Performance doubles by only changing one line of code

[xiefan46]

Hi xiandong,

Thanks for providing this amazing tutorial! Recently I am working on reduce0 and I found that I can double the performance of reduce_v0_baseline.cu kernel by simply changing a blockDim.x into THREAD_PER_BLOCK in the for loop

before

profile result:

after

profile result:

I guess this is because of loop unrolling? It's quite interesting that a simple change makes a big difference

[Liu-xiandong]

Hi xiandong,

Thanks for providing this amazing tutorial! Recently I am working on reduce0 and I found that I can double the performance of reduce_v0_baseline.cu kernel by simply changing a blockDim.x into THREAD_PER_BLOCK in the for loop

before

profile result:

after

profile result:

I guess this is because of loop unrolling? It's quite interesting that a simple change makes a big difference

Yes，If it is blockDim.x, the compiler cannot determine this value, so the loop will not be unrolled.