CUDA and HIP: Thread Organization, Memory Model, Warps, and Wavefronts¶

1. Which variables in CUDA and HIP provide the dimensions of the grid?

• A. blockDim.x/y/z
• B. threadIdx.x/y/z
• C. gridDim.x/y/z
• D. warpSize

Memory Model in CUDA and HIP¶

2. Which type of memory in CUDA and HIP is accessible to all threads across blocks and the host?

• A. Register
• B. Shared
• C. Local
• D. Global

Warps in CUDA and Wavefronts in HIP¶

3. How many threads are in a warp in CUDA?

• A. 16
• B. 32
• C. 64
• D. 128

4. True or False: In HIP, a wavefront consists of 64 threads, while in CUDA, a warp consists of 32 threads.

Thread Synchronization¶

5. Which function ensures all threads within a block reach the same execution point in CUDA and HIP?

• A. cudaDeviceSynchronize
• B. __syncthreads
• C. hipDeviceSynchronize
• D. printf

Memory Access and Performance¶

6. What is a common cause of performance loss due to thread divergence in CUDA and HIP?

• A. Using printf in kernels
• B. Running all threads in a single warp
• C. Conditional branching (e.g., if statements) within warps or wavefronts
• D. Using shared memory exclusively