Function get_R2C_fftwf_output

src/callback-client/callback-client.cpp:520–576

Compute reference output using fftw for float type

// Compute reference output using fftw for float type
fftwf_complex* get_R2C_fftwf_output(size_t* lengths, size_t fftbatchLength, int batch_size,
									clfftLayout in_layout, clfftDim dim)
{
	//In FFTW last dimension has the fastest changing index
	int fftwLengths[3] = {(int)lengths[2], (int)lengths[1], (int)lengths[0]};
	int inembed[3] = {(int)lengths[2], (int)lengths[1], (int)(lengths[0] + 2)};
	int outembed[3] = {(int)lengths[2], (int)lengths[1], (int)(lengths[0]/2 + 1)};

	fftwf_plan refPlan;
		
	int infftVectorLength = inembed[0] * inembed[1] * inembed[2];
	int outfftVectorLength = outembed[0] * outembed[1] * outembed[2];

	float *refin = (float*) malloc(sizeof(float)*fftbatchLength);
	fftwf_complex *refout = (fftwf_complex*)fftwf_malloc(sizeof(fftwf_complex)*outfftVectorLength*batch_size);

	refPlan = fftwf_plan_many_dft_r2c(dim, &fftwLengths[3 - dim], batch_size, 
									refin, &inembed[3 - dim], 1, infftVectorLength,
									refout, &outembed[3 - dim], 1, outfftVectorLength, FFTW_ESTIMATE);
	
	uint24_t* in24bitData = (uint24_t*)malloc(sizeof(uint24_t) * fftbatchLength);

	//Initialize Data
	srand(1);
	for (size_t idx = 0; idx < fftbatchLength; ++idx)
	{
		int randomVal = (int)rand();

		in24bitData[idx][0] = (randomVal >> 16) & 0xFF;
        in24bitData[idx][1] = (randomVal >> 8) & 0xFF;
        in24bitData[idx][2] = randomVal & 0xFF; 
	}

	float val; 
	
	for( size_t i = 0; i < fftbatchLength; i++)
	{
		val = (float)(in24bitData[i][0] << 16 | in24bitData[i][1] << 8 | in24bitData[i][2]) ;
		
		refin[i] = val;
	}

	fftwf_execute(refPlan);

	free(refin);

	fftwf_destroy_plan(refPlan);

	//Execute post-process code
	for (size_t idx = 0; idx < (outfftVectorLength*batch_size); ++idx)
	{
		float magnitude = sqrtf(pow(refout[idx][0], 2) + pow(refout[idx][1], 2));
		refout[idx][0] = magnitude;
	}

	return refout;
}