| 1 |
/* |
| 2 |
* Advanced Simulation Library <http://asl.org.il> |
| 3 |
* |
| 4 |
* Copyright 2015 Avtech Scientific <http://avtechscientific.com> |
| 5 |
* |
| 6 |
* |
| 7 |
* This file is part of Advanced Simulation Library (ASL). |
| 8 |
* |
| 9 |
* ASL is free software: you can redistribute it and/or modify it |
| 10 |
* under the terms of the GNU Affero General Public License as |
| 11 |
* published by the Free Software Foundation, version 3 of the License. |
| 12 |
* |
| 13 |
* ASL is distributed in the hope that it will be useful, |
| 14 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 |
* GNU Affero General Public License for more details. |
| 17 |
* |
| 18 |
* You should have received a copy of the GNU Affero General Public License |
| 19 |
* along with ASL. If not, see <http://www.gnu.org/licenses/>. |
| 20 |
* |
| 21 |
*/ |
| 22 |
|
| 23 |
|
| 24 |
#include "aclReductionAlgGenerator.h" |
| 25 |
#include "aclVectorOfElementsOperations.h" |
| 26 |
#include <acl/acl.h> |
| 27 |
#include <acl/DataTypes/aclArray.h> |
| 28 |
#include <acl/aclGenerators.h> |
| 29 |
#include <acl/Kernels/aclKernel.h> |
| 30 |
|
| 31 |
namespace acl |
| 32 |
{ |
| 33 |
|
| 34 |
inline const unsigned int nLocalGPU(unsigned int length, unsigned int nGroups) |
| 35 |
{ |
| 36 |
return std::max(1u,std::min(8u, length/nGroups)); |
| 37 |
} |
| 38 |
inline const unsigned int nLocalCPU(unsigned int length, unsigned int nGroups) |
| 39 |
{ |
| 40 |
return 1; |
| 41 |
} |
| 42 |
|
| 43 |
unsigned int getLPerUnit(unsigned int length, unsigned int nUnits) |
| 44 |
{ |
| 45 |
return ceil((double)length/nUnits); |
| 46 |
} |
| 47 |
|
| 48 |
unsigned int getNSaturatedUnits(unsigned int length, unsigned int nUnits) |
| 49 |
{ |
| 50 |
return length/getLPerUnit(length,nUnits); |
| 51 |
} |
| 52 |
|
| 53 |
unsigned int getLLastUnit(unsigned int length, unsigned int nUnits) |
| 54 |
{ |
| 55 |
return length % getLPerUnit(length,nUnits); |
| 56 |
} |
| 57 |
|
| 58 |
template <typename ResType, enum ReductionOperatorType Op> class GroupReduction |
| 59 |
{ |
| 60 |
}; |
| 61 |
|
| 62 |
template <typename ResType> class GroupReduction<ResType, ROT_SUM> |
| 63 |
{ |
| 64 |
public: |
| 65 |
static void compute(std::vector<std::vector<ResType>> v, unsigned int l, asl::AVec<ResType> & res) |
| 66 |
{ |
| 67 |
for(unsigned int i(0); i<v.size(); ++i) |
| 68 |
{ |
| 69 |
vector<ResType> & vv(v[i]); |
| 70 |
ResType r(vv[0]); |
| 71 |
for(unsigned int j(1); j<l; ++j) |
| 72 |
r+=vv[j]; |
| 73 |
res[i]=r; |
| 74 |
} |
| 75 |
} |
| 76 |
static inline VectorOfElements op(VectorOfElements res, VectorOfElements v, acl::TypeID t) |
| 77 |
{ |
| 78 |
return res+=v; |
| 79 |
} |
| 80 |
}; |
| 81 |
// template void GroupReduction<double, ROT_SUM>::compute(std::vector<std::vector<double>> v, |
| 82 |
// asl::AVec<double> & res); |
| 83 |
// template void GroupReduction<float, ROT_SUM>::compute(std::vector<std::vector<float>> v, |
| 84 |
// asl::AVec<float> & res); |
| 85 |
|
| 86 |
template <typename ResType> class GroupReduction<ResType, ROT_PRODUCT> |
| 87 |
{ |
| 88 |
public: |
| 89 |
static void compute(std::vector<std::vector<ResType>> v, unsigned int l, asl::AVec<ResType> & res) |
| 90 |
{ |
| 91 |
for(unsigned int i(0); i<v.size(); ++i) |
| 92 |
{ |
| 93 |
