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 "ioctl_wrappers.h"
  41 #include "drmtest.h"
  42 #include "intel_chipset.h"
  43 #include "intel_bufmgr.h"
  44 #include "intel_batchbuffer.h"
  45 #include "intel_io.h"
  46
  47 #define OBJECT_SIZE 16384
  48
  49 #define COPY_BLT_CMD            (2<<29|0x53<<22)
  50 #define BLT_WRITE_ALPHA         (1<<21)
  51 #define BLT_WRITE_RGB           (1<<20)
  52 #define BLT_SRC_TILED           (1<<15)
  53 #define BLT_DST_TILED           (1<<11)
  54
  55 uint32_t is_64bit;
  56 uint32_t exec_flags;
  57
  58 static inline void build_batch(uint32_t *batch, int len, uint32_t *batch_len)
  59 {
  60         unsigned int i = 0;
  61
  62         batch[i++] = COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB | (is_64bit ? 8 : 6);
  63         batch[i++] = 0xcc << 16 | 1 << 25 | 1 << 24 | len;
  64         batch[i++] = 0;
  65         batch[i++] = 1 << 16 | (len / 4);
  66         batch[i++] = 0; /* dst */
  67         if (is_64bit)
  68                 batch[i++] = 0;
  69         batch[i++] = 0;
  70         batch[i++] = len;
  71         batch[i++] = 0; /* src */
  72         if (is_64bit)
  73                 batch[i++] = 0;
  74         batch[i++] = MI_BATCH_BUFFER_END;
  75         batch[i++] = 0;
  76
  77         *batch_len = i * 4;
  78 }
  79
  80 #define BUILD_EXEC \
  81         uint32_t batch[12]; \
  82         struct drm_i915_gem_relocation_entry reloc[] = { \
  83                 { dst, 0, 4*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, I915_GEM_DOMAIN_RENDER }, \
  84                 { src, 0, (is_64bit ? 8 : 7)*sizeof(uint32_t), 0, I915_GEM_DOMAIN_RENDER, 0 }, \
  85         }; \
  86         struct drm_i915_gem_exec_object2 exec[] = { \
  87                 { src }, \
  88                 { dst }, \
  89                 { gem_create(fd, 4096), 2, (uintptr_t)reloc } \
  90         }; \
  91         struct drm_i915_gem_execbuffer2 execbuf = { \
  92                 (uintptr_t)exec, 3, \
  93                 0, 0, \
  94                 0, 0, 0, 0, \
  95                 exec_flags, \
  96         }; \
  97         build_batch(batch, len, &execbuf.batch_len); \
  98         gem_write(fd, exec[2].handle, 0, batch, execbuf.batch_len);
  99
 100
 101 static void copy(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
 102 {
 103         BUILD_EXEC;
 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 *src_ptr, *dst_ptr;
 117         BUILD_EXEC;
 118
 119         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 120         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 121
 122         while (loops--) {
 123                 gem_set_domain(fd, src,
 124                                I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 125                 memcpy(src_ptr, buf, len);
 126
 127                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 128                 gem_set_domain(fd, dst,
 129                                I915_GEM_DOMAIN_GTT, 0);
 130                 memcpy(buf, dst_ptr, len);
 131         }
 132
 133         munmap(dst_ptr, len);
 134         munmap(src_ptr, len);
 135         gem_close(fd, exec[2].handle);
 136 }
 137
 138
 139 static void as_cpu_mmap(int fd, uint32_t src, uint32_t dst, void *buf, int len, int loops)
 140 {
 141         uint32_t *src_ptr, *dst_ptr;
 142         BUILD_EXEC;
 143
 144         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 145         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 146
 147         while (loops--) {
 148                 gem_set_domain(fd, src,
 149                                I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 150                 memcpy(src_ptr, buf, len);
 151
 152                 do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 153                 gem_set_domain(fd, dst,
 154                                I915_GEM_DOMAIN_CPU, 0);
 155                 memcpy(buf, dst_ptr, len);
 156         }
 157
 158         munmap(dst_ptr, len);
 159         munmap(src_ptr, len);
 160         gem_close(fd, exec[2].handle);
 161 }
 162
 163 static void test_copy(int fd, uint32_t src, uint32_t dst, uint32_t *buf, int len)
 164 {
 165         int i;
 166         BUILD_EXEC;
 167
 168         for (i = 0; i < len/4; i++)
 169                 buf[i] = i;
 170
 171         gem_write(fd, src, 0, buf, len);
 172         memset(buf, 0, len);
 173
 174         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 175         gem_read(fd, dst, 0, buf, len);
 176
 177         gem_close(fd, exec[2].handle);
 178
 179         for (i = 0; i < len/4; i++)
 180                 igt_assert(buf[i] == i);
 181 }
 182
 183 static void test_as_gtt_mmap(int fd, uint32_t src, uint32_t dst, int len)
