|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
|                                                                                   ROCM TRACER (ACTIVITY API)                                                                                   |
|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
|                                                          LABEL                                                           | COUNT  | DEPTH  |  METRIC   | UNITS  |   SUM    |   MEAN   | % SELF |
|--------------------------------------------------------------------------------------------------------------------------|--------|--------|-----------|--------|----------|----------|--------|
| |0>>> hipSetDevice                                                                                                       |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipGetDevice                                                                                                       |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipGetDeviceProperties                                                                                             |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipMemGetInfo                                                                                                      |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipMalloc                                                                                                          |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipMemset                                                                                                          |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipDeviceSynchronize                                                                                               |      2 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipLaunchKernel                                                                                                    |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipGetLastError                                                                                                    |    496 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      1 |      1 | roctracer | sec    | 0.000180 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*)                                                     |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 2u>(float, float*)                                                           |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      1 |      1 | roctracer | sec    | 0.000295 | 0.000295 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000294 | 0.000294 |  100.0 |
| |0>>> hipEventCreate                                                                                                     |    990 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*)                                                     |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 1u>(double, double*)                                                        |      2 |      1 | roctracer | sec    | 0.000360 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*)                                                     |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      1 |      1 | roctracer | sec    | 0.000181 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 2u>(double, double*)                                                        |      2 |      1 | roctracer | sec    | 0.000357 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*)                                                     |      3 |      1 | roctracer | sec    | 0.000276 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 2u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000276 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 3u>(float, float*)                                                           |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      3 |      1 | roctracer | sec    | 0.000541 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 3u>(double, double*)                                                        |      3 |      1 | roctracer | sec    | 0.000538 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*)                                                     |      3 |      1 | roctracer | sec    | 0.000276 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 3u>(int, int*)                                                                 |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 4u>(float, float*)                                                           |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      3 |      1 | roctracer | sec    | 0.000541 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 4u>(double, double*)                                                        |      3 |      1 | roctracer | sec    | 0.000539 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*)                                                     |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 4u>(int, int*)                                                                 |      2 |      1 | roctracer | sec    | 0.000185 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 5u>(float, float*)                                                           |      3 |      1 | roctracer | sec    | 0.000275 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      3 |      1 | roctracer | sec    | 0.000540 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 5u>(double, double*)                                                        |      1 |      1 | roctracer | sec    | 0.000179 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*)                                                     |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 5u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000279 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 6u>(float, float*)                                                           |      3 |      1 | roctracer | sec    | 0.000276 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      2 |      1 | roctracer | sec    | 0.000359 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 6u>(double, double*)                                                        |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*)                                                     |      3 |      1 | roctracer | sec    | 0.000277 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 6u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000278 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 7u>(float, float*)                                                           |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      2 |      1 | roctracer | sec    | 0.000360 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 7u>(double, double*)                                                        |      3 |      1 | roctracer | sec    | 0.000539 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*)                                                     |      3 |      1 | roctracer | sec    | 0.000277 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 7u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000279 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 8u>(float, float*)                                                           |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      2 |      1 | roctracer | sec    | 0.000360 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 8u>(double, double*)                                                        |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*)                                                     |      2 |      1 | roctracer | sec    | 0.000185 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 8u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000278 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 9u>(float, float*)                                                           |      3 |      1 | roctracer | sec    | 0.000276 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      3 |      1 | roctracer | sec    | 0.000538 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 9u>(double, double*)                                                        |      3 |      1 | roctracer | sec    | 0.000535 | 0.000178 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*)                                                     |      2 |      1 | roctracer | sec    | 0.000186 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 9u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000280 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 10u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000277 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000534 | 0.000178 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 10u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000359 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000277 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 10u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000292 | 0.000097 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 11u>(float, float*)                                                          |      2 |      1 | roctracer | sec    | 0.000185 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000539 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 11u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000358 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000279 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 11u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000304 | 0.000101 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 12u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000277 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000542 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 12u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*)                                                    |      2 |      1 | roctracer | sec    | 0.000186 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 12u>(int, int*)                                                                |      2 |      1 | roctracer | sec    | 0.000211 | 0.000106 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 13u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000278 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000540 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 13u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000359 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000278 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 13u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000333 | 0.000111 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 14u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000278 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      2 |      1 | roctracer | sec    | 0.000362 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 14u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000361 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000280 