vector<ResType> & vv(v[i]); |
| 94 |
ResType r(vv[0]); |
| 95 |
for(unsigned int j(1); j<l; ++j) |
| 96 |
r*=vv[j]; |
| 97 |
res[i]=r; |
| 98 |
} |
| 99 |
} |
| 100 |
static inline VectorOfElements op(VectorOfElements res, VectorOfElements v, acl::TypeID t) |
| 101 |
{ |
| 102 |
return res*=v; |
| 103 |
} |
| 104 |
}; |
| 105 |
// template void GroupReduction<double, ROT_PRODUCT>::compute(std::vector<std::vector<double>> v, |
| 106 |
// asl::AVec<double> & res); |
| 107 |
// template void GroupReduction<float, ROT_PRODUCT>::compute(std::vector<std::vector<float>> v, |
| 108 |
// asl::AVec<float> & res); |
| 109 |
|
| 110 |
template <typename ResType> class GroupReduction<ResType, ROT_MINIMUM> |
| 111 |
{ |
| 112 |
public: |
| 113 |
static void compute(std::vector<std::vector<ResType>> v, unsigned int l, asl::AVec<ResType> & res) |
| 114 |
{ |
| 115 |
for(unsigned int i(0); i<v.size(); ++i) |
| 116 |
{ |
| 117 |
vector<ResType> & vv(v[i]); |
| 118 |
ResType r(vv[0]); |
| 119 |
for(unsigned int j(1); j<l; ++j) |
| 120 |
r=std::min(r,vv[j]); |
| 121 |
res[i]=r; |
| 122 |
} |
| 123 |
} |
| 124 |
static inline VectorOfElements op(VectorOfElements res, VectorOfElements v, acl::TypeID t) |
| 125 |
{ |
| 126 |
return res=min(res,v,t); |
| 127 |
} |
| 128 |
}; |
| 129 |
// template void GroupReduction<double, ROT_MINIMUM>::compute(std::vector<std::vector<double>> v, |
| 130 |
// asl::AVec<double> & res); |
| 131 |
// template void GroupReduction<float, ROT_MINIMUM>::compute(std::vector<std::vector<float>> v, |
| 132 |
// asl::AVec<float> & res); |
| 133 |
|
| 134 |
template <typename ResType> class GroupReduction<ResType, ROT_MAXIMUM> |
| 135 |
{ |
| 136 |
public: |
| 137 |
static void compute(std::vector<std::vector<ResType>> v, unsigned int l, asl::AVec<ResType> & res) |
| 138 |
{ |
| 139 |
for(unsigned int i(0); i<v.size(); ++i) |
| 140 |
{ |
| 141 |
vector<ResType> & vv(v[i]); |
| 142 |
ResType r(vv[0]); |
| 143 |
for(unsigned int j(1); j<l; ++j) |
| 144 |
r=std::max(r,vv[j]); |
| 145 |
res[i]=r; |
| 146 |
} |
| 147 |
} |
| 148 |
static inline VectorOfElements op(VectorOfElements res, VectorOfElements v, acl::TypeID t) |
| 149 |
{ |
| 150 |
return res=max(res,v,t); |
| 151 |
} |
| 152 |
}; |
| 153 |
// template void GroupReduction<double, ROT_MAXIMUM>::compute(std::vector<std::vector<double>> v, |
| 154 |
// asl::AVec<double> & res); |
| 155 |
// template void GroupReduction<float, ROT_MAXIMUM>::compute(std::vector<std::vector<float>> v, |
| 156 |
// asl::AVec<float> & res); |
| 157 |
|
| 158 |
template <typename ResType, enum ReductionOperatorType Op> |
| 159 |
void ReductionAlgGenerator<ResType, Op>::compute() |
| 160 |
{ |
| 161 |
if(kernel) |
| 162 |
kernel->compute(); |
| 163 |
unsigned int nC(ve.size()); |
| 164 |
unsigned int veLength(ve[0]->getSize()); |
| 165 |
unsigned int nSaturatedUnits(getNSaturatedUnits(veLength, nGroups*nLocal)); |
| 166 |
unsigned int l(std::min(nSaturatedUnits+1,nGroups*nLocal)); |
| 167 |
for (unsigned int i(0); i < nC; ++i) |
| 168 |
acl::copy(groupResACL[i], groupRes[i]); |
| 169 |
GroupReduction<ResType, Op>::compute(groupRes, l, res.v()); |
| 170 |
}; |
| 171 |
template void ReductionAlgGenerator<double, ROT_SUM>::compute(); |
| 172 |
template void ReductionAlgGenerator<float, ROT_SUM>::compute(); |
