libcm is a C development framework with an emphasis on audio signal processing applications.
Du kannst nicht mehr als 25 Themen auswählen Themen müssen mit entweder einem Buchstaben oder einer Ziffer beginnen. Sie können Bindestriche („-“) enthalten und bis zu 35 Zeichen lang sein.

cmVectOpsTemplateCode.h 78KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763764765766767768769770771772773774775776777778779780781782783784785786787788789790791792793794795796797798799800801802803804805806807808809810811812813814815816817818819820821822823824825826827828829830831832833834835836837838839840841842843844845846847848849850851852853854855856857858859860861862863864865866867868869870871872873874875876877878879880881882883884885886887888889890891892893894895896897898899900901902903904905906907908909910911912913914915916917918919920921922923924925926927928929930931932933934935936937938939940941942943944945946947948949950951952953954955956957958959960961962963964965966967968969970971972973974975976977978979980981982983984985986987988989990991992993994995996997998999100010011002100310041005100610071008100910101011101210131014101510161017101810191020102110221023102410251026102710281029103010311032103310341035103610371038103910401041104210431044104510461047104810491050105110521053105410551056105710581059106010611062106310641065106610671068106910701071107210731074107510761077107810791080108110821083108410851086108710881089109010911092109310941095109610971098109911001101110211031104110511061107110811091110111111121113111411151116111711181119112011211122112311241125112611271128112911301131113211331134113511361137113811391140114111421143114411451146114711481149115011511152115311541155115611571158115911601161116211631164116511661167116811691170117111721173117411751176117711781179118011811182118311841185118611871188118911901191119211931194119511961197119811991200120112021203120412051206120712081209121012111212121312141215121612171218121912201221122212231224122512261227122812291230123112321233123412351236123712381239124012411242124312441245124612471248124912501251125212531254125512561257125812591260126112621263126412651266126712681269127012711272127312741275127612771278127912801281128212831284128512861287128812891290129112921293129412951296129712981299130013011302130313041305130613071308130913101311131213131314131513161317131813191320132113221323132413251326132713281329133013311332133313341335133613371338133913401341134213431344134513461347134813491350135113521353135413551356135713581359136013611362136313641365136613671368136913701371137213731374137513761377137813791380138113821383138413851386138713881389139013911392139313941395139613971398139914001401140214031404140514061407140814091410141114121413141414151416141714181419142014211422142314241425142614271428142914301431143214331434143514361437143814391440144114421443144414451446144714481449145014511452145314541455145614571458145914601461146214631464146514661467146814691470147114721473147414751476147714781479148014811482148314841485148614871488148914901491149214931494149514961497149814991500150115021503150415051506150715081509151015111512151315141515151615171518151915201521152215231524152515261527152815291530153115321533153415351536153715381539154015411542154315441545154615471548154915501551155215531554155515561557155815591560156115621563156415651566156715681569157015711572157315741575157615771578157915801581158215831584158515861587158815891590159115921593159415951596159715981599160016011602160316041605160616071608160916101611161216131614161516161617161816191620162116221623162416251626162716281629163016311632163316341635163616371638163916401641164216431644164516461647164816491650165116521653165416551656165716581659166016611662166316641665166616671668166916701671167216731674167516761677167816791680168116821683168416851686168716881689169016911692169316941695169616971698169917001701170217031704170517061707170817091710171117121713171417151716171717181719172017211722172317241725172617271728172917301731173217331734173517361737173817391740174117421743174417451746174717481749175017511752175317541755175617571758175917601761176217631764176517661767176817691770177117721773177417751776177717781779178017811782178317841785178617871788178917901791179217931794179517961797179817991800180118021803180418051806180718081809181018111812181318141815181618171818181918201821182218231824182518261827182818291830183118321833183418351836183718381839184018411842184318441845184618471848184918501851185218531854185518561857185818591860186118621863186418651866186718681869187018711872187318741875187618771878187918801881188218831884188518861887188818891890189118921893189418951896189718981899190019011902190319041905190619071908190919101911191219131914191519161917191819191920192119221923192419251926192719281929193019311932193319341935193619371938193919401941194219431944194519461947194819491950195119521953195419551956195719581959196019611962196319641965196619671968196919701971197219731974197519761977197819791980198119821983198419851986198719881989199019911992199319941995199619971998199920002001200220032004200520062007200820092010201120122013201420152016201720182019202020212022202320242025202620272028202920302031203220332034203520362037203820392040204120422043204420452046204720482049205020512052205320542055205620572058205920602061206220632064206520662067206820692070207120722073207420752076207720782079208020812082208320842085208620872088208920902091209220932094209520962097209820992100210121022103210421052106210721082109211021112112211321142115211621172118211921202121212221232124212521262127212821292130213121322133213421352136213721382139214021412142214321442145214621472148214921502151215221532154215521562157215821592160216121622163216421652166216721682169217021712172217321742175217621772178217921802181218221832184218521862187218821892190219121922193219421952196219721982199220022012202220322042205220622072208220922102211221222132214221522162217221822192220222122222223222422252226222722282229223022312232223322342235223622372238223922402241224222432244224522462247224822492250225122522253225422552256225722582259226022612262226322642265226622672268226922702271227222732274227522762277227822792280228122822283228422852286228722882289229022912292229322942295229622972298229923002301230223032304230523062307230823092310231123122313231423152316231723182319232023212322232323242325232623272328232923302331233223332334233523362337233823392340234123422343234423452346234723482349235023512352235323542355235623572358235923602361236223632364236523662367236823692370237123722373237423752376237723782379238023812382238323842385238623872388238923902391239223932394239523962397239823992400240124022403240424052406240724082409241024112412241324142415241624172418241924202421242224232424242524262427242824292430243124322433243424352436243724382439244024412442244324442445244624472448244924502451245224532454245524562457245824592460246124622463246424652466246724682469247024712472247324742475247624772478247924802481248224832484248524862487248824892490249124922493249424952496249724982499250025012502250325042505250625072508250925102511251225132514251525162517251825192520252125222523252425252526252725282529253025312532253325342535253625372538253925402541254225432544254525462547254825492550255125522553255425552556255725582559256025612562256325642565256625672568256925702571257225732574257525762577257825792580258125822583258425852586258725882589259025912592259325942595259625972598259926002601260226032604260526062607260826092610261126122613261426152616261726182619262026212622262326242625262626272628262926302631263226332634263526362637263826392640264126422643264426452646264726482649265026512652265326542655265626572658265926602661266226632664266526662667266826692670267126722673267426752676267726782679268026812682268326842685268626872688268926902691269226932694269526962697269826992700270127022703270427052706270727082709271027112712271327142715271627172718271927202721272227232724272527262727272827292730273127322733273427352736273727382739274027412742274327442745274627472748274927502751275227532754275527562757275827592760276127622763276427652766276727682769277027712772277327742775277627772778277927802781278227832784278527862787278827892790279127922793279427952796279727982799280028012802280328042805280628072808280928102811281228132814281528162817281828192820282128222823282428252826282728282829283028312832283328342835283628372838283928402841284228432844284528462847284828492850285128522853285428552856285728582859286028612862286328642865286628672868286928702871287228732874287528762877287828792880288128822883288428852886288728882889289028912892289328942895289628972898289929002901290229032904290529062907290829092910291129122913291429152916291729182919292029212922292329242925292629272928292929302931293229332934293529362937293829392940294129422943294429452946294729482949295029512952295329542955295629572958295929602961296229632964296529662967296829692970297129722973297429752976297729782979298029812982298329842985298629872988298929902991299229932994299529962997299829993000300130023003300430053006300730083009301030113012301330143015301630173018301930203021302230233024302530263027302830293030303130323033303430353036303730383039304030413042304330443045304630473048304930503051305230533054305530563057305830593060306130623063306430653066306730683069307030713072307330743075307630773078307930803081308230833084308530863087308830893090309130923093309430953096309730983099310031013102310331043105310631073108310931103111311231133114311531163117311831193120312131223123312431253126312731283129313031313132313331343135313631373138313931403141314231433144314531463147314831493150315131523153315431553156315731583159316031613162316331643165316631673168316931703171317231733174317531763177317831793180318131823183318431853186318731883189319031913192319331943195319631973198319932003201320232033204
  1. #ifdef cmVectOpsTemplateCode_h
  2. void VECT_OP_FUNC(VPrint)( cmRpt_t* rpt, const char* fmt, ... )
  3. {
  4. va_list vl;
  5. va_start(vl,fmt);
  6. if( rpt != NULL )
  7. cmRptVPrintf(rpt,fmt,vl);
  8. else
  9. vprintf(fmt,vl);
  10. va_end(vl);
  11. }
  12. void VECT_OP_FUNC(Printf)( cmRpt_t* rpt, unsigned rowCnt, unsigned colCnt, const VECT_OP_TYPE* sbp, unsigned fieldWidth, unsigned decPlCnt, const char* fmt, unsigned flags )
  13. {
  14. unsigned cci;
  15. unsigned outColCnt = 10;
  16. if( fieldWidth < 0 )
  17. fieldWidth = 10;
  18. if( decPlCnt < 0 )
  19. decPlCnt = 4;
  20. if( outColCnt == -1 )
  21. outColCnt = colCnt;
