CUDA

Allocating Device Memory

cudaMalloc(LOCATION, SIZE)

• LOCATION: Memory location on Device to allocate memory, an address in the GPU's memory

• SIZE: number of bytes to allocate

De-Allocate: cudaFree()

Copy Data between Host and Device

cudaMemory(DST, SRC, NUM_BYTES, DIRECTION)

• DST: An address of the memory to copy into

• SRC: An address of the memory to copy from

• NUM_BYTES: N * sizeof(type)

• DIRECTION:

  • cudaMemcpyHostToDevice

  • cudaMemcpyDeviceToHost

Define the Kernel

__global__ void kernel(int *d_out, int *d_in) {
    d_out[0] = d_in[0];
}

Thread Index

In kernel definition, built-in variable threadIdx is accessible to get thread index within the thread block for each thread.

It has 3 dimensions: threadIdx.x, threadIdx.y and threadIdx.z.

Block Index

Index of a block: blockIdx.x, blockIdx.y and blockIdx.z.

Indexing within Grid

i = threadIdx.x + blockIdx.x * blockDim.x;

__syncthreads

To explicitly synchronize all threads (adding barriers), use __syncthreads.

int temp = a[i + 1];
__syncthreads;
a[i] = temp;
__syncthreads;

Launch the Kernel

int *h_c, *d_c;

// Allocate memeory on the device
cudaMalloc((void**)&d_c, sizeof(int));
// Copy content from host to device
cudaMemcpy(d_c, h_c, sizeof(int), cudaMemcpyHostToDevice);

// Launch the kernel
dim3 grid_size(1);
dim3 block_size(1);
kernel<<<grid_size, block_size>>>(...);

// Force host to wait on the completion of the kernel
cudaDeviceSynchronize();

// Copy data back to host
cudaMemcpy(h_c, d_c, sizeof(int), cudaMemcpyDeviceToHost);

// Clean up memory
cudaFree(d_c);
free(h_c);