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)
  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 is_64bit;
  55 uint32_t exec_flags;
  56
  57 static inline void build_batch(uint32_t *batch, int len, uint32_t *batch_len)
  58 {
  59         unsigned int i = 0;
  60
  61         batch[i++] = COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB | (is_64bit ? 8 : 6);
  62         batch[i++] = 0xcc << 16 | 1 << 25 | 1 << 24 | len;
  63         batch[i++] = 0;
  64         batch[i++] = 1 << 16 | (len / 4);
  65         batch[i++] = 0; /* dst */
  66         if (is_64bit)
  67                 batch[i++] = 0; /* FIXME */
  68         batch[i++] = 0;
  69         batch[i++] = len;
  70         batch[i++] = 0; /* src */
  71         if (is_64bit)
  72                 batch[i++] = 0; /* FIXME */
  73         batch[i++] = MI_BATCH_BUFFER_END;
  74         batch[i++] = 0;
  75
  76         *batch_len = i * 4;
  77 }
  78
  79 #define GPP_BATCH_SIZE (12 * 4)
  80
  81 static void copy(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
  82 {
  83         uint32_t batch[GPP_BATCH_SIZE] = {0};
  84
  85         struct drm_i915_gem_relocation_entry reloc[] = {
  86                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
  87                 { src, 0, (is_64bit ? 8 : 7)*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
  88         };
  89         struct drm_i915_gem_exec_object2 exec[] = {
  90                 { src },
  91                 { dst },
  92                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
  93         };
  94         struct drm_i915_gem_execbuffer2 execbuf = {
  95                 (uintptr_t)exec, 3,
  96                 0, 0,
  97                 0, 0, 0, 0,
  98                 exec_flags,
  99         };
 100
 101         build_batch(batch, len, &execbuf.batch_len);
 102
 103         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 104
 105         while (loops--) {
 106                 gem_write(fd, src, 0, buf, len);
 107                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 108                 gem_read(fd, dst, 0, buf, len);
 109         }
 110
 111         gem_close(fd, exec[2].handle);
 112 }
 113
 114 static void as_gtt_mmap(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
 115 {
 116         uint32_t batch[GPP_BATCH_SIZE] = {0};
 117
 118         struct drm_i915_gem_relocation_entry reloc[] = {
 119                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 120                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 121         };
 122         struct drm_i915_gem_exec_object2 exec[] = {
 123                 { src },
 124                 { dst },
 125                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 126         };
 127         struct drm_i915_gem_execbuffer2 execbuf = {
 128                 (uintptr_t)exec, 3,
 129                 0, GPP_BATCH_SIZE,
 130                 0, 0, 0, 0,
 131                 exec_flags,
 132         };
 133         uint32_t *src_ptr, *dst_ptr;
 134
 135         build_batch(batch, len, &execbuf.batch_len);
 136
 137         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 138
 139         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 140         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 141
 142         while (loops--) {
 143                 gem_set_domain(fd, src,
 144                                I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 145                 memcpy(src_ptr, buf, len);
 146
 147                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 148                 gem_set_domain(fd, dst,
 149                                I915_GEM_DOMAIN_GTT, 0);
 150                 memcpy(buf, dst_ptr, len);
 151         }
 152
 153         munmap(dst_ptr, len);
 154         munmap(src_ptr, len);
 155         gem_close(fd, exec[2].handle);
 156 }
 157
 158
 159 static void as_cpu_mmap(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
 160 {
 161         uint32_t batch[GPP_BATCH_SIZE] = {0};
 162
 163         struct drm_i915_gem_relocation_entry reloc[] = {
 164                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 165                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 166         };
 167         struct drm_i915_gem_exec_object2 exec[] = {
 168                 { src },
 169                 { dst },
 170                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 171         };
