tests/gem_exec_blt.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_io.h"
  44
  45 #define OBJECT_SIZE 16384
  46
  47 #define COPY_BLT_CMD            (2<<29|0x53<<22|0x6)
  48 #define BLT_WRITE_ALPHA         (1<<21)
  49 #define BLT_WRITE_RGB           (1<<20)
  50 #define BLT_SRC_TILED           (1<<15)
  51 #define BLT_DST_TILED           (1<<11)
  52
  53 static int gem_linear_blt(int fd,
  54                           uint32_t *batch,
  55                           uint32_t src,
  56                           uint32_t dst,
  57                           uint32_t length,
  58                           struct drm_i915_gem_relocation_entry *reloc)
  59 {
  60         uint32_t *b = batch;
  61         int height = length / (16 * 1024);
  62
  63         igt_assert(height <= 1<<16);
  64
  65         if (height) {
  66                 int i = 0;
  67                 b[i++] = COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB;
  68                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
  69                         b[i-1]+=2;
  70                 b[i++] = 0xcc << 16 | 1 << 25 | 1 << 24 | (16*1024);
  71                 b[i++] = 0;
  72                 b[i++] = height << 16 | (4*1024);
  73                 b[i++] = 0;
  74                 reloc->offset = (b-batch+4) * sizeof(uint32_t);
  75                 reloc->delta = 0;
  76                 reloc->target_handle = dst;
  77                 reloc->read_domains = I915_GEM_DOMAIN_RENDER;
  78                 reloc->write_domain = I915_GEM_DOMAIN_RENDER;
  79                 reloc->presumed_offset = 0;
  80                 reloc++;
  81                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
  82                         b[i++] = 0; /* FIXME */
  83
  84                 b[i++] = 0;
  85                 b[i++] = 16*1024;
  86                 b[i++] = 0;
  87                 reloc->offset = (b-batch+7) * sizeof(uint32_t);
  88                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
  89                         reloc->offset += sizeof(uint32_t);
  90                 reloc->delta = 0;
  91                 reloc->target_handle = src;
  92                 reloc->read_domains = I915_GEM_DOMAIN_RENDER;
  93                 reloc->write_domain = 0;
  94                 reloc->presumed_offset = 0;
  95                 reloc++;
  96                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
  97                         b[i++] = 0; /* FIXME */
  98
  99                 b += i;
 100                 length -= height * 16*1024;
 101         }
 102
 103         if (length) {
 104                 int i = 0;
 105                 b[i++] = COPY_BLT_CMD | BLT_WRITE_ALPHA | BLT_WRITE_RGB;
 106                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
 107                         b[i-1]+=2;
 108                 b[i++] = 0xcc << 16 | 1 << 25 | 1 << 24 | (16*1024);
 109                 b[i++] = height << 16;
 110                 b[i++] = (1+height) << 16 | (length / 4);
 111                 b[i++] = 0;
 112                 reloc->offset = (b-batch+4) * sizeof(uint32_t);
 113                 reloc->delta = 0;
 114                 reloc->target_handle = dst;
 115                 reloc->read_domains = I915_GEM_DOMAIN_RENDER;
 116                 reloc->write_domain = I915_GEM_DOMAIN_RENDER;
 117                 reloc->presumed_offset = 0;
 118                 reloc++;
 119                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
 120                         b[i++] = 0; /* FIXME */
 121
 122                 b[i++] = height << 16;
 123                 b[i++] = 16*1024;
 124                 b[i++] = 0;
 125                 reloc->offset = (b-batch+7) * sizeof(uint32_t);
 126                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
 127                         reloc->offset += sizeof(uint32_t);
 128                 reloc->delta = 0;
 129                 reloc->target_handle = src;
 130                 reloc->read_domains = I915_GEM_DOMAIN_RENDER;
 131                 reloc->write_domain = 0;
 132                 reloc->presumed_offset = 0;
 133                 reloc++;
 134                 if (intel_gen(intel_get_drm_devid(fd)) >= 8)
 135                         b[i++] = 0; /* FIXME */
 136
 137                 b += i;
 138         }
 139
 140         b[0] = MI_BATCH_BUFFER_END;
 141         b[1] = 0;
 142
