xref: /dpdk/lib/acl/acl_run_altivec.h (revision e9fd1ebf981f361844aea9ec94e17f4bda5e1479)
1 /*
2  * SPDX-License-Identifier: BSD-3-Clause
3  * Copyright (C) IBM Corporation 2016.
4  */
5 
6 #include <stdalign.h>
7 
8 #include "acl_run.h"
9 #include "acl_vect.h"
10 
11 alignas(RTE_CACHE_LINE_SIZE) struct _altivec_acl_const {
12 	rte_xmm_t xmm_shuffle_input;
13 	rte_xmm_t xmm_index_mask;
14 	rte_xmm_t xmm_ones_16;
15 	rte_xmm_t range_base;
16 } altivec_acl_const = {
17 	{
18 		.u32 = {0x00000000, 0x04040404, 0x08080808, 0x0c0c0c0c}
19 	},
20 	{
21 		.u32 = {RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX,
22 		RTE_ACL_NODE_INDEX, RTE_ACL_NODE_INDEX}
23 	},
24 	{
25 		.u16 = {1, 1, 1, 1, 1, 1, 1, 1}
26 	},
27 	{
28 		.u32 = {0xffffff00, 0xffffff04, 0xffffff08, 0xffffff0c}
29 	},
30 };
31 
32 /*
33  * Resolve priority for multiple results (altivec version).
34  * This consists comparing the priority of the current traversal with the
35  * running set of results for the packet.
36  * For each result, keep a running array of the result (rule number) and
37  * its priority for each category.
38  */
39 static inline void
resolve_priority_altivec(uint64_t transition,int n,const struct rte_acl_ctx * ctx,struct parms * parms,const struct rte_acl_match_results * p,uint32_t categories)40 resolve_priority_altivec(uint64_t transition, int n,
41 	const struct rte_acl_ctx *ctx, struct parms *parms,
42 	const struct rte_acl_match_results *p, uint32_t categories)
43 {
44 	uint32_t x;
45 	xmm_t results, priority, results1, priority1;
46 	__vector __bool int selector;
47 	xmm_t *saved_results, *saved_priority;
48 
49 	for (x = 0; x < categories; x += RTE_ACL_RESULTS_MULTIPLIER) {
50 
51 		saved_results = (xmm_t *)(&parms[n].cmplt->results[x]);
52 		saved_priority =
53 			(xmm_t *)(&parms[n].cmplt->priority[x]);
54 
55 		/* get results and priorities for completed trie */
56 		results = *(const xmm_t *)&p[transition].results[x];
57 		priority = *(const xmm_t *)&p[transition].priority[x];
58 
59 		/* if this is not the first completed trie */
60 		if (parms[n].cmplt->count != ctx->num_tries) {
61 
62 			/* get running best results and their priorities */
63 			results1 = *saved_results;
64 			priority1 = *saved_priority;
65 
66 			/* select results that are highest priority */
67 			selector = vec_cmpgt(priority1, priority);
68 			results = vec_sel(results, results1, selector);
69 			priority = vec_sel(priority, priority1,
70 				selector);
71 		}
72 
73 		/* save running best results and their priorities */
74 		*saved_results = results;
75 		*saved_priority = priority;
76 	}
77 }
78 
79 /*
80  * Check for any match in 4 transitions
81  */
82 static __rte_always_inline uint32_t
check_any_match_x4(uint64_t val[])83 check_any_match_x4(uint64_t val[])
84 {
85 	return (val[0] | val[1] | val[2] | val[3]) & RTE_ACL_NODE_MATCH;
86 }
87 
88 static __rte_always_inline void
acl_match_check_x4(int slot,const struct rte_acl_ctx * ctx,struct parms * parms,struct acl_flow_data * flows,uint64_t transitions[])89 acl_match_check_x4(int slot, const struct rte_acl_ctx *ctx, struct parms *parms,
90 	struct acl_flow_data *flows, uint64_t transitions[])
91 {
92 	while (check_any_match_x4(transitions)) {
93 		transitions[0] = acl_match_check(transitions[0], slot, ctx,
94 			parms, flows, resolve_priority_altivec);
95 		transitions[1] = acl_match_check(transitions[1], slot + 1, ctx,
96 			parms, flows, resolve_priority_altivec);
97 		transitions[2] = acl_match_check(transitions[2], slot + 2, ctx,
98 			parms, flows, resolve_priority_altivec);
99 		transitions[3] = acl_match_check(transitions[3], slot + 3, ctx,