| 173 |
template void ReductionAlgGenerator<double, ROT_PRODUCT>::compute(); |
| 174 |
template void ReductionAlgGenerator<float, ROT_PRODUCT>::compute(); |
| 175 |
template void ReductionAlgGenerator<double, ROT_MINIMUM>::compute(); |
| 176 |
template void ReductionAlgGenerator<float, ROT_MINIMUM>::compute(); |
| 177 |
template void ReductionAlgGenerator<double, ROT_MAXIMUM>::compute(); |
| 178 |
template void ReductionAlgGenerator<float, ROT_MAXIMUM>::compute(); |
| 179 |
|
| 180 |
template <typename ResType, enum ReductionOperatorType Op> |
| 181 |
void generateKernelCPU(VectorOfElements ve, VectorOfElements groupResACL, Kernel & k) |
| 182 |
{ |
| 183 |
auto nGroups(k.getGroupsNumber()); |
| 184 |
int veLength(ve[0]->getSize()); |
| 185 |
unsigned int nLocal(nLocalCPU(veLength,nGroups)); |
| 186 |
unsigned int gSize(std::max(k.getSize(),nLocal)); |
| 187 |
auto type(getElementType(ve)); |
| 188 |
auto typeI(TYPE_SELECT[type]); |
| 189 |
auto lVal(acl::generateVEPrivateVariable(ve.size(),type)); |
| 190 |
auto counter(acl::generateVEPrivateVariable(1,typeI)); |
| 191 |
auto counterEnd(acl::generateVEPrivateVariable(1,typeI)); |
| 192 |
int lPerGroup(getLPerUnit(veLength,nGroups)); |
| 193 |
int lLastGroup(getLLastUnit(veLength, nGroups)); |
| 194 |
int nSaturatedGroups(getNSaturatedUnits(veLength, nGroups)); |
| 195 |
|
| 196 |
k<<(counterEnd = select(generateVEConstant(lPerGroup), |
| 197 |
generateVEConstant(lLastGroup), |
| 198 |
generateVEGroupID()==nSaturatedGroups, |
| 199 |
type)); |
| 200 |
k<<(lVal = select(excerpt(ve,lPerGroup*generateVEGroupID()), counterEnd>0,type)); |
| 201 |
// this line ensuers that the conputations will take place on the 1st item only |
| 202 |
k<<(counterEnd = select(counterEnd, generateVEIndex(gSize)==0, type)); |
| 203 |
vector<Element> body; |
| 204 |
body<<(GroupReduction<ResType, Op>:: |
| 205 |
op(lVal,excerpt(ve, |
| 206 |
int(lPerGroup)*generateVEGroupID() + |
| 207 |
counter),type)); |
| 208 |
|
| 209 |
auto loop(elementOperators::forLoop((counter = generateVEConstant(1))[0], |
| 210 |
(counter < counterEnd)[0], |
| 211 |
(counter += generateVEConstant(1))[0], |
| 212 |
body)); |
| 213 |
k.addExpression(loop); |
| 214 |
k<<(excerpt(groupResACL,generateVEGroupID()) = lVal); |
| 215 |
} |
| 216 |
|
| 217 |
template <typename ResType, enum ReductionOperatorType Op> |
| 218 |
void generateKernelGPU(VectorOfElements ve, VectorOfElements groupResACL, Kernel & k) |
| 219 |
{ |
| 220 |
auto nGroups(k.getGroupsNumber()); |
| 221 |
int veLength(ve[0]->getSize()); |
| 222 |
unsigned int nLocal(nLocalGPU(veLength,nGroups)); |
| 223 |
unsigned int gSize(std::max(k.getSize(),nLocal)); |
| 224 |
auto type(getElementType(ve)); |
| 225 |
auto typeI(TYPE_SELECT[type]); |
| 226 |
auto lVal(acl::generateVEPrivateVariable(ve.size(),type)); |
| 227 |
auto counter(acl::generateVEPrivateVariable(1,typeI)); |
| 228 |
auto counterEnd(acl::generateVEPrivateVariable(1,typeI)); |
| 229 |
auto nUnits(nGroups*nLocal); |
| 230 |
int lPerUnit(getLPerUnit(veLength,nUnits)); |
| 231 |
int lLastUnit(getLLastUnit(veLength, nUnits)); |
| 232 |
unsigned int nSaturatedUnits(getNSaturatedUnits(veLength, nUnits)); |
| 233 |
auto unitID(generateVEGroupID()*nLocal +generateVEIndex()); |
| 234 |
|
| 235 |
if(nSaturatedUnits<nUnits) |