  22. for(cci=0; cci<colCnt; cci+=outColCnt)
  23. {
  24. unsigned ci0 = cci;
  25. unsigned cn = cci + outColCnt;
  26. unsigned ri;
  27. if(cn > colCnt)
  28. cn = colCnt;
  29. if( colCnt > outColCnt )
  30. {
  31. if( cmIsFlag(flags,cmPrintMatlabLabelsFl) )
  32. VECT_OP_FUNC(VPrint)(rpt,"Columns:%i to %i\n",ci0,cn-1);
  33. else
  34. if( cmIsFlag(flags,cmPrintShortLabelsFl) )
  35. VECT_OP_FUNC(VPrint)(rpt,"%3i: ",ci0);
  36. }
  37. if( rowCnt > 1 )
  38. VECT_OP_FUNC(VPrint)(rpt,"\n");
  39. for(ri=0; ri<rowCnt; ++ri)
  40. {
  41. unsigned ci;
  42. for(ci=ci0; ci<cn; ++ci )
  43. VECT_OP_FUNC(VPrint)(rpt,fmt,fieldWidth,decPlCnt,sbp[ (ci*rowCnt) + ri ]);
  44. if( cn > 0 )
  45. VECT_OP_FUNC(VPrint)(rpt,"\n");
  46. }
  47. }
  48. }
  49. void VECT_OP_FUNC(Print)( cmRpt_t* rpt, unsigned rn, unsigned cn, const VECT_OP_TYPE* sbp )
  50. { VECT_OP_FUNC(Printf)(rpt,rn,cn,sbp,cmDefaultFieldWidth,cmDefaultDecPlCnt,"%*.*f ",cmPrintShortLabelsFl); }
  51. void VECT_OP_FUNC(PrintE)( cmRpt_t* rpt, unsigned rn, unsigned cn, const VECT_OP_TYPE* sbp )
  52. { VECT_OP_FUNC(Printf)(rpt,rn,cn,sbp,cmDefaultFieldWidth,cmDefaultDecPlCnt,"%*.*e ",cmPrintShortLabelsFl); }
  53. void VECT_OP_FUNC(PrintLf)( const char* label, cmRpt_t* rpt, unsigned rn, unsigned cn, const VECT_OP_TYPE* dbp, unsigned fieldWidth, unsigned decPlCnt, const char* fmt )
  54. {
  55. VECT_OP_FUNC(VPrint)( rpt, "%s\n", label );
  56. VECT_OP_FUNC(Printf)( rpt, rn, cn, dbp, fieldWidth, decPlCnt,fmt,cmPrintShortLabelsFl );
  57. }
  58. void VECT_OP_FUNC(PrintL)( const char* label, cmRpt_t* rpt, unsigned rn, unsigned cn, const VECT_OP_TYPE* dbp )
  59. {
  60. VECT_OP_FUNC(VPrint)( rpt, "%s\n", label );
  61. VECT_OP_FUNC(Printf)( rpt, rn, cn, dbp, cmDefaultFieldWidth,cmDefaultDecPlCnt,"%*.*f ",cmPrintShortLabelsFl );
  62. }
  63. void VECT_OP_FUNC(PrintLE)( const char* label, cmRpt_t* rpt, unsigned rn, unsigned cn, const VECT_OP_TYPE* dbp )
  64. {
  65. VECT_OP_FUNC(VPrint)( rpt, "%s\n", label );
  66. VECT_OP_FUNC(Printf)( rpt, rn, cn, dbp, cmDefaultFieldWidth,cmDefaultDecPlCnt,"%*.*e ",cmPrintShortLabelsFl );
  67. }
  68. VECT_OP_TYPE* VECT_OP_FUNC(NormalizeProbabilityVV)(VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp)
  69. {
  70. VECT_OP_TYPE sum = VECT_OP_FUNC(Sum)(sbp,dn);
  71. if( sum == 0 )
  72. sum = 1;
  73. return VECT_OP_FUNC(DivVVS)(dbp,dn,sbp,sum);
  74. }
  75. VECT_OP_TYPE* VECT_OP_FUNC(NormalizeProbability)(VECT_OP_TYPE* dbp, unsigned dn)
  76. { return VECT_OP_FUNC(NormalizeProbabilityVV)(dbp,dn,dbp); }
  77. VECT_OP_TYPE* VECT_OP_FUNC(NormalizeProbabilityN)(VECT_OP_TYPE* dbp, unsigned dn, unsigned stride)
  78. {
  79. VECT_OP_TYPE sum = VECT_OP_FUNC(SumN)(dbp,dn,stride);
  80. if( sum == 0 )
  81. return dbp;
  82. VECT_OP_TYPE* dp = dbp;
  83. VECT_OP_TYPE* ep = dp + (dn*stride);
  84. for(; dp < ep; dp+=stride )
  85. *dp /= sum;
  86. return dbp;
  87. }
  88. VECT_OP_TYPE* VECT_OP_FUNC(StandardizeRows)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, VECT_OP_TYPE* uV, VECT_OP_TYPE* sdV )
  89. {
  90. bool uFl = false;
  91. bool sFl = false;
  92. unsigned i;
  93. if( uV == NULL )
  94. {
  95. uV = cmMemAllocZ(VECT_OP_TYPE,drn);
  96. uFl = true;
  97. }
  98. if( sdV == NULL )
  99. {
  100. sdV = cmMemAllocZ(VECT_OP_TYPE,drn);
  101. sFl = true;
  102. }
  103. VECT_OP_FUNC(MeanM)(uV, dbp, drn, dcn, 1 );
  104. VECT_OP_FUNC(VarianceM)(sdV, dbp, drn, dcn, uV, 1 );
  105. for(i=0; i<dcn; ++i)
  106. {
  107. VECT_OP_FUNC(SubVV)(dbp + i * drn, drn, uV );
  108. VECT_OP_FUNC(DivVV)(dbp + i * drn, drn, sdV );
  109. }
  110. if(uFl)
  111. cmMemFree(uV);
  112. if(sFl)
  113. cmMemFree(sdV);
  114. return dbp;
  115. }
  116. VECT_OP_TYPE* VECT_OP_FUNC(StandardizeCols)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, VECT_OP_TYPE* uV, VECT_OP_TYPE* sdV )
  117. {
  118. bool uFl = false;
  119. bool sFl = false;
  120. unsigned i;
  121. if( uV == NULL )
  122. {
  123. uV = cmMemAllocZ(VECT_OP_TYPE,dcn);
  124. uFl = true;
  125. }
  126. if( sdV == NULL )
  127. {
  128. sdV = cmMemAllocZ(VECT_OP_TYPE,dcn);
  129. sFl = true;
  130. }
  131. VECT_OP_FUNC(MeanM)(uV, dbp, drn, dcn, 0 );
  132. VECT_OP_FUNC(VarianceM)(sdV, dbp, drn, dcn, uV, 0 );
  133. for(i=0; i<drn; ++i)
  134. {
  135. VECT_OP_FUNC(SubVVNN)(dbp + i, dcn, drn, uV, 1 );
  136. VECT_OP_FUNC(DivVVNN)(dbp + i, dcn, drn, sdV, 1 );
  137. }
  138. if(uFl)
  139. cmMemFree(uV);
  140. if(sFl)
  141. cmMemFree(sdV);
  142. return dbp;
  143. }
  144. VECT_OP_TYPE* VECT_OP_FUNC(HalfWaveRectify)(VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp )
  145. {
  146. VECT_OP_TYPE* dp = dbp;
  147. VECT_OP_TYPE* ep = dbp + dn;
  148. for(; dp < ep; ++dp,++sp )
  149. *dp = *sp < 0 ? 0 : *sp;
  150. return dbp;
  151. }
  152. VECT_OP_TYPE* VECT_OP_FUNC(CumSum)(VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp)
  153. {
  154. VECT_OP_TYPE* dep = dbp + dn;
  155. VECT_OP_TYPE* rp = dbp;
  156. VECT_OP_TYPE sum = 0;
  157. while( dbp < dep )
  158. {
  159. sum += *sbp++;
  160. *dbp++ = sum;
  161. }
  162. return rp;
  163. }
  164. VECT_OP_TYPE VECT_OP_FUNC(Mean)( const VECT_OP_TYPE* bp, unsigned n )
  165. { return VECT_OP_FUNC(Sum)(bp,n)/n; }
  166. VECT_OP_TYPE VECT_OP_FUNC(MeanN)( const VECT_OP_TYPE* bp, unsigned n, unsigned stride )
  167. { return VECT_OP_FUNC(SumN)(bp,n,stride)/n; }
  168. VECT_OP_TYPE* VECT_OP_FUNC(MeanM)( VECT_OP_TYPE* dp, const VECT_OP_TYPE* sp, unsigned srn, unsigned scn, unsigned dim )
  169. {
  170. unsigned i;
  171. unsigned cn = dim == 0 ? scn : srn;
  172. unsigned rn = dim == 0 ? srn : scn;
  173. unsigned inc = dim == 0 ? srn : 1;
  174. unsigned stride = dim == 0 ? 1 : srn;
  175. unsigned d0 = 0;
  176. for(i=0; i<cn; ++i, d0+=inc)
  177. dp[i] = VECT_OP_FUNC(MeanN)(sp + d0, rn, stride );
  178. return dp;
  179. }
  180. VECT_OP_TYPE* VECT_OP_FUNC(Mean2)( VECT_OP_TYPE* dp, const VECT_OP_TYPE* (*srcFuncPtr)(void* arg, unsigned idx ), unsigned D, unsigned N, void* argPtr )
  181. {
  182. unsigned i,n;
  183. const VECT_OP_TYPE* sp;
  184. VECT_OP_FUNC(Zero)(dp,D);
  185. if( N > 1 )
  186. {
  187. n = 0;
  188. for(i=0; i<N; ++i)
  189. if((sp = srcFuncPtr(argPtr,i)) != NULL )
  190. {
  191. VECT_OP_FUNC(AddVV)(dp,D,sp);
  192. ++n;
  193. }
  194. VECT_OP_FUNC(DivVS)(dp,D,n);
  195. }
  196. return dp;
  197. }
  198. VECT_OP_TYPE VECT_OP_FUNC(Variance)( const VECT_OP_TYPE* sp, unsigned sn, const VECT_OP_TYPE* avgPtr )
  199. { return VECT_OP_FUNC(VarianceN)(sp,sn,1,avgPtr); }
  200. VECT_OP_TYPE VECT_OP_FUNC(VarianceN)( const VECT_OP_TYPE* sp, unsigned sn, unsigned stride, const VECT_OP_TYPE* meanPtr )
  201. {
  202. VECT_OP_TYPE mean = 0;
  203. if( sn <= 1 )
  204. return 0;
  205. if( meanPtr == NULL )
  206. mean = VECT_OP_FUNC(MeanN)( sp, sn, stride );
  207. else
  208. mean = *meanPtr;
  209. const VECT_OP_TYPE* ep = sp + (sn*stride);
  210. VECT_OP_TYPE sum = 0;
  211. for(; sp < ep; sp += stride )
  212. sum += (*sp-mean) * (*sp-mean);
  213. return sum / (sn-1);
  214. }
  215. VECT_OP_TYPE* VECT_OP_FUNC(VarianceM)(VECT_OP_TYPE* dp, const VECT_OP_TYPE* sp, unsigned srn, unsigned scn, const VECT_OP_TYPE* avgPtr, unsigned dim )
  216. {
  217. unsigned i;
  218. unsigned cn = dim == 0 ? scn : srn;
  219. unsigned rn = dim == 0 ? srn : scn;
  220. unsigned inc = dim == 0 ? srn : 1;
  221. unsigned stride = dim == 0 ? 1 : srn;
  222. unsigned d0 = 0;
  223. for(i=0; i<cn; ++i, d0+=inc)
  224. dp[i] = VECT_OP_FUNC(VarianceN)(sp + d0, rn, stride, avgPtr==NULL ? NULL : avgPtr+i );
  225. return dp;
  226. }
  227. unsigned VECT_OP_FUNC(NormToMax)( VECT_OP_TYPE* dp, unsigned dn )
  228. {
  229. unsigned i = VECT_OP_FUNC(MaxIndex)(dp,dn,1);
  230. if( i != cmInvalidIdx )
  231. {
  232. VECT_OP_TYPE v = dp[i];
  233. VECT_OP_FUNC(DivVS)(dp,dn,v);
  234. }
  235. return i;
  236. }
  237. VECT_OP_TYPE VECT_OP_FUNC(AlphaNorm)(const VECT_OP_TYPE* sp, unsigned sn, VECT_OP_TYPE alpha )
  238. {
  239. double sum = 0;
  240. const VECT_OP_TYPE* bp = sp;
  241. const VECT_OP_TYPE* ep = sp + sn;
  242. while( bp < ep )
  243. sum += pow(fabs(*bp++),alpha);
  244. return (VECT_OP_TYPE)pow(sum/sn,1.0/alpha);
  245. }
  246. void VECT_OP_FUNC(GaussCovariance)(VECT_OP_TYPE* yM, unsigned D, const VECT_OP_TYPE* xM, unsigned xN, const VECT_OP_TYPE* uV, const unsigned* selIdxV, unsigned selKey )
  247. {
  248. unsigned i,j,k,n = 0;
  249. VECT_OP_TYPE tV[ D ];
  250. VECT_OP_FUNC(Fill)(yM,D*D,0);
  251. // if the mean was not given - then calculate it
  252. if( uV == NULL )
  253. {
  254. VECT_OP_FUNC(Fill)(tV,D,0);
  255. // sum each row of xM[] into uM[]
  256. for(i=0; i<D; ++i)
  257. {
  258. n = 0;
  259. for(j=0; j<xN; ++j)
  260. if( selIdxV==NULL || selIdxV[j]==selKey )
  261. {
  262. tV[i] += xM[ (j*D) + i ];
  263. ++n;
  264. }
  265. }
  266. // form an average from the sum in tV[]
  267. VECT_OP_FUNC(DivVS)(tV,D,n);
  268. uV = tV;
  269. }
  270. for(i=0; i<D; ++i)
  271. for(j=i; j<D; ++j)
  272. {
  273. n = 0;
  274. for(k=0; k<xN; ++k)
  275. if( selIdxV==NULL || selIdxV[k]==selKey)
  276. {
  277. unsigned yi = (i*D)+j;
  278. yM[ yi ] += ((xM[ (k*D)+j ]-uV[j]) * (xM[ (k*D) + i ]-uV[i]));
  279. if( i != j )
  280. yM[ (j*D)+i ] = yM[ yi ];
  281. ++n;
  282. }
  283. }
  284. if( n>1 )
  285. VECT_OP_FUNC(DivVS)( yM, D*D, n-1 );
  286. }
  287. void VECT_OP_FUNC(GaussCovariance2)(VECT_OP_TYPE* yM, unsigned D, const VECT_OP_TYPE* (*srcFunc)(void* userPtr, unsigned idx), unsigned xN, void* userPtr, const VECT_OP_TYPE* uV, const unsigned* selIdxV, unsigned selKey )
  288. {
  289. unsigned i,j,k = 0,n;
  290. VECT_OP_TYPE tV[ D ];
  291. const VECT_OP_TYPE* sp;
  292. VECT_OP_FUNC(Fill)(yM,D*D,0);
  293. // if the mean was not given - then calculate it
  294. if( uV == NULL )
  295. {
  296. VECT_OP_FUNC(Fill)(tV,D,0);
  297. n = 0;
  298. // sum each row of xM[] into uM[]
  299. for(i=0; i<xN; ++i)
  300. if( (selIdxV==NULL || selIdxV[i]==selKey) && ((sp=srcFunc(userPtr,i))!=NULL) )
  301. {
  302. VECT_OP_FUNC(AddVV)(tV,D,sp);
  303. ++n;
  304. }
  305. // form an average from the sum in tV[]
  306. VECT_OP_FUNC(DivVS)(tV,D,n);
  307. uV = tV;
  308. }
  309. for(i=0; i<xN; ++i)
  310. if( selIdxV==NULL || selIdxV[i]==selKey )
  311. {
  312. // get a pointer to the ith data point
  313. const VECT_OP_TYPE* sV = srcFunc(userPtr,i);
  314. // note: this algorithm works because when a data point element (scalar)
  315. // is multiplied by another data point element those two elements
  316. // are always part of the same data point (vector). Two elements