 184 {
 185         uint32_t *src_ptr, *dst_ptr;
 186         int i;
 187         BUILD_EXEC;
 188
 189         src_ptr = gem_mmap__gtt(fd, src, OBJECT_SIZE, PROT_WRITE);
 190         dst_ptr = gem_mmap__gtt(fd, dst, OBJECT_SIZE, PROT_READ);
 191
 192         gem_set_domain(fd, src, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
 193         for (i = 0; i < len/4; i++)
 194                 src_ptr[i] = i;
 195
 196         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 197         gem_close(fd, exec[2].handle);
 198
 199         gem_set_domain(fd, dst, I915_GEM_DOMAIN_GTT, 0);
 200         for (i = 0; i < len/4; i++)
 201                 igt_assert(dst_ptr[i] == i);
 202
 203         munmap(dst_ptr, len);
 204         munmap(src_ptr, len);
 205 }
 206
 207 static void test_as_cpu_mmap(int fd, uint32_t src, uint32_t dst, int len)
 208 {
 209         uint32_t *src_ptr, *dst_ptr;
 210         int i;
 211         BUILD_EXEC;
 212
 213         src_ptr = gem_mmap__cpu(fd, src, OBJECT_SIZE, PROT_WRITE);
 214         dst_ptr = gem_mmap__cpu(fd, dst, OBJECT_SIZE, PROT_READ);
 215
 216         gem_set_domain(fd, src, I915_GEM_DOMAIN_CPU, I915_GEM_DOMAIN_CPU);
 217         for (i = 0; i < len/4; i++)
 218                 src_ptr[i] = i;
 219
 220         do_or_die(drmIoctl(fd, DRM_IOCTL_I915_GEM_EXECBUFFER2, &execbuf));
 221         gem_close(fd, exec[2].handle);
 222
 223         gem_set_domain(fd, dst, I915_GEM_DOMAIN_CPU, 0);
 224         for (i = 0; i < len/4; i++)
 225                 igt_assert(dst_ptr[i] == i);
 226
 227         munmap(dst_ptr, len);
 228         munmap(src_ptr, len);
 229 }
 230
 231 static double elapsed(const struct timeval *start,
 232                       const struct timeval *end,
 233                       int loop)
 234 {
 235         return (1e6*(end->tv_sec - start->tv_sec) + (end->tv_usec - start->tv_usec))/loop;
 236 }
 237
 238 static const char *bytes_per_sec(char *buf, double v)
 239 {
 240         const char *order[] = {
 241                 "",
 242                 "KiB",
 243                 "MiB",
 244                 "GiB",
 245                 "TiB",
 246                 NULL,
 247         }, **o = order;
 248
 249         while (v > 1000 && o[1]) {
 250                 v /= 1000;
 251                 o++;
 252         }
 253         sprintf(buf, "%.1f%s/s", v, *o);
 254         return buf;
 255 }
 256
 257 uint32_t *tmp, src, dst;
 258 int fd;
 259
 260 int main(int argc, char **argv)
 261 {
 262         int object_size = 0;
 263         uint32_t buf[20];
 264         int count;
 265
 266         igt_subtest_init(argc, argv);
 267         igt_skip_on_simulation();
 268
 269         if (argc > 1)
 270                 object_size = atoi(argv[1]);
 271         if (object_size == 0)
 272                 object_size = OBJECT_SIZE;
 273         object_size = (object_size + 3) & -4;
 274
 275         igt_fixture {
 276                 uint32_t devid;
 277
 278                 fd = drm_open_any();
 279
 280                 dst = gem_create(fd, object_size);
 281                 src = gem_create(fd, object_size);
 282                 tmp = malloc(object_size);
 283
 284                 gem_set_caching(fd, src, 0);
 285                 gem_set_caching(fd, dst, 0);
 286
 287                 devid = intel_get_drm_devid(fd);
 288                 is_64bit = intel_gen(devid) >= 8;
 289                 exec_flags = HAS_BLT_RING(devid) ? I915_EXEC_BLT : 0;
 290         }
 291
 292         igt_subtest("uncached-copy-correctness")
 293                 test_copy(fd, src, dst, tmp, object_size);
 294         igt_subtest("uncached-copy-performance") {
 295                 for (count = 1; count <= 1<<17; count <<= 1) {
 296                         struct timeval start, end;
 297
 298                         gettimeofday(&start, NULL);
 299                         copy(fd, src, dst, tmp, object_size, count);
 300                         gettimeofday(&end, NULL);
 301                         igt_info("Time to uncached copy %d bytes x %6d: %7.3fµs, %s\n",
 302                                  object_size, count,
 303                                  elapsed(&start, &end, count),
 304                                  bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 305                         fflush(stdout);
 306                 }
 307         }
 308
 309         igt_subtest("uncached-pwrite-blt-gtt_mmap-correctness")
 310                 test_as_gtt_mmap(fd, src, dst, object_size);
 311         igt_subtest("uncached-pwrite-blt-gtt_mmap-performance") {
 312                 for (count = 1; count <= 1<<17; count <<= 1) {
 313                         struct timeval start, end;
 314