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 14u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000352 | 0.000117 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 15u>(float, float*)                                                          |      2 |      1 | roctracer | sec    | 0.000185 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      2 |      1 | roctracer | sec    | 0.000362 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 15u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000280 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 15u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000369 | 0.000123 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 16u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000278 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000536 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 16u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000536 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 16u>(int, int*)                                                                |      1 |      1 | roctracer | sec    | 0.000128 | 0.000128 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000538 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 17u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000535 | 0.000178 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000280 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 17u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000404 | 0.000135 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 18u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000280 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000536 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 18u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000283 | 0.000094 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 18u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000422 | 0.000141 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 20u>(float, float*)                                                          |      2 |      1 | roctracer | sec    | 0.000187 | 0.000094 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 20u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000296 | 0.000099 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 20u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000461 | 0.000154 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 22u>(float, float*)                                                          |      2 |      1 | roctracer | sec    | 0.000187 | 0.000094 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000537 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 22u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000535 | 0.000178 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000103 | 0.000103 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 22u>(int, int*)                                                                |      2 |      1 | roctracer | sec    | 0.000334 | 0.000167 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 24u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000281 | 0.000094 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000536 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 24u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000359 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*)                                                    |      2 |      1 | roctracer | sec    | 0.000217 | 0.000108 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 24u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000536 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 28u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000300 | 0.000100 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      2 |      1 | roctracer | sec    | 0.000362 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 28u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000362 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000363 | 0.000121 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 28u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000612 | 0.000204 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 32u>(float, float*)                                                          |      2 |      1 | roctracer | sec    | 0.000215 | 0.000108 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      2 |      1 | roctracer | sec    | 0.000365 | 0.000183 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 32u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000544 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000396 | 0.000132 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 32u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.000689 | 0.000230 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 40u>(float, float*)                                                          |      2 |      1 | roctracer | sec    | 0.000249 | 0.000124 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000567 | 0.000189 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 40u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000572 | 0.000191 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*)                                                    |      2 |      1 | roctracer | sec    | 0.000315 | 0.000158 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 40u>(int, int*)                                                                |      2 |      1 | roctracer | sec    | 0.000562 | 0.000281 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000625 | 0.000208 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000550 | 0.000183 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 48u>(int, int*)                                                                |      2 |      1 | roctracer | sec    | 0.000663 | 0.000332 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 56u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000489 | 0.000163 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      3 |      1 | roctracer | sec    | 0.000687 | 0.000229 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 56u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000457 | 0.000228 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000629 | 0.000210 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 56u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.001148 | 0.000383 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 64u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000548 | 0.000183 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      2 |      1 | roctracer | sec    | 0.000502 | 0.000251 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 64u>(double, double*)                                                       |      3 |      1 | roctracer | sec    | 0.000732 | 0.000244 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.000704 | 0.000235 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 64u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.001344 | 0.000448 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 80u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000220 | 0.000220 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 80u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000299 | 0.000299 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*)                                                    |      2 |      1 | roctracer | sec    | 0.000588 | 0.000294 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 80u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.001655 | 0.000552 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 96u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000262 | 0.000262 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<fl... |      1 |      1 | roctracer | sec    | 0.000460 | 0.000460 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 256u>(double, double*)                                                      |      2 |      1 | roctracer | sec    | 0.001747 | 0.000873 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*)                                                   |      3 |      1 | roctracer | sec    | 0.002560 | 0.000853 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 256u>(int, int*)                                                               |      3 |      1 | roctracer | sec    | 0.005075 | 0.001692 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 512u>(float, float*)                                                         |      3 |      1 | roctracer | sec    | 0.004976 | 0.001659 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<fl... |      1 |      1 | roctracer | sec    | 0.001720 | 0.001720 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 512u>(double, double*)                                                      |      3 |      1 | roctracer | sec    | 0.005126 | 0.001709 |  100.0 |
| |0>>> hipExtLaunchKernel                                                                                                 |    495 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 0u>(float, float*)                                                           |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 0u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000275 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 5u>(double, double*)                                                        |      1 |      1 | roctracer | sec    | 0.000182 | 0.000182 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 11u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000179 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 13u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000180 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 15u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      1 |      1 | roctracer | sec    | 0.000179 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 17u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000279 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000103 | 0.000103 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      1 |      1 | roctracer | sec    | 0.000182 | 0.000182 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 28u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000180 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 48u>(float, float*)                                                          |      3 |      1 | roctracer | sec    | 0.000434 | 0.000145 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 48u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000208 | 0.000208 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 48u>(int, int*)                                                                |      1 |      1 | roctracer | sec    | 0.000331 | 0.000331 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 80u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000223 | 0.000223 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      2 |      1 | roctracer | sec    | 0.000593 | 0.000297 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 80u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000589 | 0.000294 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 96u>(float, float*)                                                          |      2 |      1 | roctracer | sec    | 0.000522 | 0.000261 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      2 |      1 | roctracer | sec    | 0.000695 | 0.000348 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<fl... |      2 |      1 | roctracer | sec    | 0.003435 | 0.001718 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*)                                                   |      1 |      1 | roctracer | sec    | 0.001669 | 0.001669 |  100.0 |
