CUDA branch divergence doesn't make any differences - cuda

Question

I'm trying to learn CUDA by myself, and I'm now into the issue of branch divergence. As far as I understand, this is the name given to the problem that arises when several threads in a block are said to take a branch (due to if or switch statements, for example), but others in that block don't have to take it.
In order to investigate a little bit further this phenomena and its consequences, I've written a little file with a couple of CUDA functions. One of them is supposed to take lots of time, since the threads are stopped for much more time (9999... iterations) than in the other one (in which they're only stopped for an assignation).
However, when I run the code, I'm getting very similar times. Furthermore, even measuring the time that running both of them takes I get a time similar to running only one. Did I code anything wrong, or is there a logical explanation for this?
Code:
#include <stdio.h>
#include <stdlib.h>
#include <cutil.h>
#define ITERATIONS 9999999999999999999
#define BLOCK_SIZE 16
unsigned int hTimer;
void checkCUDAError (const char *msg)
{
cudaError_t err = cudaGetLastError();
if (cudaSuccess != err)
{
fprintf(stderr, "Cuda error: %s: %s.\n", msg,cudaGetErrorString( err) );
getchar();
exit(EXIT_FAILURE);
}
}
__global__ void divergence(float *A, float *B){
float result = 0;
if(threadIdx.x % 2 == 0)
{
for(int i=0;i<ITERATIONS;i++){
result+=A[threadIdx.x]*A[threadIdx.x];
}
} else
for(int i=0;i<ITERATIONS;i++){
result+=A[threadIdx.x]*B[threadIdx.x];
}
}
__global__ void betterDivergence(float *A, float *B){
float result = 0;
float *aux;
//This structure should not affect performance that much
if(threadIdx.x % 2 == 0)
aux = A;
else
aux = B;
for(int i=0;i<ITERATIONS;i++){
result+=A[threadIdx.x]*aux[threadIdx.x];
}
}
// ------------------------
// MAIN function
// ------------------------
int main(int argc, char ** argv){
float* d_a;
float* d_b;
float* d_result;
float *elementsA;
float *elementsB;
elementsA = (float *)malloc(BLOCK_SIZE*sizeof(float));
elementsB = (float *)malloc(BLOCK_SIZE*sizeof(float));
//"Randomly" filling the arrays
for(int x=0;x<BLOCK_SIZE;x++){
elementsA[x] = (x%2==0)?2:1;
elementsB[x] = (x%2==0)?1:3;
}
cudaMalloc((void**) &d_a, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_b, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_result, sizeof(float));
cudaMemcpy(d_a, elementsA, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_b, elementsB, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
CUT_SAFE_CALL(cutCreateTimer(&hTimer));
CUT_CHECK_ERROR("cudaCreateTimer\n");
CUT_SAFE_CALL( cutResetTimer(hTimer) );
CUT_CHECK_ERROR("reset timer\n");
CUT_SAFE_CALL( cutStartTimer(hTimer) );
CUT_CHECK_ERROR("start timer\n");
float timerValue;
dim3 dimBlock(BLOCK_SIZE,BLOCK_SIZE);
dim3 dimGrid(32/dimBlock.x, 32/dimBlock.y);
divergence<<<dimBlock, dimGrid>>>(d_a, d_b);
betterDivergence<<<dimBlock, dimGrid>>>(d_a, d_b);
checkCUDAError("kernel invocation");
cudaThreadSynchronize();
CUT_SAFE_CALL(cutStopTimer(hTimer));
CUT_CHECK_ERROR("stop timer\n");
timerValue = cutGetTimerValue(hTimer);
printf("kernel execution time (secs): %f s\n", timerValue);
return 0;
}

