Method Kernel

src/library/generator.stockham.cpp:2942–3145

    public:
        Kernel( const FFTKernelGenKeyParams &paramsVal) :
			params(paramsVal), r2c2r(false)

        {
			length = params.fft_N[0];
			workGroupSize = params.fft_SIMD;
			numTrans = (workGroupSize * params.fft_R) / length;

			r2c = false;
			c2r = false;
			// Check if it is R2C or C2R transform
			if(params.fft_inputLayout == CLFFT_REAL)  r2c = true;
			if(params.fft_outputLayout == CLFFT_REAL) c2r = true;
			r2c2r = (r2c || c2r);

			if(r2c)
			{
				rcFull = (	(params.fft_outputLayout == CLFFT_COMPLEX_INTERLEAVED) ||
							(params.fft_outputLayout == CLFFT_COMPLEX_PLANAR) ) ? true : false;
			}
			if(c2r)
			{
				rcFull = (	(params.fft_inputLayout  == CLFFT_COMPLEX_INTERLEAVED) ||
							(params.fft_inputLayout  == CLFFT_COMPLEX_PLANAR) ) ? true : false;
			}

			rcSimple = params.fft_RCsimple;

			halfLds = true;
			linearRegs = true;

			realSpecial = params.fft_realSpecial;

			blockCompute = params.blockCompute;
			blockComputeType = params.blockComputeType;
			// Make sure we can utilize all Lds if we are going to
			// use blocked columns to compute FFTs
			if(blockCompute)
			{
				assert(length <= 256);  // 256 parameter comes from prototype experiments
										// largest length at which block column possible given 32KB LDS limit
										// if LDS limit is different this number need to be changed appropriately
				halfLds = false;
				linearRegs = true;
			}

			assert( ((length*numTrans)%workGroupSize) == 0 );
			cnPerWI = (numTrans * length) / workGroupSize;
			workGroupSizePerTrans = workGroupSize/numTrans;

			// !!!! IMPORTANT !!!! Keep these assertions unchanged, algorithm depend on these to be true
			assert( (cnPerWI * workGroupSize) == (numTrans * length) );
			assert( cnPerWI <= length ); // Don't do more than 1 fft per work-item

			// Breakdown into passes

			size_t LS = 1;
			size_t L;