app/test-compress-perf/comp_perf_test_benchmark.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <rte_malloc.h>
   6 #include <rte_eal.h>
   7 #include <rte_log.h>
   8 #include <rte_cycles.h>
   9 #include <rte_compressdev.h>
  10
  11 #include "comp_perf_test_benchmark.h"
  12
  13 void
  14 cperf_benchmark_test_destructor(void *arg)
  15 {
  16         if (arg) {
  17                 comp_perf_free_memory(
  18                                 &((struct cperf_benchmark_ctx *)arg)->ver.mem);
  19                 rte_free(arg);
  20         }
  21 }
  22
  23 void *
  24 cperf_benchmark_test_constructor(uint8_t dev_id, uint16_t qp_id,
  25                 struct comp_test_data *options)
  26 {
  27         struct cperf_benchmark_ctx *ctx = NULL;
  28
  29         ctx = rte_malloc(NULL, sizeof(struct cperf_benchmark_ctx), 0);
  30
  31         if (ctx == NULL)
  32                 return NULL;
  33
  34         ctx->ver.mem.dev_id = dev_id;
  35         ctx->ver.mem.qp_id = qp_id;
  36         ctx->ver.options = options;
  37         ctx->ver.silent = 1; /* ver. part will be silent */
  38
  39         if (!comp_perf_allocate_memory(ctx->ver.options, &ctx->ver.mem)
  40                         && !prepare_bufs(ctx->ver.options, &ctx->ver.mem))
  41                 return ctx;
  42
  43         cperf_benchmark_test_destructor(ctx);
  44         return NULL;
  45 }
  46
  47 static int
  48 main_loop(struct cperf_benchmark_ctx *ctx, enum rte_comp_xform_type type)
  49 {
  50         struct comp_test_data *test_data = ctx->ver.options;
  51         struct cperf_mem_resources *mem = &ctx->ver.mem;
  52         uint8_t dev_id = mem->dev_id;
  53         uint32_t i, iter, num_iter;
  54         struct rte_comp_op **ops, **deq_ops;
  55         void *priv_xform = NULL;
  56         struct rte_comp_xform xform;
  57         struct rte_mbuf **input_bufs, **output_bufs;
  58         int res = 0;
  59         int allocated = 0;
  60         uint32_t out_seg_sz;
  61
  62         if (test_data == NULL || !test_data->burst_sz) {
  63                 RTE_LOG(ERR, USER1,
  64                         "Unknown burst size\n");
  65                 return -1;
  66         }
  67
  68         ops = rte_zmalloc_socket(NULL,
  69                 2 * mem->total_bufs * sizeof(struct rte_comp_op *),
  70                 0, rte_socket_id());
  71
  72         if (ops == NULL) {
  73                 RTE_LOG(ERR, USER1,
  74                         "Can't allocate memory for ops strucures\n");
  75                 return -1;
  76         }
  77
  78         deq_ops = &ops[mem->total_bufs];
  79
  80         if (type == RTE_COMP_COMPRESS) {
  81                 xform = (struct rte_comp_xform) {
  82                         .type = RTE_COMP_COMPRESS,
  83                         .compress = {
  84                                 .algo = RTE_COMP_ALGO_DEFLATE,
  85                                 .deflate.huffman = test_data->huffman_enc,
  86                                 .level = test_data->level,
  87                                 .window_size = test_data->window_sz,
  88                                 .chksum = RTE_COMP_CHECKSUM_NONE,
  89                                 .hash_algo = RTE_COMP_HASH_ALGO_NONE
  90                         }
  91                 };
  92                 input_bufs = mem->decomp_bufs;
  93                 output_bufs = mem->comp_bufs;
  94                 out_seg_sz = test_data->out_seg_sz;
  95         } else {
  96                 xform = (struct rte_comp_xform) {
  97                         .type = RTE_COMP_DECOMPRESS,
  98                         .decompress = {
  99                                 .algo = RTE_COMP_ALGO_DEFLATE,
 100                                 .chksum = RTE_COMP_CHECKSUM_NONE,
 101                                 .window_size = test_data->window_sz,
 102                                 .hash_algo = RTE_COMP_HASH_ALGO_NONE
 103                         }
 104                 };
 105                 input_bufs = mem->comp_bufs;
 106                 output_bufs = mem->decomp_bufs;
 107                 out_seg_sz = test_data->seg_sz;
 108         }
 109
 110         /* Create private xform */
 111         if (rte_compressdev_private_xform_create(dev_id, &xform,
 112                         &priv_xform) < 0) {
 113                 RTE_LOG(ERR, USER1, "Private xform could not be created\n");
 114                 res = -1;
 115                 goto end;
 116         }
 117
 118         uint64_t tsc_start, tsc_end, tsc_duration;
 119
 120         num_iter = test_data->num_iter;
 121         tsc_start = tsc_end = tsc_duration = 0;
 122         tsc_start = rte_rdtsc_precise();
 123
 124         for (iter = 0; iter < num_iter; iter++) {
 125                 uint32_t total_ops = mem->total_bufs;
 126                 uint32_t remaining_ops = mem->total_bufs;
 127                 uint32_t total_deq_ops = 0;
 128                 uint32_t total_enq_ops = 0;
 129                 uint16_t ops_unused = 0;
 130                 uint16_t num_enq = 0;
 131                 uint16_t num_deq = 0;
 132
 133                 while (remaining_ops > 0) {
 134                         uint16_t num_ops = RTE_MIN(remaining_ops,
 135                                                    test_data->burst_sz);
 136                         uint16_t ops_needed = num_ops - ops_unused;
 137
 138                         /*
 139                          * Move the unused operations from the previous
 140                          * enqueue_burst call to the front, to maintain order
 141                          */
 142                         if ((ops_unused > 0) && (num_enq > 0)) {
 143                                 size_t nb_b_to_mov =
