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
|
#include <cub/block/block_reduce.cuh>
#include "caffe2/core/context_gpu.h"
#include "caffe2/operators/one_hot_ops.h"
namespace caffe2 {
__global__ void OneHotOpKernel(
const int64_t batch_size,
const int64_t index_size,
const int64_t* indices,
float* output) {
CUDA_1D_KERNEL_LOOP(i, batch_size) {
output[i * index_size + indices[i]] = 1.;
}
}
template <>
void OneHotOp<CUDAContext>::DoOneHotOp(
int64_t batch_size,
int64_t index_size,
const Tensor& indices,
Tensor* output) {
float* output_ptr = output->template mutable_data<float>();
math::Set<float, CUDAContext>(output->numel(), 0., output_ptr, &context_);
OneHotOpKernel<<<
CAFFE_GET_BLOCKS(batch_size),
CAFFE_CUDA_NUM_THREADS,
0,
context_.cuda_stream()>>>(
batch_size, index_size, indices.data<int64_t>(), output_ptr);
}
REGISTER_CUDA_OPERATOR(OneHot, OneHotOp<CUDAContext>);
} // namespace
|
