| 1 | /* | 
|---|
| 2 | * $Id: methods.cc,v 1.1.1.1 1999-11-26 16:37:07 ansari Exp $ | 
|---|
| 3 | * | 
|---|
| 4 | * $Log: not supported by cvs2svn $ | 
|---|
| 5 | // Revision 1.1.1.1  1999/04/09  17:59:03  ansari | 
|---|
| 6 | // Creation module DPC/Blitz (blitz 0.4) Reza 09/04/99 | 
|---|
| 7 | // | 
|---|
| 8 | * Revision 1.4  1998/03/14 00:04:47  tveldhui | 
|---|
| 9 | * 0.2-alpha-05 | 
|---|
| 10 | * | 
|---|
| 11 | * Revision 1.3  1997/08/18 19:13:08  tveldhui | 
|---|
| 12 | * Just prior to implementing fastRead() optimization for array | 
|---|
| 13 | * expression evaluation. | 
|---|
| 14 | * | 
|---|
| 15 | * Revision 1.2  1997/08/15 21:14:10  tveldhui | 
|---|
| 16 | * Just prior to loop-collapse change | 
|---|
| 17 | * | 
|---|
| 18 | */ | 
|---|
| 19 |  | 
|---|
| 20 | #ifndef BZ_ARRAYMETHODS_CC | 
|---|
| 21 | #define BZ_ARRAYMETHODS_CC | 
|---|
| 22 |  | 
|---|
| 23 | #ifndef BZ_ARRAY_H | 
|---|
| 24 | #error <blitz/array/methods.cc> must be included via <blitz/array.h> | 
|---|
| 25 | #endif | 
|---|
| 26 |  | 
|---|
| 27 | #include <blitz/minmax.h>  // Needed for resizeAndPreserve() | 
|---|
| 28 |  | 
|---|
| 29 | BZ_NAMESPACE(blitz) | 
|---|
| 30 |  | 
|---|
| 31 | template<class P_numtype, int N_rank> template<class T_expr> | 
|---|
| 32 | Array<P_numtype,N_rank>::Array(_bz_ArrayExpr<T_expr> expr) | 
|---|
| 33 | { | 
|---|
| 34 | BZ_NOT_IMPLEMENTED(); | 
|---|
| 35 |  | 
|---|
| 36 | // Obtain storage order from an operand in the expression | 
|---|
| 37 | // (if possible).  Probably best to assume C-style storage, | 
|---|
| 38 | // then pass the storage object to the expression for possible | 
|---|
| 39 | // modification. | 
|---|
| 40 |  | 
|---|
| 41 | // Obtain ubounds/lbounds from array operands.  Precondition | 
|---|
| 42 | // failure if any bounds missing. | 
|---|
| 43 |  | 
|---|
| 44 | // Size array. | 
|---|
| 45 |  | 
|---|
| 46 | // assignment of expression. | 
|---|
| 47 | } | 
|---|
| 48 |  | 
|---|
| 49 | template<class T_numtype, int N_rank> | 
|---|
| 50 | Array<T_numtype,N_rank>::Array(const TinyVector<int, N_rank>& lbounds, | 
|---|
| 51 | const TinyVector<int, N_rank>& extent, | 
|---|
| 52 | const GeneralArrayStorage<N_rank>& storage) | 
|---|
| 53 | : storage_(storage) | 
|---|
| 54 | { | 
|---|
| 55 | length_ = extent; | 
|---|
| 56 | storage_.setBase(lbounds); | 
|---|
| 57 | setupStorage(N_rank - 1); | 
|---|
| 58 | } | 
|---|
| 59 |  | 
|---|
| 60 |  | 
|---|
| 61 | /* | 
|---|
| 62 | * This routine takes the storage information for the array | 
|---|
| 63 | * (ascendingFlag_[], base_[], and ordering_[]) and the size | 
|---|
| 64 | * of the array (length_[]) and computes the stride vector | 
|---|
| 65 | * (stride_[]) and the zero offset (see explanation in array.h). | 
