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 <assert.h>
  34 #include <fcntl.h>
  35 #include <inttypes.h>
  36 #include <errno.h>
  37 #include <sys/stat.h>
  38 #include <sys/ioctl.h>
  39 #include <sys/mman.h>
  40 #include <sys/time.h>
  41 #include "drm.h"
  42 #include "i915_drm.h"
  43 #include "drmtest.h"
  44 #include "intel_bufmgr.h"
  45 #include "intel_batchbuffer.h"
  46 #include "intel_gpu_tools.h"
  47
  48 #define OBJECT_SIZE 16384
  49
  50 #define COPY_BLT_CMD            (2<<29|0x53<<22|0x6)
  51 #define BLT_WRITE_ALPHA         (1<<21)
  52 #define BLT_WRITE_RGB           (1<<20)
  53 #define BLT_SRC_TILED           (1<<15)
  54 #define BLT_DST_TILED           (1<<11)
  55
  56 static void copy(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
  57 {
  58         struct drm_i915_gem_relocation_entry reloc[] = {
  59                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
  60                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
  61         };
  62         struct drm_i915_gem_exec_object2 exec[] = {
  63                 { src },
  64                 { dst },
  65                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
  66         };
  67         uint32_t batch[] = {
  68                 COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB,
  69                 0xcc << 16 | 1 << 25 | 1 << 24 | len,
  70                 0,
  71                 1 << 16 | (len / 4),
  72                 0, /* dst */
  73                 0,
  74                 len,
  75                 0, /* src */
  76                 MI_BATCH_BUFFER_END,
  77                 0
  78         };
  79         struct drm_i915_gem_execbuffer2 execbuf = {
  80                 (uintptr_t)exec, 3,
  81                 0, sizeof(batch),
  82                 0, 0, 0, 0,
  83                 HAS_BLT_RING(intel_get_drm_devid(fd)) ? I915_EXEC_BLT : 0,
  84         };
  85         gem_write(fd, exec[2].handle, 0, batch, sizeof(batch));
  86
  87         while (loops--) {
  88                 gem_write(fd, src, 0, buf, len);
  89                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
  90                 gem_read(fd, dst, 0, buf, len);
  91         }
  92
  93         gem_close(fd, exec[2].handle);
  94 }
  95
  96 static void as_gtt_mmap(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
  97 {
  98         struct drm_i915_gem_relocation_entry reloc[] = {
  99                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 100                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 101         };
 102         struct drm_i915_gem_exec_object2 exec[] = {
 103                 { src },
 104                 { dst },
 105                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 106         };
 107         uint32_t batch[] = {
 108                 COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB,
 109                 0xcc << 16 | 1 << 25 | 1 << 24 | len,
 110                 0,
 111                 1 << 16 | (len / 4),
 112                 0, /* dst */
 113                 0,
 114                 len,
 115                 0, /* src */
 116                 MI_BATCH_BUFFER_END,
 117                 0
 118         };
 119         struct drm_i915_gem_execbuffer2 execbuf = {
 120                 (uintptr_t)exec, 3,
 121                 0, sizeof(batch),
 122                 0, 0, 0, 0,
 123                 HAS_BLT_RING(intel_get_drm_devid(fd)) ? I915_EXEC_BLT : 0,
 124         };
 125         uint32_t *src_ptr, *dst_ptr;
 126
 127         gem_write(fd, exec[2].handle, 0, batch, sizeof(batch));
 128
 129         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 130         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 131
 132         while (loops--) {
 133                 gem_set_domain(fd, src,
 134                                I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 135                 memcpy(src_ptr, buf, len);
 136
 137                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 138                 gem_set_domain(fd, dst,
 139                                I915_GEM_DOMAIN_GTT, 0);
 140                 memcpy(buf, dst_ptr, len);
 141         }
 142
 143         munmap(dst_ptr, len);
 144         munmap(src_ptr, len);
 145         gem_close(fd, exec[2].handle);
 146 }
 147
 148
 149 static void as_cpu_mmap(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
 150 {
 151         struct drm_i915_gem_relocation_entry reloc[] = {
 152                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 153                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 154         };
 155         struct drm_i915_gem_exec_object2 exec[] = {
 156                 { src },
 157                 { dst },
 158                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 159         };
 160         uint32_t batch[] = {
 161                 COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB,
 162                 0xcc << 16 | 1 << 25 | 1 << 24 | len,
 163                 0,
 164                 1 << 16 | (len / 4),