100 			parms, flows, resolve_priority_altivec);
101 	}
102 }
103 
104 /*
105  * Process 4 transitions (in 2 XMM registers) in parallel
106  */
107 static __rte_always_inline xmm_t
transition4(xmm_t next_input,const uint64_t * trans,xmm_t * indices1,xmm_t * indices2)108 transition4(xmm_t next_input, const uint64_t *trans,
109 	xmm_t *indices1, xmm_t *indices2)
110 {
111 	xmm_t addr, tr_lo, tr_hi;
112 	xmm_t in, node_type, r, t;
113 	xmm_t dfa_ofs, quad_ofs;
114 	xmm_t *index_mask, *tp;
115 	__vector __bool int dfa_msk;
116 	__vector signed char zeroes = {};
117 	union {
118 		uint64_t d64[2];
119 		uint32_t d32[4];
120 	} v;
121 
122 	/* Move low 32 into tr_lo and high 32 into tr_hi */
123 	tr_lo = (xmm_t){(*indices1)[0], (*indices1)[2],
124 			(*indices2)[0], (*indices2)[2]};
125 	tr_hi = (xmm_t){(*indices1)[1], (*indices1)[3],
126 			(*indices2)[1], (*indices2)[3]};
127 
128 	 /* Calculate the address (array index) for all 4 transitions. */
129 	index_mask = (xmm_t *)&altivec_acl_const.xmm_index_mask.u32;
130 	t = vec_xor(*index_mask, *index_mask);
131 	in = vec_perm(next_input, (xmm_t){},
132 		*(__vector unsigned char *)&altivec_acl_const.xmm_shuffle_input);
133 
134 	/* Calc node type and node addr */
135 	node_type = vec_and(vec_nor(*index_mask, *index_mask), tr_lo);
136 	addr = vec_and(tr_lo, *index_mask);
137 
138 	/* mask for DFA type(0) nodes */
139 	dfa_msk = vec_cmpeq(node_type, t);
140 
141 	/* DFA calculations. */
142 	r = vec_sr(in, (__vector unsigned int){30, 30, 30, 30});
143 	tp = (xmm_t *)&altivec_acl_const.range_base.u32;
144 	r = vec_add(r, *tp);
145 	t = vec_sr(in, (__vector unsigned int){24, 24, 24, 24});
146 	r = vec_perm(tr_hi, (xmm_t){(uint16_t)0 << 16},
147 		(__vector unsigned char)r);
148 
149 	dfa_ofs = vec_sub(t, r);
150 
151 	/* QUAD/SINGLE calculations. */
152 	t = (xmm_t)vec_cmpgt((__vector signed char)in, (__vector signed char)tr_hi);
153 	t = (xmm_t)vec_sel(
154 		vec_sel(
155 			(__vector signed char)vec_sub(
156 				zeroes, (__vector signed char)t),
157 			(__vector signed char)t,
158 			vec_cmpgt((__vector signed char)t, zeroes)),
159 		zeroes,
160 		vec_cmpeq((__vector signed char)t, zeroes));
161 
162 	t = (xmm_t)vec_msum((__vector signed char)t,
163 		(__vector unsigned char)t, (xmm_t){});
164 	quad_ofs = (xmm_t)vec_msum((__vector signed short)t,
165 		*(__vector signed short *)&altivec_acl_const.xmm_ones_16.u16,
166 		(xmm_t){});
167 
168 	/* blend DFA and QUAD/SINGLE. */
169 	t = vec_sel(quad_ofs, dfa_ofs, dfa_msk);
170 
171 	/* calculate address for next transitions. */
172 	addr = vec_add(addr, t);
173 
174 	v.d64[0] = (uint64_t)trans[addr[0]];
175 	v.d64[1] = (uint64_t)trans[addr[1]];
176 	*indices1 = (xmm_t){v.d32[0], v.d32[1], v.d32[2], v.d32[3]};
177 	v.d64[0] = (uint64_t)trans[addr[2]];
178 	v.d64[1] = (uint64_t)trans[addr[3]];
179 	*indices2 = (xmm_t){v.d32[0], v.d32[1], v.d32[2], v.d32[3]};
180 
181 	return vec_sr(next_input,
182 		(__vector unsigned int){CHAR_BIT, CHAR_BIT, CHAR_BIT, CHAR_BIT});
183 }
184 
185 /*
186  * Execute trie traversal with 8 traversals in parallel
187  */
188 static inline int
search_altivec_8(const struct rte_acl_ctx * ctx,const uint8_t ** data,uint32_t * results,uint32_t total_packets,uint32_t categories)189 search_altivec_8(const struct rte_acl_ctx *ctx, const uint8_t **data,
190 	uint32_t *results, uint32_t total_packets, uint32_t categories)
191 {
192 	int n;
193 	struct acl_flow_data flows;
194 	uint64_t index_array[MAX_SEARCHES_ALTIVEC8];
195 	struct completion cmplt[MAX_SEARCHES_ALTIVEC8];
196 	struct parms parms[MAX_SEARCHES_ALTIVEC8];
197 	xmm_t input0, input1;
198 