|---|
| 66 | */ | 
|---|
| 67 | template<class P_numtype, int N_rank> | 
|---|
| 68 | _bz_inline2 void Array<P_numtype, N_rank>::computeStrides() | 
|---|
| 69 | { | 
|---|
| 70 | if (N_rank > 1) | 
|---|
| 71 | { | 
|---|
| 72 | int stride = 1; | 
|---|
| 73 |  | 
|---|
| 74 | // This flag simplifies the code in the loop, encouraging | 
|---|
| 75 | // compile-time computation of strides through constant folding. | 
|---|
| 76 | _bz_bool allAscending = storage_.allRanksStoredAscending(); | 
|---|
| 77 |  | 
|---|
| 78 | // BZ_OLD_FOR_SCOPING | 
|---|
| 79 | int n; | 
|---|
| 80 | for (n=0; n < N_rank; ++n) | 
|---|
| 81 | { | 
|---|
| 82 | int strideSign = +1; | 
|---|
| 83 |  | 
|---|
| 84 | // If this rank is stored in descending order, then the stride | 
|---|
| 85 | // will be negative. | 
|---|
| 86 | if (!allAscending) | 
|---|
| 87 | { | 
|---|
| 88 | if (!isRankStoredAscending(ordering(n))) | 
|---|
| 89 | strideSign = -1; | 
|---|
| 90 | } | 
|---|
| 91 |  | 
|---|
| 92 | // The stride for this rank is the product of the lengths of | 
|---|
| 93 | // the ranks minor to it. | 
|---|
| 94 | stride_[ordering(n)] = stride * strideSign; | 
|---|
| 95 |  | 
|---|
| 96 | stride *= length_[ordering(n)]; | 
|---|
| 97 | } | 
|---|
| 98 | } | 
|---|
| 99 | else { | 
|---|
| 100 | // Specialization for N_rank == 1 | 
|---|
| 101 | // This simpler calculation makes it easier for the compiler | 
|---|
| 102 | // to propagate stride values. | 
|---|
| 103 |  | 
|---|
| 104 | if (isRankStoredAscending(0)) | 
|---|
| 105 | stride_[0] = 1; | 
|---|
| 106 | else | 
|---|
| 107 | stride_[0] = -1; | 
|---|
| 108 | } | 
|---|
| 109 |  | 
|---|
| 110 | calculateZeroOffset(); | 
|---|
| 111 | } | 
|---|
| 112 |  | 
|---|
| 113 | template<class T_numtype, int N_rank> | 
|---|
| 114 | void Array<T_numtype, N_rank>::calculateZeroOffset() | 
|---|
| 115 | { | 
|---|
| 116 | // Calculate the offset of (0,0,...,0) | 
|---|
| 117 | zeroOffset_ = 0; | 
|---|
| 118 |  | 
|---|
| 119 | // zeroOffset_ = - sum(where(ascendingFlag_, stride_ * base_, | 
|---|
| 120 | //     (length_ - 1 + base_) * stride_)) | 
|---|
| 121 | for (int n=0; n < N_rank; ++n) | 
|---|
| 122 | { | 
|---|
| 123 | if (!isRankStoredAscending(n)) | 
|---|
| 124 | zeroOffset_ -= (length_[n] - 1 + base(n)) * stride_[n]; | 
|---|
| 125 | else | 
|---|
| 126 | zeroOffset_ -= stride_[n] * base(n); | 
|---|
| 127 | } | 
|---|
| 128 | } | 
|---|
| 129 |  | 
|---|
| 130 |  | 
|---|
| 131 |  | 
|---|
| 132 | template<class P_numtype, int N_rank> | 
|---|
| 133 | void Array<P_numtype, N_rank>::dumpStructureInformation(ostream& os) const | 
|---|
| 134 | { | 
|---|
| 135 | os << "Dump of Array<" << BZ_DEBUG_TEMPLATE_AS_STRING_LITERAL(P_numtype) | 
|---|
