively a popcnt of bytes that are greater than the
266 temp = MM_CMPGT8(in, range);
268 /* convert -1 to 1 (bytes greater than input byte */
269 temp = MM_SIGN8(temp, temp);
271 /* horizontal add pairs of bytes into words */
272 temp = MM_MADD8(temp, temp);
274 /* horizontal add pairs of words into dwords */
275 quad_ofs = MM_MADD16(temp, ones_16);
277 /* mask to range type nodes */
278 temp = _mm_blendv_epi8(quad_ofs, dfa_ofs, dfa_msk);
280 /* add index into node position */
281 return MM_ADD32(addr, temp);
285 * Process 4 transitions (in 2 SIMD registers) in parallel
287 static inline __attribute__((always_inline)) xmm_t
288 transition4(xmm_t next_input, const uint64_t *trans,
289 xmm_t *indices1, xmm_t *indices2)
292 uint64_t trans0, trans2;
294 /* Calculate the address (array index) for all 4 transitions. */
296 addr = calc_addr_sse(xmm_index_mask.x, next_input, xmm_shuffle_input.x,
297 xmm_ones_16.x, *indices1, *indices2);
299 /* Gather 64 bit transitions and pack back into 2 registers. */
301 trans0 = trans[MM_CVT32(addr)];
305 /* {x0, x1, x2, x3} -> {x2, x1, x2, x3} */
306 addr = MM_SHUFFLE32(addr, SHUFFLE32_SLOT2);
307 trans2 = trans[MM_CVT32(addr)];
311 /* {x2, x1, x2, x3} -> {x1, x1, x2, x3} */
312 addr = MM_SHUFFLE32(addr, SHUFFLE32_SLOT1);
313 *indices1 = MM_SET64(trans[MM_CVT32(addr)], trans0);
317 /* {x1, x1, x2, x3} -> {x3, x1, x2, x3} */
318 addr = MM_SHUFFLE32(addr, SHUFFLE32_SLOT3);
319 *indices2 = MM_SET64(trans[MM_CVT32(addr)], trans2);
321 return MM_SRL32(next_input, CHAR_BIT);
325 * Execute trie traversal with 8 traversals in parallel
328 search_sse_8(const struct rte_acl_ctx *ctx, const uint8_t **data,
329 uint32_t *results, uint32_t total_packets, uint32_t categories)
332 struct acl_flow_data flows;
333 uint64_t index_array[MAX_SEARCHES_SSE8];
334 struct completion cmplt[MAX_SEARCHES_SSE8];
335 struct parms parms[MAX_SEARCHES_SSE8];
336 xmm_t input0, input1;
337 xmm_t indices1, indices2, indices3, indices4;
339 acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
340 total_packets, categories, ctx->trans_table);
342 for (n = 0; n < MAX_SEARCHES_SSE8; n++) {
344 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
348 * indices1 contains index_array[0,1]
349 * indices2 contains index_array[2,3]
350 * indices3 contains index_array[4,5]
351 * indices4 contains index_array[6,7]
354 indices1 = MM_LOADU((xmm_t *) &index_array[0]);
355 indices2 = MM_LOADU((xmm_t *) &index_array[2]);
357 indices3 = MM_LOADU((xmm_t *) &index_array[4]);
358 indices4 = MM_LOADU((xmm_t *) &index_array[6]);
360 /* Check for any matches. */
361 acl_match_check_x4(0, ctx, parms, &flows,
362 &indices1, &indices2, xmm_match_mask.x);
363 acl_match_check_x4(4, ctx, parms, &flows,
364 &indices3, &indices4, xmm_match_mask.x);
366 while (flows.started > 0) {
368 /* Gather 4 bytes of input data for each stream. */
369 input0 = _mm_cvtsi32_si128(GET_NEXT_4BYTES(parms, 0));
370 input1 = _mm_cvtsi32_si128(GET_NEXT_4BYTES(parms, 4));
372 input0 = MM_INSERT32(input0, GET_NEXT_4BYTES(parms, 1), 1);
373 input1 = MM_INSERT32(input1, GET_NEXT_4BYTES(parms, 5), 1);
375 input0 = MM_INSERT32(input0, GET_NEXT_4BYTES(parms, 2), 2);
376 input1 = MM_INSERT32(input1, GET_NEXT_4BYTES(parms, 6), 2);
378 input0 = MM_INSERT32(input0, GET_NEXT_4BYTES(parms, 3), 3);
379 input1 = MM_INSERT32(input1, GET_NEXT_4BYTES(parms, 7), 3);
381 /* Process the 4 bytes of input on each stream. */
383 input0 = transition4(input0, flows.trans,
384 &indices1, &indices2);
385 input1 = transition4(input1, flows.trans,
386 &indices3, &indices4);
388 input0 = transition4(input0, flows.trans,
389 &indices1, &indices2);
390 input1 = transition4(input1, flows.trans,
391 &indices3, &indices4);
393 input0 = transition4(input0, flows.trans,
394 &indices1, &indices2);
395 input1 = transition4(input1, flows.trans,
396 &indices3, &indices4);
