tests/gem_pwrite_pread.c

   1 /*
   2  * Copyright © 2011 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Chris Wilson <chris@chris-wilson.co.uk>
  25  *
  26  */
  27
  28 #include <unistd.h>
  29 #include <stdlib.h>
  30 #include <stdint.h>
  31 #include <stdio.h>
  32 #include <string.h>
  33 #include <fcntl.h>
  34 #include <inttypes.h>
  35 #include <errno.h>
  36 #include <sys/stat.h>
  37 #include <sys/ioctl.h>
  38 #include <sys/time.h>
  39 #include "drm.h"
  40 #include "i915_drm.h"
  41 #include "drmtest.h"
  42 #include "intel_bufmgr.h"
  43 #include "intel_batchbuffer.h"
  44 #include "intel_gpu_tools.h"
  45
  46 #define OBJECT_SIZE 16384
  47
  48 #define COPY_BLT_CMD            (2<<29|0x53<<22|0x6)
  49 #define BLT_WRITE_ALPHA         (1<<21)
  50 #define BLT_WRITE_RGB           (1<<20)
  51 #define BLT_SRC_TILED           (1<<15)
  52 #define BLT_DST_TILED           (1<<11)
  53
  54 uint32_t devid;
  55
  56 static inline void build_batch(uint32_t *batch, int len, uint32_t *batch_len)
  57 {
  58         unsigned int i = 0;
  59
  60         batch[i++] = COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB;
  61         batch[i++] = 0xcc << 16 | 1 << 25 | 1 << 24 | len;
  62         batch[i++] = 0;
  63         batch[i++] = 1 << 16 | (len / 4);
  64         batch[i++] = 0; /* dst */
  65         batch[i++] = 0;
  66         batch[i++] = len;
  67         batch[i++] = 0; /* src */
  68         batch[i++] = MI_BATCH_BUFFER_END;
  69         batch[i++] = 0;
  70
  71         *batch_len = i * 4;
  72 }
  73
  74 #define GPP_BATCH_SIZE (10 * 4)
  75
  76 static void copy(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
  77 {
  78         uint32_t batch[GPP_BATCH_SIZE] = {0};
  79
  80         struct drm_i915_gem_relocation_entry reloc[] = {
  81                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
  82                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
  83         };
  84         struct drm_i915_gem_exec_object2 exec[] = {
  85                 { src },
  86                 { dst },
  87                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
  88         };
  89         struct drm_i915_gem_execbuffer2 execbuf = {
  90                 (uintptr_t)exec, 3,
  91                 0, 0,
  92                 0, 0, 0, 0,
  93                 HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0,
  94         };
  95
  96         build_batch(batch, len, &execbuf.batch_len);
  97
  98         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
  99
 100         while (loops--) {
 101                 gem_write(fd, src, 0, buf, len);
 102                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 103                 gem_read(fd, dst, 0, buf, len);
 104         }
 105
 106         gem_close(fd, exec[2].handle);
 107 }
 108
 109 static void as_gtt_mmap(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
 110 {
 111         uint32_t batch[GPP_BATCH_SIZE] = {0};
 112
 113         struct drm_i915_gem_relocation_entry reloc[] = {
 114                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 115                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 116         };
 117         struct drm_i915_gem_exec_object2 exec[] = {
 118                 { src },
 119                 { dst },
 120                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 121         };
 122         struct drm_i915_gem_execbuffer2 execbuf = {
 123                 (uintptr_t)exec, 3,
 124                 0, GPP_BATCH_SIZE,
 125                 0, 0, 0, 0,
 126                 HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0,
 127         };
 128         uint32_t *src_ptr, *dst_ptr;
 129
 130         build_batch(batch, len, &execbuf.batch_len);
 131
 132         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 133
 134         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 135         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 136
 137         while (loops--) {
 138                 gem_set_domain(fd, src,
 139                                I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 140                 memcpy(src_ptr, buf, len);
 141
 142                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 143                 gem_set_domain(fd, dst,
 144                                I915_GEM_DOMAIN_GTT, 0);
 145                 memcpy(buf, dst_ptr, len);
 146         }
 147
 148         munmap(dst_ptr, len);
 149         munmap(src_ptr, len);
 150         gem_close(fd, exec[2].handle);
 151 }
 152
 153
 154 static void as_cpu_mmap(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
