FFT Implementation on GPU

GPU_Kernel.h

/*******************************************************************************************
 * My work in progress Code for FFT on GPU
 * 
 *******************************************************************************************/

//every shader needs folllowing lines
#define GL3ES_VERSION_SEPC "#version 300 es        \n"
#define FP_PREC_SPEC   "precision mediump float;   \n"
#define FS_LAYOUT_LINE "layout(location = 0) out vec4 fragColor; \n"
#define L_(STMT) "  " #STMT "  \n"

static const char VERTEX_SHADER[] =
	GL3ES_VERSION_SEPC
    L_(layout(location = 0) in vec2 vPosition;     	)
	L_(void main() {                               	)
	L_(  gl_Position = vec4(vPosition, 0.0, 1.0);	)
	L_(}  											);
static const char FRAGMENT_SHADER[] =
    GL3ES_VERSION_SEPC
    FP_PREC_SPEC
    L_(layout(location = 0) out vec4 fragColor;		)
    L_(void main() { 								)
   	L_(  fragColor = vec4(0.95, 0.95, 0, 0); 		)
    L_(} 											);

static const char VERTEX_FILL_SHADER[] =
	GL3ES_VERSION_SEPC
    L_(layout(location = 0) in vec4 vPosition;			)
	L_(out vec2 texCoord; 								)
	L_(void main() { 									)
	L_(  gl_Position = vec4(vPosition.xy, 0.0, 1.0);	)
	L_(  texCoord = vPosition.zw; 						)
	L_(} 												);

static const char FRAGMENT_FILL_SHADER[] =
	GL3ES_VERSION_SEPC
	FP_PREC_SPEC
    FS_LAYOUT_LINE
    L_(in vec2 texCoord;														)
    L_(uniform sampler2D floatArray;											)
    L_(void main() { 															)
    L_(  ivec2 lookupInd = ivec2(gl_FragCoord.xy);                      		)
    L_(  fragColor = vec4(texelFetch(floatArray, lookupInd, 0).rg, 0.0, 1.0);	)
    L_(}                                                             			);

/* for simplicity first assume we only use 1D Texture arrays, so GL_FragCoord only take care one dimensional data */
static const char DECIMATION_IN_TIME_SHADER[] =
	GL3ES_VERSION_SEPC
	FP_PREC_SPEC
	FS_LAYOUT_LINE
	L_(uniform sampler2D floatArray; 										)
  L_(uniform int butterflyIndex;              							)
  L_(const float PI = 3.141592654;										)
  L_(vec2 mult_twiddleFactor(vec2 in, float kn, float sum) {									)
  L_(  vec2 out = vec(0);			)
  L_(  out.x = in.x*cos(-2.0*PI*kn/sum) - in.y*sin(-2.0*PI*kn/sum);				)
  L_(  out.y = in.y*cos(-2.0*PI*kn/sum) + in.x*sin(-2.0*PI*kn/sum);				)
  L_(  return out;  																)
  L_(} 																			)
	L_(void main() {														)		
	L_(  ivec2 lookupInd = ivec2(gl_FragCoord.xy);   						)
	L_(  ivec2 lookupInd_c = lookupInd ^ butterflyIndex;							)
	L_(  int signess = -((lookupInd.x & butterflyIndex) * 2 - 1);           )	
	L_(  vec2 complex_pair_1 = float(signess) * texelFetch(floatArray, lookupInd, 0).rg;		)
	L_(  vec2 complex_pair_2 = texelFetch(floatArray, lookupInd_c, 0).rg;		)
	L_(  complex_pair_1 += complex_pair_2;        								)
	L_(  if (((butterflyIndex << 1) & lookupInd.x) != 0) {						)
	L_(    complex_pair_1 = mult_twiddleFactor(complex_pair_1, float(lookupInd.x&((butterflyIndex<<1)-1)), butterflyIndex<<2); 	)
	L_(  })
	L_(  fragColor = vec4(complex_pair_1, 0.0, 1.0);							)
	L_(});