bioem_algorithm.h 11 KB
Newer Older
qon's avatar
qon committed
1 2
#ifndef BIOEM_ALGORITHM_H
#define BIOEM_ALGORITHM_H
3 4 5 6 7 8 9 10 11 12
//#include <boost/iterator/iterator_concepts.hpp>

#ifndef BIOEM_GPUCODE
//#define SSECODE //Explicit SSE code, not correct yet since loop counter is assumed multiple of 4, anyway not faster than autovectorized code.
#endif

#ifdef SSECODE
#include <emmintrin.h>
#include <smmintrin.h>
#endif
qon's avatar
qon committed
13 14 15 16 17 18 19 20 21 22 23 24

template <int GPUAlgo, class RefT>
__device__ static inline void compareRefMap(const int iRefMap, const int iOrient, const int iConv, const bioem_map& Mapconv, bioem_Probability* pProb, const bioem_param_device& param, const RefT& RefMap,
	const int cent_x, const int cent_y, const int myShift = 0, const int nShifts2 = 0, const int myRef = 0)
{
    /**************************************************************************************/
    /**********************  Calculating BioEM Probability ********************************/
    /************************* Loop of center displacement here ***************************/

    // Taking into account the center displacement

    /*** Inizialzing crosscorrelations of calculated projected convolutions ***/
25 26 27 28 29
#ifdef SSECODE
	myfloat_t sum, sumsquare, crossproMapConv;
	__m128 sum_v = _mm_setzero_ps(), sumsquare_v = _mm_setzero_ps(), cross_v = _mm_setzero_ps(), d1, d2;
#else
	myfloat_t sum=0.0;
qon's avatar
qon committed
30 31
    myfloat_t sumsquare=0.0;
    myfloat_t crossproMapConv=0.0;
32
#endif
qon's avatar
qon committed
33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59
    /****** Loop over Pixels to calculate dot product and cross-correlations of displaced Ref Conv. Map***/
	if (GPUAlgo != 2 || iRefMap < RefMap.ntotRefMap) 
	{
		int iStart, jStart, iEnd, jEnd;
		if (cent_x < 0)
		{
			iStart = -cent_x;
			iEnd = param.NumberPixels;
		}
		else
		{
			iStart = 0;
			iEnd = param.NumberPixels - cent_x;
		}
		if (cent_y < 0)
		{
			jStart = -cent_y;
			jEnd = param.NumberPixels;
		}
		else
		{
			jStart = 0;
			jEnd = param.NumberPixels - cent_y;
		}

		for (int i = iStart; i < iEnd; i += 1)
		{
60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75
#ifdef SSECODE
			const float* ptr1 = &Mapconv.points[i + cent_x][jStart + cent_y];
			const float* ptr2 = RefMap.getp(iRefMap, i, jStart);
			int j;
			const int count = jEnd - jStart;
			for (j = 0;j <= count - 4;j += 4)
			{
				d1 = _mm_loadu_ps(ptr1);
				d2 = _mm_loadu_ps(ptr2);
				sum_v = _mm_add_ps(sum_v, d1);
				sumsquare_v = _mm_add_ps(sumsquare_v, _mm_mul_ps(d1, d1));
				cross_v = _mm_add_ps(cross_v, _mm_mul_ps(d1, d2));
				ptr1 += 4;
				ptr2 += 4;
			}
#else
qon's avatar
qon committed
76 77
			for (int j = jStart; j < jEnd; j += 1)
			{
78
				const myfloat_t pointMap = Mapconv.points[i + cent_x][j + cent_y];
qon's avatar
qon committed
79 80 81 82 83 84 85
				const myfloat_t pointRefMap = RefMap.get(iRefMap, i, j);
				crossproMapConv += pointMap * pointRefMap;
				// Crosscorrelation of calculated displaced map
				sum += pointMap;
				// Calculate Sum of pixels squared
				sumsquare += pointMap*pointMap;
			}
86
#endif
qon's avatar
qon committed
87
		}
88 89 90 91 92 93 94 95 96 97 98
#ifdef SSECODE
		sum_v = _mm_hadd_ps(sum_v, sum_v);
		sumsquare_v = _mm_hadd_ps(sumsquare_v, sumsquare_v);
		cross_v = _mm_hadd_ps(cross_v, cross_v);
		sum_v = _mm_hadd_ps(sum_v, sum_v);
		sumsquare_v = _mm_hadd_ps(sumsquare_v, sumsquare_v);
		cross_v = _mm_hadd_ps(cross_v, cross_v);
		sum = _mm_cvtss_f32(sum_v);
		sumsquare = _mm_cvtss_f32(sumsquare_v);
		crossproMapConv = _mm_cvtss_f32(cross_v);
#endif
qon's avatar
qon committed
99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328
	}
	
