Method getWorkSizes

src/library/action.transpose.cpp:339–518

clfftStatus FFTGeneratedTransposeNonSquareAction::getWorkSizes(std::vector< size_t >& globalWS, std::vector< size_t >& localWS)
{

    size_t wg_slice;
    size_t smaller_dim = (this->signature.fft_N[0] < this->signature.fft_N[1]) ? this->signature.fft_N[0] : this->signature.fft_N[1];
	size_t bigger_dim = (this->signature.fft_N[0] >= this->signature.fft_N[1]) ? this->signature.fft_N[0] : this->signature.fft_N[1];
	size_t dim_ratio = bigger_dim / smaller_dim;
    size_t global_item_size;

    if (this->signature.nonSquareKernelType == NON_SQUARE_TRANS_TRANSPOSE_BATCHED_LEADING)
    {
        if (smaller_dim % (16 * reShapeFactor) == 0)
            wg_slice = smaller_dim / 16 / reShapeFactor;
        else
            wg_slice = (smaller_dim / (16 * reShapeFactor)) + 1;

        global_item_size = wg_slice*(wg_slice + 1) / 2 * 16 * 16 * this->plan->batchsize;

        for (int i = 2; i < this->signature.fft_DataDim - 1; i++)
        {
            global_item_size *= this->signature.fft_N[i];
        }

        /*Push the data required for the transpose kernels*/
        globalWS.clear();
		if(this->signature.nonSquareKernelType == NON_SQUARE_TRANS_TRANSPOSE_BATCHED_LEADING)
			globalWS.push_back(global_item_size * dim_ratio);
		else if (this->signature.nonSquareKernelType == NON_SQUARE_TRANS_TRANSPOSE_BATCHED)
			globalWS.push_back(global_item_size);


        localWS.clear();
        localWS.push_back(lwSize);
    }
	else if (this->signature.nonSquareKernelType == NON_SQUARE_TRANS_TRANSPOSE_BATCHED)
	{
		if (smaller_dim % (16 * reShapeFactor) == 0)
			wg_slice = smaller_dim / 16 / reShapeFactor;
		else
			wg_slice = (smaller_dim / (16 * reShapeFactor)) + 1;

		global_item_size = wg_slice*(wg_slice + 1) / 2 * 16 * 16 * this->plan->batchsize;

		for (int i = 2; i < this->plan->length.size(); i++)
		{
			global_item_size *= this->plan->length[i];
		}

		/*Push the data required for the transpose kernels*/
		globalWS.clear();
		globalWS.push_back(global_item_size);


		localWS.clear();
		localWS.push_back(lwSize);
	}
    else
    {