 165                 0, /* dst */
 166                 0,
 167                 len,
 168                 0, /* src */
 169                 MI_BATCH_BUFFER_END,
 170                 0
 171         };
 172         struct drm_i915_gem_execbuffer2 execbuf = {
 173                 (uintptr_t)exec, 3,
 174                 0, sizeof(batch),
 175                 0, 0, 0, 0,
 176                 HAS_BLT_RING(intel_get_drm_devid(fd)) ? I915_EXEC_BLT : 0,
 177         };
 178         uint32_t *src_ptr, *dst_ptr;
 179
 180         gem_write(fd, exec[2].handle, 0, batch, sizeof(batch));
 181
 182         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 183         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 184
 185         while (loops--) {
 186                 gem_set_domain(fd, src,
 187                                I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 188                 memcpy(src_ptr, buf, len);
 189
 190                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 191                 gem_set_domain(fd, dst,
 192                                I915_GEM_DOMAIN_CPU, 0);
 193                 memcpy(buf, dst_ptr, len);
 194         }
 195
 196         munmap(dst_ptr, len);
 197         munmap(src_ptr, len);
 198         gem_close(fd, exec[2].handle);
 199 }
 200
 201 static void test_copy(int fd, uint32_t src, uint32_t dst, uint32_t *buf, int len)
 202 {
 203         struct drm_i915_gem_relocation_entry reloc[] = {
 204                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 205                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 206         };
 207         struct drm_i915_gem_exec_object2 exec[] = {
 208                 { src },
 209                 { dst },
 210                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 211         };
 212         uint32_t batch[] = {
 213                 COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB,
 214                 0xcc << 16 | 1 << 25 | 1 << 24 | len,
 215                 0,
 216                 1 << 16 | (len / 4),
 217                 0, /* dst */
 218                 0,
 219                 len,
 220                 0, /* src */
 221                 MI_BATCH_BUFFER_END,
 222                 0
 223         };
 224         struct drm_i915_gem_execbuffer2 execbuf = {
 225                 (uintptr_t)exec, 3,
 226                 0, sizeof(batch),
 227                 0, 0, 0, 0,
 228                 HAS_BLT_RING(intel_get_drm_devid(fd)) ? I915_EXEC_BLT : 0,
 229         };
 230         int i;
 231
 232         gem_write(fd, exec[2].handle, 0, batch, sizeof(batch));
 233
 234         for (i = 0; i < len/4; i++)
 235                 buf[i] = i;
 236
 237         gem_write(fd, src, 0, buf, len);
 238         memset(buf, 0, len);
 239
 240         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 241         gem_read(fd, dst, 0, buf, len);
 242
 243         gem_close(fd, exec[2].handle);
 244
 245         for (i = 0; i < len/4; i++)
 246                 assert(buf[i] == i);
 247 }
 248
 249 static void test_as_gtt_mmap(int fd, uint32_t src, uint32_t dst, int len)
 250 {
 251         struct drm_i915_gem_relocation_entry reloc[] = {
 252                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 253                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 254         };
 255         struct drm_i915_gem_exec_object2 exec[] = {
 256                 { src },
 257                 { dst },
 258                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 259         };
 260         uint32_t batch[] = {
 261                 COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB,
 262                 0xcc << 16 | 1 << 25 | 1 << 24 | len,
 263                 0,
 264                 1 << 16 | (len / 4),
 265                 0, /* dst */
 266                 0,
 267                 len,
 268                 0, /* src */
 269                 MI_BATCH_BUFFER_END,
 270                 0
 271         };
 272         struct drm_i915_gem_execbuffer2 execbuf = {
 273                 (uintptr_t)exec, 3,
 274                 0, sizeof(batch),
 275                 0, 0, 0, 0,
 276                 HAS_BLT_RING(intel_get_drm_devid(fd)) ? I915_EXEC_BLT : 0,
 277         };
 278         uint32_t *src_ptr, *dst_ptr;
 279         int i;
 280
 281         gem_write(fd, exec[2].handle, 0, batch, sizeof(batch));
 282
 283         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 284         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 285
 286         gem_set_domain(fd, src, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 287         for (i = 0; i < len/4; i++)
 288                 src_ptr[i] = i;
 289
 290         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 291         gem_close(fd, exec[2].handle);
 292
 293         gem_set_domain(fd, dst, I915_GEM_DOMAIN_GTT, 0);
 294         for (i = 0; i < len/4; i++)
 295                 assert(dst_ptr[i] == i);
 296
 297         munmap(dst_ptr, len);
 298         munmap(src_ptr, len);
 299 }
 300