    /********** Calculating elements in BioEM Probability formula ********/
    // Related to Reference calculated Projection
    const myfloat_t ForLogProb = (sumsquare * param.Ntotpi - sum * sum);

    // Products of different cross-correlations (first element in formula)
    const myfloat_t firstele = param.Ntotpi * (RefMap.sumsquare_RefMap[iRefMap] * sumsquare-crossproMapConv * crossproMapConv) +
                             2 * RefMap.sum_RefMap[iRefMap] * sum * crossproMapConv - RefMap.sumsquare_RefMap[iRefMap] * sum * sum - RefMap.sum_RefMap[iRefMap] * RefMap.sum_RefMap[iRefMap] * sumsquare;

    //******* Calculating log of Prob*********/
    // As in fortran code: logpro=(3-Ntotpi)*0.5*log(firstele/pConvMap[iOrient].ForLogProbfromConv[iConv])+(Ntotpi*0.5-2)*log(Ntotpi-2)-0.5*log(pConvMap[iOrient].ForLogProbfromConv[iConv])+0.5*log(PI)+(1-Ntotpi*0.5)*(log(2*PI)+1);
    const myfloat_t logpro = (3 - param.Ntotpi) * 0.5 * log(firstele) + (param.Ntotpi * 0.5 - 2) * log((param.Ntotpi - 2) * ForLogProb);

#ifdef BIOEM_GPUCODE
	if (GPUAlgo == 2)
	{
		extern __shared__ myfloat_t buf[];
		myfloat_t* buf2 = &buf[myBlockDimX];
		myfloat_t* buf3 = &buf2[myBlockDimX + 4 * myRef];
		int* bufint = (int*) buf3;
		
		buf[myThreadIdxX] = logpro;
		if (myShift == 0)
		{
			bufint[0] = 0;
		}
		__syncthreads();
		
		if (nShifts2 == CUDA_MAX_SHIFT_REDUCE) // 1024
		{
			if (myShift < 512) if (buf[myThreadIdxX + 512] > buf[myThreadIdxX]) buf[myThreadIdxX] = buf[myThreadIdxX + 512];
			__syncthreads();
		}
		
		if (nShifts2 >= 512)
		{
			if (myShift < 256) if (buf[myThreadIdxX + 256] > buf[myThreadIdxX]) buf[myThreadIdxX] = buf[myThreadIdxX + 256];
			__syncthreads();
		}

		if (nShifts2 >= 256)
		{
			if (myShift < 128) if (buf[myThreadIdxX + 128] > buf[myThreadIdxX]) buf[myThreadIdxX] = buf[myThreadIdxX + 128];
			__syncthreads();
		}

		if (nShifts2 >= 128)
		{
			if (myShift < 64) if (buf[myThreadIdxX + 64] > buf[myThreadIdxX]) buf[myThreadIdxX] = buf[myThreadIdxX + 64];
			__syncthreads();
		}