| 236 |
k<<(counterEnd = select(generateVEConstant(lPerUnit), |
| 237 |
generateVEConstant(lLastUnit), |
| 238 |
unitID==nSaturatedUnits, |
| 239 |
typeI)); |
| 240 |
if((nSaturatedUnits+1)<nUnits) |
| 241 |
k<<(counterEnd = select(counterEnd, |
| 242 |
unitID<=nSaturatedUnits, |
| 243 |
typeI)); |
| 244 |
|
| 245 |
// this line ensuers that the conputations will take place on first item only |
| 246 |
k<<(counterEnd = select(counterEnd, generateVEIndex(gSize)<nLocal, type)); |
| 247 |
k<<(lVal = select(excerpt(ve,lPerUnit*unitID), counterEnd>0,type)); |
| 248 |
vector<Element> body; |
| 249 |
body<<(GroupReduction<ResType, Op>:: |
| 250 |
op(lVal,excerpt(ve, lPerUnit*unitID + counter),type)); |
| 251 |
|
| 252 |
auto loop(elementOperators::forLoop((counter = generateVEConstant(1))[0], |
| 253 |
(counter < counterEnd)[0], |
| 254 |
(counter += generateVEConstant(1))[0], |
| 255 |
body)); |
| 256 |
k.addExpression(loop); |
| 257 |
k<<(excerpt(groupResACL,unitID) = lVal); |
| 258 |
} |
| 259 |
|
| 260 |
template <typename ResType, enum ReductionOperatorType Op> |
| 261 |
void ReductionAlgGenerator<ResType, Op>::generateAlg(Kernel & k) |
| 262 |
{ |
| 263 |
if(!k.getConfiguration().local) |
| 264 |
asl::errorMessage("ReductionAlgGenerator::generateAlg: The kernel should be local"); |
| 265 |
|
| 266 |
nGroups= k.getGroupsNumber(); |
| 267 |
unsigned int veLength(ve[0]->getSize()); |
| 268 |
nLocal=(getDeviceType(ve[0]->getQueue())== CL_DEVICE_TYPE_CPU ? |
| 269 |
nLocalCPU(veLength,nGroups) : nLocalGPU(veLength,nGroups)); |
| 270 |
for(unsigned int i(0); i<ve.size(); ++i) |
| 271 |
groupRes[i]=std::vector<ResType>(nGroups*nLocal); |
| 272 |
copy(generateVEData<ResType>(nGroups*nLocal,ve.size()),groupResACL); |
| 273 |
|
| 274 |
|
| 275 |
switch (getDeviceType(ve[0]->getQueue())) |
| 276 |
{ |
| 277 |
case CL_DEVICE_TYPE_CPU: |
| 278 |
generateKernelCPU<ResType, Op>(ve,groupResACL,k); |
| 279 |
k.setup(); |
| 280 |
break; |
| 281 |
case CL_DEVICE_TYPE_GPU: |
| 282 |
generateKernelGPU<ResType, Op>(ve,groupResACL,k); |
| 283 |
k.setup(); |
| 284 |
break; |
| 285 |
default: errorMessage("ReductionAlgGenerator: device type " + |
| 286 |
numToStr(getDeviceType(k.getQueue())) + " is unknown!"); |
| 287 |
} |
| 288 |
} |
| 289 |
|
| 290 |
template void ReductionAlgGenerator<double, ROT_SUM>::generateAlg(Kernel & k); |
| 291 |
template void ReductionAlgGenerator<float, ROT_SUM>::generateAlg(Kernel & k); |
| 292 |
template void ReductionAlgGenerator<double, ROT_PRODUCT>::generateAlg(Kernel & k); |
| 293 |
template void ReductionAlgGenerator<float, ROT_PRODUCT>::generateAlg(Kernel & k); |
| 294 |
template void ReductionAlgGenerator<double, ROT_MINIMUM>::generateAlg(Kernel & k); |
| 295 |
template void ReductionAlgGenerator<float, ROT_MINIMUM>::generateAlg(Kernel & k); |
| 296 |
template void ReductionAlgGenerator<double, ROT_MAXIMUM>::generateAlg(Kernel & k); |
| 297 |
template void ReductionAlgGenerator<float, ROT_MAXIMUM>::generateAlg(Kernel & k); |
| 298 |
|
| 299 |
template <typename ResType, enum ReductionOperatorType Op> |
| 300 |
void ReductionAlgGenerator<ResType, Op>::generateAlg() |
| 301 |
{ |
| 302 |
KernelConfiguration kConf(KERNEL_BASIC); |
| 303 |
kConf.local = true; |
| 304 |
|
| 305 |
kernel=std::make_shared<Kernel>(kConf); |
