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 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 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 54 exit_do_move: 55 return r; 56 } 57 58 59 static void amdgpu_benchmark_log_results(struct amdgpu_device *adev, 60 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 67 dev_info(adev->dev, "amdgpu: %s %u bo moves of %u kB from" 68 " %d to %d in %u ms, throughput: %u Mb/s or %u MB/s\n", 69 kind, n, size >> 10, sdomain, ddomain, time, 70 throughput * 8, throughput); 71 } 72 73 static int amdgpu_benchmark_move(struct amdgpu_device *adev, unsigned size, 74 unsigned sdomain, unsigned ddomain) 75 { 76 struct amdgpu_bo *dobj = NULL; 77 struct amdgpu_bo *sobj = NULL; 78 struct amdgpu_bo_param bp; 79 uint64_t saddr, daddr; 80 int r, n; 81 82 memset(&bp, 0, sizeof(bp)); 83 bp.size = size; 84 bp.byte_align = PAGE_SIZE; 85 bp.domain = sdomain; 86 bp.flags = 0; 87 bp.type = ttm_bo_type_kernel; 88 bp.resv = NULL; 89 bp.bo_ptr_size = sizeof(struct amdgpu_bo); 90 91 n = AMDGPU_BENCHMARK_ITERATIONS; 92 r = amdgpu_bo_create(adev, &bp, &sobj); 93 if (r) { 94 goto out_cleanup; 95 } 96 r = amdgpu_bo_reserve(sobj, false); 97 if (unlikely(r != 0)) 98 goto out_cleanup; 99 r = amdgpu_bo_pin(sobj, sdomain); 100 if (r) { 101 amdgpu_bo_unreserve(sobj); 102 goto out_cleanup; 103 } 104 r = amdgpu_ttm_alloc_gart(&sobj->tbo); 105 amdgpu_bo_unreserve(sobj); 106 if (r) { 107 goto out_cleanup; 108 } 109 saddr = amdgpu_bo_gpu_offset(sobj); 110 bp.domain = ddomain; 111 r = amdgpu_bo_create(adev, &bp, &dobj); 112 if (r) { 113 goto out_cleanup; 114 } 115 r = amdgpu_bo_reserve(dobj, false); 116 if (unlikely(r != 0)) 117 goto out_cleanup; 118 r = amdgpu_bo_pin(dobj, ddomain); 119 if (r) { 120 amdgpu_bo_unreserve(sobj); 121 goto out_cleanup; 122 } 123 r = amdgpu_ttm_alloc_gart(&dobj->tbo); 124 amdgpu_bo_unreserve(dobj); 125 if (r) { 126 goto out_cleanup; 127 } 128 daddr = amdgpu_bo_gpu_offset(dobj); 129 130 if (adev->mman.buffer_funcs) { 131 r = amdgpu_benchmark_do_move(adev, size, saddr, daddr, n); 132 if (r < 0) 133 goto out_cleanup; 134 if (r > 0) 135 amdgpu_benchmark_log_results(adev, n, size, r, 136 sdomain, ddomain, "dma"); 137 } 138 139 out_cleanup: 140 /* Check error value now. The value can be overwritten when clean up.*/ 141 if (r < 0) 142 dev_info(adev->dev, "Error while benchmarking BO move.\n"); 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 return r; 161 } 162 163 int amdgpu_benchmark(struct amdgpu_device *adev, int test_number) 164 { 165 int i, r; 166 static const int common_modes[AMDGPU_BENCHMARK_COMMON_MODES_N] = { 167 640 * 480 * 4, 168 720 * 480 * 4, 169 800 * 600 * 4, 170 848 * 480 * 4, 171 1024 * 768 * 4, 172 1152 * 768 * 4, 173 1280 * 720 * 4, 174 1280 * 800 * 4, 175 1280 * 854 * 4, 176 1280 * 960 * 4, 177 1280 * 1024 * 4, 178 1440 * 900 * 4, 179 1400 * 1050 * 4, 180 1680 * 1050 * 4, 181 1600 * 1200 * 4, 182 1920 * 1080 * 4, 183 1920 * 1200 * 4 184 }; 185 186 mutex_lock(&adev->benchmark_mutex); 187 switch (test_number) { 188 case 1: 189 dev_info(adev->dev, 190 "benchmark test: %d (simple test, VRAM to GTT and GTT to VRAM)\n", 191 test_number); 