#include "butterfly.h"
#ifdef __APPLE__
#include <OpenCL/opencl.h>
#else
#include <CL/cl.h>
#endif
#include "Func.h"
SchedDefineTask1(butterfly,butterfly);

static int
butterfly(SchedTask* s,void* rbuf,void* wbuf)
{
    cl_float2* x_in = (cl_float2*)s->get_inputAddr(0);
    cl_float2* w = (cl_float2*)s->get_inputAddr(1);
    int* n = (int*)s->get_inputAddr(3);
    unsigned int* flag = (unsigned int*)s->get_inputAddr(4);
    
    cl_float2* x_out = (cl_float2*)s->get_outputAddr(0);

    unsigned long gid = (unsigned long)s->get_param(0);
    unsigned long nid = (unsigned long)s->get_param(1);
    long iter = (long)s->get_param(2) + 1;

    int butterflySize = 1 << (iter-1);
    int butterflyGrpDist = 1 << iter;
    int butterflyGrpNum = n[0] >> iter;
    int butterflyGrpBase = (gid >> (iter-1))*(butterflyGrpDist);
    int butterflyGrpOffset = gid & (butterflySize-1);

    int a = nid * n[0] + butterflyGrpBase + butterflyGrpOffset;
    int b = a + butterflySize;

    int l = butterflyGrpNum * butterflyGrpOffset;

    cl_float2 xa, xb, xbxx, xbyy, wab, wayx, wbyx, resa, resb;

    xa = x_in[a];
    xb = x_in[b];
    xbxx.x = xbxx.y = xb.x;
    xbyy.x = xbyy.y = xb.y;
    
    wab.x = (cl_float)((cl_uint)w[l].x ^ 0x0);
    wab.y = (cl_float)((cl_uint)w[l].y ^ flag[0]);

    wayx.x = (cl_float)((cl_uint)wab.y ^ 0x80000000);
    wayx.y = (cl_float)((cl_uint)wab.x ^ 0x0);

    wbyx.x = (cl_float)((cl_uint)wab.y ^ 0x0);
    wbyx.y = (cl_float)((cl_uint)wab.x ^ 0x80000000);

    resa.x = xa.x + xbxx.x*wab.x + xbyy.x*wayx.x;
    resa.y = xa.y + xbxx.y*wab.y + xbyy.y*wayx.y;

    resb.x = xa.x - xbxx.x*wab.x + xbyy.x*wbyx.x;
    resb.y = xa.y - xbxx.y*wab.y + xbyy.y*wbyx.y;

    x_out[a] = resa;
    x_out[b] = resb;

    return 0;
}