LXR linux/drivers/gpu/drm/amd/amdgpu/amdgpu

   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
  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
  31static 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 dma_fence *fence;
  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                                       false, false, false);
  44                if (r)
  45                        goto exit_do_move;
  46                r = dma_fence_wait(fence, false);
  47                dma_fence_put(fence);
  48                if (r)
  49                        goto exit_do_move;
  50        }
  51        end_jiffies = jiffies;
  52        r = jiffies_to_msecs(end_jiffies - start_jiffies);
  53
  54exit_do_move:
  55        return r;
  56}
  57
  58
  59static void amdgpu_benchmark_log_results(int n, unsigned size,
  60                                         unsigned int time,
  61                                         unsigned sdomain, unsigned ddomain,
  62                                         char *kind)
  63{
  64        unsigned int throughput = (n * (size >> 10)) / time;
  65        DRM_INFO("amdgpu: %s %u bo moves of %u kB from"
  66                 " %d to %d in %u ms, throughput: %u Mb/s or %u MB/s\n",
  67                 kind, n, size >> 10, sdomain, ddomain, time,
  68                 throughput * 8, throughput);
  69}
  70
  71static void amdgpu_benchmark_move(struct amdgpu_device *adev, unsigned size,
  72                                  unsigned sdomain, unsigned ddomain)
  73{
  74        struct amdgpu_bo *dobj = NULL;
  75        struct amdgpu_bo *sobj = NULL;
  76        struct amdgpu_bo_param bp;
  77        uint64_t saddr, daddr;
  78        int r, n;
  79        int time;
  80
  81        memset(&bp, 0, sizeof(bp));
  82        bp.size = size;
  83        bp.byte_align = PAGE_SIZE;
  84        bp.domain = sdomain;
  85        bp.flags = 0;
  86        bp.type = ttm_bo_type_kernel;
  87        bp.resv = NULL;
  88        bp.bo_ptr_size = sizeof(struct amdgpu_bo);
  89
  90        n = AMDGPU_BENCHMARK_ITERATIONS;
  91        r = amdgpu_bo_create(adev, &bp, &sobj);
  92        if (r) {
  93                goto out_cleanup;
  94        }
  95        r = amdgpu_bo_reserve(sobj, false);
  96        if (unlikely(r != 0))
  97                goto out_cleanup;
  98        r = amdgpu_bo_pin(sobj, sdomain);
  99        if (r) {
 100                amdgpu_bo_unreserve(sobj);
 101                goto out_cleanup;
 102        }
 103        r = amdgpu_ttm_alloc_gart(&sobj->tbo);
 104        amdgpu_bo_unreserve(sobj);
 105        if (r) {
 106                goto out_cleanup;
 107        }
 108        saddr = amdgpu_bo_gpu_offset(sobj);
 109        bp.domain = ddomain;
 110        r = amdgpu_bo_create(adev, &bp, &dobj);
 111        if (r) {
 112                goto out_cleanup;
 113        }
 114        r = amdgpu_bo_reserve(dobj, false);
 115        if (unlikely(r != 0))
 116                goto out_cleanup;
 117        r = amdgpu_bo_pin(dobj, ddomain);
 118        if (r) {
 119                amdgpu_bo_unreserve(sobj);
 120                goto out_cleanup;
 121        }
 122        r = amdgpu_ttm_alloc_gart(&dobj->tbo);
 123        amdgpu_bo_unreserve(dobj);
 124        if (r) {
 125                goto out_cleanup;
 126        }
 127        daddr = amdgpu_bo_gpu_offset(dobj);
 128
 129        if (adev->mman.buffer_funcs) {
 130                time = amdgpu_benchmark_do_move(adev, size, saddr, daddr, n);
 131                if (time < 0)
 132                        goto out_cleanup;
 133                if (time > 0)
 134                        amdgpu_benchmark_log_results(n, size, time,
 135                                                     sdomain, ddomain, "dma");
 136        }
 137
 138out_cleanup:
 139        /* Check error value now. The value can be overwritten when clean up.*/
 140        if (r) {
 141                DRM_ERROR("Error while benchmarking BO move.\n");
 142        }
 143
 144        if (sobj) {
 145                r = amdgpu_bo_reserve(sobj, true);
 146                if (likely(r == 0)) {
 147                        amdgpu_bo_unpin(sobj);
 148                        amdgpu_bo_unreserve(sobj);
 149                }
 150                amdgpu_bo_unref(&sobj);
 151        }
 152        if (dobj) {