  317. // from different data points are never multiplied.
  318. if( sV != NULL )
  319. for(j=0; j<D; ++j)
  320. for(k=j; k<D; ++k)
  321. yM[j + k*D] += (sV[j]-uV[j]) * (sV[k]-uV[k]);
  322. }
  323. if( n > 1 )
  324. VECT_OP_FUNC(DivVS)( yM, D*D, n-1 );
  325. // fill in the lower triangle
  326. for(j=0; j<D; ++j)
  327. for(k=j; k<D; ++k)
  328. yM[k + j*D] = yM[j + k*D];
  329. }
  330. bool VECT_OP_FUNC(Equal)( const VECT_OP_TYPE* s0p, const VECT_OP_TYPE* s1p, unsigned sn )
  331. {
  332. const VECT_OP_TYPE* ep = s0p + sn;
  333. while( s0p < ep )
  334. if( *s0p++ != *s1p++ )
  335. return false;
  336. return true;
  337. }
  338. bool VECT_OP_FUNC(IsNormal)( const VECT_OP_TYPE* sp, unsigned sn )
  339. {
  340. const VECT_OP_TYPE* ep = sp + sn;
  341. for(; sp<ep; ++sp)
  342. if( !isnormal(*sp) )
  343. return false;
  344. return true;
  345. }
  346. bool VECT_OP_FUNC(IsNormalZ)(const VECT_OP_TYPE* sp, unsigned sn )
  347. {
  348. const VECT_OP_TYPE* ep = sp + sn;
  349. for(; sp<ep; ++sp)
  350. if( (*sp != 0) && (!isnormal(*sp)) )
  351. return false;
  352. return true;
  353. }
  354. unsigned VECT_OP_FUNC(FindNonNormal)( unsigned* dp, unsigned dn, const VECT_OP_TYPE* sbp )
  355. {
  356. const VECT_OP_TYPE* sp = sbp;
  357. const VECT_OP_TYPE* ep = sp + dn;
  358. unsigned n = 0;
  359. for(; sp<ep; ++sp)
  360. if( !isnormal(*sp) )
  361. dp[n++] = sp - sbp;
  362. return n;
  363. }
  364. unsigned VECT_OP_FUNC(FindNonNormalZ)( unsigned* dp, unsigned dn, const VECT_OP_TYPE* sbp )
  365. {
  366. const VECT_OP_TYPE* sp = sbp;
  367. const VECT_OP_TYPE* ep = sp + dn;
  368. unsigned n = 0;
  369. for(; sp<ep; ++sp)
  370. if( (*sp!=0) && (!isnormal(*sp)) )
  371. dp[n++] = sp - sbp;
  372. return n;
  373. }
  374. unsigned VECT_OP_FUNC(ZeroCrossCount)( const VECT_OP_TYPE* bp, unsigned bn, VECT_OP_TYPE* delaySmpPtr)
  375. {
  376. unsigned n = delaySmpPtr != NULL ? ((*delaySmpPtr >= 0) != (*bp >= 0)) : 0 ;
  377. const VECT_OP_TYPE* ep = bp + bn;
  378. for(; bp<ep-1; ++bp)
  379. if( (*bp >= 0) != (*(bp+1) >= 0) )
  380. ++n;
  381. if( delaySmpPtr != NULL )
  382. *delaySmpPtr = *bp;
  383. return n;
  384. }
  385. VECT_OP_TYPE VECT_OP_FUNC(RMS)( const VECT_OP_TYPE* bp, unsigned bn, unsigned wndSmpCnt )
  386. {
  387. const VECT_OP_TYPE* ep = bp + bn;
  388. if( bn==0 )
  389. return 0;
  390. assert( bn <= wndSmpCnt );
  391. double sum = 0;
  392. for(; bp < ep; ++bp )
  393. sum += *bp * *bp;
  394. return (VECT_OP_TYPE)sqrt(sum/wndSmpCnt);
  395. }
  396. VECT_OP_TYPE* VECT_OP_FUNC(RmsV)( VECT_OP_TYPE* dp, unsigned dn, const VECT_OP_TYPE* sp, unsigned sn, unsigned wndSmpCnt, unsigned hopSmpCnt )
  397. {
  398. const VECT_OP_TYPE* dep = dp + dn;
  399. const VECT_OP_TYPE* sep = sp + sn;
  400. VECT_OP_TYPE* rp = dp;
  401. for(; dp<dep && sp<sep; sp+=hopSmpCnt)
  402. *dp++ = VECT_OP_FUNC(RMS)( sp, cmMin(wndSmpCnt,sep-sp), wndSmpCnt );
  403. VECT_OP_FUNC(Zero)(dp,dep-dp);
  404. return rp;
  405. }
  406. VECT_OP_TYPE VECT_OP_FUNC(EuclidNorm)( const VECT_OP_TYPE* sp, unsigned sn )
  407. { return (VECT_OP_TYPE)sqrt( VECT_OP_FUNC(MultSumVV)(sp,sp,sn)); }
  408. /*
  409. From:http://www.ee.ic.ac.uk/hp/staff/dmb/voicebox/doc/voicebox/distitpf.html
  410. [nf1,p2]=size(pf1);
  411. p1=p2-1;
  412. nf2=size(pf2,1);
  413. nx= min(nf1,nf2);
  414. r = pf1(1:nx,:)./pf2(1:nx,:);
  415. q = r-log(r);
  416. s = sum( q(:,2:p1),2) + 0.5 * (q(:,1)+q(:,p2))
  417. d= s/p1-1;
  418. */
  419. VECT_OP_TYPE VECT_OP_FUNC(ItakuraDistance)( const VECT_OP_TYPE* s0p, const VECT_OP_TYPE* s1p, unsigned sn )
  420. {
  421. VECT_OP_TYPE d = 0;
  422. VECT_OP_TYPE r[ sn ];
  423. VECT_OP_TYPE q[ sn ];
  424. // r = pf1(1:nx,:)./pf2(1:nx,:);
  425. VECT_OP_FUNC(DivVVV)(r,sn,s0p,s1p);
  426. //q=log(r);
  427. VECT_OP_FUNC(LogV)(q,sn,r);
  428. //r = r - q = r - log(r)
  429. VECT_OP_FUNC(SubVV)(r,sn,q);
  430. //r = r - sn = r - log(r) - 1
  431. VECT_OP_FUNC(SubVS)(r,sn,sn);
  432. // d = sum(r);
  433. d = VECT_OP_FUNC(Sum)(r,sn);
  434. return (VECT_OP_TYPE)(d / sn);
  435. //d = log( VECT_OP_FUNC(Sum)(r,sn) /sn );
  436. //d -= VECT_OP_FUNC(Sum)(q,sn)/sn;
  437. return d;
  438. }
  439. VECT_OP_TYPE VECT_OP_FUNC(CosineDistance)( const VECT_OP_TYPE* s0p, const VECT_OP_TYPE* s1p, unsigned sn )
  440. {
  441. VECT_OP_TYPE d0 = VECT_OP_FUNC(EuclidNorm)(s0p,sn);
  442. VECT_OP_TYPE d1 = VECT_OP_FUNC(EuclidNorm)(s1p,sn);
  443. if( d0 == 0 )
  444. d0 = cmReal_MIN;
  445. if( d1 == 0 )
  446. d1 = cmReal_MIN;
  447. return (VECT_OP_TYPE)(VECT_OP_FUNC(MultSumVV)(s0p,s1p,sn) / (d0 * d1));
  448. }
  449. VECT_OP_TYPE VECT_OP_FUNC(EuclidDistance)( const VECT_OP_TYPE* s0p, const VECT_OP_TYPE* s1p, unsigned sn )
  450. {
  451. double d = 0;
  452. const VECT_OP_TYPE* sep = s0p + sn;
  453. for(; s0p<sep; ++s0p,++s1p)
  454. d += (*s0p - *s1p) * (*s0p - *s1p);
  455. return (VECT_OP_TYPE)(sqrt(d));
  456. }
  457. VECT_OP_TYPE VECT_OP_FUNC(L1Distance)( const VECT_OP_TYPE* s0p, const VECT_OP_TYPE* s1p, unsigned sn )
  458. {
  459. double d = 0;
  460. const VECT_OP_TYPE* sep = s0p + sn;
  461. for(; s0p<sep; ++s0p,++s1p)
  462. d += (VECT_OP_TYPE)fabs(*s0p - *s1p);
  463. return d;
  464. }
  465. VECT_OP_TYPE VECT_OP_FUNC(MahalanobisDistance)( const VECT_OP_TYPE* x, unsigned D, const VECT_OP_TYPE* u, const VECT_OP_TYPE* invCovM )
  466. {
  467. VECT_OP_TYPE t[ D ];
  468. VECT_OP_TYPE d[ D ];
  469. // t[] = x[] - u[];
  470. VECT_OP_FUNC(SubVVV)(t,D,x,u);
  471. // d[1,D] = t[1,D] * covM[D,D]
  472. VECT_OP_FUNC(MultVVM)( d, D, t, D, invCovM );
  473. // d = sum(d[].*t[])
  474. VECT_OP_TYPE dist = VECT_OP_FUNC(MultSumVV)(d,t,D);
  475. return (VECT_OP_TYPE)sqrt(dist);
  476. }
  477. VECT_OP_TYPE VECT_OP_FUNC(KL_Distance)( const VECT_OP_TYPE* up, const VECT_OP_TYPE* sp, unsigned sn )
  478. {
  479. VECT_OP_TYPE v[ sn ];
  480. VECT_OP_FUNC(DivVVV)(v,sn,up,sp); // v = up ./ sp
  481. VECT_OP_FUNC(LogV)(v,sn,v); // v = log(v)
  482. VECT_OP_FUNC(MultVV)(v,sn,up); // v *= up;
  483. return VECT_OP_FUNC(Sum)(v,sn); // sum(v)
  484. }
  485. VECT_OP_TYPE VECT_OP_FUNC(KL_Distance2)( const VECT_OP_TYPE* up, const VECT_OP_TYPE* sp, unsigned sn )
  486. {
  487. VECT_OP_TYPE v0[ sn ];
  488. VECT_OP_TYPE v1[ sn ];
  489. VECT_OP_FUNC(NormalizeProbabilityVV)(v0,sn,up);
  490. VECT_OP_FUNC(NormalizeProbabilityVV)(v1,sn,sp);
  491. return VECT_OP_FUNC(KL_Distance)(v0,v1,sn);
  492. }
  493. /// If dv[scn] is non NULL then return the Euclidean distance from sv[scn] to each column of sm[srn,scn].
  494. /// The function returns the index of the closest data point (column) in sm[].
  495. unsigned VECT_OP_FUNC(EuclidDistanceVM)( VECT_OP_TYPE* dv, const VECT_OP_TYPE* sv, const VECT_OP_TYPE* sm, unsigned srn, unsigned scn )
  496. {
  497. unsigned minIdx = cmInvalidIdx;
  498. VECT_OP_TYPE minDist = 0;
  499. unsigned i = 0;
  500. for(; i<scn; ++i )
  501. {
  502. VECT_OP_TYPE dist = VECT_OP_FUNC(EuclidDistance)(sv, sm + (i*srn), srn );
  503. if( dv != NULL )
  504. *dv++ = dist;
  505. if( dist < minDist || minIdx == cmInvalidIdx )
  506. {
  507. minIdx = i;
  508. minDist = dist;
  509. }
  510. }
  511. return minIdx;
  512. }
  513. void VECT_OP_FUNC(DistVMM)( VECT_OP_TYPE* dM, VECT_OP_TYPE* mvV, unsigned* miV, unsigned rn, const VECT_OP_TYPE* s0M, unsigned s0cn, const VECT_OP_TYPE* s1M, unsigned s1cn, VECT_OP_TYPE (*distFunc)( void* userPtr, const VECT_OP_TYPE* s0V, const VECT_OP_TYPE* s1V, unsigned sn ), void* userPtr )
  514. {
  515. unsigned i,j,k;
  516. // for each col in s0M[];
  517. for(i=0,k=0; i<s0cn; ++i)
  518. {
  519. VECT_OP_TYPE min_val = VECT_OP_MAX;
  520. unsigned min_idx = cmInvalidIdx;
  521. // for each col in s1M[]
  522. for(j=0; j<s1cn; ++j,++k)
  523. {
  524. // v = distance(s0M[:,i],s1M[:,j]
  525. VECT_OP_TYPE v = distFunc( userPtr, s1M + (j*rn), s0M + (i*rn), rn );
  526. if( dM != NULL )
  527. dM[k] = v; // store distance
  528. // track closest col in s1M[]
  529. if( v < min_val || min_idx==cmInvalidIdx )
  530. {
  531. min_val = v;
  532. min_idx = j;
  533. }
  534. }
  535. if( mvV != NULL )
  536. mvV[i] = min_val;
  537. if( miV != NULL )
  538. miV[i] = min_idx;
  539. }
  540. }
  541. void VECT_OP_FUNC(SelectRandom) ( VECT_OP_TYPE *dM, unsigned* selIdxV, unsigned selIdxN, const VECT_OP_TYPE* sM, unsigned srn, unsigned scn )
  542. {
  543. bool freeFl = false;
  544. unsigned i;
  545. assert( selIdxN != 0 );
  546. // if no selIdxV[] was given then create one
  547. if( selIdxV == NULL )
  548. {
  549. selIdxV = cmMemAlloc( unsigned, selIdxN );
  550. freeFl = true;
  551. }
  552. // select datapoints at random
  553. cmVOU_UniqueRandom(selIdxV,selIdxN,scn);
  554. // copy the data points into the output matrix
  555. if( dM != NULL )
  556. for(i=0; i<selIdxN; ++i)
  557. {
  558. assert( selIdxV[i] < scn );
  559. VECT_OP_FUNC(Copy)( dM + (i*srn), srn, sM + selIdxV[i]*srn );
  560. }
  561. if( freeFl )
  562. cmMemPtrFree(&selIdxV);
  563. }
  564. void VECT_OP_FUNC(_SelectDist)( VECT_OP_TYPE *dM, unsigned* selIdxV, unsigned selIdxN, const VECT_OP_TYPE* sM, unsigned srn, unsigned scn, VECT_OP_TYPE (*distFunc)( void* userPtr, const VECT_OP_TYPE* s0V, const VECT_OP_TYPE* s1V, unsigned sn ), void* userPtr, bool avgFl )
  565. {
  566. unsigned i;
  567. unsigned dcn = 0;
  568. bool freeFl = false;
  569. assert( selIdxN > 0 );
  570. if( dM == NULL )
  571. {
  572. dM = cmMemAllocZ( VECT_OP_TYPE, srn*selIdxN );
  573. freeFl = true;
  574. }
  575. // allocate distM[scn,selIdxN] to hold the distances from each selected column to all columns in sM[]
  576. VECT_OP_TYPE* distM = cmMemAllocZ( VECT_OP_TYPE, scn*selIdxN );
  577. // sumV[] is a temp vector to hold the summed distances to from the selected columns to each column in sM[]
  578. VECT_OP_TYPE* sumV = cmMemAllocZ( VECT_OP_TYPE, scn );
  579. // select a random point from sM[] and copy it to the first column of dM[]
  580. cmVOU_Random(&i,1,scn);
  581. VECT_OP_FUNC(Copy)(dM, srn, sM + (i*srn));
  582. if( selIdxV != NULL )
  583. selIdxV[0] = i;
  584. for(dcn=1; dcn<selIdxN; ++dcn)
  585. {
  586. // set distM[scn,dcn] with the dist from dM[dcn,srn] to each column in sM[]
  587. VECT_OP_FUNC(DistVMM)( distM, NULL, NULL, srn, dM, dcn, sM, scn, distFunc, userPtr );
  588. // sum the rows of distM[ scn, dcn ] into sumV[scn]
  589. VECT_OP_FUNC(SumMN)( distM, scn, dcn, sumV );
  590. if( avgFl )
  591. VECT_OP_FUNC(DivVS)( sumV, scn, dcn );
  592. // find the point in sM[] which has the greatest combined distance to all previously selected points.
