- Cuda Toolkit
- Unix (Tested on Ubuntu)
You may need to modify the make file to point to the right CUDA and CUPTI directories
- git clone https://github.com/rafsalas19/cupti_profile.git
- cd cupti_profile
- make
This will generate a shared library /build/libcuProfile.so It will also generate an executable /build/cupti_profile (this is still in the development phase but can list avalable metrics).
The shared libray must be preloaded when launching the executable you would like to profile i.e.
env LD_PRELOAD=./build/libcuProfile.so LD_LIBRARY_PATH=${LD_LIBRARY_PATH}:`pwd` <myexe>
I have provided a run.sh file that can be used as an example or as a launching script:
./run.sh <myexe>
Some modification to file paths may be required
Set these environmental variables to control profiling
- "CUPTI_METRIC_CODES" - Provide comma separated metric codes for the applicable metrics you would like to see.
- "INJECTION_KERNEL_COUNT" - Provide the number of kernel launches you would like to profile per CUDA context.
| Code | Metric | Code | Metric | Code | Metric |
|---|---|---|---|---|---|
| 1000 | "achieved_occupancy" | 1001 | "branch_efficiency" | 1002 | "cf_executed" |
| 1003 | "double_precision_fu_utilization" | 1004 | "dram_read_bytes" | 1005 | "dram_read_throughput" |
| 1006 | "dram_read_transactions" | 1007 | "dram_utilization" | 1008 | "dram_write_bytes" |
| 1010 | "dram_write_transactions" | 1011 | "eligible_warps_per_cycle" | 1012 | "flop_count_dp_add" |
| 1013 | "flop_count_dp_fma" | 1014 | "flop_count_dp_mul" | 1015 | "flop_count_hp_add" |
| 1016 | "flop_count_hp_fma" | 1017 | "flop_count_hp_mul" | 1018 | "flop_count_sp_add" |
| 1019 | "flop_count_sp_fma" | 1020 | "flop_count_sp_mul" | 1021 | "flop_dp_efficiency" |
| 1022 | "flop_hp_efficiency" | 1023 | "flop_sp_efficiency" | 1024 | "gld_efficiency" |
| 1025 | "gld_throughput" | 1026 | "gld_transactions" | 1027 | "gld_transactions_per_request" |
| 1028 | "global_atomic_requests" | 1029 | "global_load_requests" | 1030 | "global_reduction_requests" |
| 1031 | "global_store_requests" | 1032 | "gst_efficiency" | 1033 | "gst_throughput" |
| 1034 | "gst_transactions" | 1035 | "gst_transactions_per_request" | 1036 | "half_precision_fu_utilization" |
| 1037 | "inst_bit_convert" | 1038 | "inst_compute_ld_st" | 1039 | "inst_control" |
| 1040 | "inst_executed" | 1041 | "inst_executed_global_atomics" | 1042 | "inst_executed_global_loads" |
| 1043 | "inst_executed_global_reductions" | 1044 | "inst_executed_global_stores" | 1045 | "inst_executed_local_loads" |
| 1046 | "inst_executed_local_stores" | 1047 | "inst_executed_shared_loads" | 1048 | "inst_executed_shared_stores" |
| 1049 | "inst_executed_surface_atomics" | 1050 | "inst_executed_surface_reductions" | 1051 | "inst_executed_surface_stores" |
| 1052 | "inst_executed_tex_ops" | 1053 | "inst_fp_16" | 1054 | "inst_fp_32" |
| 1055 | "inst_fp_64" | 1056 | "inst_integer" | 1057 | "inst_inter_thread_communication" |
| 1058 | "inst_issued" | 1059 | "inst_misc" | 1060 | "inst_per_warp" |
| 1061 | "ipc" | 1062 | "issue_slot_utilization" | 1063 | "issue_slots" |
| 1064 | "issued_ipc" | 1065 | "l1_sm_lg_utilization" | 1066 | "l2_atomic_throughput" |
| 1067 | "l2_atomic_transactions" | 1068 | "l2_global_atomic_store_bytes" | 1069 | "l2_global_load_bytes" |
| 1070 | "l2_local_load_bytes" | 1071 | "l2_read_throughput" | 1072 | "l2_read_transactions" |
| 1073 | "l2_surface_load_bytes" | 1074 | "l2_surface_store_bytes" | 1075 | "l2_tex_hit_rate" |
| 1076 | "l2_tex_read_hit_rate" | 1077 | "l2_tex_read_throughput" | 1078 | "l2_tex_read_transactions" |
| 1079 | "l2_tex_write_hit_rate" | 1080 | "l2_tex_write_throughput" | 1081 | "l2_tex_write_transactions" |
| 1082 | "l2_utilization" | 1083 | "l2_write_throughput" | 1084 | "l2_write_transactions" |
| 1085 | "ldst_fu_utilization" | 1086 | "local_load_requests" | 1087 | "local_load_throughput" |
| 1088 | "local_load_transactions" | 1089 | "local_load_transactions_per_request" | 1090 | "local_store_requests" |
| 1091 | "local_store_throughput" | 1092 | "local_store_transactions" | 1093 | "local_store_transactions_per_request" |
| 1094 | "pcie_total_data_received" | 1095 | "pcie_total_data_transmitted" | 1096 | "shared_efficiency" |
| 1097 | "shared_load_throughput" | 1098 | "shared_load_transactions" | 1099 | "shared_store_throughput" |
| 1100 | "shared_store_transactions" | 1101 | "shared_utilization" | 1102 | "single_precision_fu_utilization" |
| 1103 | "sm_efficiency" | 1104 | "sm_tex_utilization" | 1105 | "special_fu_utilization" |
| 1106 | "stall_constant_memory_dependency" | 1107 | "stall_inst_fetch" | 1108 | "stall_memory_dependency" |
| 1109 | "stall_not_selected" | 1110 | "stall_sleeping" | 1111 | "stall_texture" |
| 1112 | "surface_atomic_requests" | 1113 | "surface_load_requests" | 1114 | "surface_reduction_requests" |
| 1115 | "surface_store_requests" | 1116 | "sysmem_read_bytes" | 1117 | "sysmem_read_throughput" |
| 1118 | "sysmem_read_transactions" | 1119 | "sysmem_write_bytes" | 1120 | "sysmem_write_throughput" |
| 1121 | "sysmem_write_transactions" | 1122 | "tensor_precision_fu_utilization" | 1123 | "tex_cache_hit_rate" |
| 1124 | "tex_fu_utilization" | 1125 | "tex_sm_tex_utilization" | 1126 | "tex_sm_utilization" |
| 1127 | "texture_load_requests" | 1128 | "warp_execution_efficiency" | 1129 | "warp_nonpred_execution_efficiency" |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent