optimize occupancy

Author: AndrewBoessen

main.cpp

@@ -13,7 +13,7 @@ bool isSorted(int *arr, int size) {
 }
 
 int main() {
-  const int SIZE = 1024; // Must be a multiple of 32 for this example
+  const int SIZE = 1048576; // Must be a multiple of 32 for this example
 
   // Allocate and initialize host array
   int *h_arr = new int[SIZE];

smem_bitonic_sort.cu

@@ -99,7 +99,7 @@ __global__ void smemBitonicSort(int *arr, int size) {
 }
 
 void launchBitonicSort(int *arr, int size) {
-  const int BLOCK_SIZE = 1024;
+  const int BLOCK_SIZE = 512;
   smemBitonicSort<<<(size + BLOCK_SIZE - 1) / BLOCK_SIZE, BLOCK_SIZE,
                     BLOCK_SIZE * sizeof(int)>>>(arr, size);
 }

warp_bitonic_sort.cu

@@ -77,6 +77,6 @@ __global__ void warpBitonicSort(int *arr, int size) {
 }
 
 void launchBitonicSort(int *arr, int size) {
-  const int BLOCK_SIZE = 256;
+  const int BLOCK_SIZE = 512;
   warpBitonicSort<<<size / BLOCK_SIZE, BLOCK_SIZE>>>(arr, size);
 }