| 306 |
unsigned int groupsNumber = getNComputeUnits(ve[0]->getQueue()); |
| 307 |
kernel->setGroupsNumber(groupsNumber); |
| 308 |
|
| 309 |
generateAlg(*kernel); |
| 310 |
}; |
| 311 |
template void ReductionAlgGenerator<double, ROT_SUM>::generateAlg(); |
| 312 |
template void ReductionAlgGenerator<float, ROT_SUM>::generateAlg(); |
| 313 |
template void ReductionAlgGenerator<double, ROT_PRODUCT>::generateAlg(); |
| 314 |
template void ReductionAlgGenerator<float, ROT_PRODUCT>::generateAlg(); |
| 315 |
template void ReductionAlgGenerator<double, ROT_MINIMUM>::generateAlg(); |
| 316 |
template void ReductionAlgGenerator<float, ROT_MINIMUM>::generateAlg(); |
| 317 |
template void ReductionAlgGenerator<double, ROT_MAXIMUM>::generateAlg(); |
| 318 |
template void ReductionAlgGenerator<float, ROT_MAXIMUM>::generateAlg(); |
| 319 |
|
| 320 |
template <typename ResType, enum ReductionOperatorType Op> |
| 321 |
ReductionAlgGenerator<ResType, Op>::ReductionAlgGenerator(VectorOfElements v): |
| 322 |
ve(v), |
| 323 |
res(asl::AVec<ResType>(ve.size())), |
| 324 |
groupRes(ve.size()) |
| 325 |
{ |
| 326 |
} |
| 327 |
|
| 328 |
template <typename ResType> std::shared_ptr<ReductionAlgGenerator<ResType, ROT_SUM>> |
| 329 |
generateSumAlg(VectorOfElements v) |
| 330 |
{ |
| 331 |
return make_shared<ReductionAlgGenerator<ResType, ROT_SUM>>(v); |
| 332 |
} |
| 333 |
|
| 334 |
template std::shared_ptr<ReductionAlgGenerator<double, ROT_SUM>> |
| 335 |
generateSumAlg<double>(VectorOfElements v); |
| 336 |
template std::shared_ptr<ReductionAlgGenerator<float, ROT_SUM>> |
| 337 |
generateSumAlg<float>(VectorOfElements v); |
| 338 |
|
| 339 |
template <typename ResType> std::shared_ptr<ReductionAlgGenerator<ResType, ROT_MINIMUM>> |
| 340 |
generateMinAlg(VectorOfElements v) |
| 341 |
{ |
| 342 |
return make_shared<ReductionAlgGenerator<ResType, ROT_MINIMUM>>(v); |
| 343 |
} |
| 344 |
|
| 345 |
template std::shared_ptr<ReductionAlgGenerator<double, ROT_MINIMUM>> |
| 346 |
generateMinAlg<double>(VectorOfElements v); |
| 347 |
template std::shared_ptr<ReductionAlgGenerator<float, ROT_MINIMUM>> |
| 348 |
generateMinAlg<float>(VectorOfElements v); |
| 349 |
|
| 350 |
template <typename ResType> std::shared_ptr<ReductionAlgGenerator<ResType, ROT_MAXIMUM>> |
| 351 |
generateMaxAlg(VectorOfElements v) |
| 352 |
{ |
| 353 |
return make_shared<ReductionAlgGenerator<ResType, ROT_MAXIMUM>>(v); |
| 354 |
} |
| 355 |
|
| 356 |
template std::shared_ptr<ReductionAlgGenerator<double, ROT_MAXIMUM>> |
| 357 |
generateMaxAlg<double>(VectorOfElements v); |
| 358 |
template std::shared_ptr<ReductionAlgGenerator<float, ROT_MAXIMUM>> |
| 359 |
generateMaxAlg<float>(VectorOfElements v); |
| 360 |
|
| 361 |
template <typename ResType> std::shared_ptr<ReductionAlgGenerator<ResType, ROT_PRODUCT>> |
| 362 |
generateProductAlg(VectorOfElements v) |
| 363 |
{ |
| 364 |
return make_shared<ReductionAlgGenerator<ResType, ROT_PRODUCT>>(v); |
| 365 |
} |
| 366 |
|
| 367 |
template std::shared_ptr<ReductionAlgGenerator<double, ROT_PRODUCT>> |
| 368 |
generateProductAlg<double>(VectorOfElements v); |
| 369 |
template std::shared_ptr<ReductionAlgGenerator<float, ROT_PRODUCT>> |
| 370 |
generateProductAlg<float>(VectorOfElements v); |
| 371 |
|
| 372 |
} // namespace acl |