drivers/gpu/drm/amd/amdgpu/amdgpu_benchmark.c

   1 /*
   2  * Copyright 2009 Jerome Glisse.
   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 shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Jerome Glisse
  23  */
  24 #include <drm/drmP.h>
  25 #include <drm/amdgpu_drm.h>
  26 #include "amdgpu.h"
  27
  28 #define AMDGPU_BENCHMARK_ITERATIONS 1024
  29 #define AMDGPU_BENCHMARK_COMMON_MODES_N 17
  30
  31 static int amdgpu_benchmark_do_move(struct amdgpu_device *adev, unsigned size,
  32                                     uint64_t saddr, uint64_t daddr, int n)
  33 {
  34         unsigned long start_jiffies;
  35         unsigned long end_jiffies;
  36         struct fence *fence = NULL;
  37         int i, r;
  38
  39         start_jiffies = jiffies;
  40         for (i = 0; i < n; i++) {
  41                 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
  42                 r = amdgpu_copy_buffer(ring, saddr, daddr, size, NULL, &fence);
  43                 if (r)
  44                         goto exit_do_move;
  45                 r = fence_wait(fence, false);
  46                 if (r)
  47                         goto exit_do_move;
  48                 fence_put(fence);
  49         }
  50         end_jiffies = jiffies;
  51         r = jiffies_to_msecs(end_jiffies - start_jiffies);
  52
  53 exit_do_move:
  54         if (fence)
  55                 fence_put(fence);
  56         return r;
  57 }
  58
  59
  60 static void amdgpu_benchmark_log_results(int n, unsigned size,
  61                                          unsigned int time,
  62                                          unsigned sdomain, unsigned ddomain,
  63                                          char *kind)
  64 {
  65         unsigned int throughput = (n * (size >> 10)) / time;
  66         DRM_INFO("amdgpu: %s %u bo moves of %u kB from"
  67                  " %d to %d in %u ms, throughput: %u Mb/s or %u MB/s\n",
  68                  kind, n, size >> 10, sdomain, ddomain, time,
  69                  throughput * 8, throughput);
  70 }
  71
  72 static void amdgpu_benchmark_move(struct amdgpu_device *adev, unsigned size,
  73                                   unsigned sdomain, unsigned ddomain)
  74 {
  75         struct amdgpu_bo *dobj = NULL;
  76         struct amdgpu_bo *sobj = NULL;
  77         uint64_t saddr, daddr;
  78         int r, n;
  79         int time;
  80
  81         n = AMDGPU_BENCHMARK_ITERATIONS;
  82         r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, sdomain, 0, NULL, &sobj);
  83         if (r) {
  84                 goto out_cleanup;
  85         }
  86         r = amdgpu_bo_reserve(sobj, false);
  87         if (unlikely(r != 0))
  88                 goto out_cleanup;
  89         r = amdgpu_bo_pin(sobj, sdomain, &saddr);
  90         amdgpu_bo_unreserve(sobj);
  91         if (r) {
  92                 goto out_cleanup;
  93         }
  94         r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, ddomain, 0, NULL, &dobj);
  95         if (r) {
  96                 goto out_cleanup;
  97         }
  98         r = amdgpu_bo_reserve(dobj, false);
  99         if (unlikely(r != 0))
 100                 goto out_cleanup;
 101         r = amdgpu_bo_pin(dobj, ddomain, &daddr);
 102         amdgpu_bo_unreserve(dobj);
 103         if (r) {
 104                 goto out_cleanup;
 105         }
 106
 107         if (adev->mman.buffer_funcs) {
 108                 time = amdgpu_benchmark_do_move(adev, size, saddr, daddr, n);
 109                 if (time < 0)
 110                         goto out_cleanup;
 111                 if (time > 0)
 112                         amdgpu_benchmark_log_results(n, size, time,
 113                                                      sdomain, ddomain, "dma");
 114         }
 115
 116 out_cleanup:
 117         if (sobj) {
 118                 r = amdgpu_bo_reserve(sobj, false);
 119                 if (likely(r == 0)) {
 120                         amdgpu_bo_unpin(sobj);
 121                         amdgpu_bo_unreserve(sobj);
 122                 }
 123                 amdgpu_bo_unref(&sobj);
 124         }
 125         if (dobj) {
 126                 r = amdgpu_bo_reserve(dobj, false);
 127                 if (likely(r == 0)) {
 128                         amdgpu_bo_unpin(dobj);
 129                         amdgpu_bo_unreserve(dobj);
