gpu_info_rocm.c 6.4 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191
  1. #ifndef __APPLE__
  2. #include "gpu_info_rocm.h"
  3. #include <string.h>
  4. void rocm_init(char *rocm_lib_path, rocm_init_resp_t *resp) {
  5. rsmi_status_t ret;
  6. resp->err = NULL;
  7. const int buflen = 256;
  8. char buf[buflen + 1];
  9. int i;
  10. struct lookup {
  11. char *s;
  12. void **p;
  13. } l[] = {
  14. {"rsmi_init", (void *)&resp->rh.rsmi_init},
  15. {"rsmi_shut_down", (void *)&resp->rh.rsmi_shut_down},
  16. {"rsmi_dev_memory_total_get", (void *)&resp->rh.rsmi_dev_memory_total_get},
  17. {"rsmi_dev_memory_usage_get", (void *)&resp->rh.rsmi_dev_memory_usage_get},
  18. {"rsmi_version_get", (void *)&resp->rh.rsmi_version_get},
  19. {"rsmi_num_monitor_devices", (void*)&resp->rh.rsmi_num_monitor_devices},
  20. {"rsmi_dev_id_get", (void*)&resp->rh.rsmi_dev_id_get},
  21. {"rsmi_dev_name_get", (void *)&resp->rh.rsmi_dev_name_get},
  22. {"rsmi_dev_brand_get", (void *)&resp->rh.rsmi_dev_brand_get},
  23. {"rsmi_dev_vendor_name_get", (void *)&resp->rh.rsmi_dev_vendor_name_get},
  24. {"rsmi_dev_vram_vendor_get", (void *)&resp->rh.rsmi_dev_vram_vendor_get},
  25. {"rsmi_dev_serial_number_get", (void *)&resp->rh.rsmi_dev_serial_number_get},
  26. {"rsmi_dev_subsystem_name_get", (void *)&resp->rh.rsmi_dev_subsystem_name_get},
  27. {"rsmi_dev_vbios_version_get", (void *)&resp->rh.rsmi_dev_vbios_version_get},
  28. {NULL, NULL},
  29. };
  30. resp->rh.handle = LOAD_LIBRARY(rocm_lib_path, RTLD_LAZY);
  31. if (!resp->rh.handle) {
  32. char *msg = LOAD_ERR();
  33. snprintf(buf, buflen,
  34. "Unable to load %s library to query for Radeon GPUs: %s\n",
  35. rocm_lib_path, msg);
  36. free(msg);
  37. resp->err = strdup(buf);
  38. return;
  39. }
  40. // TODO once we've squashed the remaining corner cases remove this log
  41. LOG(resp->rh.verbose, "wiring rocm management library functions in %s\n", rocm_lib_path);
  42. for (i = 0; l[i].s != NULL; i++) {
  43. // TODO once we've squashed the remaining corner cases remove this log
  44. LOG(resp->rh.verbose, "dlsym: %s\n", l[i].s);
  45. *l[i].p = LOAD_SYMBOL(resp->rh.handle, l[i].s);
  46. if (!l[i].p) {
  47. resp->rh.handle = NULL;
  48. char *msg = LOAD_ERR();
  49. LOG(resp->rh.verbose, "dlerr: %s\n", msg);
  50. UNLOAD_LIBRARY(resp->rh.handle);
  51. snprintf(buf, buflen, "symbol lookup for %s failed: %s", l[i].s,
  52. msg);
  53. free(msg);
  54. resp->err = strdup(buf);
  55. return;
  56. }
  57. }
  58. ret = (*resp->rh.rsmi_init)(0);
  59. if (ret != RSMI_STATUS_SUCCESS) {
  60. LOG(resp->rh.verbose, "rsmi_init err: %d\n", ret);
  61. UNLOAD_LIBRARY(resp->rh.handle);
  62. resp->rh.handle = NULL;
  63. snprintf(buf, buflen, "rocm vram init failure: %d", ret);
  64. resp->err = strdup(buf);
  65. }
  66. return;
  67. }
  68. void rocm_check_vram(rocm_handle_t h, mem_info_t *resp) {
  69. resp->err = NULL;
  70. uint64_t totalMem = 0;
  71. uint64_t usedMem = 0;
  72. rsmi_status_t ret;
  73. const int buflen = 256;
  74. char buf[buflen + 1];
  75. int i;
  76. if (h.handle == NULL) {
  77. resp->err = strdup("rocm handle not initialized");
  78. return;
  79. }
  80. ret = (*h.rsmi_num_monitor_devices)(&resp->count);
  81. if (ret != RSMI_STATUS_SUCCESS) {
  82. snprintf(buf, buflen, "unable to get device count: %d", ret);
  83. resp->err = strdup(buf);
  84. return;
  85. }
  86. LOG(h.verbose, "discovered %d ROCm GPU Devices\n", resp->count);