 172         struct drm_i915_gem_execbuffer2 execbuf = {
 173                 (uintptr_t)exec, 3,
 174                 0, GPP_BATCH_SIZE,
 175                 0, 0, 0, 0,
 176                 exec_flags,
 177         };
 178         uint32_t *src_ptr, *dst_ptr;
 179
 180         build_batch(batch, len, &execbuf.batch_len);
 181
 182         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 183
 184         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 185         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 186
 187         while (loops--) {
 188                 gem_set_domain(fd, src,
 189                                I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 190                 memcpy(src_ptr, buf, len);
 191
 192                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 193                 gem_set_domain(fd, dst,
 194                                I915_GEM_DOMAIN_CPU, 0);
 195                 memcpy(buf, dst_ptr, len);
 196         }
 197
 198         munmap(dst_ptr, len);
 199         munmap(src_ptr, len);
 200         gem_close(fd, exec[2].handle);
 201 }
 202
 203 static void test_copy(int fd, uint32_t src, uint32_t dst, uint32_t *buf, int len)
 204 {
 205         uint32_t batch[GPP_BATCH_SIZE] = {0};
 206
 207         struct drm_i915_gem_relocation_entry reloc[] = {
 208                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 209                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 210         };
 211         struct drm_i915_gem_exec_object2 exec[] = {
 212                 { src },
 213                 { dst },
 214                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 215         };
 216         struct drm_i915_gem_execbuffer2 execbuf = {
 217                 (uintptr_t)exec, 3,
 218                 0, GPP_BATCH_SIZE,
 219                 0, 0, 0, 0,
 220                 exec_flags,
 221         };
 222         int i;
 223
 224         build_batch(batch, len, &execbuf.batch_len);
 225
 226         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 227
 228         for (i = 0; i < len/4; i++)
 229                 buf[i] = i;
 230
 231         gem_write(fd, src, 0, buf, len);
 232         memset(buf, 0, len);
 233
 234         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 235         gem_read(fd, dst, 0, buf, len);
 236
 237         gem_close(fd, exec[2].handle);
 238
 239         for (i = 0; i < len/4; i++)
 240                 igt_assert(buf[i] == i);
 241 }
 242
 243 static void test_as_gtt_mmap(int fd, uint32_t src, uint32_t dst, int len)
 244 {
 245         uint32_t batch[GPP_BATCH_SIZE] = {0};
 246
 247         struct drm_i915_gem_relocation_entry reloc[] = {
 248                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 249                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 250         };
 251         struct drm_i915_gem_exec_object2 exec[] = {
 252                 { src },
 253                 { dst },
 254                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 255         };
 256         struct drm_i915_gem_execbuffer2 execbuf = {
 257                 (uintptr_t)exec, 3,
 258                 0, GPP_BATCH_SIZE,
 259                 0, 0, 0, 0,
 260                 exec_flags,
 261         };
 262         uint32_t *src_ptr, *dst_ptr;
 263         int i;
 264
 265         build_batch(batch, len, &execbuf.batch_len);
 266
 267         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 268
 269         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 270         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 271
 272         gem_set_domain(fd, src, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 273         for (i = 0; i < len/4; i++)
 274                 src_ptr[i] = i;
 275
 276         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 277         gem_close(fd, exec[2].handle);
 278
 279         gem_set_domain(fd, dst, I915_GEM_DOMAIN_GTT, 0);
 280         for (i = 0; i < len/4; i++)
 281                 igt_assert(dst_ptr[i] == i);
 282
 283         munmap(dst_ptr, len);
 284         munmap(src_ptr, len);
 285 }
 286
 287 static void test_as_cpu_mmap(int fd, uint32_t src, uint32_t dst, int len)
 288 {
 289         uint32_t batch[GPP_BATCH_SIZE] = {0};
 290
 291         struct drm_i915_gem_relocation_entry reloc[] = {
 292                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 293                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 294         };
