A 2-D grid launch that over-provisions threads for the image size

This kernel processes a width x height image with width = height = 1000, launched as process<<<dim3(32,32), dim3(32,32)>>>(img, 1000, 1000) (so grid and block dims give 1024x1024 = 1,048,576 total threads for a 1,000,000-pixel image). There is no bounds check before the write. What happens?

kernel.cu
__global__ void process(float* img, int width, int height) {
    int col = blockIdx.x * blockDim.x + threadIdx.x;
    int row = blockIdx.y * blockDim.y + threadIdx.y;
    int idx = row * width + col;
    img[idx] = img[idx] * 2.0f;   // no "if (row >= height || col >= width) return;" guard
}

2 of 46 in GPU / CUDA