| 136 | << ", " << N_rank << ">:" << endl | 
|---|
| 137 | << "ordering_      = " << storage_.ordering() << endl | 
|---|
| 138 | << "ascendingFlag_ = " << storage_.ascendingFlag() << endl | 
|---|
| 139 | << "base_          = " << storage_.base() << endl | 
|---|
| 140 | << "length_        = " << length_ << endl | 
|---|
| 141 | << "stride_        = " << stride_ << endl | 
|---|
| 142 | << "zeroOffset_    = " << zeroOffset_ << endl | 
|---|
| 143 | << "numElements()  = " << numElements() << endl | 
|---|
| 144 | << "storageContiguous = " << storageContiguous_ << endl; | 
|---|
| 145 | } | 
|---|
| 146 |  | 
|---|
| 147 | /* | 
|---|
| 148 | * Make this array a view of another array's data. | 
|---|
| 149 | */ | 
|---|
| 150 | template<class P_numtype, int N_rank> | 
|---|
| 151 | void Array<P_numtype, N_rank>::reference(Array<P_numtype, N_rank>& array) | 
|---|
| 152 | { | 
|---|
| 153 | storage_ = array.storage_; | 
|---|
| 154 | length_ = array.length_; | 
|---|
| 155 | stride_ = array.stride_; | 
|---|
| 156 | zeroOffset_ = array.zeroOffset_; | 
|---|
| 157 | storageContiguous_ = array.storageContiguous_; | 
|---|
| 158 |  | 
|---|
| 159 | MemoryBlockReference<P_numtype>::changeBlock(array, array.zeroOffset_); | 
|---|
| 160 |  | 
|---|
| 161 | data_ = array.data_; | 
|---|
| 162 | } | 
|---|
| 163 |  | 
|---|
| 164 | /* | 
|---|
| 165 | * This method is called to allocate memory for a new array. | 
|---|
| 166 | */ | 
|---|
| 167 | template<class P_numtype, int N_rank> | 
|---|
| 168 | _bz_inline2 void Array<P_numtype, N_rank>::setupStorage(int lastRankInitialized) | 
|---|
| 169 | { | 
|---|
| 170 | TAU_TYPE_STRING(p1, "Array<T,N>::setupStorage() [T=" | 
|---|
| 171 | + CT(P_numtype) + ",N=" + CT(N_rank) + "]"); | 
|---|
| 172 | TAU_PROFILE(" ", p1, TAU_BLITZ); | 
|---|
| 173 |  | 
|---|
| 174 | /* | 
|---|
| 175 | * If the length of some of the ranks was unspecified, fill these | 
|---|
| 176 | * in using the last specified value. | 
|---|
| 177 | * | 
|---|
| 178 | * e.g. Array<int,3> A(40) results in a 40x40x40 array. | 
|---|
| 179 | */ | 
|---|
| 180 | for (int i=lastRankInitialized + 1; i < N_rank; ++i) | 
|---|
| 181 | { | 
|---|
| 182 | storage_.setBase(i, storage_.base(lastRankInitialized)); | 
|---|
| 183 | length_[i] = length_[lastRankInitialized]; | 
|---|
| 184 | } | 
|---|
| 185 |  | 
|---|
| 186 | // Compute strides | 
|---|
| 187 | computeStrides(); | 
|---|
| 188 |  | 
|---|
| 189 | // Allocate a block of memory | 
|---|
| 190 | MemoryBlockReference<P_numtype>::newBlock(numElements()); | 
|---|
| 191 |  | 
|---|
| 192 | // Adjust the base of the array to account for non-zero base | 
|---|
| 193 | // indices and reversals | 
|---|
| 194 | data_ += zeroOffset_; | 
|---|
| 195 |  | 
|---|