 144                                       ops_unused * sizeof(struct rte_comp_op *);
 145
 146                                 memmove(ops, &ops[num_enq], nb_b_to_mov);
 147                         }
 148
 149                         /* Allocate compression operations */
 150                         if (ops_needed && !rte_comp_op_bulk_alloc(
 151                                                 mem->op_pool,
 152                                                 &ops[ops_unused],
 153                                                 ops_needed)) {
 154                                 RTE_LOG(ERR, USER1,
 155                                       "Could not allocate enough operations\n");
 156                                 res = -1;
 157                                 goto end;
 158                         }
 159                         allocated += ops_needed;
 160
 161                         for (i = 0; i < ops_needed; i++) {
 162                                 /*
 163                                  * Calculate next buffer to attach to operation
 164                                  */
 165                                 uint32_t buf_id = total_enq_ops + i +
 166                                                 ops_unused;
 167                                 uint16_t op_id = ops_unused + i;
 168                                 /* Reset all data in output buffers */
 169                                 struct rte_mbuf *m = output_bufs[buf_id];
 170
 171                                 m->pkt_len = out_seg_sz * m->nb_segs;
 172                                 while (m) {
 173                                         m->data_len = m->buf_len - m->data_off;
 174                                         m = m->next;
 175                                 }
 176                                 ops[op_id]->m_src = input_bufs[buf_id];
 177                                 ops[op_id]->m_dst = output_bufs[buf_id];
 178                                 ops[op_id]->src.offset = 0;
 179                                 ops[op_id]->src.length =
 180                                         rte_pktmbuf_pkt_len(input_bufs[buf_id]);
 181                                 ops[op_id]->dst.offset = 0;
 182                                 ops[op_id]->flush_flag = RTE_COMP_FLUSH_FINAL;
 183                                 ops[op_id]->input_chksum = buf_id;
 184                                 ops[op_id]->private_xform = priv_xform;
 185                         }
 186
 187                         num_enq = rte_compressdev_enqueue_burst(dev_id,
 188                                                                 mem->qp_id, ops,
 189                                                                 num_ops);
 190                         if (num_enq == 0) {
 191                                 struct rte_compressdev_stats stats;
 192
 193                                 rte_compressdev_stats_get(dev_id, &stats);
 194                                 if (stats.enqueue_err_count) {
 195                                         res = -1;
 196                                         goto end;
 197                                 }
 198                         }
 199
 200                         ops_unused = num_ops - num_enq;
 201                         remaining_ops -= num_enq;
 202                         total_enq_ops += num_enq;
 203
 204                         num_deq = rte_compressdev_dequeue_burst(dev_id,
 205                                                            mem->qp_id,
 206                                                            deq_ops,
 207                                                            test_data->burst_sz);
 208                         total_deq_ops += num_deq;
 209
 210                         if (iter == num_iter - 1) {
 211                                 for (i = 0; i < num_deq; i++) {
 212                                         struct rte_comp_op *op = deq_ops[i];
 213
 214                                         if (op->status !=
 215                                                 RTE_COMP_OP_STATUS_SUCCESS) {
 216                                                 RTE_LOG(ERR, USER1,
 217                                        "Some operations were not successful\n");
 218                                                 goto end;
 219                                         }
 220
 221                                         struct rte_mbuf *m = op->m_dst;
 222
 223                                         m->pkt_len = op->produced;
 224                                         uint32_t remaining_data = op->produced;
 225                                         uint16_t data_to_append;
 226
 227                                         while (remaining_data > 0) {
 228                                                 data_to_append =
 229                                                         RTE_MIN(remaining_data,
 230                                                              out_seg_sz);
 231                                                 m->data_len = data_to_append;
 232                                                 remaining_data -=
 233                                                                 data_to_append;
 234                                                 m = m->next;
 235                                         }
 236                                 }
 237                         }
 238                         rte_mempool_put_bulk(mem->op_pool,
 239                                              (void **)deq_ops, num_deq);
 240                         allocated -= num_deq;
 241                 }
 242
 243                 /* Dequeue the last operations */
 244                 while (total_deq_ops < total_ops) {
 245                         num_deq = rte_compressdev_dequeue_burst(dev_id,