  87. resp->total = 0;
  88. resp->free = 0;
  89. for (i = 0; i < resp->count; i++) {
  90. if (h.verbose) {
  91. // When in verbose mode, report more information about
  92. // the card we discover, but don't fail on error
  93. ret = (*h.rsmi_dev_name_get)(i, buf, buflen);
  94. if (ret != RSMI_STATUS_SUCCESS) {
  95. LOG(h.verbose, "rsmi_dev_name_get failed: %d\n", ret);
  96. } else {
  97. LOG(h.verbose, "[%d] ROCm device name: %s\n", i, buf);
  98. }
  99. ret = (*h.rsmi_dev_brand_get)(i, buf, buflen);
  100. if (ret != RSMI_STATUS_SUCCESS) {
  101. LOG(h.verbose, "rsmi_dev_brand_get failed: %d\n", ret);
  102. } else {
  103. LOG(h.verbose, "[%d] ROCm brand: %s\n", i, buf);
  104. }
  105. ret = (*h.rsmi_dev_vendor_name_get)(i, buf, buflen);
  106. if (ret != RSMI_STATUS_SUCCESS) {
  107. LOG(h.verbose, "rsmi_dev_vendor_name_get failed: %d\n", ret);
  108. } else {
  109. LOG(h.verbose, "[%d] ROCm vendor: %s\n", i, buf);
  110. }
  111. ret = (*h.rsmi_dev_vram_vendor_get)(i, buf, buflen);
  112. if (ret != RSMI_STATUS_SUCCESS) {
  113. LOG(h.verbose, "rsmi_dev_vram_vendor_get failed: %d\n", ret);
  114. } else {
  115. LOG(h.verbose, "[%d] ROCm VRAM vendor: %s\n", i, buf);
  116. }
  117. ret = (*h.rsmi_dev_serial_number_get)(i, buf, buflen);
  118. if (ret != RSMI_STATUS_SUCCESS) {
  119. LOG(h.verbose, "rsmi_dev_serial_number_get failed: %d\n", ret);
  120. } else {
  121. LOG(h.verbose, "[%d] ROCm S/N: %s\n", i, buf);
  122. }
  123. ret = (*h.rsmi_dev_subsystem_name_get)(i, buf, buflen);
  124. if (ret != RSMI_STATUS_SUCCESS) {
  125. LOG(h.verbose, "rsmi_dev_subsystem_name_get failed: %d\n", ret);
  126. } else {
  127. LOG(h.verbose, "[%d] ROCm subsystem name: %s\n", i, buf);
  128. }
  129. ret = (*h.rsmi_dev_vbios_version_get)(i, buf, buflen);
  130. if (ret != RSMI_STATUS_SUCCESS) {
  131. LOG(h.verbose, "rsmi_dev_vbios_version_get failed: %d\n", ret);
  132. } else {
  133. LOG(h.verbose, "[%d] ROCm vbios version: %s\n", i, buf);
  134. }
  135. }
  136. // Get total memory - used memory for available memory
  137. ret = (*h.rsmi_dev_memory_total_get)(i, RSMI_MEM_TYPE_VRAM, &totalMem);
  138. if (ret != RSMI_STATUS_SUCCESS) {
  139. snprintf(buf, buflen, "rocm total mem lookup failure: %d", ret);
  140. resp->err = strdup(buf);
  141. return;
  142. }
  143. ret = (*h.rsmi_dev_memory_usage_get)(i, RSMI_MEM_TYPE_VRAM, &usedMem);
  144. if (ret != RSMI_STATUS_SUCCESS) {
  145. snprintf(buf, buflen, "rocm usage mem lookup failure: %d", ret);
  146. resp->err = strdup(buf);
  147. return;
  148. }
  149. LOG(h.verbose, "[%d] ROCm totalMem %ld\n", i, totalMem);
  150. LOG(h.verbose, "[%d] ROCm usedMem %ld\n", i, usedMem);
  151. resp->total += totalMem;
  152. resp->free += totalMem - usedMem;
  153. }
  154. }
  155. void rocm_get_version(rocm_handle_t h, rocm_version_resp_t *resp) {
  156. const int buflen = 256;
  157. char buf[buflen + 1];
  158. if (h.handle == NULL) {
  159. resp->str = strdup("nvml handle not initialized");
  160. resp->status = 1;
  161. return;
  162. }
  163. rsmi_version_t ver;
  164. rsmi_status_t ret;
  165. ret = h.rsmi_version_get(&ver);
  166. if (ret != RSMI_STATUS_SUCCESS) {
  167. snprintf(buf, buflen, "unexpected response on version lookup %d", ret);
  168. resp->status = 1;
  169. } else {
  170. snprintf(buf, buflen, "%d", ver.major);
  171. resp->status = 0;
  172. }
  173. resp->str = strdup(buf);
  174. }
  175. #endif // __APPLE__