  593. unsigned maxIdx = VECT_OP_FUNC(MaxIndex)(sumV, scn, 1 );
  594. // copy the point into dM[]
  595. VECT_OP_FUNC(Copy)(dM + (dcn*srn), srn, sM + (maxIdx*srn));
  596. if( selIdxV != NULL )
  597. selIdxV[dcn] = maxIdx;
  598. }
  599. cmMemPtrFree(&distM);
  600. cmMemPtrFree(&sumV);
  601. if( freeFl )
  602. cmMemPtrFree(&dM);
  603. }
  604. void VECT_OP_FUNC(SelectMaxDist)( VECT_OP_TYPE *dM, unsigned* selIdxV, unsigned selIdxN, const VECT_OP_TYPE* sM, unsigned srn, unsigned scn, VECT_OP_TYPE (*distFunc)( void* userPtr, const VECT_OP_TYPE* s0V, const VECT_OP_TYPE* s1V, unsigned sn ), void* userPtr )
  605. { VECT_OP_FUNC(_SelectDist)(dM,selIdxV,selIdxN,sM,srn,scn,distFunc,userPtr,false); }
  606. void VECT_OP_FUNC(SelectMaxAvgDist)( VECT_OP_TYPE *dM, unsigned* selIdxV, unsigned selIdxN, const VECT_OP_TYPE* sM, unsigned srn, unsigned scn, VECT_OP_TYPE (*distFunc)( void* userPtr, const VECT_OP_TYPE* s0V, const VECT_OP_TYPE* s1V, unsigned sn ), void* userPtr )
  607. { VECT_OP_FUNC(_SelectDist)(dM,selIdxV,selIdxN,sM,srn,scn,distFunc,userPtr,true); }
  608. #ifdef CM_VECTOP
  609. VECT_OP_TYPE VECT_OP_FUNC(MultSumVV)( const VECT_OP_TYPE* s0p, const VECT_OP_TYPE* s1p, unsigned sn )
  610. { return VECT_OP_BLAS_FUNC(dot)(sn, s0p, 1, s1p, 1); }
  611. #else
  612. VECT_OP_TYPE VECT_OP_FUNC(MultSumVV)( const VECT_OP_TYPE* s0p, const VECT_OP_TYPE* s1p, unsigned sn )
  613. {
  614. VECT_OP_TYPE sum = 0;
  615. const VECT_OP_TYPE* sep = s0p + sn;
  616. while(s0p<sep)
  617. sum += *s0p++ * *s1p++;
  618. return sum;
  619. }
  620. #endif
  621. VECT_OP_TYPE VECT_OP_FUNC(MultSumVS)( const VECT_OP_TYPE* s0p, unsigned sn, VECT_OP_TYPE s1 )
  622. {
  623. VECT_OP_TYPE sum = 0;
  624. const VECT_OP_TYPE* sep = s0p + sn;
  625. while(s0p<sep)
  626. sum += *s0p++ * s1;
  627. return sum;
  628. }
  629. #ifdef CM_VECTOP
  630. VECT_OP_TYPE* VECT_OP_FUNC(MultVMV)( VECT_OP_TYPE* dbp, unsigned mrn, const VECT_OP_TYPE* mp, unsigned mcn, const VECT_OP_TYPE* vp )
  631. {
  632. VECT_OP_BLAS_FUNC(gemv)( CblasColMajor, CblasNoTrans, mrn, mcn, 1.0, mp, mrn, vp, 1, 0.0, dbp, 1 );
  633. return dbp;
  634. }
  635. #else
  636. VECT_OP_TYPE* VECT_OP_FUNC(MultVMV)( VECT_OP_TYPE* dbp, unsigned mrn, const VECT_OP_TYPE* mp, unsigned mcn, const VECT_OP_TYPE* vp )
  637. {
  638. const VECT_OP_TYPE* dep = dbp + mrn;
  639. VECT_OP_TYPE* dp = dbp;
  640. const VECT_OP_TYPE* vep = vp + mcn;
  641. // for each dest element
  642. for(; dbp < dep; ++dbp )
  643. {
  644. const VECT_OP_TYPE* vbp = vp;
  645. const VECT_OP_TYPE* mbp = mp++;
  646. *dbp = 0;
  647. // for each source vector row and src mtx col
  648. while( vbp < vep )
  649. {
  650. *dbp += *mbp * *vbp++;
  651. mbp += mrn;
  652. }
  653. }
  654. return dp;
  655. }
  656. #endif
  657. #ifdef CM_VECTOP
  658. VECT_OP_TYPE* VECT_OP_FUNC(MultVVM)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* vp, unsigned vn, const VECT_OP_TYPE* mp )
  659. {
  660. VECT_OP_BLAS_FUNC(gemv)( CblasColMajor, CblasTrans, vn, dn, 1.0, mp, vn, vp, 1, 0.0, dbp, 1 );
  661. return dbp;
  662. }
  663. #else
  664. VECT_OP_TYPE* VECT_OP_FUNC(MultVVM)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* vp, unsigned vn, const VECT_OP_TYPE* mp )
  665. {
  666. unsigned i;
  667. for(i=0; i<dn; ++i)
  668. dbp[i] = VECT_OP_FUNC(MultSumVV)(vp,mp + (i*vn),vn);
  669. return dbp;
  670. }
  671. #endif
  672. #ifdef CM_VECTOP
  673. VECT_OP_TYPE* VECT_OP_FUNC(MultVMtV)( VECT_OP_TYPE* dbp, unsigned mcn, const VECT_OP_TYPE* mp, unsigned mrn, const VECT_OP_TYPE* vp )
  674. {
  675. VECT_OP_BLAS_FUNC(gemv)( CblasColMajor, CblasTrans, mrn, mcn, 1.0, mp, mrn, vp, 1, 0.0, dbp, 1 );
  676. return dbp;
  677. }
  678. #else
  679. VECT_OP_TYPE* VECT_OP_FUNC(MultVMtV)( VECT_OP_TYPE* dbp, unsigned mcn, const VECT_OP_TYPE* mp, unsigned mrn, const VECT_OP_TYPE* vp )
  680. {
  681. const VECT_OP_TYPE* dep = dbp + mcn;
  682. VECT_OP_TYPE* dp = dbp;
  683. const VECT_OP_TYPE* vep = vp + mrn;
  684. // for each dest element
  685. for(; dbp < dep; ++dbp )
  686. {
  687. const VECT_OP_TYPE* vbp = vp;
  688. *dbp = 0;
  689. // for each source vector row and src mtx col
  690. while( vbp < vep )
  691. *dbp += *mp++ * *vbp++;
  692. }
  693. return dp;
  694. }
  695. #endif
  696. VECT_OP_TYPE* VECT_OP_FUNC(MultDiagVMV)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* mp, unsigned mcn, const VECT_OP_TYPE* vp )
  697. {
  698. VECT_OP_TYPE* rp = dbp;
  699. const VECT_OP_TYPE* mep = mp + (dn*mcn);
  700. // for each dest element
  701. for(; mp < mep; mp += dn+1 )
  702. *dbp++ = *vp++ * *mp;
  703. return rp;
  704. }
  705. /*
  706. Fortran Doc: http://www.netlib.org/blas/cgemm.f
  707. C Doc: http://techpubs.sgi.com/library/tpl/cgi-bin/getdoc.cgi?cmd=getdoc&coll=0650&db=man&fname=3%20INTRO_CBLAS
  708. C = alpha * op(A) * op(B) + beta * C
  709. cblas_Xgemm(
  710. order, enum CBLAS_ORDER {CblasRowMajor=101, CblasColMajor=102};
  711. transposeA, enum CBLAS_TRANSPOSE { CblasNoTrans, CblasTrans, CBlasConjTrans }
  712. transposeB,
  713. M, row op(A) and rows C (i.e. rows of A 'after' optional transpose)
  714. N, col op(B) and cols C (i.e. rows of B 'after' optional transpose)
  715. K, col op(A) and rows op(B)
  716. alpha, A scalar
  717. A, pointer to source matrix A
  718. lda, number of rows in A as it is stored in memory (assuming col major order)
  719. B, pointer to source matrix B
  720. ldb, number of rows in B as it is stored in memory (assuming col major order)
  721. beta C scalar
  722. C, pointer to destination matrix C
  723. ldc number of rows in C as it is stored in memory (assuming col major order)
  724. )
  725. */
  726. #ifdef CM_VECTOP
  727. VECT_OP_TYPE* VECT_OP_FUNC(MultMMM1)(VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, VECT_OP_TYPE alpha, const VECT_OP_TYPE* m0, const VECT_OP_TYPE* m1, unsigned n, VECT_OP_TYPE beta, unsigned flags )
  728. {
  729. bool t0fl = cmIsFlag(flags,kTransposeM0Fl);
  730. bool t1fl = cmIsFlag(flags,kTransposeM1Fl);
  731. VECT_OP_BLAS_FUNC(gemm)(
  732. CblasColMajor,
  733. t0fl ? CblasTrans : CblasNoTrans,
  734. t1fl ? CblasTrans : CblasNoTrans,
  735. drn, dcn, n,
  736. alpha,
  737. m0, t0fl ? n : drn,
  738. m1, t1fl ? dcn : n,
  739. beta,
  740. dbp, drn );
  741. return dbp;
  742. }
  743. #else
  744. // Not implemented.
  745. #endif
  746. #ifdef CM_VECTOP
  747. VECT_OP_TYPE* VECT_OP_FUNC(MultMMM2)(VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, VECT_OP_TYPE alpha, const VECT_OP_TYPE* m0, const VECT_OP_TYPE* m1, unsigned n, VECT_OP_TYPE beta, unsigned flags, unsigned dprn, unsigned m0prn, unsigned m1prn )
  748. {
  749. VECT_OP_BLAS_FUNC(gemm)(
  750. CblasColMajor,
  751. cmIsFlag(flags,kTransposeM0Fl) ? CblasTrans : CblasNoTrans,
  752. cmIsFlag(flags,kTransposeM1Fl) ? CblasTrans : CblasNoTrans,
  753. drn, dcn, n,
  754. alpha,
  755. m0, m0prn,
  756. m1, m1prn,
  757. beta,
  758. dbp, dprn );
  759. return dbp;
  760. }
  761. #else
  762. // Not implemented.
  763. #endif
  764. #ifdef CM_VECTOP
  765. VECT_OP_TYPE* VECT_OP_FUNC(MultMMM)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* m0, const VECT_OP_TYPE* m1, unsigned n )
  766. {
  767. VECT_OP_BLAS_FUNC(gemm)(
  768. CblasColMajor,
  769. CblasNoTrans, CblasNoTrans,
  770. drn, dcn, n,
  771. 1.0, m0, drn,
  772. m1, n,
  773. 0.0, dbp, drn );
  774. return dbp;
  775. }
  776. #else
  777. VECT_OP_TYPE* VECT_OP_FUNC(MultMMM)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* m0, const VECT_OP_TYPE* m1, unsigned m0cn_m1rn )
  778. {
  779. unsigned i;
  780. for(i=0; i<dcn; ++i)
  781. VECT_OP_FUNC(MultVMV)(dbp+(i*drn),drn,m0,m0cn_m1rn,m1+(i*m0cn_m1rn));
  782. return dbp;
  783. }
  784. #endif
  785. #ifdef CM_VECTOP
  786. VECT_OP_TYPE* VECT_OP_FUNC(MultMMMt)(VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* m0, const VECT_OP_TYPE* m1, unsigned m0cn_m1rn )
  787. {
  788. VECT_OP_BLAS_FUNC(gemm)( CblasColMajor, CblasNoTrans, CblasTrans,
  789. drn, dcn, m0cn_m1rn,
  790. 1.0, m0, drn,
  791. m1, dcn,
  792. 0.0, dbp, drn );
  793. return dbp;
  794. }
  795. #else
  796. VECT_OP_TYPE* VECT_OP_FUNC(MultMMMt)(VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* m0, const VECT_OP_TYPE* m1, unsigned m0cn_m1rn )
  797. {
  798. unsigned i,j,k;
  799. VECT_OP_FUNC(Zero)(dbp,drn*dcn);
  800. for(i=0; i<dcn; ++i)
  801. for(j=0; j<drn; ++j)
  802. for(k=0; k<m0cn_m1rn; ++k)
  803. dbp[ i*drn + j ] += m0[ k*drn + j ] * m1[ k*dcn + i ];
  804. return dbp;
  805. }
  806. #endif
  807. VECT_OP_TYPE* VECT_OP_FUNC(PowVS)( VECT_OP_TYPE* dbp, unsigned dn, VECT_OP_TYPE expo )
  808. {
  809. VECT_OP_TYPE* dp = dbp;
  810. VECT_OP_TYPE* ep = dp + dn;
  811. for(; dp < ep; ++dp )
  812. *dp = (VECT_OP_TYPE)pow(*dp,expo);
  813. return dbp;
  814. }
  815. VECT_OP_TYPE* VECT_OP_FUNC(PowVVS)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp, VECT_OP_TYPE expo )
  816. {
  817. VECT_OP_TYPE* dp = dbp;
  818. VECT_OP_TYPE* ep = dp + dn;
  819. for(; dp < ep; ++dp,++sp )
  820. *dp = (VECT_OP_TYPE)pow(*sp,expo);
  821. return dbp;
  822. }
  823. VECT_OP_TYPE* VECT_OP_FUNC(LogV)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp )
  824. {
  825. VECT_OP_TYPE* dp = dbp;
  826. VECT_OP_TYPE* ep = dp + dn;
  827. for(; dp <ep; ++dp,++sbp)
  828. *dp = (VECT_OP_TYPE)log(*sbp);
  829. return dbp;
  830. }
  831. VECT_OP_TYPE* VECT_OP_FUNC(AmplToDbVV)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp, VECT_OP_TYPE minDb )
  832. {
  833. VECT_OP_TYPE minVal = pow(10.0,minDb/20.0);
  834. VECT_OP_TYPE* dp = dbp;
  835. VECT_OP_TYPE* ep = dp + dn;
  836. for(; dp<ep; ++dp,++sbp)
  837. *dp = *sbp<minVal ? minDb : 20.0 * log10(*sbp);
  838. return dbp;
  839. }
  840. VECT_OP_TYPE* VECT_OP_FUNC(DbToAmplVV)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp)
  841. {
  842. VECT_OP_TYPE* dp = dbp;
  843. VECT_OP_TYPE* ep = dp + dn;
  844. for(; dp<ep; ++dp,++sbp)
  845. *dp = pow(10.0,*sbp/20.0);
  846. return dbp;
  847. }
  848. VECT_OP_TYPE* VECT_OP_FUNC(PowToDbVV)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp, VECT_OP_TYPE minDb )
  849. {
  850. VECT_OP_TYPE minVal = pow(10.0,minDb/10.0);
  851. VECT_OP_TYPE* dp = dbp;
  852. VECT_OP_TYPE* ep = dp + dn;
  853. for(; dp<ep; ++dp,++sbp)
  854. *dp = *sbp<minVal ? minDb : 10.0 * log10(*sbp);
  855. return dbp;
  856. }
  857. VECT_OP_TYPE* VECT_OP_FUNC(DbToPowVV)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp)
  858. {
  859. VECT_OP_TYPE* dp = dbp;
  860. VECT_OP_TYPE* ep = dp + dn;
  861. for(; dp<ep; ++dp,++sbp)
  862. *dp = pow(10.0,*sbp/10.0);
  863. return dbp;
  864. }
  865. VECT_OP_TYPE* VECT_OP_FUNC(RandSymPosDef)( VECT_OP_TYPE* dbp, unsigned dn, VECT_OP_TYPE* t )
  866. {
  867. unsigned i,j;
  868. bool fl = t == NULL;
  869. if( fl )
  870. t = cmMemAlloc( VECT_OP_TYPE , dn*dn );
  871. do
  872. {
  873. // intialize t[] as a square symetric matrix with random values
  874. for(i=0; i<dn; ++i)
  875. for(j=i; j<dn; ++j)
  876. {
  877. VECT_OP_TYPE v = (VECT_OP_TYPE)rand()/RAND_MAX;
  878. t[ (i*dn) + j ] = v;
  879. if( i != j )
  880. t[ (j*dn) + i ] = v;
  881. }
  882. // square t[] to force the eigenvalues to be positive
  883. VECT_OP_FUNC(MultMMM)(dbp,dn,dn,t,t,dn);
  884. VECT_OP_FUNC(Copy)(t,dn*dn,dbp);
  885. // test that func is positive definite
  886. }while( VECT_OP_FUNC(Chol)(t,dn)==NULL );
  887. if( fl )
  888. cmMemFree(t);
  889. return dbp;
  890. }
  891. // Calculate the determinant of a matrix previously factored by
  892. // the lapack function dgetrf_()
  893. VECT_OP_TYPE VECT_OP_FUNC(LUDet)( const VECT_OP_TYPE* lu, const int_lap_t* ipiv, int rn )
  894. {
  895. VECT_OP_TYPE det1 = 1;
  896. int det2 = 0;
  897. int i;
  898. for(i=0; i<rn; ++i)
  899. {
  900. if( ipiv != NULL && ipiv[i] != (i+1) )
  901. det1 = -det1;
  902. det1 = lu[ (i*rn) + i ] * det1;
  903. if( det1 == 0 )
  904. break;
  905. while( fabs(det1) <= 1 )
  906. {
  907. det1 *= 10;
  908. det2 -= 1;
  909. }
  910. //continue;
  911. while( fabs(det1) >= 10 )
  912. {
  913. det1 /= 10;
  914. det2 += 1;
  915. }
  916. }
  917. // Here's where underflow or overflow might happen.
  918. // Enable floating point exception handling to trap.
  919. det1 *= pow(10.0,det2);
  920. return det1;
  921. }
  922. // take the inverse of a matrix factored via lapack dgetrf_()
  923. VECT_OP_TYPE* VECT_OP_FUNC(LUInverse)(VECT_OP_TYPE* dp, int_lap_t* ipiv, int drn )
  924. {
  925. int_lap_t ispec = 1;
  926. int_lap_t rn = drn;
  927. int_lap_t n1 = drn;
  928. int_lap_t n2 = drn;
  929. int_lap_t n3 = drn;
  930. int_lap_t n4 = drn;
  931. char funcNameStr[] = {"DGETRI"};
  932. // Calculate the NB factor for LWORK -
  933. // The two args are length of string args 'funcNameStr' and ' '.
