Chapter 4

1. CPU timers will include time overhead of thread latency in OS and scheduling in OS, among many other factors. The time measured using CPU will also depend on the availability of high precision CPU timer. The host is frequently performing asynchronous computation while GPU kernel is running, and hence CPU timers may not give correct time for kernel executions.
2. Open Nvidia Visual profiler from C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v9.0\libnvvp. Then, go to -> New Session and Select .exe file for matrix multiplication example. You can visualize the performance of your code.
3. Divide by zero, incorrect variable types or sizes, nonexistent variables, subscripts out of range etc are examples of semantic errors.
4. An example of thread divergence can be given as follows:

__global__ void gpuCube(float *d_in, float *d_out) 
{
     int tid = threadIdx.x; 
if(tid%2 =...