mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
4f34cd6d1e
Avoid exposing defines that conflict with google logging, since this blocks external usage of libtorch in certain cases. All the 'interesting' changes should be in these two files, and the rest should just be mechanical changes via sed. c10/util/logging_is_not_google_glog.h c10/util/logging_is_google_glog.h Fixes https://github.com/pytorch/pytorch/issues/81415 cc @miladm @malfet Pull Request resolved: https://github.com/pytorch/pytorch/pull/82032 Approved by: https://github.com/soumith, https://github.com/miladm
46 lines
1.2 KiB
Plaintext
46 lines
1.2 KiB
Plaintext
#include "caffe2/core/context_gpu.h"
|
|
#include "caffe2/operators/perplexity_op.h"
|
|
#include "caffe2/utils/math.h"
|
|
#include <thrust/device_vector.h>
|
|
#include <thrust/transform_reduce.h>
|
|
#include <thrust/system/cuda/execution_policy.h>
|
|
|
|
namespace caffe2 {
|
|
|
|
struct perplexity_function
|
|
{
|
|
perplexity_function(float p) : pow(p) {}
|
|
__host__ __device__ float operator()(float x) const
|
|
{
|
|
return powf(1.0f/x, pow);
|
|
}
|
|
float pow;
|
|
};
|
|
|
|
template <>
|
|
bool PerplexityOp<float, CUDAContext>::RunOnDevice() {
|
|
auto& X = Input(0);
|
|
|
|
TORCH_DCHECK_EQ(X.dim(), 1);
|
|
int N = X.dim32(0);
|
|
|
|
auto* Y = Output(0, vector<int64_t>(), at::dtype<float>());
|
|
float* Ydata = Y->template mutable_data<float>();
|
|
const float* Xdata = X.data<float>();
|
|
|
|
float perplexity = thrust::transform_reduce(
|
|
#if THRUST_VERSION >= 100800
|
|
thrust::cuda::par.on(context_.cuda_stream()),
|
|
#endif // THRUST_VERSION >= 100800
|
|
Xdata, Xdata + N,
|
|
perplexity_function(1.0f/N),
|
|
1.0f,
|
|
thrust::multiplies<float>());
|
|
|
|
math::Set<float, CUDAContext>(1, perplexity, Ydata, &context_);
|
|
return true;
|
|
}
|
|
|
|
REGISTER_CUDA_OPERATOR(Perplexity, PerplexityOp<float, CUDAContext>);
|
|
} // namespace caffe2
|