Class CopyKernel

src/library/generator.copy.cpp:94–508

    // Copy kernel
    template <Precision PR>
    class CopyKernel
    {
        size_t N;
		size_t Nt;
		const FFTGeneratedCopyAction::Signature & params;
		bool h2c, c2h;
		bool general;

		inline std::string OffsetCalc(const std::string &off, bool input = true)
		{
			std::string str;

			const size_t *pStride = input ? params.fft_inStride : params.fft_outStride;

			str += "\t"; str += off; str += " = ";
			std::string nextBatch = "batch";
			for(size_t i=(params.fft_DataDim - 1); i>1; i--)
			{
				size_t currentLength = 1;
				for(int j=1; j<i; j++) currentLength *= params.fft_N[j];

				str += "("; str += nextBatch; str += "/"; str += SztToStr(currentLength);
				str += ")*"; str += SztToStr(pStride[i]); str += " + ";

				nextBatch = "(" + nextBatch + "%" + SztToStr(currentLength) + ")";
			}

			str += nextBatch; str += "*"; str += SztToStr(pStride[1]); str += ";\n";

			return str;
		}

    public:
        CopyKernel( const FFTGeneratedCopyAction::Signature &paramsVal) :
					params(paramsVal)

        {
			N = params.fft_N[0];
			Nt = 1 + N/2;

			h2c = (	(params.fft_inputLayout == CLFFT_HERMITIAN_PLANAR) ||
					(params.fft_inputLayout == CLFFT_HERMITIAN_INTERLEAVED) ) ? true : false;
			c2h = (	(params.fft_outputLayout == CLFFT_HERMITIAN_PLANAR) ||
					(params.fft_outputLayout == CLFFT_HERMITIAN_INTERLEAVED) ) ? true : false;

			general = !(h2c || c2h);

			// We only do out-of-place copies at this point
			assert(params.fft_placeness == CLFFT_OUTOFPLACE);
		}

        void GenerateKernel(std::string &str)
		{
			std::string rType  = RegBaseType<PR>(1);
			std::string r2Type  = RegBaseType<PR>(2);

			bool inIlvd; // Input is interleaved format
			bool outIlvd; // Output is interleaved format