Hi,
Another CUDA Newbie Question :
Can some one shed light ,if there is a scope of parallel reduction in the below mentioned scenario :
for (j=0;j<numiteration;j++)
{
if (supreme<fabs(d_b[j]-d_a[j]))
supreme=fabs(d_b[j]-d_a[j]);
}
Can I safely try the typical parallel implementation (like the sum/max reduction), recursively halve threads, use shared memory to hold supreme ?
Any pointers to similar examples for reduction will be of great help ?
Thanks much.
The inner_product function in Thrust makes this easy to do:
[codebox]include <thrust/inner_product.h>
include <thrust/functional.h>
include <thrust/device_vector.h>
// this example computes the maximum absolute difference
// between the elements of two vectors
struct abs_diff
{
template <typename T>
__host__ __device__
T operator()(const T x, const T y)
{
return fabs(b - a);
}
};
int main(void)
{
thrust::device_vector<float> d_a(4);
thrust::device_vector<float> d_b(4);
d_a[0] = 1.0; d_b[0] = 2.0;
d_a[1] = 2.0; d_b[1] = 4.0;
d_a[2] = 3.0; d_b[2] = 3.0;
d_a[3] = 4.0; d_b[3] = 0.0;
// initial value of the reduction
float init = 0;
// binary operations
thrust::maximum<float> binary_op1;
abs_diff binary_op2;
float max_abs_diff = thrust::inner_product(d_a.begin(), d_a.end(), d_b.begin(), init, binary_op1, binary_op2);
std::cout << "maximum absolute difference: " << max_abs_diff << std::endl;
return 0;
}[/codebox]
This example uses Thrust’s device_vector container, but that’s not necessary. You can wrap a “raw” pointer with thrust::device_ptr as this example shows.
Thanks so much…made it so lucid !!