  934. // It is not clear how many 'n' args are requred so all are passed set to 'drn'
  935. int nb = ilaenv_(&ispec, funcNameStr, " ", &n1,&n2,&n3,&n4, strlen(funcNameStr), 1 );
  936. VECT_OP_TYPE w[drn * nb]; // allocate working memory
  937. int_lap_t info;
  938. // calculate inv(A) base on LU factorization
  939. VECT_OP_LAP_FUNC(getri_)(&rn,dp,&rn,ipiv,w,&rn,&info);
  940. assert(info==0);
  941. return info ==0 ? dp : NULL;
  942. }
  943. VECT_OP_TYPE VECT_OP_FUNC(DetM)( const VECT_OP_TYPE* sp, unsigned srn )
  944. {
  945. int_lap_t arn = srn;
  946. VECT_OP_TYPE A[ arn * arn ];
  947. int_lap_t ipiv[ arn ];
  948. int_lap_t info;
  949. VECT_OP_FUNC(Copy)(A,arn*arn,sp);
  950. // PLU factor
  951. VECT_OP_LAP_FUNC(getrf_)(&arn,&arn,A,&arn,ipiv,&info);
  952. if( info == 0 )
  953. return VECT_OP_FUNC(LUDet)(A,ipiv,arn);
  954. return 0;
  955. }
  956. VECT_OP_TYPE VECT_OP_FUNC(DetDiagM)( const VECT_OP_TYPE* sp, unsigned srn )
  957. { return VECT_OP_FUNC(LUDet)(sp,NULL,srn); }
  958. VECT_OP_TYPE VECT_OP_FUNC(LogDetM)( const VECT_OP_TYPE* sp, unsigned srn )
  959. {
  960. cmReal_t det = 0;
  961. unsigned ne2 = srn * srn;
  962. VECT_OP_TYPE U[ne2];
  963. const VECT_OP_TYPE* up = U;
  964. const VECT_OP_TYPE* ep = up + ne2;
  965. VECT_OP_FUNC(Copy)(U,ne2,sp);
  966. VECT_OP_FUNC(Chol)(U,srn);
  967. for(; up<ep; up += (srn+1) )
  968. det += log(*up);
  969. return 2*det;
  970. }
  971. VECT_OP_TYPE VECT_OP_FUNC(LogDetDiagM)( const VECT_OP_TYPE* sp, unsigned srn )
  972. { return log(VECT_OP_FUNC(DetDiagM)(sp,srn)); }
  973. VECT_OP_TYPE* VECT_OP_FUNC(InvM)( VECT_OP_TYPE* dp, unsigned drn )
  974. {
  975. int_lap_t rn = drn;
  976. int_lap_t ipiv[ rn ];
  977. int_lap_t info;
  978. // PLU factor
  979. VECT_OP_LAP_FUNC(getrf_)(&rn,&rn,dp,&rn,ipiv,&info);
  980. if( info == 0 )
  981. return VECT_OP_FUNC(LUInverse)(dp,ipiv,rn );
  982. return NULL;
  983. }
  984. VECT_OP_TYPE* VECT_OP_FUNC(InvDiagM)( VECT_OP_TYPE* dp, unsigned drn )
  985. {
  986. const VECT_OP_TYPE* dep = dp + (drn*drn);
  987. VECT_OP_TYPE* rp = dp;
  988. for(; dp < dep; dp += drn+1 )
  989. {
  990. *dp = 1.0 / *dp;
  991. // if any element on the diagonal is zero then the
  992. // determinant is zero and the matrix is not invertable
  993. if( *dp == 0 )
  994. break;
  995. }
  996. return dp < dep ? NULL : rp;
  997. }
  998. VECT_OP_TYPE* VECT_OP_FUNC(SolveLS)( VECT_OP_TYPE* A, unsigned an, VECT_OP_TYPE* B, unsigned bcn )
  999. {
  1000. int_lap_t aN = an;
  1001. int_lap_t bcN = bcn;
  1002. int_lap_t ipiv[ an ];
  1003. int_lap_t info = 0;
  1004. VECT_OP_LAP_FUNC(gesv_)(&aN,&bcN,(VECT_OP_TYPE*)A,&aN,ipiv,B,&aN,&info);
  1005. return info == 0 ? B : NULL;
  1006. }
  1007. VECT_OP_TYPE* VECT_OP_FUNC(Chol)(VECT_OP_TYPE* A, unsigned an )
  1008. {
  1009. char uplo = 'U';
  1010. int_lap_t N = an;
  1011. int_lap_t lda = an;
  1012. int_lap_t info = 0;
  1013. VECT_OP_LAP_FUNC(potrf_(&uplo,&N,(VECT_OP_TYPE*)A,&lda,&info));
  1014. return info == 0 ? A : NULL;
  1015. }
  1016. VECT_OP_TYPE* VECT_OP_FUNC(CholZ)(VECT_OP_TYPE* A, unsigned an )
  1017. {
  1018. unsigned i,j;
  1019. VECT_OP_FUNC(Chol)(A,an);
  1020. // zero the lower triangle of A
  1021. for(i=0; i<an; ++i)
  1022. for(j=i+1; j<an; ++j)
  1023. A[ (i*an) + j ] = 0;
  1024. return A;
  1025. }
  1026. VECT_OP_TYPE VECT_OP_FUNC(FracAvg)( double bi, double ei, const VECT_OP_TYPE* sbp, unsigned sn )
  1027. {
  1028. unsigned bii = cmMax(0,cmMin(sn-1,(unsigned)ceil(bi)));
  1029. unsigned eii = cmMax(0,cmMin(sn,(unsigned)floor(ei)+1));
  1030. double begW = bii - bi;
  1031. double endW = eii - floor(ei);
  1032. double cnt = eii - bii;
  1033. double sum = (double)VECT_OP_FUNC(Sum)(sbp+bii,eii-bii);
  1034. if( begW>0 && bii > 0 )
  1035. {
  1036. cnt += begW;
  1037. sum += begW * sbp[ bii-1 ];
  1038. }
  1039. if( endW>0 && eii+1 < sn )
  1040. {
  1041. cnt += endW;
  1042. sum += endW * sbp[ eii+1 ];
  1043. }
  1044. return (VECT_OP_TYPE)(sum / cnt);
  1045. }
  1046. VECT_OP_TYPE* VECT_OP_FUNC(DownSampleAvg)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp, unsigned sn )
  1047. {
  1048. const VECT_OP_TYPE* dep = dbp + dn;
  1049. VECT_OP_TYPE* rp = dbp;
  1050. unsigned i = 0;
  1051. double fact = (double)sn / dn;
  1052. assert( sn >= dn );
  1053. for(i=0; dbp < dep; ++i )
  1054. *dbp++ = VECT_OP_FUNC(FracAvg)( fact*i, fact*(i+1), sbp, sn );
  1055. return rp;
  1056. }
  1057. VECT_OP_TYPE* VECT_OP_FUNC(UpSampleInterp)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp, unsigned sn )
  1058. {
  1059. const VECT_OP_TYPE* dep = dbp + dn;
  1060. const VECT_OP_TYPE* sep = sbp + sn;
  1061. VECT_OP_TYPE* rp = dbp;
  1062. double fact = (double)sn / dn;
  1063. double phs = 0;
  1064. assert( sn <= dn );
  1065. while( dbp<dep )
  1066. {
  1067. if( sbp < sep )
  1068. *dbp++ = (VECT_OP_TYPE)((*sbp) + (phs * ((*(sbp+1)) - (*sbp))));
  1069. else
  1070. *dbp++ = (*(sep-1));
  1071. phs += fact;
  1072. while( phs > 1.0 )
  1073. {
  1074. phs -= 1.0;
  1075. sbp++;
  1076. }
  1077. }
  1078. return rp;
  1079. }
  1080. VECT_OP_TYPE* VECT_OP_FUNC(FitToSize)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp, unsigned sn )
  1081. {
  1082. if( dn == sn )
  1083. return VECT_OP_FUNC(Copy)(dbp,dn,sbp);
  1084. if( dn < sn )
  1085. return VECT_OP_FUNC(DownSampleAvg)(dbp,dn,sbp,sn);
  1086. return VECT_OP_FUNC(UpSampleInterp)(dbp,dn,sbp,sn);
  1087. }
  1088. VECT_OP_TYPE* VECT_OP_FUNC(LinearMap)(VECT_OP_TYPE* dV, unsigned dn, VECT_OP_TYPE* sV, unsigned sn )
  1089. {
  1090. if( dn == sn )
  1091. {
  1092. memcpy(dV,sV,dn*sizeof(VECT_OP_TYPE));
  1093. return dV;
  1094. }
  1095. unsigned i,j,k;
  1096. // if stretching
  1097. if( dn > sn )
  1098. {
  1099. VECT_OP_TYPE f_n = (VECT_OP_TYPE)dn/sn;
  1100. VECT_OP_TYPE f_nn = f_n;
  1101. unsigned i_n = floor(f_n);
  1102. k = 0;
  1103. i = 0;
  1104. // for each set of ceiling(dn/sn) dst values
  1105. while(1)
  1106. {
  1107. // repeat floor(dn/sn) src val into dst
  1108. for(j=0; j<i_n; ++j,++i)
  1109. dV[i] = sV[k];
  1110. if( k + 1 == sn )
  1111. break;
  1112. // interpolate between the cur and nxt source value
  1113. VECT_OP_TYPE w = f_nn - floor(f_nn);
  1114. dV[i] = sV[k] + w * (sV[k+1]-sV[k]);
  1115. ++i;
  1116. ++k;
  1117. i_n = floor(f_n - (1.0-w));
  1118. f_nn += f_n;
  1119. }
  1120. }
  1121. else // if shrinking
  1122. {
  1123. VECT_OP_TYPE f_n = (VECT_OP_TYPE)sn/dn;
  1124. VECT_OP_TYPE f_nn = f_n;
  1125. unsigned i_n = floor(f_n);
  1126. k = 0;
  1127. i = 0;
  1128. VECT_OP_TYPE acc = 0;
  1129. // for each seq of ceil(sn/dn) src values
  1130. while(1)
  1131. {
  1132. // accum first floor(sn/dn) src values
  1133. for(j=0; j<i_n; ++j,++i)
  1134. acc += sV[i];
  1135. if( k == dn-1 )
  1136. {
  1137. dV[k] = acc/f_n;
  1138. break;
  1139. }
  1140. // interpolate frac of last src value
  1141. VECT_OP_TYPE w = f_nn - floor(f_nn);
  1142. // form avg
  1143. dV[k] = (acc + (w*sV[i]))/f_n;
  1144. // reload acc with inverse frac of src value
  1145. acc = (1.0-w) * sV[i];
  1146. ++i;
  1147. ++k;
  1148. i_n = floor(f_n-(1.0-w));
  1149. f_nn += f_n;
  1150. }
  1151. }
  1152. return dV;
  1153. }
  1154. VECT_OP_TYPE* VECT_OP_FUNC(Random)( VECT_OP_TYPE* dbp, unsigned dn, VECT_OP_TYPE minVal, VECT_OP_TYPE maxVal )
  1155. {
  1156. const VECT_OP_TYPE* dep = dbp + dn;
  1157. VECT_OP_TYPE* dp =dbp;
  1158. double fact = (maxVal - minVal)/RAND_MAX;
  1159. while( dbp < dep )
  1160. *dbp++ = fact * rand() + minVal;
  1161. return dp;
  1162. }
  1163. unsigned* VECT_OP_FUNC(WeightedRandInt)( unsigned *dbp, unsigned dn, const VECT_OP_TYPE* wp, unsigned wn )
  1164. {
  1165. unsigned i,j;
  1166. VECT_OP_TYPE a[ wn ];
  1167. // form bin boundaries by taking a cum. sum of the weight values.
  1168. VECT_OP_FUNC(CumSum)(a,wn,wp);
  1169. for(j=0; j<dn; ++j)
  1170. {
  1171. // gen a random number from a uniform distribution betwen 0 and the max value from the cumsum.