 153                r = amdgpu_bo_reserve(dobj, true);
 154                if (likely(r == 0)) {
 155                        amdgpu_bo_unpin(dobj);
 156                        amdgpu_bo_unreserve(dobj);
 157                }
 158                amdgpu_bo_unref(&dobj);
 159        }
 160}
 161
 162void amdgpu_benchmark(struct amdgpu_device *adev, int test_number)
 163{
 164        int i;
 165        static const int common_modes[AMDGPU_BENCHMARK_COMMON_MODES_N] = {
 166                640 * 480 * 4,
 167                720 * 480 * 4,
 168                800 * 600 * 4,
 169                848 * 480 * 4,
 170                1024 * 768 * 4,
 171                1152 * 768 * 4,
 172                1280 * 720 * 4,
 173                1280 * 800 * 4,
 174                1280 * 854 * 4,
 175                1280 * 960 * 4,
 176                1280 * 1024 * 4,
 177                1440 * 900 * 4,
 178                1400 * 1050 * 4,
 179                1680 * 1050 * 4,
 180                1600 * 1200 * 4,
 181                1920 * 1080 * 4,
 182                1920 * 1200 * 4
 183        };
 184
 185        switch (test_number) {
 186        case 1:
 187                /* simple test, VRAM to GTT and GTT to VRAM */
 188                amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_GTT,
 189                                      AMDGPU_GEM_DOMAIN_VRAM);
 190                amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_VRAM,
 191                                      AMDGPU_GEM_DOMAIN_GTT);
 192                break;
 193        case 2:
 194                /* simple test, VRAM to VRAM */
 195                amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_VRAM,
 196                                      AMDGPU_GEM_DOMAIN_VRAM);
 197                break;
 198        case 3:
 199                /* GTT to VRAM, buffer size sweep, powers of 2 */
 200                for (i = 1; i <= 16384; i <<= 1)
 201                        amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE,
 202                                              AMDGPU_GEM_DOMAIN_GTT,
 203                                              AMDGPU_GEM_DOMAIN_VRAM);
 204                break;
 205        case 4:
 206                /* VRAM to GTT, buffer size sweep, powers of 2 */
 207                for (i = 1; i <= 16384; i <<= 1)
 208                        amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE,
 209                                              AMDGPU_GEM_DOMAIN_VRAM,
 210                                              AMDGPU_GEM_DOMAIN_GTT);
 211                break;
 212        case 5:
 213                /* VRAM to VRAM, buffer size sweep, powers of 2 */
 214                for (i = 1; i <= 16384; i <<= 1)
 215                        amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE,
 216                                              AMDGPU_GEM_DOMAIN_VRAM,
 217                                              AMDGPU_GEM_DOMAIN_VRAM);
 218                break;
 219        case 6:
 220                /* GTT to VRAM, buffer size sweep, common modes */
 221                for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++)
 222                        amdgpu_benchmark_move(adev, common_modes[i],
 223                                              AMDGPU_GEM_DOMAIN_GTT,
 224                                              AMDGPU_GEM_DOMAIN_VRAM);
 225                break;
 226        case 7:
 227                /* VRAM to GTT, buffer size sweep, common modes */
 228                for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++)
 229                        amdgpu_benchmark_move(adev, common_modes[i],
 230                                              AMDGPU_GEM_DOMAIN_VRAM,
 231                                              AMDGPU_GEM_DOMAIN_GTT);
 232                break;
 233        case 8:
 234                /* VRAM to VRAM, buffer size sweep, common modes */
 235                for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++)
 236                        amdgpu_benchmark_move(adev, common_modes[i],
 237                                              AMDGPU_GEM_DOMAIN_VRAM,
 238                                              AMDGPU_GEM_DOMAIN_VRAM);
 239                break;
 240
 241        default:
 242                DRM_ERROR("Unknown benchmark\n");
 243        }
 244}
 245