File: perplexity_op.cu

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (45 lines) | stat: -rw-r--r-- 1,196 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
#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