add and fix some feature

src/parallel_radix.cu

@@ -7,7 +7,7 @@
 #include "device_launch_parameters.h"
 
 #define MAX_BLOCK_SIZE 1024
-#define ARRAY_SIZE 100000
+#define ARRAY_SIZE 200000
 #define BASE 10
 #define RNG_SEED 13516027
 
@@ -68,7 +68,7 @@ int main(int argc, char *argv[])
     cudaMemcpy(harr, darr, count * sizeof(int), cudaMemcpyDeviceToHost);
     //find max
     if(count > MAX_BLOCK_SIZE){
-        grid = count / (MAX_BLOCK_SIZE) + 1;
+        grid = count / (MAX_BLOCK_SIZE);
     }
     else{
         grid = 1;