 295         struct drm_i915_gem_exec_object2 exec[] = {
 296                 { src },
 297                 { dst },
 298                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 299         };
 300         struct drm_i915_gem_execbuffer2 execbuf = {
 301                 (uintptr_t)exec, 3,
 302                 0, GPP_BATCH_SIZE,
 303                 0, 0, 0, 0,
 304                 exec_flags,
 305         };
 306         uint32_t *src_ptr, *dst_ptr;
 307         int i;
 308
 309         build_batch(batch, len, &execbuf.batch_len);
 310
 311         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
 312
 313         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 314         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 315
 316         gem_set_domain(fd, src, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 317         for (i = 0; i < len/4; i++)
 318                 src_ptr[i] = i;
 319
 320         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 321         gem_close(fd, exec[2].handle);
 322
 323         gem_set_domain(fd, dst, I915_GEM_DOMAIN_CPU, 0);
 324         for (i = 0; i < len/4; i++)
 325                 igt_assert(dst_ptr[i] == i);
 326
 327         munmap(dst_ptr, len);
 328         munmap(src_ptr, len);
 329 }
 330
 331 static double elapsed(const struct timeval *start,
 332                       const struct timeval *end,
 333                       int loop)
 334 {
 335         return (1e6*(end->tv_sec - start->tv_sec) + (end->tv_usec - start->tv_usec))/loop;
 336 }
 337
 338 static const char *bytes_per_sec(char *buf, double v)
 339 {
 340         const char *order[] = {
 341                 "",
 342                 "KiB",
 343                 "MiB",
 344                 "GiB",
 345                 "TiB",
 346                 NULL,
 347         }, **o = order;
 348
 349         while (v > 1000 && o[1]) {
 350                 v /= 1000;
 351                 o++;
 352         }
 353         sprintf(buf, "%.1f%s/s", v, *o);
 354         return buf;
 355 }
 356
 357 uint32_t *tmp, src, dst;
 358 int fd;
 359
 360 int main(int argc, char **argv)
 361 {
 362         int object_size = 0;
 363         uint32_t buf[20];
 364         int count;
 365
 366         igt_subtest_init(argc, argv);
 367         igt_skip_on_simulation();
 368
 369         if (argc > 1)
 370                 object_size = atoi(argv[1]);
 371         if (object_size == 0)
 372                 object_size = OBJECT_SIZE;
 373         object_size = (object_size + 3) & -4;
 374
 375         igt_fixture {
 376                 uint32_t devid;
 377
 378                 fd = drm_open_any();
 379
 380                 dst = gem_create(fd, object_size);
 381                 src = gem_create(fd, object_size);
 382                 tmp = malloc(object_size);
 383
 384                 gem_set_caching(fd, src, 0);
 385                 gem_set_caching(fd, dst, 0);
 386
 387                 devid = intel_get_drm_devid(fd);
 388                 is_64bit = intel_gen(devid) >= 8;
 389                 exec_flags = HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0;
 390         }
 391
 392         igt_subtest("uncached-copy-correctness")
 393                 test_copy(fd, src, dst, tmp, object_size);
 394         igt_subtest("uncached-copy-performance") {
 395                 for (count = 1; count <= 1<<17; count <<= 1) {
 396                         struct timeval start, end;
 397
 398                         gettimeofday(&start, NULL);
 399                         copy(fd, src, dst, tmp, object_size, count);
 400                         gettimeofday(&end, NULL);
 401                         printf("Time to uncached copy %d bytes x %6d:   %7.3fµs, %s\n",
 402                                object_size, count,
 403                                elapsed(&start, &end, count),
 404                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 405                         fflush(stdout);
 406                 }
 407         }
 408
 409         igt_subtest("uncached-pwrite-blt-gtt_mmap-correctness")
 410                 test_as_gtt_mmap(fd, src, dst, object_size);
 411         igt_subtest("uncached-pwrite-blt-gtt_mmap-performance") {
 412                 for (count = 1; count <= 1<<17; count <<= 1) {
 413                         struct timeval start, end;
 414
 415                         gettimeofday(&start, NULL);
 416                         as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 417                         gettimeofday(&end, NULL);
