mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-21 05:34:18 +08:00
6ba60ec9b0
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/23837 This is a temporary workaround to an issue in MKL-DNN's Convolution backwards implementation: https://github.com/pytorch/pytorch/issues/23825 It is only used to enable testing quantization Test Plan: Imported from OSS Differential Revision: D16659081 Pulled By: jamesr66a fbshipit-source-id: de18ebe98dec2a042f28b23373e20da2b44a42a2
61 lines
2.3 KiB
C++
61 lines
2.3 KiB
C++
#include <ATen/core/ivalue.h>
|
|
#include <torch/csrc/utils/init.h>
|
|
#include <torch/csrc/utils/throughput_benchmark.h>
|
|
|
|
#include <ATen/native/Convolution.h>
|
|
|
|
#include <pybind11/functional.h>
|
|
|
|
namespace torch {
|
|
namespace throughput_benchmark {
|
|
|
|
void initThroughputBenchmarkBindings(PyObject* module) {
|
|
auto m = py::handle(module).cast<py::module>();
|
|
using namespace torch::throughput_benchmark;
|
|
py::class_<BenchmarkConfig>(m, "BenchmarkConfig")
|
|
.def(py::init<>())
|
|
.def_readwrite(
|
|
"num_calling_threads", &BenchmarkConfig::num_calling_threads)
|
|
.def_readwrite("num_worker_threads", &BenchmarkConfig::num_worker_threads)
|
|
.def_readwrite("num_warmup_iters", &BenchmarkConfig::num_warmup_iters)
|
|
.def_readwrite("num_iters", &BenchmarkConfig::num_iters);
|
|
|
|
py::class_<BenchmarkExecutionStats>(m, "BenchmarkExecutionStats")
|
|
.def_readonly("latency_avg_ms", &BenchmarkExecutionStats::latency_avg_ms)
|
|
.def_readonly("num_iters", &BenchmarkExecutionStats::num_iters);
|
|
|
|
py::class_<ThroughputBenchmark>(m, "ThroughputBenchmark", py::dynamic_attr())
|
|
.def(py::init<jit::script::Module>())
|
|
.def(py::init<py::object>())
|
|
.def(
|
|
"add_input",
|
|
[](ThroughputBenchmark& self, py::args args, py::kwargs kwargs) {
|
|
self.addInput(std::move(args), std::move(kwargs));
|
|
})
|
|
.def(
|
|
"run_once",
|
|
[](ThroughputBenchmark& self, py::args args, py::kwargs kwargs) {
|
|
// Depending on this being ScriptModule of nn.Module we will release
|
|
// the GIL or not further down in the stack
|
|
return self.runOnce(std::move(args), std::move(kwargs));
|
|
})
|
|
.def("benchmark", [](ThroughputBenchmark& self, BenchmarkConfig config) {
|
|
// The benchmark always runs without the GIL. GIL will be used where
|
|
// needed. This will happen only in the nn.Module mode when manipulating
|
|
// inputs and running actual inference
|
|
AutoNoGIL no_gil_guard;
|
|
return self.benchmark(config);
|
|
});
|
|
|
|
|
|
m.def("_enable_mkldnn_conv", []() {
|
|
at::native::disable_mkldnn_conv.exchange(false);
|
|
});
|
|
m.def("_disable_mkldnn_conv", []() {
|
|
at::native::disable_mkldnn_conv.exchange(true);
|
|
});
|
|
}
|
|
|
|
} // namespace throughput_benchmark
|
|
} // namespace torch
|