Matrix Multiplication: CPU vs. GPU Implementation¶

1. In CPU matrix multiplication, which element of the matrices is computed by each loop iteration?

• A. A single row
• B. A single column
• C. A single element (dot product result)
• D. The entire matrix

GPU Matrix Multiplication with CUDA¶

2. Which CUDA function is used to allocate memory for matrices on the GPU?

• A. malloc
• B. cudaMalloc
• C. free
• D. memcpy

3. Why do we use cudaMemcpy(d_a, a, sizeof(float) * (N * N), cudaMemcpyHostToDevice); in GPU programs?

• A. To transfer data from GPU to CPU
• B. To allocate memory on the GPU
• C. To transfer data from CPU to GPU
• D. To initialize device memory

Thread and Block Configuration¶

4. What is the purpose of defining dim3 dimBlock(32, 32, 1); for GPU matrix multiplication?

• A. It sets the matrix dimensions
• B. It defines the number of threads per block
• C. It specifies the memory allocation size
• D. It configures CPU threads

5. True or False: The calculation dim3 dimGrid((N + blockSize - 1) / blockSize, (N + blockSize - 1) / blockSize, 1); ensures complete coverage of the matrix, even if N is not a multiple of blockSize.

Kernel Execution and Indexing¶

6. In the CUDA kernel matrix_mul, what does int row = blockIdx.x * blockDim.x + threadIdx.x; compute?

• A. The matrix size
• B. The global row index for each thread
• C. The column size of the matrix
• D. The shared memory size

7. Why do we check if row < width && col < width in the CUDA kernel?

• A. To initialize the matrix
• B. To avoid accessing out-of-bounds memory
• C. To synchronize threads
• D. To allocate memory

8. After executing a CUDA kernel, which function do we use to transfer results from the GPU back to the CPU?

• A. cudaMemcpy(d_a, a, ..., cudaMemcpyDeviceToHost);
• B. cudaFree
• C. cudaMemcpy(d_c, c, ..., cudaMemcpyHostToDevice);
• D. cudaMemcpy(d_c, c, ..., cudaMemcpyDeviceToHost);