		if (myShift < 32) //Warp Size is 32, threads are synched automatically
		{
			volatile float* vbuf = buf; //Mem must be volatile such that memory access is not reordered
			if (nShifts2 >= 64 && vbuf[myThreadIdxX + 32] > vbuf[myThreadIdxX]) vbuf[myThreadIdxX] = vbuf[myThreadIdxX + 32];
			if (nShifts2 >= 32 && vbuf[myThreadIdxX + 16] > vbuf[myThreadIdxX]) vbuf[myThreadIdxX] = vbuf[myThreadIdxX + 16];
			if (nShifts2 >= 16 && vbuf[myThreadIdxX + 8] > vbuf[myThreadIdxX]) vbuf[myThreadIdxX] = vbuf[myThreadIdxX + 8];
			if (nShifts2 >= 8 && vbuf[myThreadIdxX + 4] > vbuf[myThreadIdxX]) vbuf[myThreadIdxX] = vbuf[myThreadIdxX + 4];
			if (nShifts2 >= 4 && vbuf[myThreadIdxX + 2] > vbuf[myThreadIdxX]) vbuf[myThreadIdxX] = vbuf[myThreadIdxX + 2];
			if (nShifts2 >= 2 && vbuf[myThreadIdxX + 1] > vbuf[myThreadIdxX]) vbuf[myThreadIdxX] = vbuf[myThreadIdxX + 1];
			if (myShift == 0 && iRefMap < RefMap.ntotRefMap)
			{
				const myfloat_t logpro_max = vbuf[myThreadIdxX];
				if(pProb[iRefMap].Constoadd < logpro_max)
				{
					pProb[iRefMap].Total = pProb[iRefMap].Total * exp(-logpro_max + pProb[iRefMap].Constoadd);
					pProb[iRefMap].Constoadd = logpro_max;
				}
				if(pProb[iRefMap].ConstAngle[iOrient] < logpro_max)
				{
					pProb[iRefMap].forAngles[iOrient] = pProb[iRefMap].forAngles[iOrient] * exp(-logpro_max + pProb[iRefMap].ConstAngle[iOrient]);
					pProb[iRefMap].ConstAngle[iOrient] = logpro_max;
				}
				if(pProb[iRefMap].max_prob < logpro_max)
				{
					pProb[iRefMap].max_prob = logpro_max;
					pProb[iRefMap].max_prob_orient = iOrient;
					pProb[iRefMap].max_prob_conv = iConv;
					bufint[0] = 1;
					buf3[1] = logpro_max;
				}
			}
		}
		
		__syncthreads();
		if (bufint[0] == 1 && buf3[1] == logpro && iRefMap < RefMap.ntotRefMap && atomicAdd(&bufint[0], 1) == 1)
		{
			pProb[iRefMap].max_prob_cent_x = cent_x;
			pProb[iRefMap].max_prob_cent_y = cent_y;
		}
		
		__syncthreads();
		
		if (iRefMap < RefMap.ntotRefMap)
		{
			buf[myThreadIdxX] = exp(logpro - pProb[iRefMap].Constoadd);
			buf2[myThreadIdxX] = exp(logpro - pProb[iRefMap].ConstAngle[iOrient]);
		}
		__syncthreads();
		
		if (nShifts2 == CUDA_MAX_SHIFT_REDUCE) // 1024
		{
			if (myShift < 512)
			{
				buf[myThreadIdxX] += buf[myThreadIdxX + 512];
				buf2[myThreadIdxX] += buf2[myThreadIdxX + 512];
			}
			__syncthreads();
		}
		
		if (nShifts2 >= 512)
		{
			if (myShift < 256)
			{
				buf[myThreadIdxX] += buf[myThreadIdxX + 256];
				buf2[myThreadIdxX] += buf2[myThreadIdxX + 256];
			}
			__syncthreads();
		}

		if (nShifts2 >= 256)
		{
			if (myShift < 128)
			{
				buf[myThreadIdxX] += buf[myThreadIdxX + 128];
				buf2[myThreadIdxX] += buf2[myThreadIdxX + 128];
			}
			__syncthreads();
		}

		if (nShifts2 >= 128)
		{
			if (myShift < 64)
			{
				buf[myThreadIdxX] += buf[myThreadIdxX + 64];
				buf2[myThreadIdxX] += buf2[myThreadIdxX + 64];
			}
			__syncthreads();
		}