 301 static void test_as_cpu_mmap(int fd, uint32_t src, uint32_t dst, int len)
 302 {
 303         struct drm_i915_gem_relocation_entry reloc[] = {
 304                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER },
 305                 { src, 0, 7*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 },
 306         };
 307         struct drm_i915_gem_exec_object2 exec[] = {
 308                 { src },
 309                 { dst },
 310                 { gem_create(fd, 4096), 2, (uintptr_t)reloc }
 311         };
 312         uint32_t batch[] = {
 313                 COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB,
 314                 0xcc << 16 | 1 << 25 | 1 << 24 | len,
 315                 0,
 316                 1 << 16 | (len / 4),
 317                 0, /* dst */
 318                 0,
 319                 len,
 320                 0, /* src */
 321                 MI_BATCH_BUFFER_END,
 322                 0
 323         };
 324         struct drm_i915_gem_execbuffer2 execbuf = {
 325                 (uintptr_t)exec, 3,
 326                 0, sizeof(batch),
 327                 0, 0, 0, 0,
 328                 HAS_BLT_RING(intel_get_drm_devid(fd)) ? I915_EXEC_BLT : 0,
 329         };
 330         uint32_t *src_ptr, *dst_ptr;
 331         int i;
 332
 333         gem_write(fd, exec[2].handle, 0, batch, sizeof(batch));
 334
 335         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 336         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 337
 338         gem_set_domain(fd, src, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 339         for (i = 0; i < len/4; i++)
 340                 src_ptr[i] = i;
 341
 342         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 343         gem_close(fd, exec[2].handle);
 344
 345         gem_set_domain(fd, dst, I915_GEM_DOMAIN_CPU, 0);
 346         for (i = 0; i < len/4; i++)
 347                 assert(dst_ptr[i] == i);
 348
 349         munmap(dst_ptr, len);
 350         munmap(src_ptr, len);
 351 }
 352
 353 static double elapsed(const struct timeval *start,
 354                       const struct timeval *end,
 355                       int loop)
 356 {
 357         return (1e6*(end->tv_sec - start->tv_sec) + (end->tv_usec - start->tv_usec))/loop;
 358 }
 359
 360 static const char *bytes_per_sec(char *buf, double v)
 361 {
 362         const char *order[] = {
 363                 "",
 364                 "KiB",
 365                 "MiB",
 366                 "GiB",
 367                 "TiB",
 368                 NULL,
 369         }, **o = order;
 370
 371         while (v > 1000 && o[1]) {
 372                 v /= 1000;
 373                 o++;
 374         }
 375         sprintf(buf, "%.1f%s/s", v, *o);
 376         return buf;
 377 }
 378
 379 int main(int argc, char **argv)
 380 {
 381         int object_size = 0;
 382         uint32_t buf[20];
 383         uint32_t *tmp, src, dst;
 384         int fd, count;
 385
 386         drmtest_subtest_init(argc, argv);
 387         drmtest_skip_on_simulation();
 388
 389         if (argc > 1)
 390                 object_size = atoi(argv[1]);
 391         if (object_size == 0)
 392                 object_size = OBJECT_SIZE;
 393         object_size = (object_size + 3) & -4;
 394
 395         fd = drm_open_any();
 396
 397         dst = gem_create(fd, object_size);
 398         src = gem_create(fd, object_size);
 399         tmp = malloc(object_size);
 400
 401         if (gem_set_cacheing(fd, src, 0) == 0 &&
 402             gem_set_cacheing(fd, dst, 0) == 0) {
 403                 if (drmtest_run_subtest("uncached-copy-correctness"))
 404                         test_copy(fd, src, dst, tmp, object_size);
 405                 if (drmtest_run_subtest("uncached-copy-performance")) {
 406                         for (count = 1; count <= 1<<17; count <<= 1) {
 407                                 struct timeval start, end;
 408
 409                                 gettimeofday(&start, NULL);
 410                                 copy(fd, src, dst, tmp, object_size, count);
 411                                 gettimeofday(&end, NULL);
 412                                 printf("Time to uncached copy %d bytes x %6d:   %7.3fµs, %s\n",
 413                                        object_size, count,
 414                                        elapsed(&start, &end, count),
 415                                        bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 416                                 fflush(stdout);
 417                         }
 418                 }
 419
 420                 if (drmtest_run_subtest("uncached-pwrite-blt-gtt_mmap-correctness"))
 421                         test_as_gtt_mmap(fd, src, dst, object_size);
 422                 if (drmtest_run_subtest("uncached-pwrite-blt-gtt_mmap-performance")) {
 423                         for (count = 1; count <= 1<<17; count <<= 1) {
 424                                 struct timeval start, end;
 425
 426                                 gettimeofday(&start, NULL);
 427                                 as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 428                                 gettimeofday(&end, NULL);