 155 {
 156         uint32_t batch[GPP_BATCH_SIZE] = {0};
 157
 158         struct drm_i915_gem_relocation_entry reloc[] = {
 159                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 160                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 161         };
 162         struct drm_i915_gem_exec_object2 exec[] = {
 163                 { src },
 164                 { dst },
 165                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 166         };
 167         struct drm_i915_gem_execbuffer2 execbuf = {
 168                 (uintptr_t)exec, 3,
 169                 0, GPP_BATCH_SIZE,
 170                 0, 0, 0, 0,
 171                 HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0,
 172         };
 173         uint32_t *src_ptr, *dst_ptr;
 174
 175         build_batch(batch, len, &execbuf.batch_len);
 176
 177         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 178
 179         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 180         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 181
 182         while (loops--) {
 183                 gem_set_domain(fd, src,
 184                                I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 185                 memcpy(src_ptr, buf, len);
 186
 187                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 188                 gem_set_domain(fd, dst,
 189                                I915_GEM_DOMAIN_CPU, 0);
 190                 memcpy(buf, dst_ptr, len);
 191         }
 192
 193         munmap(dst_ptr, len);
 194         munmap(src_ptr, len);
 195         gem_close(fd, exec[2].handle);
 196 }
 197
 198 static void test_copy(int fd, uint32_t src, uint32_t dst, uint32_t *buf, int len)
 199 {
 200         uint32_t batch[GPP_BATCH_SIZE] = {0};
 201
 202         struct drm_i915_gem_relocation_entry reloc[] = {
 203                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 204                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 205         };
 206         struct drm_i915_gem_exec_object2 exec[] = {
 207                 { src },
 208                 { dst },
 209                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 210         };
 211         struct drm_i915_gem_execbuffer2 execbuf = {
 212                 (uintptr_t)exec, 3,
 213                 0, GPP_BATCH_SIZE,
 214                 0, 0, 0, 0,
 215                 HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0,
 216         };
 217         int i;
 218
 219         build_batch(batch, len, &execbuf.batch_len);
 220
 221         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 222
 223         for (i = 0; i < len/4; i++)
 224                 buf[i] = i;
 225
 226         gem_write(fd, src, 0, buf, len);
 227         memset(buf, 0, len);
 228
 229         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 230         gem_read(fd, dst, 0, buf, len);
 231
 232         gem_close(fd, exec[2].handle);
 233
 234         for (i = 0; i < len/4; i++)
 235                 igt_assert(buf[i] == i);
 236 }
 237
 238 static void test_as_gtt_mmap(int fd, uint32_t src, uint32_t dst, int len)
 239 {
 240         uint32_t batch[GPP_BATCH_SIZE] = {0};
 241
 242         struct drm_i915_gem_relocation_entry reloc[] = {
 243                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 244                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 245         };
 246         struct drm_i915_gem_exec_object2 exec[] = {
 247                 { src },
 248                 { dst },
 249                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 250         };
 251         struct drm_i915_gem_execbuffer2 execbuf = {
 252                 (uintptr_t)exec, 3,
 253                 0, GPP_BATCH_SIZE,
 254                 0, 0, 0, 0,
 255                 HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0,
 256         };
 257         uint32_t *src_ptr, *dst_ptr;
 258         int i;
 259
 260         build_batch(batch, len, &execbuf.batch_len);
 261
 262         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 263
 264         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 265         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 266
 267         gem_set_domain(fd, src, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 268         for (i = 0; i < len/4; i++)
 269                 src_ptr[i] = i;
 270
 271         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 272         gem_close(fd, exec[2].handle);
 273
 274         gem_set_domain(fd, dst, I915_GEM_DOMAIN_GTT, 0);
 275         for (i = 0; i < len/4; i++)