 130                 }
 131                 amdgpu_bo_unref(&dobj);
 132         }
 133
 134         if (r) {
 135                 DRM_ERROR("Error while benchmarking BO move.\n");
 136         }
 137 }
 138
 139 void amdgpu_benchmark(struct amdgpu_device *adev, int test_number)
 140 {
 141         int i;
 142         int common_modes[AMDGPU_BENCHMARK_COMMON_MODES_N] = {
 143                 640 * 480 * 4,
 144                 720 * 480 * 4,
 145                 800 * 600 * 4,
 146                 848 * 480 * 4,
 147                 1024 * 768 * 4,
 148                 1152 * 768 * 4,
 149                 1280 * 720 * 4,
 150                 1280 * 800 * 4,
 151                 1280 * 854 * 4,
 152                 1280 * 960 * 4,
 153                 1280 * 1024 * 4,
 154                 1440 * 900 * 4,
 155                 1400 * 1050 * 4,
 156                 1680 * 1050 * 4,
 157                 1600 * 1200 * 4,
 158                 1920 * 1080 * 4,
 159                 1920 * 1200 * 4
 160         };
 161
 162         switch (test_number) {
 163         case 1:
 164                 /* simple test, VRAM to GTT and GTT to VRAM */
 165                 amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_GTT,
 166                                       AMDGPU_GEM_DOMAIN_VRAM);
 167                 amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_VRAM,
 168                                       AMDGPU_GEM_DOMAIN_GTT);
 169                 break;
 170         case 2:
 171                 /* simple test, VRAM to VRAM */
 172                 amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_VRAM,
 173                                       AMDGPU_GEM_DOMAIN_VRAM);
 174                 break;
 175         case 3:
 176                 /* GTT to VRAM, buffer size sweep, powers of 2 */
 177                 for (i = 1; i <= 16384; i <<= 1)
 178                         amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE,
 179                                               AMDGPU_GEM_DOMAIN_GTT,
 180                                               AMDGPU_GEM_DOMAIN_VRAM);
 181                 break;
 182         case 4:
 183                 /* VRAM to GTT, buffer size sweep, powers of 2 */
 184                 for (i = 1; i <= 16384; i <<= 1)
 185                         amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE,
 186                                               AMDGPU_GEM_DOMAIN_VRAM,
 187                                               AMDGPU_GEM_DOMAIN_GTT);
 188                 break;
 189         case 5:
 190                 /* VRAM to VRAM, buffer size sweep, powers of 2 */
 191                 for (i = 1; i <= 16384; i <<= 1)
 192                         amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE,
 193                                               AMDGPU_GEM_DOMAIN_VRAM,
 194                                               AMDGPU_GEM_DOMAIN_VRAM);
 195                 break;
 196         case 6:
 197                 /* GTT to VRAM, buffer size sweep, common modes */
 198                 for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++)
 199                         amdgpu_benchmark_move(adev, common_modes[i],
 200                                               AMDGPU_GEM_DOMAIN_GTT,
 201                                               AMDGPU_GEM_DOMAIN_VRAM);
 202                 break;
 203         case 7:
 204                 /* VRAM to GTT, buffer size sweep, common modes */
 205                 for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++)
 206                         amdgpu_benchmark_move(adev, common_modes[i],
 207                                               AMDGPU_GEM_DOMAIN_VRAM,
 208                                               AMDGPU_GEM_DOMAIN_GTT);
 209                 break;
 210         case 8:
 211                 /* VRAM to VRAM, buffer size sweep, common modes */
 212                 for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++)
 213                         amdgpu_benchmark_move(adev, common_modes[i],
 214                                               AMDGPU_GEM_DOMAIN_VRAM,
 215                                               AMDGPU_GEM_DOMAIN_VRAM);
 216                 break;
 217
 218         default:
 219                 DRM_ERROR("Unknown benchmark\n");
 220         }
 221 }