API in readme, simplify required includes

This commit is contained in:
Carl Pearson
2019-09-20 08:48:26 -05:00
parent a61b3d7fe7
commit 72c39c1dd5
11 changed files with 121 additions and 12 deletions

110
README.md
View File

@@ -17,11 +17,119 @@ CPU/GPU performance control library for benchmarking
- [x] Disable GPU turbo (nvidia)
- [x] Flush addresses from cache (amd64, POWER)
## API
## Installing
### CMake
Ensure you have CMake 3.13+.
Add the source tree to your project and then use add_subdirectory
```
git submodule add git@github.com:cwpearson/perfect.git thirdparty/perfect
```
`CMakeLists.txt`
```
...
add_subdirectory(thirdparty/perfect)
...
target_link_libraries(your-target perfect)
```
### Without CMake
Download the source and add the include directory to your includes and `nvidia-ml` to your link flags
```
g++ code_using_perfect.cpp -I perfect/include -l nvidia-ml
nvcc code_using_perfect.cu -I perfect/include
```
## Usage
The `perfect` functions all return a `perfect::Result`, which is defined in [include/perfect/result.hpp].
When things are working, it will be `perfect::Result::SUCCESS`.
A `PERFECT` macro is also defined, which will terminate with an error message unless the `perfect::Result` is `perfect::Result::SUCCESS`.
```c++
perfect::CpuTurboState state;
PERFECT(perfect::get_cpu_turbo_state(&state));
```
### CPU Turbo
`perfect` can enable and disable CPU boost through the Intel p-state mechanism or the ACPI cpufreq mechanism.
See [examples/cpu_turbo.cpp].
```c++
#include "perfect/cpu_turbo.hpp"
```
* `Result get_cpu_turbo_state(CpuTurboState *state)`
* `Result set_cpu_turbo_state(CpuTurboState *state)`
* `Result disable_cpu_turbo()`
* `Result enable_cpu_turbo()`
* `bool is_turbo_enabled(CpuTurboState state)`
### OS Performance
`perfect` can control the OS governor on linux.
See [examples/os_perf.cpp].
```c++
#include "perfect/os_perf.hpp"
```
* `Result get_os_perf_state(OsPerfState *state, const int cpu)`: Save the current OS governor mode for CPU `cpu`.
* `Result os_perf_state_maximum(const int cpu)`: Set the OS governor to it's maximum performance mode.
* `Result set_os_perf_state(const int cpu, OsPerfState state)`: Restore a previously-saved OS governor mode.
### GPU Turbo
`perfect` can enable/disable GPU turbo boost.
See [examples/gpu_turbo.cu]
```c++
#include "perfect/gpu_turbo.hpp"
```
* `Result get_gpu_turbo_state(GpuTurboState *state, unsigned int idx)`: Get the current turbo state for GPU `idx`, useful to restore later.
* `bool is_turbo_enabled(GpuTurboState state)`: Check if turbo is enabled.
* `Result set_gpu_turbo_state(GpuTurboState state, unsigned int idx)`: Set a previously saved turbo state.
* `Result disable_gpu_turbo(unsigned int idx)`: Disable GPU `idx` turbo.
* `Result enable_gpu_turbo(unsigned int idx)`: Enable GPU `idx` turbo.
### GPU Clocks
`perfect` can lock GPU clocks to their maximum values.
See [examples/gpu_clocks.cu]
```c++
#include "perfect/gpu_clocks.hpp`
```
* `Result set_max_gpu_clocks(unsigned int idx)`: Set GPU `idx` clocks to their maximum reported values.
* `Result reset_gpu_clocks(unsigned int idx)`: Unset GPU `idx` clocks.
### CPU Cache
`perfect` can flush data from CPU caches. Unlike the other APIs, these do not return a `Result` because they do not fail.
See [examples/cpu_cache.cpp].
```c++
#include "perfect/cpu_cache.hpp`
```
* `void flush_all(void *p, const size_t n)`: Flush all cache lines starting at `p` for `n` bytes.
## Wish List
- [ ] Make CUDA Optional
- [ ] Nvidia GPU power monitoring
- [ ] Nivida GPU utilization monitoring