  1172. VECT_OP_TYPE rv = (VECT_OP_TYPE)rand() * a[wn-1] / RAND_MAX;
  1173. // find the bin the rv falls into
  1174. for(i=0; i<wn-1; ++i)
  1175. if( rv <= a[i] )
  1176. {
  1177. dbp[j] = i;
  1178. break;
  1179. }
  1180. if(i==wn-1)
  1181. dbp[j]= wn-1;
  1182. }
  1183. return dbp;
  1184. }
  1185. VECT_OP_TYPE* VECT_OP_FUNC(RandomGauss)( VECT_OP_TYPE* dbp, unsigned dn, VECT_OP_TYPE mean, VECT_OP_TYPE var )
  1186. {
  1187. const VECT_OP_TYPE* dep = dbp + dn;
  1188. VECT_OP_TYPE* rp = dbp;
  1189. // The code below implements the Box-Muller uniform to
  1190. // Gaussian distribution transformation. In rectangular
  1191. // coordinates this transform is defined as:
  1192. // y1 = sqrt( - 2.0 * log(x1) ) * cos( 2.0*M_PI*x2 )
  1193. // y2 = sqrt( - 2.0 * log(x1) ) * sin( 2.0*M_PI*x2 )
  1194. //
  1195. while( dbp < dep )
  1196. *dbp++ = sqrt( -2.0 * log((VECT_OP_TYPE)rand()/RAND_MAX)) * cos(2.0*M_PI*((VECT_OP_TYPE)rand()/RAND_MAX)) * var + mean;
  1197. return rp;
  1198. }
  1199. VECT_OP_TYPE* VECT_OP_FUNC(RandomGaussV)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* meanV, const VECT_OP_TYPE* varV )
  1200. {
  1201. VECT_OP_TYPE* rp = dbp;
  1202. const VECT_OP_TYPE* dep = dbp + dn;
  1203. while( dbp < dep )
  1204. VECT_OP_FUNC(RandomGauss)( dbp++, 1, *meanV++, *varV++ );
  1205. return rp;
  1206. }
  1207. VECT_OP_TYPE* VECT_OP_FUNC(RandomGaussM)( VECT_OP_TYPE* dbp, unsigned rn, unsigned cn, const VECT_OP_TYPE* meanV, const VECT_OP_TYPE* varV )
  1208. {
  1209. unsigned i;
  1210. for(i=0; i<cn; ++i)
  1211. VECT_OP_FUNC(RandomGaussV)( dbp+(i*rn), rn, meanV, varV );
  1212. return dbp;
  1213. }
  1214. VECT_OP_TYPE* VECT_OP_FUNC(RandomGaussDiagM)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* meanV, const VECT_OP_TYPE* covarM )
  1215. {
  1216. unsigned i,j;
  1217. for(i=0; i<dcn; ++i)
  1218. for(j=0; j<drn; ++j)
  1219. VECT_OP_FUNC(RandomGauss)(dbp + (i*drn)+j, 1, meanV[j], covarM[ (j*drn) + j]);
  1220. return dbp;
  1221. }
  1222. VECT_OP_TYPE* VECT_OP_FUNC(RandomGaussNonDiagM)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* meanV, const VECT_OP_TYPE* covarM, VECT_OP_TYPE* t )
  1223. {
  1224. bool fl = t == NULL;
  1225. if( fl )
  1226. t = cmMemAlloc(VECT_OP_TYPE, drn * drn );
  1227. VECT_OP_FUNC(Copy)(t,drn*drn,covarM);
  1228. if( VECT_OP_FUNC(CholZ)(t,drn) == NULL )
  1229. {
  1230. // Cholesky decomposition failed - should try eigen analysis next
  1231. // From octave mvnrnd.m
  1232. // [E,Lambda]=eig(Sigma);
  1233. // if (min(diag(Lambda))<0),error('Sigma must be positive semi-definite.'),end
  1234. // U = sqrt(Lambda)*E';
  1235. assert(0);
  1236. }
  1237. /*
  1238. unsigned i,j;
  1239. for(i=0; i<drn; ++i)
  1240. {
  1241. for(j=0; j<drn; ++j)
  1242. printf("%f ",t[ (j*drn) + i]);
  1243. printf("\n");
  1244. }
  1245. */
  1246. VECT_OP_FUNC(RandomGaussNonDiagM2)(dbp,drn,dcn,meanV,t);
  1247. if(fl)
  1248. cmMemFree(t);
  1249. return dbp;
  1250. }
  1251. VECT_OP_TYPE* VECT_OP_FUNC(RandomGaussNonDiagM2)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* meanV, const VECT_OP_TYPE* uM )
  1252. {
  1253. unsigned i;
  1254. for(i=0; i<dcn; ++i)
  1255. {
  1256. VECT_OP_TYPE r[ drn ];
  1257. VECT_OP_FUNC(RandomGauss)(r,drn,0,1); // r = randn(drn,1);
  1258. VECT_OP_FUNC(MultVVM)( dbp+(i*drn),drn,r,drn,uM); // dbp[:i] = r * uM;
  1259. VECT_OP_FUNC(AddVV)( dbp+(i*drn),drn,meanV); // dbp[:,i] += meanV;
  1260. }
  1261. return dbp;
  1262. }
  1263. VECT_OP_TYPE* VECT_OP_FUNC(RandomGaussMM)( VECT_OP_TYPE* dbp, unsigned drn, unsigned dcn, const VECT_OP_TYPE* meanM, const VECT_OP_TYPE* varM, unsigned K )
  1264. {
  1265. unsigned k;
  1266. unsigned D = drn;
  1267. unsigned N = dcn/K;
  1268. for(k=0; k<K; ++k)
  1269. VECT_OP_FUNC(RandomGaussM)( dbp + (k*N*D), drn, N, meanM + (k*D), varM + (k*D) );
  1270. return dbp;
  1271. }
  1272. VECT_OP_TYPE* VECT_OP_FUNC(CircleCoords)( VECT_OP_TYPE* dbp, unsigned dn, VECT_OP_TYPE x, VECT_OP_TYPE y, VECT_OP_TYPE varX, VECT_OP_TYPE varY )
  1273. {
  1274. unsigned i;
  1275. for(i=0; i<dn; ++i)
  1276. {
  1277. double a = 2.0*M_PI*i/(dn-1);
  1278. dbp[ i ] = (VECT_OP_TYPE)(varX * cos(a) + x);
  1279. dbp[ i+dn ] = (VECT_OP_TYPE)(varY * sin(a) + y);
  1280. }
  1281. return dbp;
  1282. }
  1283. unsigned VECT_OP_FUNC(SynthSine)( VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz )
  1284. {
  1285. const VECT_OP_TYPE* dep = dbp + dn;
  1286. double rps = 2.0*M_PI*hz/srate;
  1287. while( dbp < dep )
  1288. *dbp++ = (VECT_OP_TYPE)sin( rps * phase++ );
  1289. return phase;
  1290. }
  1291. unsigned VECT_OP_FUNC(SynthCosine)( VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz )
  1292. {
  1293. const VECT_OP_TYPE* dep = dbp + dn;
  1294. double rps = 2.0*M_PI*hz/srate;
  1295. while( dbp < dep )
  1296. *dbp++ = (VECT_OP_TYPE)cos( rps * phase++ );
  1297. return phase;
  1298. }
  1299. unsigned VECT_OP_FUNC(SynthSquare)( VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz, unsigned otCnt )
  1300. {
  1301. const VECT_OP_TYPE* dep = dbp + dn;
  1302. if( otCnt > 0 )
  1303. {
  1304. unsigned i;
  1305. // initialize the buffer with the fundamental
  1306. VECT_OP_FUNC(SynthSine)( dbp, dn, phase, srate, hz );
  1307. otCnt *= 2;
  1308. // sum in each additional harmonic
  1309. for(i=3; i<otCnt; i+=2)
  1310. {
  1311. VECT_OP_TYPE* dp = dbp;
  1312. double rps = 2.0 * M_PI * i * hz / srate;
  1313. unsigned phs = phase;
  1314. double g = 1.0/i;
  1315. while( dp < dep )
  1316. *dp++ += (VECT_OP_TYPE)(g * sin( rps * phs++ ));
  1317. }
  1318. }
  1319. return phase + (dep - dbp);
  1320. }
  1321. unsigned VECT_OP_FUNC(SynthTriangle)( VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz, unsigned otCnt )
  1322. {
  1323. const VECT_OP_TYPE* dep = dbp + dn;
  1324. if( otCnt > 0 )
  1325. {
  1326. unsigned i;
  1327. // initialize the buffer with the fundamental
  1328. VECT_OP_FUNC(SynthCosine)( dbp, dn, phase, srate, hz );
  1329. otCnt *= 2;
  1330. // sum in each additional harmonic
  1331. for(i=3; i<otCnt; i+=2)
  1332. {
  1333. VECT_OP_TYPE* dp = dbp;
  1334. double rps = 2.0 * M_PI * i * hz / srate;
  1335. unsigned phs = phase;
  1336. double g = 1.0/(i*i);
  1337. while( dp < dep )
  1338. *dp++ += (VECT_OP_TYPE)(g * cos( rps * phs++ ));
  1339. }
  1340. }
  1341. return phase + (dep - dbp);
  1342. }
  1343. unsigned VECT_OP_FUNC(SynthSawtooth)( VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz, unsigned otCnt )
  1344. {
  1345. const VECT_OP_TYPE* dep = dbp + dn;
  1346. if( otCnt > 0 )
  1347. {
  1348. unsigned i;
  1349. // initialize the buffer with the fundamental
  1350. VECT_OP_FUNC(SynthSine)( dbp, dn, phase, srate, hz );
  1351. // sum in each additional harmonic
  1352. for(i=2; i<otCnt; ++i)
  1353. {
  1354. VECT_OP_TYPE* dp = dbp;
  1355. double rps = 2.0 * M_PI * i * hz / srate;
  1356. unsigned phs = phase;
  1357. double g = 1.0/i;
  1358. while( dp < dep )
  1359. *dp++ += (VECT_OP_TYPE)(g * sin( rps * phs++ ));
  1360. }
  1361. VECT_OP_FUNC(MultVS)(dbp,dn,2.0/M_PI);
  1362. }
  1363. return phase + (dep - dbp);
  1364. }
  1365. unsigned VECT_OP_FUNC(SynthPulseCos)( VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz, unsigned otCnt )
  1366. {
  1367. const VECT_OP_TYPE* dep = dbp + dn;
  1368. if( otCnt > 0 )
  1369. {
  1370. unsigned i;
  1371. // initialize the buffer with the fundamental
  1372. VECT_OP_FUNC(SynthCosine)( dbp, dn, phase, srate, hz );
  1373. // sum in each additional harmonic
  1374. for(i=1; i<otCnt; ++i)
  1375. {
  1376. VECT_OP_TYPE* dp = dbp;
  1377. double rps = 2.0 * M_PI * i * hz / srate;
  1378. unsigned phs = phase;
  1379. while( dp < dep )
  1380. *dp++ += (VECT_OP_TYPE)cos( rps * phs++ );
  1381. }
  1382. VECT_OP_FUNC(MultVS)(dbp,dn,1.0/otCnt);
  1383. }
  1384. return phase + (dep - dbp);
  1385. }
  1386. unsigned VECT_OP_FUNC(SynthImpulse)( VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz )
  1387. {
  1388. const VECT_OP_TYPE* dep = dbp + dn;
  1389. double pi2 = 2.0*M_PI;
  1390. double rps = pi2*hz/srate;
  1391. double v0,v1 = fmod( rps * phase, pi2);
  1392. if( dbp == dep )
  1393. return phase;
  1394. // the phase is set to zero when the first output should be a 1
  1395. if( phase == 0 )
  1396. {
  1397. *dbp++ = 1;
  1398. ++phase;
  1399. }
  1400. while( dbp < dep )
  1401. {
  1402. // the phase vector will always be increasing
  1403. // the modulus of the phase vector will wrap with frequency 'hz'
  1404. v0 = fmod( rps * phase++, pi2 );
  1405. // notice when wrapping occurs
  1406. *dbp++ = (VECT_OP_TYPE)(v0 < v1);
  1407. v1 = v0;
  1408. }
  1409. // check if the next output should be a 1
  1410. // (this eliminates the problem of not having access to v1 on the next call to this function
  1411. if( fmod( rps * phase, pi2 ) < v1 )
  1412. phase = 0;
  1413. return phase;
  1414. }
  1415. VECT_OP_TYPE VECT_OP_FUNC(SynthPinkNoise)( VECT_OP_TYPE* dbp, unsigned n, VECT_OP_TYPE delaySmp )
  1416. {
  1417. const VECT_OP_TYPE* dep = dbp + n;
  1418. VECT_OP_TYPE tmp[ n ];
  1419. VECT_OP_FUNC(Random)(tmp,n,-1.0,1.0);
  1420. VECT_OP_TYPE* sp = tmp;
  1421. VECT_OP_TYPE reg = delaySmp;
  1422. for(; dbp < dep; ++sp)
  1423. {
  1424. *dbp++ = (*sp + reg)/2.0;
  1425. reg = *sp;
  1426. }
  1427. return *sp;
  1428. }
  1429. VECT_OP_TYPE* VECT_OP_FUNC(LinearToDb)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp, VECT_OP_TYPE mult )
  1430. {
  1431. const VECT_OP_TYPE* dep = dbp + dn;
  1432. VECT_OP_TYPE* rp = dbp;
  1433. while( dbp < dep )
  1434. *dbp++ = (VECT_OP_TYPE)(mult * log10( VECT_OP_EPSILON + *sp++ ));
  1435. return rp;
  1436. }
  1437. VECT_OP_TYPE* VECT_OP_FUNC(dBToLinear)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp, VECT_OP_TYPE mult )
  1438. {
  1439. const VECT_OP_TYPE* dep = dbp + dn;
  1440. VECT_OP_TYPE* rp = dbp;
  1441. while( dbp < dep )
  1442. *dbp++ = (VECT_OP_TYPE)pow(10.0, *sp++ / mult );
  1443. return rp;
  1444. }
  1445. VECT_OP_TYPE* VECT_OP_FUNC(AmplitudeToDb)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp )
  1446. { return VECT_OP_FUNC(LinearToDb)(dbp,dn,sp,20.0); }
  1447. VECT_OP_TYPE* VECT_OP_FUNC(PowerToDb)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp )
  1448. { return VECT_OP_FUNC(LinearToDb)(dbp,dn,sp,10.0); }
  1449. VECT_OP_TYPE* VECT_OP_FUNC(dBToAmplitude)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp )
  1450. { return VECT_OP_FUNC(dBToLinear)( dbp,dn,sp,20); }
  1451. VECT_OP_TYPE* VECT_OP_FUNC(dBToPower)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sp )
  1452. { return VECT_OP_FUNC(dBToLinear)( dbp,dn,sp,10); }
  1453. unsigned VECT_OP_FUNC(SynthPhasor)(VECT_OP_TYPE* dbp, unsigned dn, unsigned phase, double srate, double hz )
  1454. {
  1455. const VECT_OP_TYPE* dep = dbp + dn;
  1456. while( dbp < dep )
  1457. *dbp++ = (VECT_OP_TYPE)fmod( (hz * phase++)/srate, 1.0 );
  1458. return phase;
  1459. }
  1460. VECT_OP_TYPE VECT_OP_FUNC(KaiserBetaFromSidelobeReject)( double sidelobeRejectDb )
  1461. {
  1462. double beta;
  1463. if( sidelobeRejectDb < 13.26 )
  1464. sidelobeRejectDb = 13.26;
  1465. else
  1466. if( sidelobeRejectDb > 120.0)
  1467. sidelobeRejectDb = 120.0;
  1468. if( sidelobeRejectDb < 60.0 )
  1469. beta = (0.76609 * pow(sidelobeRejectDb - 13.26,0.4)) + (0.09834*(sidelobeRejectDb-13.26));