192 /* simple test, VRAM to GTT and GTT to VRAM */ 193 r = amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_GTT, 194 AMDGPU_GEM_DOMAIN_VRAM); 195 if (r) 196 goto done; 197 r = amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_VRAM, 198 AMDGPU_GEM_DOMAIN_GTT); 199 if (r) 200 goto done; 201 break; 202 case 2: 203 dev_info(adev->dev, 204 "benchmark test: %d (simple test, VRAM to VRAM)\n", 205 test_number); 206 /* simple test, VRAM to VRAM */ 207 r = amdgpu_benchmark_move(adev, 1024*1024, AMDGPU_GEM_DOMAIN_VRAM, 208 AMDGPU_GEM_DOMAIN_VRAM); 209 if (r) 210 goto done; 211 break; 212 case 3: 213 dev_info(adev->dev, 214 "benchmark test: %d (GTT to VRAM, buffer size sweep, powers of 2)\n", 215 test_number); 216 /* GTT to VRAM, buffer size sweep, powers of 2 */ 217 for (i = 1; i <= 16384; i <<= 1) { 218 r = amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE, 219 AMDGPU_GEM_DOMAIN_GTT, 220 AMDGPU_GEM_DOMAIN_VRAM); 221 if (r) 222 goto done; 223 } 224 break; 225 case 4: 226 dev_info(adev->dev, 227 "benchmark test: %d (VRAM to GTT, buffer size sweep, powers of 2)\n", 228 test_number); 229 /* VRAM to GTT, buffer size sweep, powers of 2 */ 230 for (i = 1; i <= 16384; i <<= 1) { 231 r = amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE, 232 AMDGPU_GEM_DOMAIN_VRAM, 233 AMDGPU_GEM_DOMAIN_GTT); 234 if (r) 235 goto done; 236 } 237 break; 238 case 5: 239 dev_info(adev->dev, 240 "benchmark test: %d (VRAM to VRAM, buffer size sweep, powers of 2)\n", 241 test_number); 242 /* VRAM to VRAM, buffer size sweep, powers of 2 */ 243 for (i = 1; i <= 16384; i <<= 1) { 244 r = amdgpu_benchmark_move(adev, i * AMDGPU_GPU_PAGE_SIZE, 245 AMDGPU_GEM_DOMAIN_VRAM, 246 AMDGPU_GEM_DOMAIN_VRAM); 247 if (r) 248 goto done; 249 } 250 break; 251 case 6: 252 dev_info(adev->dev, 253 "benchmark test: %d (GTT to VRAM, buffer size sweep, common modes)\n", 254 test_number); 255 /* GTT to VRAM, buffer size sweep, common modes */ 256 for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++) { 257 r = amdgpu_benchmark_move(adev, common_modes[i], 258 AMDGPU_GEM_DOMAIN_GTT, 259 AMDGPU_GEM_DOMAIN_VRAM); 260 if (r) 261 goto done; 262 } 263 break; 264 case 7: 265 dev_info(adev->dev, 266 "benchmark test: %d (VRAM to GTT, buffer size sweep, common modes)\n", 267 test_number); 268 /* VRAM to GTT, buffer size sweep, common modes */ 269 for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++) { 270 r = amdgpu_benchmark_move(adev, common_modes[i], 271 AMDGPU_GEM_DOMAIN_VRAM, 272 AMDGPU_GEM_DOMAIN_GTT); 273 if (r) 274 goto done; 275 } 276 break; 277 case 8: 278 dev_info(adev->dev, 279 "benchmark test: %d (VRAM to VRAM, buffer size sweep, common modes)\n", 280 test_number); 281 /* VRAM to VRAM, buffer size sweep, common modes */ 282 for (i = 0; i < AMDGPU_BENCHMARK_COMMON_MODES_N; i++) { 283 r = amdgpu_benchmark_move(adev, common_modes[i], 284 AMDGPU_GEM_DOMAIN_VRAM, 285 AMDGPU_GEM_DOMAIN_VRAM); 286 if (r) 287 goto done; 288 } 289 break; 290 291 default: 292 dev_info(adev->dev, "Unknown benchmark %d\n", test_number); 293 r = -EINVAL; 294 break; 295 } 296 297 done: 298 mutex_unlock(&adev->benchmark_mutex); 299 300 return r; 301 } 302