Question

Im looking at this implementation of DCT using cuda: http://www.cse.nd.edu/courses/cse60881/www/source_code/dct8x8/dct8x8_kernel1.cu
The part in question is here:

__shared__ float CurBlockLocal1[BLOCK_SIZE2];

__global__ void CUDAkernel1DCT(float *Dst, int ImgWidth, int OffsetXBlocks, int OffsetYBlocks)
{
    // Block index
    const int bx = blockIdx.x + OffsetXBlocks;
    const int by = blockIdx.y + OffsetYBlocks;

    // Thread index (current coefficient)
    const int tx = threadIdx.x;
    const int ty = threadIdx.y;

    // Texture coordinates
    const float tex_x = (float)( (bx << BLOCK_SIZE_LOG2) + tx ) + 0.5f;
    const float tex_y = (float)( (by << BLOCK_SIZE_LOG2) + ty ) + 0.5f;

    //copy current image pixel to the first block
    CurBlockLocal1[ (ty << BLOCK_SIZE_LOG2) + tx ] = tex2D(TexSrc, tex_x, tex_y);

    //synchronize threads to make sure the block is copied
    __syncthreads();

where block size is 8 so block_size_log2 is 3.

Why is the texture coordinates defined as it is? Why do we need to use texture coordinates? What is the “<<” in Cuda?

Answer 1

To answer your questions in reverse order:

1. As in standard C or C++, the << operator is the bitwise left shift operator. This implies that a << b is equivalent to a * 2^b where a and b are both positive integers. So code you are asking about is basically shorthand for integer power of two multiplication.
2. As discussed in the Appendices of the Cuda programming guide, textures are indexed using floating point coordinates which are voxel centred, which is why the read arguments in the code you posted are offset by 0.5 in each direction
3. The code you are asking about looks to have been written for an early generation of CUDA hardware which had considerably slower integer arithmetic performance than floating point. The use of bit shifting in place of power of two multiplication is mostly likely a performance optimization, and may be of no use on newer generations of CUDA hardware.

The code you have asked about could probably been written as

__shared__ float CurBlockLocal1[BLOCK_SIZE2];

__global__ void CUDAkernel1DCT(float *Dst, int ImgWidth, int OffsetXBlocks, int OffsetYBlocks)
{
    // Block index
    const int bx = blockIdx.x + OffsetXBlocks;
    const int by = blockIdx.y + OffsetYBlocks;

    // Thread index (current coefficient)
    const int tx = threadIdx.x;
    const int ty = threadIdx.y;

    // Texture coordinates
    const float tex_x = (float)( (bx * BLOCK_SIZE) + tx ) + 0.5f;
    const float tex_y = (float)( (by * BLOCK_SIZE) + ty ) + 0.5f;

    //copy current image pixel to the first block
    CurBlockLocal1[ (ty * BLOCK_SIZE) + tx ] = tex2D(TexSrc, tex_x, tex_y);

    ......
}