  1470. else
  1471. beta = 0.12438 * (sidelobeRejectDb + 6.3);
  1472. return (VECT_OP_TYPE)beta;
  1473. }
  1474. VECT_OP_TYPE VECT_OP_FUNC(KaiserFreqResolutionFactor)( double sidelobeRejectDb )
  1475. { return (6.0 * (sidelobeRejectDb + 12.0))/155.0; }
  1476. VECT_OP_TYPE* VECT_OP_FUNC(Kaiser)( VECT_OP_TYPE* dbp, unsigned n, double beta )
  1477. {
  1478. bool zeroFl = false;
  1479. int M = 0;
  1480. double den = cmBessel0(beta); // wnd func denominator
  1481. int cnt = n;
  1482. int i;
  1483. assert( n >= 3 );
  1484. // force ele cnt to be odd
  1485. if( cmIsEvenU(cnt) )
  1486. {
  1487. cnt--;
  1488. zeroFl = true;
  1489. }
  1490. // at this point cnt is odd and >= 3
  1491. // calc half the window length
  1492. M = (int)((cnt - 1.0)/2.0);
  1493. double Msqrd = M*M;
  1494. for(i=0; i<cnt; i++)
  1495. {
  1496. double v0 = (double)(i - M);
  1497. double num = cmBessel0(beta * sqrt(1.0 - ((v0*v0)/Msqrd)));
  1498. dbp[i] = (VECT_OP_TYPE)(num/den);
  1499. }
  1500. if( zeroFl )
  1501. dbp[cnt] = 0.0; // zero the extra element in the output array
  1502. return dbp;
  1503. }
  1504. VECT_OP_TYPE* VECT_OP_FUNC(Gaussian)( VECT_OP_TYPE* dbp, unsigned dn, double mean, double variance )
  1505. {
  1506. int M = dn-1;
  1507. double sqrt2pi = sqrt(2.0*M_PI);
  1508. unsigned i;
  1509. for(i=0; i<dn; i++)
  1510. {
  1511. double arg = ((((double)i/M) - 0.5) * M);
  1512. arg = pow( (double)(arg-mean), 2.0);
  1513. arg = exp( -arg / (2.0*variance));
  1514. dbp[i] = (VECT_OP_TYPE)(arg / (sqrt(variance) * sqrt2pi));
  1515. }
  1516. return dbp;
  1517. }
  1518. VECT_OP_TYPE* VECT_OP_FUNC(Hamming)( VECT_OP_TYPE* dbp, unsigned dn )
  1519. {
  1520. const VECT_OP_TYPE* dep = dbp + dn;
  1521. VECT_OP_TYPE* dp = dbp;
  1522. double fact = 2.0 * M_PI / (dn-1);
  1523. unsigned i;
  1524. for(i=0; dbp < dep; ++i )
  1525. *dbp++ = (VECT_OP_TYPE)(.54 - (.46 * cos(fact*i)));
  1526. return dp;
  1527. }
  1528. VECT_OP_TYPE* VECT_OP_FUNC(Hann)( VECT_OP_TYPE* dbp, unsigned dn )
  1529. {
  1530. const VECT_OP_TYPE* dep = dbp + dn;
  1531. VECT_OP_TYPE* dp = dbp;
  1532. double fact = 2.0 * M_PI / (dn-1);
  1533. unsigned i;
  1534. for(i=0; dbp < dep; ++i )
  1535. *dbp++ = (VECT_OP_TYPE)(.5 - (.5 * cos(fact*i)));
  1536. return dp;
  1537. }
  1538. VECT_OP_TYPE* VECT_OP_FUNC(HannMatlab)( VECT_OP_TYPE* dbp, unsigned dn )
  1539. {
  1540. const VECT_OP_TYPE* dep = dbp + dn;
  1541. VECT_OP_TYPE* dp = dbp;
  1542. double fact = 2.0 * M_PI / (dn+1);
  1543. unsigned i;
  1544. for(i=0; dbp < dep; ++i )
  1545. *dbp++ = (VECT_OP_TYPE)(0.5*(1.0-cos(fact*(i+1))));
  1546. return dp;
  1547. }
  1548. VECT_OP_TYPE* VECT_OP_FUNC(Triangle)( VECT_OP_TYPE* dbp, unsigned dn )
  1549. {
  1550. unsigned n = dn/2;
  1551. VECT_OP_TYPE incr = 1.0/n;
  1552. VECT_OP_FUNC(Seq)(dbp,n,0,incr);
  1553. VECT_OP_FUNC(Seq)(dbp+n,dn-n,1,-incr);
  1554. return dbp;
  1555. }
  1556. VECT_OP_TYPE* VECT_OP_FUNC(GaussWin)( VECT_OP_TYPE* dbp, unsigned dn, double arg )
  1557. {
  1558. const VECT_OP_TYPE* dep = dbp + dn;
  1559. VECT_OP_TYPE* rp = dbp;
  1560. int N = (dep - dbp) - 1;
  1561. int n = -N/2;
  1562. if( N == 0 )
  1563. *dbp = 1.0;
  1564. else
  1565. {
  1566. while( dbp < dep )
  1567. {
  1568. double a = (arg * n++) / (N/2);
  1569. *dbp++ = (VECT_OP_TYPE)exp( -(a*a)/2 );
  1570. }
  1571. }
  1572. return rp;
  1573. }
  1574. VECT_OP_TYPE* VECT_OP_FUNC(Filter)(
  1575. VECT_OP_TYPE* y,
  1576. unsigned yn,
  1577. const VECT_OP_TYPE* x,
  1578. unsigned xn,
  1579. cmReal_t b0,
  1580. const cmReal_t* b,
  1581. const cmReal_t* a,
  1582. cmReal_t* d,
  1583. unsigned dn )
  1584. {
  1585. int i,j;
  1586. VECT_OP_TYPE y0 = 0;
  1587. unsigned n = cmMin( yn, xn );
  1588. // This is a direct form II algorithm based on the MATLAB implmentation
  1589. // http://www.mathworks.com/access/helpdesk/help/techdoc/ref/filter.html#f83-1015962
  1590. for(i=0; i<n; ++i)
  1591. {
  1592. y[i] = (x[i] * b0) + d[0];
  1593. y0 = y[i];
  1594. for(j=0; j<dn; ++j)
  1595. d[j] = (b[j] * x[i]) - (a[j] * y0) + d[j+1];
  1596. }
  1597. // if fewer input samples than output samples - zero the end of the output buffer
  1598. if( yn > xn )
  1599. VECT_OP_FUNC(Fill)(y+i,yn-i,0);
  1600. return cmOkRC;
  1601. }
  1602. VECT_OP_TYPE* VECT_OP_FUNC(FilterFilter)(struct cmFilter_str* f, cmRC_t (*func)( struct cmFilter_str* f, const VECT_OP_TYPE* x, unsigned xn, VECT_OP_TYPE* y, unsigned yn ), const cmReal_t bb[], unsigned bn, const cmReal_t aa[], unsigned an, const VECT_OP_TYPE* x, unsigned xn, VECT_OP_TYPE* y, unsigned yn )
  1603. {
  1604. int i,j;
  1605. int nfilt = cmMax(bn,an);
  1606. int nfact = 3*(nfilt-1);
  1607. const cmReal_t* a = aa;
  1608. const cmReal_t* b = bb;
  1609. cmReal_t* m = NULL;
  1610. cmReal_t* p;
  1611. unsigned zn = (nfilt-1)*(nfilt-1);
  1612. unsigned mn = 2*zn; // space for mtx z0 and z1
  1613. mn += nfilt; // space for zero padded coeff vector
  1614. mn += 2*nfact; // space for begin/end sequences
  1615. if( nfact >= xn )
  1616. {
  1617. return cmOkRC;
  1618. }
  1619. m = cmMemAllocZ( cmReal_t, mn );
  1620. p = m;
  1621. cmReal_t* z0 = p;
  1622. p += zn;
  1623. cmReal_t* z1 = p;
  1624. p += zn;
  1625. cmReal_t* s0 = p;
  1626. p += nfact;
  1627. cmReal_t* s1 = p;
  1628. p += nfact;
  1629. // zero pad the shorter coeff vect
  1630. if( bn < nfilt )
  1631. {
  1632. cmVOR_Copy(p,bn,bb);
  1633. b = p;
  1634. p += nfilt;
  1635. }
  1636. else
  1637. if( an < nfilt )
  1638. {
  1639. cmVOR_Copy(p,an,aa);
  1640. a = p;
  1641. p += nfilt;
  1642. }
  1643. // z0=eye(nfilt-1)
  1644. cmVOR_Identity(z0,nfilt-1,nfilt-1);
  1645. // z1=[eye(nfilt-1,nfilt-2); zeros(1,nfilt-1)];
  1646. cmVOR_Identity(z1,nfilt-1,nfilt-2);
  1647. // z0(:,1) -= a(:)
  1648. for(i=0; i<nfilt-1; ++i)
  1649. z0[i] -= -a[i+1];
  1650. // z0(:,2:end) -= z1;
  1651. for(i=1; i<nfilt-1; ++i)
  1652. for(j=0; j<nfilt-1; ++j)
  1653. z0[ (i*(nfilt-1)) + j ] -= z1[ ((i-1)*(nfilt-1)) + j ];
  1654. // z1 = b - (a * b[0])
  1655. for(i=1; i<nfilt; ++i)
  1656. z1[i-1] = b[i] - (a[i] * b[0]);
  1657. // z1 = z0\z1
  1658. cmVOR_SolveLS(z0,nfilt-1,z1,1);
  1659. // if yn<xn then truncate x.
  1660. xn = cmMin(xn,yn);
  1661. yn = xn;
  1662. // fill in the beginning sequence
  1663. for(i=0; i<nfact; ++i)
  1664. s0[i] = 2*x[0] - x[ nfact-i ];
  1665. // fill in the ending sequence
  1666. for(i=0; i<nfact; ++i)
  1667. s1[i] = 2*x[xn-1] - x[ xn-2-i ];
  1668. cmVOR_MultVVS( z0, nfact, z1, s0[0]);
  1669. unsigned pn = cmMin(1024,xn);
  1670. //acFilter* f = cmFilterAlloc(c,NULL,b,bn,a,an,pn,z0);
  1671. cmFilterInit(f,b,bn,a,an,pn,z0);
  1672. const VECT_OP_TYPE* xx = x;
  1673. for(j=0; j<2; ++j)
  1674. {
  1675. unsigned n = pn;
  1676. // filter begining sequence
  1677. cmFilterExecR(f,s0,nfact,s0,nfact);
  1678. // filter middle sequence
  1679. for(i=0; i<xn; i+=n)
  1680. {
  1681. n = cmMin(pn,xn-i);
  1682. func(f,xx+i,n,y+i,n);
  1683. }
  1684. // filter ending sequence
  1685. cmFilterExecR(f,s1,nfact,s1,nfact);
  1686. // flip all the sequences
  1687. cmVOR_Flip(s0,nfact);
  1688. cmVOR_Flip(s1,nfact);
  1689. VECT_OP_FUNC(Flip)(y,yn);
  1690. if( j==0)
  1691. {
  1692. // swap the begin and end sequences
  1693. cmReal_t* t = s0;
  1694. s0 = s1;
  1695. s1 = t;
  1696. xx = y;
  1697. cmVOR_MultVVS( z0, nfact, z1, s0[0]);
  1698. cmFilterInit(f,b,bn,a,an,pn,z0);
  1699. }
  1700. }
  1701. //cmFilterFree(&f);
  1702. cmMemPtrFree(&m);
  1703. return y;
  1704. }
  1705. VECT_OP_TYPE* VECT_OP_FUNC(LP_Sinc)(VECT_OP_TYPE* dp, unsigned dn, double srate, double fcHz, unsigned flags )
  1706. {
  1707. VECT_OP_TYPE* rp = dp;
  1708. int dM = dn % 2; // dM is used to handle odd length windows
  1709. int M = (dn - dM)/2;
  1710. int Mi = -M;
  1711. double signFact = cmIsFlag(flags, kHighPass_LPSincFl) ? -0.5 : 0.5;
  1712. double phsFact = 2.0 * M_PI * fcHz / srate;
  1713. double sum = 0;
  1714. M += dM;
  1715. //printf("M=%i Mi=%i sign:%f phs:%f\n",M,Mi,signFact,phsFact);
  1716. for(; Mi<M; ++Mi,++dp)
  1717. {
  1718. double phs = phsFact * Mi;
  1719. *dp = Mi == 0 ? 0.5 : signFact * sin(phs)/phs;
  1720. sum += *dp;
  1721. }
  1722. if( cmIsFlag(flags,kNormalize_LPSincFl) )
  1723. VECT_OP_FUNC(DivVS)(rp,dn,sum);
  1724. return rp;
  1725. }
  1726. VECT_OP_TYPE VECT_OP_FUNC(ComplexDetect)(const VECT_OP_TYPE* mag0V, const VECT_OP_TYPE* mag1V, const VECT_OP_TYPE* phs0V, const VECT_OP_TYPE* phs1V, const VECT_OP_TYPE* phs2V, unsigned binCnt )
  1727. {
  1728. double sum = 0;
  1729. const VECT_OP_TYPE* ep = mag0V + binCnt;
  1730. unsigned i = 0;
  1731. for(; mag0V < ep; ++i )
  1732. {
  1733. // calc phase deviation from expected
  1734. double dev_rads = *phs0V++ - (2 * *phs1V++) + *phs2V++;
  1735. // map deviation into range: -pi to pi
  1736. //double dev_rads1 = mod(dev_rads0 + M_PI, -2*M_PI ) + M_PI;
  1737. while( dev_rads > M_PI)
  1738. dev_rads -= 2*M_PI;
  1739. while( dev_rads < -M_PI)
  1740. dev_rads += 2*M_PI;
  1741. // convert into rect coord's
  1742. double m1r = *mag1V++;
  1743. double m0r = *mag0V * cos(dev_rads);
  1744. double m0i = *mag0V++ * sin(dev_rads);
  1745. // calc the combined amplitude and phase deviation
  1746. // sum += hypot( m1 - (m0 * e^(-1*dev_rads)));
  1747. sum += hypot( m1r-m0r, -m0i );
  1748. }
  1749. return (VECT_OP_TYPE)sum;
  1750. }
  1751. VECT_OP_TYPE* VECT_OP_FUNC(MelMask)( VECT_OP_TYPE* maskMtx, unsigned filterCnt, unsigned binCnt, double srate, unsigned flags )
  1752. {
  1753. unsigned fi,bi;
  1754. double mxh = srate/2.0; // nyquist
  1755. double dh = mxh/(binCnt-1) ; // binHz
  1756. double mxm = 1127.0 * log( 1.0 + mxh/700.0); // max mel value in Hz
  1757. double dm = mxm / (filterCnt+1); // avg mel band hz
  1758. double sum = 0;
  1759. for(fi=0; fi<filterCnt; ++fi)
  1760. {
  1761. double m = (fi+1) * dm;
  1762. // calc min/center/max frequencies for this band
  1763. double minHz = 700.0 * (exp((m-dm)/1127.01048)-1.0);
  1764. double ctrHz = 700.0 * (exp( m /1127.01048)-1.0);
  1765. double maxHz = 700.0 * (exp((m+dm)/1127.01048)-1.0);
  1766. // shift the band min/ctr/max to the nearest bin ctr frequency
  1767. if( cmIsFlag(flags,kShiftMelFl) )
  1768. {
  1769. unsigned i;
  1770. i = (unsigned)floor(minHz/dh);
  1771. minHz = minHz - (dh*i) < dh*(i+1) - minHz ? dh*i : dh*(i+1);
  1772. i = (unsigned)floor(ctrHz/dh);
  1773. ctrHz = ctrHz - (dh*i) < dh*(i+1) - ctrHz ? dh*i : dh*(i+1);
  1774. i = (unsigned)floor(maxHz/dh);
  1775. maxHz = maxHz - (dh*i) < dh*(i+1) - maxHz ? dh*i : dh*(i+1);
  1776. }
  1777. // calc the height of the triangle - such that all bands have equal area
  1778. double a = 2.0/(maxHz - minHz);
  1779. for(bi=0; bi<binCnt; ++bi)
  1780. {
  1781. double h = bi*dh;
  1782. unsigned mi = bi*filterCnt + fi;
  1783. if( h < minHz || h > maxHz )
  1784. maskMtx[mi] = 0;
  1785. else
  1786. {
  1787. if( h <= ctrHz )
  1788. maskMtx[mi] = a * (h - minHz)/(ctrHz-minHz);
  1789. else
  1790. maskMtx[mi] = a * (maxHz - h)/(maxHz-ctrHz);
  1791. sum += maskMtx[mi];
  1792. }
  1793. }
  1794. }
  1795. if( cmIsFlag(flags,kNormalizeMelFl) )
  1796. VECT_OP_FUNC(DivVS)( maskMtx, (filterCnt*binCnt), sum );
  1797. return maskMtx;
  1798. }
