CUDA Programming and GPU Kernels Review

What CUDA function allocates memory on the GPU?

cudaMalloc()

What CUDA function copies data between CPU and GPU memory?

cudaMemcpy()

What CUDA function frees GPU memory?

cudaFree()

What keyword identifies a CUDA kernel?

global

What CUDA variable identifies the thread within a block?

threadIdx.x

What CUDA variable identifies the block?

blockIdx.x

What CUDA variable gives the number of threads per block?

blockDim.x

What does __syncthreads() do?

Synchronizes threads within a block.

What does vector addition primarily depend on: memory bandwidth or computation?

Memory bandwidth

What CUDA function waits for previously launched GPU work to finish?

cudaDeviceSynchronize()

How many threads are in each CUDA block in the vadd program?

$$512$$

What is the name of the GPU kernel for vector addition?

vadd

What keyword indicates that vadd is a GPU kernel?

global

How is the global thread index calculated in the vadd kernel?

$$threadIdx.x + blockIdx.x \times blockDim.x$$

What does each GPU thread compute in vadd?

One element: $c[i] = a[i] + b[i]$

Why does the vadd kernel check if (i < elems)?

To prevent out-of-bounds access.

How many GPU arrays are allocated in the vadd program?

$3$: d_a, d_b, d_c

What function checks for CUDA errors?

CheckCuda

What is the expected value of c[i] in the vector addition test?

2 * i +1

Does the reported compute time for vadd include copying results back to the CPU?

No, only the kernel launch and wait.

Why is cudaMemcpy(d_c, c, …) optional before launching the vadd kernel?

Every valid output element is overwritten by the kernel.

What is the name of the CUDA kernel that computes the fractal image?

fractal

How many threads per block are used in the fractal program?

$$512$$

What type of loop performs the fractal iterations?

do-while

What is the maximum number of iterations per pixel in the fractal program?

$$256$$

How many pixels does each GPU thread calculate in the fractal program?

$$1$$

How many pixels are in an image of width w?

$$w \times w$$

What data type is used in fractal_cuda.cu?

double

What data type is used in fractal_float.cu?

float

Which variable identifies the pixel's column in the fractal program?

$$col = i \bmod width$$

Which variable identifies the pixel's row in the fractal program?

$$row = i / width$$

What is the minimum allowed image width in the fractal program?

$$10$$

What condition stops the iteration in the fractal program?

$count \le 0$ or $z\_re2 + z\_im2 \ge 5.0$

What is the output array's data type in the fractal program?

unsigned char

Which function saves the BMP image in the fractal program?

bmp.save()

When does the fractal program generate a BMP file?

When $width < 1000$

Which CUDA function allocates GPU image memory in the fractal program?

cudaMalloc()

Which function copies the image from GPU to CPU in the fractal program?

cudaMemcpy()

Does the measured compute time include saving the BMP image in the fractal program?

No

What is the name of the CUDA kernel for matrix multiplication?

multiply

How many threads per block are used in the matrix multiplication program?

$$512$$

How many matrix elements does each GPU thread calculate in matrix multiplication?

$$1$$

What is the data type of the matrices in the matrix multiplication program?

float

What is the minimum allowed matrix size in the matrix multiplication program?

$$10$$

What function initializes the matrices in the matrix multiplication program?

init

What type of matrix is mat1 initialized as?

Identity matrix

What does multiplying mat1 by mat2 produce?

mat2

How is the global thread index calculated in the matrix multiplication kernel?

$$blockIdx.x \times blockDim.x + threadIdx.x$$

How is the row index calculated in the matrix multiplication kernel?

$$i = idx / size$$

How is the column index calculated in the matrix multiplication kernel?

$$j = idx \bmod size$$

What type of loop is inside the matrix multiplication CUDA kernel?

for loop

How many iterations does the matrix multiplication kernel's loop perform per output element?

$$size$$

How many GPU arrays are allocated in the matrix multiplication program?

$$3$$

Does the matrix multiplication kernel use shared memory?

No

Which function copies the matrix multiplication result back to the CPU?

cudaMemcpy

How does the program verify the matrix multiplication result?

Checks whether $mat2[i] == mat3[i]$

What function frees GPU memory in the matrix multiplication program?

cudaFree