| 196 | // A new array will always have contiguous storage | 
|---|
| 197 | storageContiguous_ = _bz_true; | 
|---|
| 198 | } | 
|---|
| 199 |  | 
|---|
| 200 | template<class T_numtype, int N_rank> | 
|---|
| 201 | Array<T_numtype, N_rank> Array<T_numtype, N_rank>::copy() const | 
|---|
| 202 | { | 
|---|
| 203 | if (numElements()) | 
|---|
| 204 | { | 
|---|
| 205 | Array<T_numtype, N_rank> z(length_, storage_); | 
|---|
| 206 | z = *this; | 
|---|
| 207 | return z; | 
|---|
| 208 | } | 
|---|
| 209 | else { | 
|---|
| 210 | // Null array-- don't bother allocating an empty block. | 
|---|
| 211 | return *this; | 
|---|
| 212 | } | 
|---|
| 213 | } | 
|---|
| 214 |  | 
|---|
| 215 | template<class T_numtype, int N_rank> | 
|---|
| 216 | void Array<T_numtype, N_rank>::makeUnique() | 
|---|
| 217 | { | 
|---|
| 218 | if (numReferences() > 1) | 
|---|
| 219 | { | 
|---|
| 220 | T_array tmp = copy(); | 
|---|
| 221 | reference(tmp); | 
|---|
| 222 | } | 
|---|
| 223 | } | 
|---|
| 224 |  | 
|---|
| 225 | template<class T_numtype, int N_rank> | 
|---|
| 226 | Array<T_numtype, N_rank> Array<T_numtype, N_rank>::transpose(int r0, int r1, | 
|---|
| 227 | int r2, int r3, int r4, int r5, int r6, int r7, int r8, int r9, int r10) | 
|---|
| 228 | { | 
|---|
| 229 | T_array B(*this); | 
|---|
| 230 | B.transposeSelf(r0,r1,r2,r3,r4,r5,r6,r7,r8,r9,r10); | 
|---|
| 231 | return B; | 
|---|
| 232 | } | 
|---|
| 233 |  | 
|---|
| 234 | template<class T_numtype, int N_rank> | 
|---|
| 235 | void Array<T_numtype, N_rank>::transposeSelf(int r0, int r1, int r2, int r3, | 
|---|
| 236 | int r4, int r5, int r6, int r7, int r8, int r9, int r10) | 
|---|
| 237 | { | 
|---|
| 238 | BZPRECHECK(r0+r1+r2+r3+r4+r5+r6+r7+r8+r9+r10 == N_rank * (N_rank-1) / 2, | 
|---|
| 239 | "Invalid array transpose() arguments." << endl | 
|---|
| 240 | << "Arguments must be a permutation of the numerals (0,...," | 
|---|
| 241 | << (N_rank - 1) << ")"); | 
|---|
| 242 |  | 
|---|
| 243 | // Create a temporary reference copy of this array | 
|---|
| 244 | Array<T_numtype, N_rank> x(*this); | 
|---|
| 245 |  | 
|---|
| 246 | // Now reorder the dimensions using the supplied permutation | 
|---|
| 247 | doTranspose(0, r0, x); | 
|---|
| 248 | doTranspose(1, r1, x); | 
|---|
| 249 | doTranspose(2, r2, x); | 
|---|
| 250 | doTranspose(3, r3, x); | 
|---|
| 251 | doTranspose(4, r4, x); | 
|---|
| 252 | doTranspose(5, r5, x); | 
|---|
| 253 | doTranspose(6, r6, x); | 
|---|
| 254 | doTranspose(7, r7, x); | 
|---|
| 255 | doTranspose(8, r8, x); | 
|---|
| 256 | doTranspose(9, r9, x); | 
|---|
| 257 | doTranspose(10, r10, x); | 
|---|
| 258 | } | 
|---|
| 259 |  | 
|---|
| 260 | template<class T_numtype, int N_rank> | 
|---|
| 261 | void Array<T_numtype, N_rank>::doTranspose(int destRank, int sourceRank, | 