 143         return (b+2 - batch) * sizeof(uint32_t);
 144 }
 145
 146 static double elapsed(const struct timeval *start,
 147                       const struct timeval *end,
 148                       int loop)
 149 {
 150         return (1e6*(end->tv_sec - start->tv_sec) + (end->tv_usec - start->tv_usec))/loop;
 151 }
 152
 153 static const char *bytes_per_sec(char *buf, double v)
 154 {
 155         const char *order[] = {
 156                 "",
 157                 "KiB",
 158                 "MiB",
 159                 "GiB",
 160                 "TiB",
 161                 "PiB",
 162                 NULL,
 163         }, **o = order;
 164
 165         while (v > 1024 && o[1]) {
 166                 v /= 1024;
 167                 o++;
 168         }
 169         sprintf(buf, "%.1f%s/s", v, *o);
 170         return buf;
 171 }
 172
 173 static void run(int object_size)
 174 {
 175         struct drm_i915_gem_execbuffer2 execbuf;
 176         struct drm_i915_gem_exec_object2 exec[3];
 177         struct drm_i915_gem_relocation_entry reloc[4];
 178         uint32_t buf[20];
 179         uint32_t handle, src, dst;
 180         int fd, len, count;
 181         int ring;
 182
 183         fd = drm_open_any();
 184         handle = gem_create(fd, 4096);
 185         src = gem_create(fd, object_size);
 186         dst = gem_create(fd, object_size);
 187
 188         len = gem_linear_blt(fd, buf, src, dst, object_size, reloc);
 189         gem_write(fd, handle, 0, buf, len);
 190
 191         exec[0].handle = src;
 192         exec[0].relocation_count = 0;
 193         exec[0].relocs_ptr = 0;
 194         exec[0].alignment = 0;
 195         exec[0].offset = 0;
 196         exec[0].flags = 0;
 197         exec[0].rsvd1 = 0;
 198         exec[0].rsvd2 = 0;
 199
 200         exec[1].handle = dst;
 201         exec[1].relocation_count = 0;
 202         exec[1].relocs_ptr = 0;
 203         exec[1].alignment = 0;
 204         exec[1].offset = 0;
 205         exec[1].flags = 0;
 206         exec[1].rsvd1 = 0;
 207         exec[1].rsvd2 = 0;
 208
 209         exec[2].handle = handle;
 210         if (intel_gen(intel_get_drm_devid(fd)) >= 8)
 211                 exec[2].relocation_count = len > 56 ? 4 : 2;
 212         else
 213                 exec[2].relocation_count = len > 40 ? 4 : 2;
 214         exec[2].relocs_ptr = (uintptr_t)reloc;
 215         exec[2].alignment = 0;
 216         exec[2].offset = 0;
 217         exec[2].flags = 0;
 218         exec[2].rsvd1 = 0;
 219         exec[2].rsvd2 = 0;
 220
 221         ring = 0;
 222         if (HAS_BLT_RING(intel_get_drm_devid(fd)))
 223                 ring = I915_EXEC_BLT;
 224
 225         execbuf.buffers_ptr = (uintptr_t)exec;
 226         execbuf.buffer_count = 3;
 227         execbuf.batch_start_offset = 0;
 228         execbuf.batch_len = len;
 229         execbuf.cliprects_ptr = 0;
 230         execbuf.num_cliprects = 0;
 231         execbuf.DR1 = 0;
 232         execbuf.DR4 = 0;
 233         execbuf.flags = ring;
 234         i915_execbuffer2_set_context_id(execbuf, 0);
 235         execbuf.rsvd2 = 0;
 236
 237         for (count = 1; count <= 1<<12; count <<= 1) {
 238                 struct timeval start, end;
 239
 240                 gettimeofday(&start, NULL);
 241                 for (int loop = 0; loop < count; loop++)
 242                         gem_execbuf(fd, &execbuf);
 243                 gem_sync(fd, handle);
 244                 gettimeofday(&end, NULL);
 245                 igt_info("Time to blt %d bytes x %6d:   %7.3fµs, %s\n",
 246                          object_size, count,
 247                          elapsed(&start, &end, count),
 248                          bytes_per_sec((char *)buf, object_size/elapsed(&start, &end, count)*1e6));
 249                 fflush(stdout);
 250         }
 251         gem_close(fd, handle);
 252
 253         close(fd);
 254 }
 255
 256 int main(int argc, char **argv)
 257 {
 258         int i;
 259
 260         igt_simple_init(argc, argv);
 261
 262         igt_skip_on_simulation();
 263
 264         if (argc > 1) {
 265                 for (i = 1; i < argc; i++) {
 266                         int object_size = atoi(argv[i]);
 267                         if (object_size)
 268                                 run((object_size + 3) & -4);
 269                 }
 270         } else
 271                 run(OBJECT_SIZE);
 272
 273         return 0;
 274 }