398 input0 = transition4(input0, flows.trans,
399 &indices1, &indices2);
400 input1 = transition4(input1, flows.trans,
401 &indices3, &indices4);
403 /* Check for any matches. */
404 acl_match_check_x4(0, ctx, parms, &flows,
405 &indices1, &indices2, xmm_match_mask.x);
406 acl_match_check_x4(4, ctx, parms, &flows,
407 &indices3, &indices4, xmm_match_mask.x);
414 * Execute trie traversal with 4 traversals in parallel
417 search_sse_4(const struct rte_acl_ctx *ctx, const uint8_t **data,
418 uint32_t *results, int total_packets, uint32_t categories)
421 struct acl_flow_data flows;
422 uint64_t index_array[MAX_SEARCHES_SSE4];
423 struct completion cmplt[MAX_SEARCHES_SSE4];
424 struct parms parms[MAX_SEARCHES_SSE4];
425 xmm_t input, indices1, indices2;
427 acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
428 total_packets, categories, ctx->trans_table);
430 for (n = 0; n < MAX_SEARCHES_SSE4; n++) {
432 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
435 indices1 = MM_LOADU((xmm_t *) &index_array[0]);
436 indices2 = MM_LOADU((xmm_t *) &index_array[2]);
438 /* Check for any matches. */
439 acl_match_check_x4(0, ctx, parms, &flows,
440 &indices1, &indices2, xmm_match_mask.x);
442 while (flows.started > 0) {
444 /* Gather 4 bytes of input data for each stream. */
445 input = _mm_cvtsi32_si128(GET_NEXT_4BYTES(parms, 0));
446 input = MM_INSERT32(input, GET_NEXT_4BYTES(parms, 1), 1);
447 input = MM_INSERT32(input, GET_NEXT_4BYTES(parms, 2), 2);
448 input = MM_INSERT32(input, GET_NEXT_4BYTES(parms, 3), 3);
450 /* Process the 4 bytes of input on each stream. */
451 input = transition4(input, flows.trans, &indices1, &indices2);
452 input = transition4(input, flows.trans, &indices1, &indices2);
453 input = transition4(input, flows.trans, &indices1, &indices2);
454 input = transition4(input, flows.trans, &indices1, &indices2);
456 /* Check for any matches. */
457 acl_match_check_x4(0, ctx, parms, &flows,
458 &indices1, &indices2, xmm_match_mask.x);
464 static inline __attribute__((always_inline)) xmm_t
465 transition2(xmm_t next_input, const uint64_t *trans, xmm_t *indices1)
468 xmm_t addr, indices2;
470 indices2 = _mm_setzero_si128();
472 addr = calc_addr_sse(xmm_index_mask.x, next_input, xmm_shuffle_input.x,
473 xmm_ones_16.x, *indices1, indices2);
475 /* Gather 64 bit transitions and pack 2 per register. */
477 t = trans[MM_CVT32(addr)];
480 addr = MM_SHUFFLE32(addr, SHUFFLE32_SLOT1);
481 *indices1 = MM_SET64(trans[MM_CVT32(addr)], t);
483 return MM_SRL32(next_input, CHAR_BIT);
487 * Execute trie traversal with 2 traversals in parallel.
490 search_sse_2(const struct rte_acl_ctx *ctx, const uint8_t **data,
491 uint32_t *results, uint32_t total_packets, uint32_t categories)
494 struct acl_flow_data flows;
495 uint64_t index_array[MAX_SEARCHES_SSE2];
496 struct completion cmplt[MAX_SEARCHES_SSE2];
497 struct parms parms[MAX_SEARCHES_SSE2];
498 xmm_t input, indices;
500 acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
501 total_packets, categories, ctx->trans_table);
503 for (n = 0; n < MAX_SEARCHES_SSE2; n++) {
505 index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
508 indices = MM_LOADU((xmm_t *) &index_array[0]);
510 /* Check for any matches. */
511 acl_match_check_x2(0, ctx, parms, &flows, &indices,
514 while (flows.started > 0) {
516 /* Gather 4 bytes of input data for each stream. */
517 input = _mm_cvtsi32_si128(GET_NEXT_4BYTES(parms, 0));
518 input = MM_INSERT32(input, GET_NEXT_4BYTES(parms, 1), 1);
520 /* Process the 4 bytes of input on each stream. */
522 input = transition2(input, flows.trans, &indices);
523 input = transition2(input, flows.trans, &indices);
524 input = transition2(input, flows.trans, &indices);
525 input = transition2(input, flows.trans, &indices);
527 /* Check for any matches. */
528 acl_match_check_x2(0, ctx, parms, &flows, &indices,