 246                                                            mem->qp_id,
 247                                                            deq_ops,
 248                                                            test_data->burst_sz);
 249                         if (num_deq == 0) {
 250                                 struct rte_compressdev_stats stats;
 251
 252                                 rte_compressdev_stats_get(dev_id, &stats);
 253                                 if (stats.dequeue_err_count) {
 254                                         res = -1;
 255                                         goto end;
 256                                 }
 257                         }
 258
 259                         total_deq_ops += num_deq;
 260
 261                         if (iter == num_iter - 1) {
 262                                 for (i = 0; i < num_deq; i++) {
 263                                         struct rte_comp_op *op = deq_ops[i];
 264
 265                                         if (op->status !=
 266                                                 RTE_COMP_OP_STATUS_SUCCESS) {
 267                                                 RTE_LOG(ERR, USER1,
 268                                        "Some operations were not successful\n");
 269                                                 goto end;
 270                                         }
 271
 272                                         struct rte_mbuf *m = op->m_dst;
 273
 274                                         m->pkt_len = op->produced;
 275                                         uint32_t remaining_data = op->produced;
 276                                         uint16_t data_to_append;
 277
 278                                         while (remaining_data > 0) {
 279                                                 data_to_append =
 280                                                 RTE_MIN(remaining_data,
 281                                                         out_seg_sz);
 282                                                 m->data_len = data_to_append;
 283                                                 remaining_data -=
 284                                                                 data_to_append;
 285                                                 m = m->next;
 286                                         }
 287                                 }
 288                         }
 289                         rte_mempool_put_bulk(mem->op_pool,
 290                                              (void **)deq_ops, num_deq);
 291                         allocated -= num_deq;
 292                 }
 293         }
 294
 295         tsc_end = rte_rdtsc_precise();
 296         tsc_duration = tsc_end - tsc_start;
 297
 298         if (type == RTE_COMP_COMPRESS)
 299                 ctx->comp_tsc_duration[test_data->level] =
 300                                 tsc_duration / num_iter;
 301         else
 302                 ctx->decomp_tsc_duration[test_data->level] =
 303                                 tsc_duration / num_iter;
 304
 305 end:
 306         rte_mempool_put_bulk(mem->op_pool, (void **)ops, allocated);
 307         rte_compressdev_private_xform_free(dev_id, priv_xform);
 308         rte_free(ops);
 309         return res;
 310 }
 311
 312 int
 313 cperf_benchmark_test_runner(void *test_ctx)
 314 {
 315         struct cperf_benchmark_ctx *ctx = test_ctx;
 316         struct comp_test_data *test_data = ctx->ver.options;
 317         uint32_t lcore = rte_lcore_id();
 318         static rte_atomic16_t display_once = RTE_ATOMIC16_INIT(0);
 319
 320         ctx->ver.mem.lcore_id = lcore;
 321         int i, ret = EXIT_SUCCESS;
 322
 323         /*
 324          * First the verification part is needed
 325          */
 326         if (cperf_verify_test_runner(&ctx->ver)) {
 327                 ret =  EXIT_FAILURE;
 328                 goto end;
 329         }
 330
 331         /*
 332          * Run the tests twice, discarding the first performance
 333          * results, before the cache is warmed up
 334          */
 335         for (i = 0; i < 2; i++) {
 336                 if (main_loop(ctx, RTE_COMP_COMPRESS) < 0) {
 337                         ret = EXIT_FAILURE;
 338                         goto end;
 339                 }
 340         }
 341
 342         for (i = 0; i < 2; i++) {
 343                 if (main_loop(ctx, RTE_COMP_DECOMPRESS) < 0) {
 344                         ret = EXIT_FAILURE;
 345                         goto end;
 346                 }
 347         }
 348
 349         ctx->comp_tsc_byte =
 350                         (double)(ctx->comp_tsc_duration[test_data->level]) /
 351                                         test_data->input_data_sz;
 352
 353         ctx->decomp_tsc_byte =
 354                         (double)(ctx->decomp_tsc_duration[test_data->level]) /
 355                                         test_data->input_data_sz;
 356
 357         ctx->comp_gbps = rte_get_tsc_hz() / ctx->comp_tsc_byte * 8 /
 358                         1000000000;
 359
 360         ctx->decomp_gbps = rte_get_tsc_hz() / ctx->decomp_tsc_byte * 8 /
 361                         1000000000;
 362
 363         if (rte_atomic16_test_and_set(&display_once)) {
 364                 printf("%12s%6s%12s%17s%15s%16s\n",
 365                         "lcore id", "Level", "Comp size", "Comp ratio [%]",
 366                         "Comp [Gbps]", "Decomp [Gbps]");
 367         }
 368
 369         printf("%12u%6u%12zu%17.2f%15.2f%16.2f\n",
 370                 ctx->ver.mem.lcore_id,
 371                 test_data->level, ctx->ver.comp_data_sz, ctx->ver.ratio,
 372                 ctx->comp_gbps,
 373                 ctx->decomp_gbps);
 374
 375 end:
 376         return ret;
 377 }