 429                                 printf("** mmap uncached copy %d bytes x %6d:   %7.3fµs, %s\n",
 430                                        object_size, count,
 431                                        elapsed(&start, &end, count),
 432                                        bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 433                                 fflush(stdout);
 434                         }
 435                 }
 436         }
 437
 438         if (gem_set_cacheing(fd, src, 1) == 0 &&
 439             gem_set_cacheing(fd, dst, 1) == 0) {
 440                 if (drmtest_run_subtest("snooped-copy-correctness"))
 441                         test_copy(fd, src, dst, tmp, object_size);
 442                 if (drmtest_run_subtest("snooped-copy-performance")) {
 443                         for (count = 1; count <= 1<<17; count <<= 1) {
 444                                 struct timeval start, end;
 445
 446                                 gettimeofday(&start, NULL);
 447                                 copy(fd, src, dst, tmp, object_size, count);
 448                                 gettimeofday(&end, NULL);
 449                                 printf("Time to snooped copy %d bytes x %6d:    %7.3fµs, %s\n",
 450                                        object_size, count,
 451                                        elapsed(&start, &end, count),
 452                                        bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 453                                 fflush(stdout);
 454                         }
 455                 }
 456
 457                 if (drmtest_run_subtest("snooped-pwrite-blt-cpu_mmap-correctness"))
 458                         test_as_cpu_mmap(fd, src, dst, object_size);
 459                 if (drmtest_run_subtest("snooped-pwrite-blt-cpu_mmap-performance")) {
 460                         for (count = 1; count <= 1<<17; count <<= 1) {
 461                                 struct timeval start, end;
 462
 463                                 gettimeofday(&start, NULL);
 464                                 as_cpu_mmap(fd, src, dst, tmp, object_size, count);
 465                                 gettimeofday(&end, NULL);
 466                                 printf("** mmap snooped copy %d bytes x %6d:    %7.3fµs, %s\n",
 467                                        object_size, count,
 468                                        elapsed(&start, &end, count),
 469                                        bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 470                                 fflush(stdout);
 471                         }
 472                 }
 473         }
 474
 475         if (gem_set_cacheing(fd, src, 2) == 0 &&
 476             gem_set_cacheing(fd, dst, 2) == 0) {
 477                 if (drmtest_run_subtest("display-copy-correctness"))
 478                         test_copy(fd, src, dst, tmp, object_size);
 479                 if (drmtest_run_subtest("display-copy-performance")) {
 480                         for (count = 1; count <= 1<<17; count <<= 1) {
 481                                 struct timeval start, end;
 482
 483                                 gettimeofday(&start, NULL);
 484                                 copy(fd, src, dst, tmp, object_size, count);
 485                                 gettimeofday(&end, NULL);
 486                                 printf("Time to display copy %d bytes x %6d:    %7.3fµs, %s\n",
 487                                        object_size, count,
 488                                        elapsed(&start, &end, count),
 489                                        bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 490                                 fflush(stdout);
 491                         }
 492                 }
 493
 494                 if (drmtest_run_subtest("display-pwrite-blt-gtt_mmap-correctness"))
 495                         test_as_gtt_mmap(fd, src, dst, object_size);
 496                 if (drmtest_run_subtest("display-pwrite-blt-gtt_mmap-performance")) {
 497                         for (count = 1; count <= 1<<17; count <<= 1) {
 498                                 struct timeval start, end;
 499
 500                                 gettimeofday(&start, NULL);
 501                                 as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 502                                 gettimeofday(&end, NULL);
 503                                 printf("** mmap display copy %d bytes x %6d:    %7.3fµs, %s\n",
 504                                        object_size, count,
 505                                        elapsed(&start, &end, count),
 506                                        bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 507                                 fflush(stdout);
 508                         }
 509                 }
 510         }
 511
 512         free(tmp);
 513         gem_close(fd, src);
 514         gem_close(fd, dst);
 515
 516         close(fd);
 517
 518         return 0;
 519 }