Random permutation on GPU using Thrust - cuda

Question

I'm trying to write code that will permute a vector on a gpu, but I'm have great difficulty getting Thrust to cooperate. At the moment the code below compiles fine, but does nothing to the order of the vector r. Please help. Thanks!
void rng_permutation<float>(const int n, float* r){
float* order;
cudaMalloc((void**)&order, n* sizeof(float));
/*
some lines of code that generate uniform random floats between 0 and 1 that I know work
*/
thrust::device_ptr<float> order_(order);
thrust::device_vector<float> order__(order_, order_ + n);
thrust::device_ptr<float> r_(r);
thrust::device_vector<float> r__(r_, r_ + n);
thrust::sort_by_key(order__.begin(), order__.end(), r__.begin());
thrust::copy(order_, order_ + n, order__.begin());
thrust::copy(r_, r_ + n, r__.begin());
cudaFree(order);
order__.clear();
r__.clear();
thrust::device_vector<float>().swap(order__);
thrust::device_vector<float>().swap(r__);
}

Answer 1

You've got your sources and destinations backwards here:
thrust::copy(order_, order_ + n, order__.begin());
thrust::copy(r_, r_ + n, r__.begin());
The previous line of code just sorted what is in order__. You are then copying the contents of order_ over the top of that (the first parameters to thrust::copy are the source parameters, the last are the destination parameters). Makes no sense. Instead, reverse that:
thrust::copy(order__.begin(), order__.end(), order_);
thrust::copy(r__.begin(), r__.end(), r_);
And you will get sensible results:
$ cat t312.cu
#include <stdlib.h>
#include <thrust/device_vector.h>
#include <thrust/sort.h>
#include <thrust/copy.h>
#include <iostream>
#include <thrust/sequence.h>
template <typename T>
void caffe_gpu_rng_uniform(int n, T lo, T hi, T *o)
{
T *d = (T *)malloc(n*sizeof(T));
for (int i = 0; i < n; i++) d[i] = (rand()/(float)RAND_MAX)*(hi-lo) + lo;
cudaMemcpy(o, d, n*sizeof(T), cudaMemcpyHostToDevice);
free(d);
};
template <typename T>
void print_gpu_array_entries(T *o, int x , int y , int n){
thrust::copy_n(thrust::device_pointer_cast<T>(o), x, std::ostream_iterator<T>(std::cout, ","));
std::cout << std::endl;
}
void rng_permutation(const int n, float* r){
float* order;
cudaMalloc((void**)&order, n* sizeof(float));
caffe_gpu_rng_uniform<float>(n, (float)0.0, (float)1.0, order);
print_gpu_array_entries<float>(order, 10 , 1 , n);
print_gpu_array_entries<float>(r, 10 , 1 , n);
thrust::device_ptr<float> order_(order);
thrust::device_vector<float> order__(order_, order_ + n);
thrust::device_ptr<float> r_(r);
thrust::device_vector<float> r__(r_, r_ + n);
thrust::sort_by_key(order__.begin(), order__.end(), r__.begin());
thrust::copy(order__.begin(), order__.end(), order_);
thrust::copy(r__.begin(), r__.end(), r_);
print_gpu_array_entries<float>(order, 10 , 1 , n);
print_gpu_array_entries<float>(r, 10 , 1 , n);
cudaFree(order);
order__.clear();
r__.clear();
thrust::device_vector<float>().swap(order__);
thrust::device_vector<float>().swap(r__);
}
int main(){
thrust::device_vector<float> data(10);
thrust::sequence(data.begin(), data.end());
rng_permutation(10, thrust::raw_pointer_cast(data.data()));
}
$ nvcc -o t312 t312.cu
$ ./t312
0.840188,0.394383,0.783099,0.79844,0.911647,0.197551,0.335223,0.76823,0.277775,0.55397,
0,1,2,3,4,5,6,7,8,9,
0.197551,0.277775,0.335223,0.394383,0.55397,0.76823,0.783099,0.79844,0.840188,0.911647,
5,8,6,1,9,7,2,3,0,4,
$