|---|
| 262 | Array<T_numtype, N_rank>& array) | 
|---|
| 263 | { | 
|---|
| 264 | // BZ_NEEDS_WORK: precondition check | 
|---|
| 265 |  | 
|---|
| 266 | if (destRank >= N_rank) | 
|---|
| 267 | return; | 
|---|
| 268 |  | 
|---|
| 269 | length_[destRank] = array.length_[sourceRank]; | 
|---|
| 270 | stride_[destRank] = array.stride_[sourceRank]; | 
|---|
| 271 | storage_.setAscendingFlag(destRank, | 
|---|
| 272 | array.isRankStoredAscending(sourceRank)); | 
|---|
| 273 | storage_.setBase(destRank, array.base(sourceRank)); | 
|---|
| 274 |  | 
|---|
| 275 | // BZ_NEEDS_WORK: Handling the storage ordering is currently O(N^2) | 
|---|
| 276 | // but it can be done fairly easily in linear time by constructing | 
|---|
| 277 | // the appropriate permutation. | 
|---|
| 278 |  | 
|---|
| 279 | // Find sourceRank in array.storage_.ordering_ | 
|---|
| 280 | int i=0; | 
|---|
| 281 | for (; i < N_rank; ++i) | 
|---|
| 282 | if (array.storage_.ordering(i) == sourceRank) | 
|---|
| 283 | break; | 
|---|
| 284 |  | 
|---|
| 285 | storage_.setOrdering(i, destRank); | 
|---|
| 286 | } | 
|---|
| 287 |  | 
|---|
| 288 | template<class T_numtype, int N_rank> | 
|---|
| 289 | void Array<T_numtype, N_rank>::reverseSelf(int rank) | 
|---|
| 290 | { | 
|---|
| 291 | BZPRECONDITION(rank < N_rank); | 
|---|
| 292 |  | 
|---|
| 293 | storage_.setAscendingFlag(rank, !isRankStoredAscending(rank)); | 
|---|
| 294 |  | 
|---|
| 295 | int adjustment = stride_[rank] * (length_[rank] - 1); | 
|---|
| 296 | zeroOffset_ += adjustment; | 
|---|
| 297 | data_ += adjustment; | 
|---|
| 298 | stride_[rank] *= -1; | 
|---|
| 299 | } | 
|---|
| 300 |  | 
|---|
| 301 | template<class T_numtype, int N_rank> | 
|---|
| 302 | Array<T_numtype, N_rank> Array<T_numtype,N_rank>::reverse(int rank) | 
|---|
| 303 | { | 
|---|
| 304 | T_array B(*this); | 
|---|
| 305 | B.reverseSelf(rank); | 
|---|
| 306 | return B; | 
|---|
| 307 | } | 
|---|
| 308 |  | 
|---|
| 309 | template<class T_numtype, int N_rank> template<class T_numtype2> | 
|---|
| 310 | Array<T_numtype2,N_rank> Array<T_numtype,N_rank>::extractComponent(T_numtype2, | 
|---|
| 311 | int componentNumber, int numComponents) | 
|---|
| 312 | { | 
|---|
| 313 | BZPRECONDITION((componentNumber >= 0) && (componentNumber < numComponents)); | 
|---|
| 314 |  | 
|---|
| 315 | TinyVector<int,N_rank> stride2; | 
|---|
| 316 | stride2 = stride_ * numComponents; | 
|---|
| 317 | T_numtype2* dataFirst2 = ((T_numtype2*)dataFirst()) + componentNumber; | 
|---|
| 318 | return Array<T_numtype2,N_rank>(dataFirst2, length_, stride2, storage_); | 
|---|
| 319 | } | 
|---|
| 320 |  | 
|---|
| 321 | BZ_NAMESPACE_END | 
|---|
| 322 |  | 
|---|
| 323 | #endif // BZ_ARRAY_CC | 
|---|
| 324 |  | 
|---|