| commit | d2d63b69ed1b938bb85e44f1b88615fefaae67a0 | [log] [tgz] |
|---|---|---|
| author | Antonio Caggiano <antonio.caggiano@arm.com> | Fri Jul 12 14:41:52 2019 +0100 |
| committer | Antonio Caggiano <antonio.caggiano@arm.com> | Fri Jul 12 14:41:52 2019 +0100 |
| tree | 21274cce71b36c2999253348febe5fa54a55a5c2 | |
| parent | 01a31c378a36f9a1d2da6487a620e29c13f2e048 [diff] |
Add Tiles to GPU counters - Add counter for the number of physical tiles written
HWCPipe is a simple and extensible interface for reading CPU and GPU hardware counters.
The software is provided under an MIT license.
This project has a third-party dependency, which may have independent licensing:
All contributions are accepted under the same LICENSE.
To use HWCPipe, build it as a shared library in your project.
If your project uses CMake, you can add the following to your CMakeLists.txt:
add_subdirectory(hwcpipe)
Basic usage for HWCPipe is simple:
// HWCPipe performs automated platform detection for CPU/GPU counters
hwcpipe::HWCPipe h;
// Start HWCPipe once at the beginning of the profiling session
h.run();
while (main_loop) {
// Call sample() to sample counters with the frequency you need
auto measurements = h.sample();
[...]
}
// At the end of the profiling session, stop HWCPipe
h.stop();
The sample function returns a Measurements struct, which can be accessed like this:
// Check if CPU measurements are available
if (measurements.cpu)
{
// Look for a counter in the map
const auto &counter = measurements.cpu->find(CpuCounter::Cycles);
if (counter != measurements.cpu->end())
{
// Get the data stored in the counter, casted to the type you need
auto value = counter->second.get<float>();
}
}
The available counters are specified in the CpuCounter and GpuCounter enums (cpu_profiler.h and gpu_profiler.h respectively).
Platforms will support a subset of these counters, which can be queried via:
auto cpu_counters = h.cpu_profiler()->supported_counters(); auto gpu_counters = h.gpu_profiler()->supported_counters()
You can specify the counters to be enabled in the following ways:
// Enable a default set of counters
auto h = hwcpipe::HWCPipe();
// Pass sets of CPU and GPU counters to be enabled
auto h = hwcpipe::HWCPipe({CpuCounter::Cycles, CpuCounter::Instructions}, {GpuCounter::GpuCycles});
// Pass a JSON string
auto h = hwcpipe::HWCPipe(json);
The JSON string should be formatted like this:
{
"cpu": ["Cycles", "Instructions"],
"gpu": ["GpuCycles"]
}
Available counter names can be found in cpu_counter_names (cpu_profiler.h) and gpu_counter_names (gpu_profiler.h).
For more information regarding Mali counters, see Mali Performance Counters.
In order for performance data to be displayed, profiling needs to be enabled on the device. Some devices may disable it by default.
Profiling can be enabled via adb:
adb shell setprop security.perf_harden 0
If the counters provided in CpuCounter and GpuCounter are enough for the new platform, the process is simple:
CpuProfiler of GpuProfiler (you can use PmuProfiler and MaliProfiler as references).hwcpipe.cpp. For consistency in platform detection, the constructor for your platform should throw if the platform is not available.If you need to add new counters to the existing ones, you should update the following variables:
CpuCounter/GpuCounter enum.cpu_counter_names/gpu_counter_names map (necessary for JSON initialization).cpu_counter_info/gpu_counter_info map.The CpuCounter and GpuCounter enums are meant to be expanded. Platforms must not break if new counters are added.