Timings for PFAC_matchFromDeviceReduce vs PFAC_matchFromHostReduce

[GoogleCodeExporter]

What version of the product are you using? On what operating system?
PFAC 1.0, on RHEL 6

Please provide any additional information below.

I measured the time it takes for PFAC_matchFromHostReduce and the equivalent 
steps when using PFAC_matchFromDeviceReduce. Both functions take about the same 
time to complete when the size of the input string is 100MB.

Timing for PFAC_matchFromHostReduce: 56 ms

Timing for equivalent steps using PFAC_matchFromDeviceReduce:
cudaMalloc: 0.3 ms
cudaMemcpy(d_input_string, h_input_string, input_size, cudaMemcpyHostToDevice): 
18 ms
PFAC_matchFromDeviceReduce: 26 ms
cudaMemcpy of d_pos and d_match_result back to CPU: 0.3 ms
cudaFree of d_input_string, d_pos and d_match_result: 11 ms
Total: 57 ms

Original issue reported on code.google.com by [email protected] on 29 Apr 2011 at 1:49

[GoogleCodeExporter]

PFAC_matchFromHostReduce() needs to free working space d_input_string, 
d_matched_result and d_pos.
[code]
    cudaMalloc((void **) &d_input_string, n_hat*sizeof(int) );
    cudaMalloc((void **) &d_matched_result, input_size*sizeof(int) );
    cudaMalloc((void **) &d_pos, input_size*sizeof(int) );

    cudaMemcpy(d_input_string, h_input_string, input_size, cudaMemcpyHostToDevice);

    same as PFAC_matchFromDeviceReduce()

    cudaMemcpy(h_pos,          d_pos,          (*h_num_matched)*sizeof(int), cudaMemcpyDeviceToHost);
    cudaMemcpy(h_match_result, d_match_result_zip, (*h_num_matched)*sizeof(int), cudaMemcpyDeviceToHost); 

    cudaFree(d_input_string);
    cudaFree(d_matched_result);
    cudaFree(d_pos);
[/code]   

In my tests, cudaFree() needs 12ms for 100MB input string and 24ms for 200MB 
input stream.

If you are not a beginner, then I will suggest PFAC_matchFromDeviceReduce().

Original comment by [email protected] on 29 Apr 2011 at 2:18