Questions

1. Why aren't CPU timers used to measure the performance of kernel functions?
2. Try to visualize the performance of the matrix multiplication code implemented in the previous chapter using the Nvidia Visual Profiler tool.
3. Give different examples of semantic errors encountered in programs.
4. What is the drawback of thread divergence in kernel functions? Explain with an example.
5. What is the drawback of using the cudahostAlloc function to allocate memory on the host?
6. Justify the following statement: the order of operations in CUDA streams is very important to improve the performance of a program.
7. How many blocks and threads should be launched for a 1024 x 1024 image for good performance using CUDA?