Using Shared Memory for Matrix Multiplication in CUDA¶

1. Why is shared memory beneficial for matrix multiplication on GPUs?

• A. Shared memory has higher bandwidth and lower latency than global memory
• B. Shared memory is larger than global memory
• C. Shared memory persists between kernel launches
• D. Shared memory can be used by both the CPU and GPU

Basic Approaches for Using Shared Memory¶

2. In the first shared memory approach, why is only a single row of matrix $A$ stored in shared memory?

• A. To reduce memory usage on the device
• B. To allow quick access by threads working on the same row of the result matrix
• C. To ensure that each thread has unique data
• D. To avoid using global memory altogether

3. True or False: In the second shared memory approach, both a row of matrix $A$ and a column of matrix $B$ are stored in shared memory to reduce global memory accesses.

Tiled Matrix Multiplication¶

4. What is the primary purpose of dividing matrices into tiles in tiled matrix multiplication?

• A. To simplify the matrix multiplication algorithm
• B. To reduce the number of required multiplications
• C. To load data into shared memory in manageable chunks, reducing global memory accesses
• D. To improve readability of the kernel code

5. In the tiled matrix multiplication kernel, what does __syncthreads() achieve after loading each tile?

• A. It allows threads in different blocks to synchronize
• B. It ensures that all threads within the block have loaded their portion of the tile into shared memory before proceeding
• C. It prevents overwriting of shared memory
• D. It initiates data transfer from global to shared memory

6. Why do we need a loop over i in the tiled matrix multiplication kernel?

• A. To initialize shared memory before each kernel launch
• B. To load each portion (tile) of matrices $A$ and $B$ into shared memory in steps, reducing global memory loads
• C. To handle matrix dimensions that are not multiples of the tile width
• D. To compute the result matrix on the host

Benefits of Tiled Matrix Multiplication¶

7. Which of the following is NOT a benefit of using tiled matrix multiplication?

• A. Reduced number of accesses to global memory
• B. Improved use of shared memory bandwidth
• C. Easier to debug than standard matrix multiplication
• D. Enhanced performance for large matrices