 276                 igt_assert(dst_ptr[i] == i);
 277
 278         munmap(dst_ptr, len);
 279         munmap(src_ptr, len);
 280 }
 281
 282 static void test_as_cpu_mmap(int fd, uint32_t src, uint32_t dst, int len)
 283 {
 284         uint32_t batch[GPP_BATCH_SIZE] = {0};
 285
 286         struct drm_i915_gem_relocation_entry reloc[] = {
 287                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 288                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 289         };
 290         struct drm_i915_gem_exec_object2 exec[] = {
 291                 { src },
 292                 { dst },
 293                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 294         };
 295         struct drm_i915_gem_execbuffer2 execbuf = {
 296                 (uintptr_t)exec, 3,
 297                 0, GPP_BATCH_SIZE,
 298                 0, 0, 0, 0,
 299                 HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0,
 300         };
 301         uint32_t *src_ptr, *dst_ptr;
 302         int i;
 303
 304         build_batch(batch, len, &execbuf.batch_len);
 305
 306         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 307
 308         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 309         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 310
 311         gem_set_domain(fd, src, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 312         for (i = 0; i < len/4; i++)
 313                 src_ptr[i] = i;
 314
 315         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 316         gem_close(fd, exec[2].handle);
 317
 318         gem_set_domain(fd, dst, I915_GEM_DOMAIN_CPU, 0);
 319         for (i = 0; i < len/4; i++)
 320                 igt_assert(dst_ptr[i] == i);
 321
 322         munmap(dst_ptr, len);
 323         munmap(src_ptr, len);
 324 }
 325
 326 static double elapsed(const struct timeval *start,
 327                       const struct timeval *end,
 328                       int loop)
 329 {
 330         return (1e6*(end->tv_sec - start->tv_sec) + (end->tv_usec - start->tv_usec))/loop;
 331 }
 332
 333 static const char *bytes_per_sec(char *buf, double v)
 334 {
 335         const char *order[] = {
 336                 "",
 337                 "KiB",
 338                 "MiB",
 339                 "GiB",
 340                 "TiB",
 341                 NULL,
 342         }, **o = order;
 343
 344         while (v > 1000 && o[1]) {
 345                 v /= 1000;
 346                 o++;
 347         }
 348         sprintf(buf, "%.1f%s/s", v, *o);
 349         return buf;
 350 }
 351
 352 uint32_t *tmp, src, dst;
 353 int fd;
 354
 355 int main(int argc, char **argv)
 356 {
 357         int object_size = 0;
 358         uint32_t buf[20];
 359         int count;
 360
 361         igt_subtest_init(argc, argv);
 362         igt_skip_on_simulation();
 363
 364         if (argc > 1)
 365                 object_size = atoi(argv[1]);
 366         if (object_size == 0)
 367                 object_size = OBJECT_SIZE;
 368         object_size = (object_size + 3) & -4;
 369
 370         igt_fixture {
 371                 fd = drm_open_any();
 372
 373                 dst = gem_create(fd, object_size);
 374                 src = gem_create(fd, object_size);
 375                 tmp = malloc(object_size);
 376
 377                 gem_set_caching(fd, src, 0);
 378                 gem_set_caching(fd, dst, 0);
 379         }
 380
 381         devid = intel_get_drm_devid(fd);
 382
 383         igt_subtest("uncached-copy-correctness")
 384                 test_copy(fd, src, dst, tmp, object_size);
 385         igt_subtest("uncached-copy-performance") {
 386                 for (count = 1; count <= 1<<17; count <<= 1) {
 387                         struct timeval start, end;
 388
 389                         gettimeofday(&start, NULL);
 390                         copy(fd, src, dst, tmp, object_size, count);
 391                         gettimeofday(&end, NULL);
 392                         printf("Time to uncached copy %d bytes x %6d:   %7.3fµs, %s\n",
 393                                object_size, count,
 394                                elapsed(&start, &end, count),
 395                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 396                         fflush(stdout);
 397                 }
 398         }
 399
 400         igt_subtest("uncached-pwrite-blt-gtt_mmap-correctness")
 401                 test_as_gtt_mmap(fd, src, dst, object_size);
 402         igt_subtest("uncached-pwrite-blt-gtt_mmap-performance") {
 403                 for (count = 1; count <= 1<<17; count <<= 1) {