| |0>>> hipEventSynchronize                                                                                                |    495 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> |_void benchmark_func<short, 256, 8u, 0u>(short, short*)                                                           |      1 |      1 | roctracer | sec    | 0.000008 | 0.000008 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 0u>(float, float*)                                                           |      2 |      1 | roctracer | sec    | 0.000183 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      2 |      1 | roctracer | sec    | 0.000359 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 0u>(double, double*)                                                        |      3 |      1 | roctracer | sec    | 0.000538 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*)                                                     |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 1u>(float, float*)                                                           |      3 |      1 | roctracer | sec    | 0.000276 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      3 |      1 | roctracer | sec    | 0.000540 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 1u>(double, double*)                                                        |      1 |      1 | roctracer | sec    | 0.000179 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 1u>(int, int*)                                                                 |      3 |      1 | roctracer | sec    | 0.000276 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 2u>(float, float*)                                                           |      2 |      1 | roctracer | sec    | 0.000182 | 0.000091 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      1 |      1 | roctracer | sec    | 0.000180 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 80u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000220 | 0.000220 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      1 |      1 | roctracer | sec    | 0.000346 | 0.000346 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 96u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000349 | 0.000349 |  100.0 |
| |0>>> hipEventElapsedTime                                                                                                |    495 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipEventDestroy                                                                                                    |    990 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      1 |      1 | roctracer | sec    | 0.000181 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 2u>(double, double*)                                                        |      1 |      1 | roctracer | sec    | 0.000179 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 3u>(float, float*)                                                           |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 3u>(int, int*)                                                                 |      2 |      1 | roctracer | sec    | 0.000185 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 4u>(float, float*)                                                           |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*)                                                     |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 4u>(int, int*)                                                                 |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 5u>(double, double*)                                                        |      1 |      1 | roctracer | sec    | 0.000179 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*)                                                     |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      1 |      1 | roctracer | sec    | 0.000181 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 7u>(float, float*)                                                           |      2 |      1 | roctracer | sec    | 0.000184 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      1 |      1 | roctracer | sec    | 0.000180 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 8u>(float, float*)                                                           |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<floa... |      1 |      1 | roctracer | sec    | 0.000180 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*)                                                     |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*)                                                     |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 10u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000179 | 0.000179 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 11u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000092 | 0.000092 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 12u>(int, int*)                                                                |      1 |      1 | roctracer | sec    | 0.000106 | 0.000106 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      1 |      1 | roctracer | sec    | 0.000181 | 0.000181 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 14u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000178 | 0.000178 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*)                                                    |      2 |      1 | roctracer | sec    | 0.000187 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 16u>(int, int*)                                                                |      2 |      1 | roctracer | sec    | 0.000258 | 0.000129 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 20u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 22u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000093 | 0.000093 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000103 | 0.000103 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 22u>(int, int*)                                                                |      1 |      1 | roctracer | sec    | 0.000166 | 0.000166 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 24u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000178 | 0.000178 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000108 | 0.000108 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 32u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000107 | 0.000107 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      1 |      1 | roctracer | sec    | 0.000180 | 0.000180 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 40u>(float, float*)                                                          |      1 |      1 | roctracer | sec    | 0.000125 | 0.000125 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*)                                                    |      1 |      1 | roctracer | sec    | 0.000158 | 0.000158 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 40u>(int, int*)                                                                |      1 |      1 | roctracer | sec    | 0.000280 | 0.000280 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 48u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000409 | 0.000204 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 56u>(double, double*)                                                       |      1 |      1 | roctracer | sec    | 0.000226 | 0.000226 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<flo... |      1 |      1 | roctracer | sec    | 0.000249 | 0.000249 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 96u>(double, double*)                                                       |      2 |      1 | roctracer | sec    | 0.000692 | 0.000346 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*)                                                    |      3 |      1 | roctracer | sec    | 0.001032 | 0.000344 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 96u>(int, int*)                                                                |      3 |      1 | roctracer | sec    | 0.001966 | 0.000655 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 128u>(float, float*)                                                         |      3 |      1 | roctracer | sec    | 0.001327 | 0.000442 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<fl... |      2 |      1 | roctracer | sec    | 0.000919 | 0.000459 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 128u>(double, double*)                                                      |      3 |      1 | roctracer | sec    | 0.001369 | 0.000456 |  100.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*)                                                   |      3 |      1 | roctracer | sec    | 0.001338 | 0.000446 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 128u>(int, int*)                                                               |      3 |      1 | roctracer | sec    | 0.002589 | 0.000863 |  100.0 |
| |0>>> |_void benchmark_func<float, 256, 8u, 256u>(float, float*)                                                         |      3 |      1 | roctracer | sec    | 0.002544 | 0.000848 |  100.0 |
| |0>>> |_void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<fl... |      3 |      1 | roctracer | sec    | 0.002637 | 0.000879 |  100.0 |
| |0>>> |_void benchmark_func<double, 256, 8u, 256u>(double, double*)                                                      |      1 |      1 | roctracer | sec    | 0.000875 | 0.000875 |  100.0 |
| |0>>> hipMemcpy                                                                                                          |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> |_void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*)                                                   |      2 |      1 | roctracer | sec    | 0.003334 | 0.001667 |  100.0 |
| |0>>> |_void benchmark_func<int, 256, 8u, 512u>(int, int*)                                                               |      2 |      1 | roctracer | sec    | 0.006699 | 0.003350 |  100.0 |
| |0>>> hipFree                                                                                                            |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
| |0>>> hipDeviceReset                                                                                                     |      1 |      0 | roctracer | sec    | 0.000000 | 0.000000 |    0.0 |
|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|