199 	acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
200 		total_packets, categories, ctx->trans_table);
201 
202 	for (n = 0; n < MAX_SEARCHES_ALTIVEC8; n++) {
203 		cmplt[n].count = 0;
204 		index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
205 	}
206 
207 	 /* Check for any matches. */
208 	acl_match_check_x4(0, ctx, parms, &flows, (uint64_t *)&index_array[0]);
209 	acl_match_check_x4(4, ctx, parms, &flows, (uint64_t *)&index_array[4]);
210 
211 	while (flows.started > 0) {
212 
213 		/* Gather 4 bytes of input data for each stream. */
214 		input0 = (xmm_t){GET_NEXT_4BYTES(parms, 0),
215 				GET_NEXT_4BYTES(parms, 1),
216 				GET_NEXT_4BYTES(parms, 2),
217 				GET_NEXT_4BYTES(parms, 3)};
218 
219 		input1 = (xmm_t){GET_NEXT_4BYTES(parms, 4),
220 				GET_NEXT_4BYTES(parms, 5),
221 				GET_NEXT_4BYTES(parms, 6),
222 				GET_NEXT_4BYTES(parms, 7)};
223 
224 		 /* Process the 4 bytes of input on each stream. */
225 
226 		input0 = transition4(input0, flows.trans,
227 			(xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
228 		input1 = transition4(input1, flows.trans,
229 			(xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
230 
231 		input0 = transition4(input0, flows.trans,
232 			(xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
233 		input1 = transition4(input1, flows.trans,
234 			(xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
235 
236 		input0 = transition4(input0, flows.trans,
237 			(xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
238 		input1 = transition4(input1, flows.trans,
239 			(xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
240 
241 		input0 = transition4(input0, flows.trans,
242 			(xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
243 		input1 = transition4(input1, flows.trans,
244 			(xmm_t *)&index_array[4], (xmm_t *)&index_array[6]);
245 
246 		 /* Check for any matches. */
247 		acl_match_check_x4(0, ctx, parms, &flows,
248 			(uint64_t *)&index_array[0]);
249 		acl_match_check_x4(4, ctx, parms, &flows,
250 			(uint64_t *)&index_array[4]);
251 	}
252 
253 	return 0;
254 }
255 
256 /*
257  * Execute trie traversal with 4 traversals in parallel
258  */
259 static inline int
search_altivec_4(const struct rte_acl_ctx * ctx,const uint8_t ** data,uint32_t * results,int total_packets,uint32_t categories)260 search_altivec_4(const struct rte_acl_ctx *ctx, const uint8_t **data,
261 	 uint32_t *results, int total_packets, uint32_t categories)
262 {
263 	int n;
264 	struct acl_flow_data flows;
265 	uint64_t index_array[MAX_SEARCHES_ALTIVEC4];
266 	struct completion cmplt[MAX_SEARCHES_ALTIVEC4];
267 	struct parms parms[MAX_SEARCHES_ALTIVEC4];
268 	xmm_t input;
269 
270 	acl_set_flow(&flows, cmplt, RTE_DIM(cmplt), data, results,
271 		total_packets, categories, ctx->trans_table);
272 
273 	for (n = 0; n < MAX_SEARCHES_ALTIVEC4; n++) {
274 		cmplt[n].count = 0;
275 		index_array[n] = acl_start_next_trie(&flows, parms, n, ctx);
276 	}
277 
278 	/* Check for any matches. */
279 	acl_match_check_x4(0, ctx, parms, &flows, index_array);
280 
281 	while (flows.started > 0) {
282 
283 		/* Gather 4 bytes of input data for each stream. */
284 		input = (xmm_t){GET_NEXT_4BYTES(parms, 0),
285 				GET_NEXT_4BYTES(parms, 1),
286 				GET_NEXT_4BYTES(parms, 2),
287 				GET_NEXT_4BYTES(parms, 3)};
288 
289 		/* Process the 4 bytes of input on each stream. */
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 		input = transition4(input, flows.trans,
297 			(xmm_t *)&index_array[0], (xmm_t *)&index_array[2]);
298 
299 		/* Check for any matches. */
300 		acl_match_check_x4(0, ctx, parms, &flows, index_array);
301 	}
302 
303 	return 0;
304 }
305