 315                         gettimeofday(&start, NULL);
 316                         as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 317                         gettimeofday(&end, NULL);
 318                         igt_info("** mmap uncached copy %d bytes x %6d: %7.3fµs, %s\n",
 319                                  object_size, count,
 320                                  elapsed(&start, &end, count),
 321                                  bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 322                         fflush(stdout);
 323                 }
 324         }
 325
 326         igt_fixture {
 327                 gem_set_caching(fd, src, 1);
 328                 gem_set_caching(fd, dst, 1);
 329         }
 330
 331         igt_subtest("snooped-copy-correctness")
 332                 test_copy(fd, src, dst, tmp, object_size);
 333         igt_subtest("snooped-copy-performance") {
 334                 for (count = 1; count <= 1<<17; count <<= 1) {
 335                         struct timeval start, end;
 336
 337                         gettimeofday(&start, NULL);
 338                         copy(fd, src, dst, tmp, object_size, count);
 339                         gettimeofday(&end, NULL);
 340                         igt_info("Time to snooped copy %d bytes x %6d:  %7.3fµs, %s\n",
 341                                  object_size, count,
 342                                  elapsed(&start, &end, count),
 343                                  bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 344                         fflush(stdout);
 345                 }
 346         }
 347
 348         igt_subtest("snooped-pwrite-blt-cpu_mmap-correctness")
 349                 test_as_cpu_mmap(fd, src, dst, object_size);
 350         igt_subtest("snooped-pwrite-blt-cpu_mmap-performance") {
 351                 for (count = 1; count <= 1<<17; count <<= 1) {
 352                         struct timeval start, end;
 353
 354                         gettimeofday(&start, NULL);
 355                         as_cpu_mmap(fd, src, dst, tmp, object_size, count);
 356                         gettimeofday(&end, NULL);
 357                         igt_info("** mmap snooped copy %d bytes x %6d:  %7.3fµs, %s\n",
 358                                  object_size, count,
 359                                  elapsed(&start, &end, count),
 360                                  bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 361                         fflush(stdout);
 362                 }
 363         }
 364
 365         igt_fixture {
 366                 gem_set_caching(fd, src, 2);
 367                 gem_set_caching(fd, dst, 2);
 368         }
 369
 370         igt_subtest("display-copy-correctness")
 371                 test_copy(fd, src, dst, tmp, object_size);
 372         igt_subtest("display-copy-performance") {
 373                 for (count = 1; count <= 1<<17; count <<= 1) {
 374                         struct timeval start, end;
 375
 376                         gettimeofday(&start, NULL);
 377                         copy(fd, src, dst, tmp, object_size, count);
 378                         gettimeofday(&end, NULL);
 379                         igt_info("Time to display copy %d bytes x %6d:  %7.3fµs, %s\n",
 380                                  object_size, count,
 381                                  elapsed(&start, &end, count),
 382                                  bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 383                         fflush(stdout);
 384                 }
 385         }
 386
 387         igt_subtest("display-pwrite-blt-gtt_mmap-correctness")
 388                 test_as_gtt_mmap(fd, src, dst, object_size);
 389         igt_subtest("display-pwrite-blt-gtt_mmap-performance") {
 390                 for (count = 1; count <= 1<<17; count <<= 1) {
 391                         struct timeval start, end;
 392
 393                         gettimeofday(&start, NULL);
 394                         as_gtt_mmap(fd, src, dst, tmp, object_size, count);
 395                         gettimeofday(&end, NULL);
 396                         igt_info("** mmap display copy %d bytes x %6d:  %7.3fµs, %s\n",
 397                                  object_size, count,
 398                                  elapsed(&start, &end, count),
 399                                  bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 400                         fflush(stdout);
 401                 }
 402         }
 403
 404         igt_fixture {
 405                 free(tmp);
 406                 gem_close(fd, src);
 407                 gem_close(fd, dst);
 408
 409                 close(fd);
 410         }
 411
 412         igt_exit();
 413 }