 418                         printf("** mmap uncached copy %d bytes x %6d:   %7.3fµs, %s\n",
 419                                object_size, count,
 420                                elapsed(&start, &end, count),
 421                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 422                         fflush(stdout);
 423                 }
 424         }
 425
 426         igt_fixture {
 427                 gem_set_caching(fd, src, 1);
 428                 gem_set_caching(fd, dst, 1);
 429         }
 430
 431         igt_subtest("snooped-copy-correctness")
 432                 test_copy(fd, src, dst, tmp, object_size);
 433         igt_subtest("snooped-copy-performance") {
 434                 for (count = 1; count <= 1<<17; count <<= 1) {
 435                         struct timeval start, end;
 436
 437                         gettimeofday(&start, NULL);
 438                         copy(fd, src, dst, tmp, object_size, count);
 439                         gettimeofday(&end, NULL);
 440                         printf("Time to snooped copy %d bytes x %6d:    %7.3fµs, %s\n",
 441                                object_size, count,
 442                                elapsed(&start, &end, count),
 443                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 444                         fflush(stdout);
 445                 }
 446         }
 447
 448         igt_subtest("snooped-pwrite-blt-cpu_mmap-correctness")
 449                 test_as_cpu_mmap(fd, src, dst, object_size);
 450         igt_subtest("snooped-pwrite-blt-cpu_mmap-performance") {
 451                 for (count = 1; count <= 1<<17; count <<= 1) {
 452                         struct timeval start, end;
 453
 454                         gettimeofday(&start, NULL);
 455                         as_cpu_mmap(fd, src, dst, tmp, object_size, count);
 456                         gettimeofday(&end, NULL);
 457                         printf("** mmap snooped copy %d bytes x %6d:    %7.3fµs, %s\n",
 458                                object_size, count,
 459                                elapsed(&start, &end, count),
 460                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 461                         fflush(stdout);
 462                 }
 463         }
 464
 465         igt_fixture {
 466                 gem_set_caching(fd, src, 2);
 467                 gem_set_caching(fd, dst, 2);
 468         }
 469
 470         igt_subtest("display-copy-correctness")
 471                 test_copy(fd, src, dst, tmp, object_size);
 472         igt_subtest("display-copy-performance") {
 473                 for (count = 1; count <= 1<<17; count <<= 1) {
 474                         struct timeval start, end;
 475
 476                         gettimeofday(&start, NULL);
 477                         copy(fd, src, dst, tmp, object_size, count);
 478                         gettimeofday(&end, NULL);
 479                         printf("Time to display copy %d bytes x %6d:    %7.3fµs, %s\n",
 480                                object_size, count,
 481                                elapsed(&start, &end, count),
 482                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 483                         fflush(stdout);
 484                 }
 485         }
 486
 487         igt_subtest("display-pwrite-blt-gtt_mmap-correctness")
 488                 test_as_gtt_mmap(fd, src, dst, object_size);
 489         igt_subtest("display-pwrite-blt-gtt_mmap-performance") {
 490                 for (count = 1; count <= 1<<17; count <<= 1) {
 491                         struct timeval start, end;
 492
 493                         gettimeofday(&start, NULL);
 494                         as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 495                         gettimeofday(&end, NULL);
 496                         printf("** mmap display copy %d bytes x %6d:    %7.3fµs, %s\n",
 497                                object_size, count,
 498                                elapsed(&start, &end, count),
 499                                bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 500                         fflush(stdout);
 501                 }
 502         }
 503
 504         igt_fixture {
 505                 free(tmp);
 506                 gem_close(fd, src);
 507                 gem_close(fd, dst);
 508
 509                 close(fd);
 510         }
 511
 512         igt_exit();
 513 }