		if (myShift < 32) //Warp Size is 32, threads are synched automatically
		{
			volatile float* vbuf = buf; //Mem must be volatile such that memory access is not reordered
			volatile float* vbuf2 = buf2;
			if (nShifts2 >= 64)
			{
				vbuf[myThreadIdxX] += vbuf[myThreadIdxX + 32];
				vbuf2[myThreadIdxX] += vbuf2[myThreadIdxX + 32];
			}
			if (nShifts2 >= 32)
			{
				vbuf[myThreadIdxX] += vbuf[myThreadIdxX + 16];
				vbuf2[myThreadIdxX] += vbuf2[myThreadIdxX + 16];
			}
			if (nShifts2 >= 16)
			{
				vbuf[myThreadIdxX] += vbuf[myThreadIdxX + 8];
				vbuf2[myThreadIdxX] += vbuf2[myThreadIdxX + 8];
			}
			if (nShifts2 >= 8)
			{
				vbuf[myThreadIdxX] += vbuf[myThreadIdxX + 4];
				vbuf2[myThreadIdxX] += vbuf2[myThreadIdxX + 4];
			}
			if (nShifts2 >= 4)
			{
				vbuf[myThreadIdxX] += vbuf[myThreadIdxX + 2];
				vbuf2[myThreadIdxX] += vbuf2[myThreadIdxX + 2];
			}
			if (nShifts2 >= 2)
			{
				vbuf[myThreadIdxX] += vbuf[myThreadIdxX + 1];
				vbuf2[myThreadIdxX] += vbuf2[myThreadIdxX + 1];
			}
			if (myShift == 0 && iRefMap < RefMap.ntotRefMap)
			{
				pProb[iRefMap].Total += vbuf[myThreadIdxX];
				pProb[iRefMap].forAngles[iOrient] += vbuf2[myThreadIdxX];
			}
		}
	}
	else
#endif

    /***** Summing & Storing total/Orientation Probabilites for each map ************/
    {
        /*******  Summing total Probabilities *************/
        /******* Need a constant because of numerical divergence*****/
        if(pProb[iRefMap].Constoadd < logpro)
        {
            pProb[iRefMap].Total = pProb[iRefMap].Total * exp(-logpro + pProb[iRefMap].Constoadd);
            pProb[iRefMap].Constoadd = logpro;
        }
        pProb[iRefMap].Total += exp(logpro - pProb[iRefMap].Constoadd);

		//Summing probabilities for each orientation
        if(pProb[iRefMap].ConstAngle[iOrient] < logpro)
        {
            pProb[iRefMap].forAngles[iOrient] = pProb[iRefMap].forAngles[iOrient] * exp(-logpro + pProb[iRefMap].ConstAngle[iOrient]);
            pProb[iRefMap].ConstAngle[iOrient] = logpro;
		}
        pProb[iRefMap].forAngles[iOrient] += exp(logpro - pProb[iRefMap].ConstAngle[iOrient]);

        /********** Getting parameters that maximize the probability ***********/
        if(pProb[iRefMap].max_prob < logpro)
        {
            pProb[iRefMap].max_prob = logpro;
            pProb[iRefMap].max_prob_cent_x = cent_x;
            pProb[iRefMap].max_prob_cent_y = cent_y;
            pProb[iRefMap].max_prob_orient = iOrient;
            pProb[iRefMap].max_prob_conv = iConv;
        }
    }
}

template <int GPUAlgo, class RefT>
__device__ static inline void compareRefMapShifted(const int iRefMap, const int iOrient, const int iConv, const bioem_map& Mapconv, bioem_Probability* pProb, const bioem_param_device& param, const RefT& RefMap)
{
    for (int cent_x = -param.maxDisplaceCenter; cent_x <= param.maxDisplaceCenter; cent_x=cent_x+param.GridSpaceCenter)
    {
        for (int cent_y = -param.maxDisplaceCenter; cent_y <= param.maxDisplaceCenter; cent_y=cent_y+param.GridSpaceCenter)
        {
			compareRefMap<GPUAlgo>(iRefMap, iOrient, iConv, Mapconv, pProb, param, RefMap, cent_x, cent_y);
        }
    }
}

#endif