 404                         struct timeval start, end;
 405
 406                         gettimeofday(&start, NULL);
 407                         as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 408                         gettimeofday(&end, NULL);
 409                         printf("** mmap uncached copy %d bytes x %6d:   %7.3fµs, %s\n",
 410                                object_size, count,
 411                                elapsed(&start, &end, count),
 412                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 413                         fflush(stdout);
 414                 }
 415         }
 416
 417         igt_fixture {
 418                 gem_set_caching(fd, src, 1);
 419                 gem_set_caching(fd, dst, 1);
 420         }
 421
 422         igt_subtest("snooped-copy-correctness")
 423                 test_copy(fd, src, dst, tmp, object_size);
 424         igt_subtest("snooped-copy-performance") {
 425                 for (count = 1; count <= 1<<17; count <<= 1) {
 426                         struct timeval start, end;
 427
 428                         gettimeofday(&start, NULL);
 429                         copy(fd, src, dst, tmp, object_size, count);
 430                         gettimeofday(&end, NULL);
 431                         printf("Time to snooped copy %d bytes x %6d:    %7.3fµs, %s\n",
 432                                object_size, count,
 433                                elapsed(&start, &end, count),
 434                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 435                         fflush(stdout);
 436                 }
 437         }
 438
 439         igt_subtest("snooped-pwrite-blt-cpu_mmap-correctness")
 440                 test_as_cpu_mmap(fd, src, dst, object_size);
 441         igt_subtest("snooped-pwrite-blt-cpu_mmap-performance") {
 442                 for (count = 1; count <= 1<<17; count <<= 1) {
 443                         struct timeval start, end;
 444
 445                         gettimeofday(&start, NULL);
 446                         as_cpu_mmap(fd, src, dst, tmp, object_size, count);
 447                         gettimeofday(&end, NULL);
 448                         printf("** mmap snooped copy %d bytes x %6d:    %7.3fµs, %s\n",
 449                                object_size, count,
 450                                elapsed(&start, &end, count),
 451                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 452                         fflush(stdout);
 453                 }
 454         }
 455
 456         igt_fixture {
 457                 gem_set_caching(fd, src, 2);
 458                 gem_set_caching(fd, dst, 2);
 459         }
 460
 461         igt_subtest("display-copy-correctness")
 462                 test_copy(fd, src, dst, tmp, object_size);
 463         igt_subtest("display-copy-performance") {
 464                 for (count = 1; count <= 1<<17; count <<= 1) {
 465                         struct timeval start, end;
 466
 467                         gettimeofday(&start, NULL);
 468                         copy(fd, src, dst, tmp, object_size, count);
 469                         gettimeofday(&end, NULL);
 470                         printf("Time to display copy %d bytes x %6d:    %7.3fµs, %s\n",
 471                                object_size, count,
 472                                elapsed(&start, &end, count),
 473                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 474                         fflush(stdout);
 475                 }
 476         }
 477
 478         igt_subtest("display-pwrite-blt-gtt_mmap-correctness")
 479                 test_as_gtt_mmap(fd, src, dst, object_size);
 480         igt_subtest("display-pwrite-blt-gtt_mmap-performance") {
 481                 for (count = 1; count <= 1<<17; count <<= 1) {
 482                         struct timeval start, end;
 483
 484                         gettimeofday(&start, NULL);
 485                         as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 486                         gettimeofday(&end, NULL);
 487                         printf("** mmap display copy %d bytes x %6d:    %7.3fµs, %s\n",
 488                                object_size, count,
 489                                elapsed(&start, &end, count),
 490                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 491                         fflush(stdout);
 492                 }
 493         }
 494
 495         igt_fixture {
 496                 free(tmp);
 497                 gem_close(fd, src);
 498                 gem_close(fd, dst);
 499
 500                 close(fd);
 501         }
 502
 503         igt_exit();
 504 }