Answer 1

1) You have no memory writes in your __global__ code except the local variable(result). I'm not sure that cuda compiler does that, but all your code can be safely removed with no side effect(and maybe the compiler had done that).
2) All your reads from device memory in __global__ functions are from one place on each iteration. Cuda will store the value in register memory and the longest operation(memory access) will be done very fast here.
3) May be the compiler had replaced your cycles with single multiplication like `result=ITERATIONS*A[threadIdx.x]*B[threadIdx.x]
4) If all the code in your functions will be executed as you wrote it, your betterDivergence is going to be approximately 2 times faster than your another function because you have the loops in if branches in slower one and no loops in branches in faster one. But there won't be any idle time in threads among the threads that execute same loop because all threads are going to execute the body of the loop each iteration.
I suggest you to write another example where you will store the result in some device memory and then copy that memory back to host and make some more unpredictable calculations to prevent possible optimizations.

Answer 2

Below is shown the final, tested, right example of a code that allows to compare the performance between CUDA code with and without branch divergence:
#include <stdio.h>
#include <stdlib.h>
#include <cutil.h>
//#define ITERATIONS 9999999999999999999
#define ITERATIONS 999999
#define BLOCK_SIZE 16
#define WARP_SIZE 32
unsigned int hTimer;
void checkCUDAError (const char *msg)
{
cudaError_t err = cudaGetLastError();
if (cudaSuccess != err)
{
fprintf(stderr, "Cuda error: %s: %s.\n", msg,cudaGetErrorString( err) );
getchar();
exit(EXIT_FAILURE);
}
}
__global__ void divergence(float *A, float *B){
int a = blockIdx.x*blockDim.x + threadIdx.x;
if (a >= ITERATIONS) return;
if(threadIdx.x > 2)
{
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]+1;
}
} else
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]-1;
}
}
__global__ void noDivergence(float *A, float *B){
int a = blockIdx.x*blockDim.x + threadIdx.x;
if (a >= ITERATIONS) return;
if(threadIdx.x > WARP_SIZE)
{
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]+1;
}
} else
for(int i=0;i<ITERATIONS;i++){
B[a]=A[a]-1;
}
}
// ------------------------
// MAIN function
// ------------------------
int main(int argc, char ** argv){
float* d_a;
float* d_b;
float* d_result;
float *elementsA;
float *elementsB;
elementsA = (float *)malloc(BLOCK_SIZE*sizeof(float));
elementsB = (float *)malloc(BLOCK_SIZE*sizeof(float));
//"Randomly" filling the arrays
for(int x=0;x<BLOCK_SIZE;x++){
elementsA[x] = (x%2==0)?2:1;
}
cudaMalloc((void**) &d_a, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_b, BLOCK_SIZE*sizeof(float));
cudaMalloc((void**) &d_result, sizeof(float));
cudaMemcpy(d_a, elementsA, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
cudaMemcpy(d_b, elementsB, BLOCK_SIZE*sizeof(float), cudaMemcpyHostToDevice);
CUT_SAFE_CALL(cutCreateTimer(&hTimer));
CUT_CHECK_ERROR("cudaCreateTimer\n");
CUT_SAFE_CALL( cutResetTimer(hTimer) );
CUT_CHECK_ERROR("reset timer\n");
CUT_SAFE_CALL( cutStartTimer(hTimer) );
CUT_CHECK_ERROR("start timer\n");
float timerValue;
dim3 dimBlock(BLOCK_SIZE,BLOCK_SIZE);
dim3 dimGrid(128/dimBlock.x, 128/dimBlock.y);
//divergence<<<dimGrid, dimBlock>>>(d_a, d_b);
noDivergence<<<dimGrid, dimBlock>>>(d_a, d_b);
checkCUDAError("kernel invocation");
cudaThreadSynchronize();
CUT_SAFE_CALL(cutStopTimer(hTimer));
CUT_CHECK_ERROR("stop timer\n");
timerValue = cutGetTimerValue(hTimer)/1000;
printf("kernel execution time (secs): %f s\n", timerValue);
cudaMemcpy(elementsB, d_b, BLOCK_SIZE*sizeof(float), cudaMemcpyDeviceToHost);
return 0;
}