  1799. unsigned VECT_OP_FUNC(BarkMap)(unsigned* binIdxV, unsigned* cntV, unsigned bandCnt, unsigned binCnt, double srate )
  1800. {
  1801. if( bandCnt == 0 )
  1802. return 0;
  1803. //zwicker & fastl: psychoacoustics 1999, page 159
  1804. double bandUprHz[] = { 100, 200, 300, 400, 510, 630, 770, 920, 1080, 1270, 1480, 1720, 2000, 2320, 2700, 3150, 3700, 4400, 5300, 6400, 7700, 9500, 12000, 15500 };
  1805. unsigned hn = sizeof(bandUprHz)/sizeof(double);
  1806. unsigned i, bi = 0;
  1807. bandCnt = cmMin(hn,bandCnt);
  1808. binIdxV[0] = 0;
  1809. cntV[0] = 1;
  1810. for(i=1; bi < bandCnt && i<binCnt; ++i)
  1811. {
  1812. double hz = srate * i / (2 * (binCnt-1));
  1813. if( hz <= bandUprHz[bi] )
  1814. cntV[bi]++;
  1815. else
  1816. {
  1817. //printf("%i %i %i %f\n",bi,binIdxV[bi],cntV[bi],bandUprHz[bi]);
  1818. ++bi;
  1819. if( bi < bandCnt )
  1820. {
  1821. binIdxV[bi] = i;
  1822. cntV[bi] = 1;
  1823. }
  1824. }
  1825. }
  1826. return bi;
  1827. }
  1828. VECT_OP_TYPE* VECT_OP_FUNC(TriangleMask)(VECT_OP_TYPE* maskMtx, unsigned bandCnt, unsigned binCnt, const VECT_OP_TYPE* ctrHzV, VECT_OP_TYPE binHz, VECT_OP_TYPE stSpread, const VECT_OP_TYPE* lfV, const VECT_OP_TYPE* hfV )
  1829. {
  1830. unsigned i,j;
  1831. VECT_OP_TYPE v0[ bandCnt ];
  1832. VECT_OP_TYPE v1[ bandCnt ];
  1833. // if no lower/upper band limits were give use a fixed semitone band width
  1834. if( lfV==NULL || hfV==NULL)
  1835. {
  1836. for(i=0; i<bandCnt; ++i)
  1837. {
  1838. v0[i] = ctrHzV[i] * pow(2.0,-stSpread/12.0);
  1839. v1[i] = ctrHzV[i] * pow(2.0, stSpread/12.0);
  1840. }
  1841. lfV = v0;
  1842. hfV = v1;
  1843. }
  1844. VECT_OP_FUNC(Zero)(maskMtx,bandCnt*binCnt);
  1845. // for each band
  1846. for(i=0; i<bandCnt; ++i)
  1847. {
  1848. // calc bin index of first possible bin in this band
  1849. // j = (unsigned)floor(lfV[i] / binHz);
  1850. double binHz_j = 0;
  1851. // for each bin whose ctr frq is <= the band upper limit
  1852. for(j=0; j<binCnt; ++j)
  1853. {
  1854. double v;
  1855. // if bin[j] is inside the lower leg of the triangle
  1856. if( lfV[i] <= binHz_j && binHz_j <= ctrHzV[i] )
  1857. v = (binHz_j - lfV[i]) / cmMax(VECT_OP_MIN, ctrHzV[i] - lfV[i] );
  1858. else
  1859. // if bin[j] is inside the upper leg of the triangle
  1860. if( ctrHzV[i] < binHz_j && binHz_j <= hfV[i] )
  1861. v = (hfV[i] - binHz_j) / cmMax(VECT_OP_MIN, hfV[i] - ctrHzV[i] );
  1862. else
  1863. v = 0;
  1864. maskMtx[ (j*bandCnt)+i ] = v;
  1865. binHz_j = binHz * (j+1);
  1866. }
  1867. }
  1868. return maskMtx;
  1869. }
  1870. VECT_OP_TYPE* VECT_OP_FUNC(BarkMask)(VECT_OP_TYPE* maskMtx, unsigned bandCnt, unsigned binCnt, double binHz )
  1871. {
  1872. // -1 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 (23+1)
  1873. VECT_OP_TYPE b[]= {0, 50,150,250,350,450,570,700,840,1000,1170,1370,1600,1850,2150,2500,2900,3400,4000,4800,5800,7000,8500,10500,13500, 15500 };
  1874. bandCnt = cmMin(bandCnt,kDefaultBarkBandCnt);
  1875. VECT_OP_FUNC(TriangleMask)(maskMtx, bandCnt, binCnt, b+1, binHz, 0, b+0, b+2 );
  1876. return maskMtx;
  1877. }
  1878. VECT_OP_TYPE* VECT_OP_FUNC(TerhardtThresholdMask)(VECT_OP_TYPE* maskV, unsigned binCnt, double srate, unsigned flags )
  1879. {
  1880. unsigned i;
  1881. double c0 = cmIsFlag(flags,kModifiedTtmFl) ? 0.6 : 1.0;
  1882. double c1 = cmIsFlag(flags,kModifiedTtmFl) ? 0.5 : 6.5;
  1883. maskV[0]=0;
  1884. for(i=0; i<binCnt; ++i)
  1885. {
  1886. double hz = srate * i / (2 * (binCnt-1));
  1887. maskV[i] = pow(pow(10,(c0 * -3.64* pow(hz/1000,-0.8) + c1 * exp(-0.6 * pow(hz/1000 - 3.3,2)) - 0.001* pow(hz/1000,4))/20),2);
  1888. }
  1889. return maskV;
  1890. }
  1891. VECT_OP_TYPE* VECT_OP_FUNC(ShroederSpreadingFunc)(VECT_OP_TYPE* m, unsigned bandCnt, double srate)
  1892. {
  1893. int fi,bi;
  1894. for(fi=0; fi<bandCnt; ++fi)
  1895. for(bi=0; bi<bandCnt; ++bi )
  1896. m[ fi + (bi*bandCnt) ] = pow(10,(15.81 + 7.5 * ((fi-bi)+0.474)-17.5*pow(1+pow((fi-bi)+0.474,2),0.5))/10);
  1897. return m;
  1898. }
  1899. VECT_OP_TYPE* VECT_OP_FUNC(DctMatrix)( VECT_OP_TYPE* dp, unsigned coeffCnt, unsigned filtCnt )
  1900. {
  1901. VECT_OP_TYPE* dbp = dp;
  1902. double c0 = 1.0/sqrt(filtCnt/2); // row 1-coeffCnt factor
  1903. double c1 = c0 * sqrt(2)/2; // row 0 factor
  1904. unsigned i,j;
  1905. // for each column
  1906. for(i=0; i<filtCnt; ++i)
  1907. // for each row
  1908. for(j=0; j<coeffCnt; ++j)
  1909. *dp++ = (j==0 ? c1 : c0) * cos( (0.5 + i) * M_PI * j / filtCnt);
  1910. return dbp;
  1911. }
  1912. unsigned VECT_OP_FUNC(PeakIndexes)( unsigned* dbp, unsigned dn, const VECT_OP_TYPE* sbp, unsigned sn, VECT_OP_TYPE threshold )
  1913. {
  1914. unsigned pkCnt = 0;
  1915. const unsigned* dep = dbp + dn;
  1916. const VECT_OP_TYPE* sep = sbp + sn;
  1917. const VECT_OP_TYPE* s2p = sbp;
  1918. const VECT_OP_TYPE* s0p = s2p++;
  1919. const VECT_OP_TYPE* s1p = s2p++;
  1920. while( dbp < dep && s2p < sep )
  1921. {
  1922. if( (*s0p < *s1p) && (*s1p > *s2p) && (*s1p >= threshold) )
  1923. {
  1924. *dbp++ = s1p - sbp;
  1925. s0p = s2p++;
  1926. s1p = s2p++;
  1927. ++pkCnt;
  1928. }
  1929. else
  1930. {
  1931. s0p = s1p;
  1932. s1p = s2p++;
  1933. }
  1934. }
  1935. return pkCnt;
  1936. }
  1937. unsigned VECT_OP_FUNC(BinIndex)( const VECT_OP_TYPE* sbp, unsigned sn, VECT_OP_TYPE v )
  1938. {
  1939. const VECT_OP_TYPE* sep = sbp + sn;
  1940. const VECT_OP_TYPE* bp = sbp;
  1941. sep--;
  1942. for(; sbp < sep; ++sbp )
  1943. if( *sbp <= v && v < *(sbp+1) )
  1944. return sbp - bp;
  1945. return cmInvalidIdx;
  1946. }
  1947. unsigned VECT_OP_FUNC(Kmeans)(
  1948. unsigned* classIdxV, // classIdxV[scn] - data point class assignments
  1949. VECT_OP_TYPE* centroidM, // centroidM[srn,K] - cluster centroids
  1950. unsigned K, // count of clusters
  1951. const VECT_OP_TYPE* sM, // sM[srn,scn] source data matrix
  1952. unsigned srn, // dimensionality of each data point
  1953. unsigned scn, // count of data points
  1954. const unsigned* selIdxV, // data subset selection id vector (optional)
  1955. unsigned selKey, // data subset selection key (optional)
  1956. bool initFromCentroidFl,// true if the starting centroids are in centroidM[]
  1957. VECT_OP_TYPE (*distFunc)( void* userPtr, const VECT_OP_TYPE* s0V, const VECT_OP_TYPE* s1V, unsigned sn ),
  1958. void* userDistPtr
  1959. )
  1960. {
  1961. unsigned D = srn; // data dimensionality
  1962. unsigned N = scn; // count of data points to cluster
  1963. unsigned iterCnt = 0;
  1964. unsigned ki;
  1965. unsigned i = 0;
  1966. unsigned selN = N;
  1967. // if a data point selection vector was given
  1968. if( selIdxV != NULL )
  1969. {
  1970. selN = 0;
  1971. for(i=0; i<N; ++i)
  1972. {
  1973. selN += selIdxV[i]==selKey;
  1974. classIdxV[i] = K;
  1975. }
  1976. }
  1977. assert(K<=selN);
  1978. // if the numer of datapoints and the number of clusters is the same
  1979. // make the datapoints the centroids and return
  1980. if( K == selN )
  1981. {
  1982. ki = 0;
  1983. for(i=0; i<N; ++i)
  1984. if( selIdxV==NULL || selIdxV[i]==selKey )
  1985. {
  1986. VECT_OP_FUNC(Copy)(centroidM+(ki*D),D,sM+(i*D));
  1987. classIdxV[ki] = ki;
  1988. ++ki;
  1989. }
  1990. return 0;
  1991. }
  1992. // if centroidM[] has not been initialized with the starting centroid vectors.
  1993. if( initFromCentroidFl == false )
  1994. {
  1995. unsigned* kiV = cmMemAlloc( unsigned, N );
  1996. // select K unique datapoints at random as the initial centroids
  1997. cmVOU_RandomSeq(kiV,N);
  1998. for(i=0,ki=0; i<N && ki<K; ++i)
  1999. {
  2000. if( selIdxV==NULL || selIdxV[ kiV[i] ]==selKey )
  2001. {
  2002. VECT_OP_FUNC(Copy)( centroidM + (ki*D), D, sM + (kiV[i]*D) );
  2003. ++ki;
  2004. }
  2005. }
  2006. cmMemPtrFree(&kiV);
  2007. }
  2008. unsigned* nV = cmMemAllocZ( unsigned,K);
  2009. while(1)
  2010. {
  2011. unsigned changeCnt = 0;
  2012. cmVOU_Zero(nV,K);
  2013. // for each data point - assign data point to a cluster
  2014. for(i=0; i<N; ++i)
  2015. if( selIdxV==NULL || selIdxV[i] == selKey )
  2016. {
  2017. // set ki with the index of the centroid closest to sM[:,i]
  2018. VECT_OP_FUNC(DistVMM)( NULL, NULL, &ki, D, sM + (i*srn), 1, centroidM, K, distFunc, userDistPtr );
  2019. assert(ki<K);
  2020. nV[ki]++;
  2021. changeCnt += ( ki != classIdxV[i] );
  2022. classIdxV[i] = ki;
  2023. }
  2024. // if no data points change classes then the centroids have converged
  2025. if( changeCnt == 0 )
  2026. break;
  2027. ++iterCnt;
  2028. // zero the centroid matrix
  2029. VECT_OP_FUNC(Fill)(centroidM, D*K, 0 );
  2030. // update the centroids
  2031. for(ki=0; ki<K; ++ki)
  2032. {
  2033. unsigned n = 0;
  2034. // sum the all datapoints belonging to class ki
  2035. for(i=0; i<N; ++i)
  2036. if( classIdxV[i] == ki )
  2037. {
  2038. VECT_OP_FUNC(AddVV)(centroidM + (ki*D), D, sM + (i*srn) );
  2039. ++n;
  2040. }
  2041. // convert the sum to a mean to form the centroid
  2042. if( n > 0 )
  2043. VECT_OP_FUNC(DivVS)(centroidM + (ki*D), D, n );
  2044. }
  2045. }
  2046. cmVOU_PrintL("class cnt:",NULL,1,K,nV);
  2047. cmMemPtrFree(&nV);
  2048. return iterCnt;
  2049. }
  2050. unsigned VECT_OP_FUNC(Kmeans2)(
  2051. unsigned* classIdxV, // classIdxV[scn] - data point class assignments
  2052. VECT_OP_TYPE* centroidM, // centroidM[srn,K] - cluster centroids
  2053. unsigned K, // count of clusters
  2054. const VECT_OP_TYPE* (*srcFunc)(void* userPtr, unsigned frmIdx ),
  2055. unsigned srn, // dimensionality of each data point
  2056. unsigned scn, // count of data points
  2057. void* userSrcPtr, // callback data for srcFunc
  2058. VECT_OP_TYPE (*distFunc)( void* userPtr, const VECT_OP_TYPE* s0V, const VECT_OP_TYPE* s1V, unsigned sn ),
  2059. void* distUserPtr,
  2060. int maxIterCnt,
  2061. int deltaStopCnt
  2062. )
  2063. {
  2064. unsigned D = srn; // data dimensionality
  2065. unsigned N = scn; // count of data points to cluster
  2066. unsigned iterCnt = 0;
  2067. unsigned ki;
  2068. unsigned i = 0;
  2069. const VECT_OP_TYPE* sp;
  2070. assert(K<N);
  2071. deltaStopCnt = cmMax(0,deltaStopCnt);
  2072. // nV[K] - class assignment vector
  2073. unsigned* nV = cmMemAllocZ( unsigned,2*K);
  2074. // roV[K] - read-only flag centroid
  2075. // centroids flagged as read-only will not be updated by the clustering routine
  2076. unsigned* roV = nV + K;
  2077. // copy the read-only flags into roV[K]
  2078. for(i=0; i<K; ++i)
  2079. roV[i] = classIdxV[i];
  2080. while(1)
  2081. {
  2082. unsigned changeCnt = 0;
  2083. cmVOU_Zero(nV,K);
  2084. // for each data point - assign data point to a cluster
  2085. for(i=0; i<N; ++i)
  2086. if((sp = srcFunc(userSrcPtr,i)) != NULL)
  2087. {
  2088. // set ki with the index of the centroid closest to sM[:,i]
  2089. VECT_OP_FUNC(DistVMM)( NULL, NULL, &ki, D, sp, 1, centroidM, K, distFunc, distUserPtr );
  2090. assert(ki<K);
  2091. // track the number of data points assigned to each centroid
  2092. nV[ki]++;
  2093. // track the number of data points which change classes
  2094. changeCnt += ( ki != classIdxV[i] );
  2095. // update the class that this data point belongs to
  2096. classIdxV[i] = ki;
  2097. }
  2098. // if the count of data points which changed classes is less than deltaStopCnt
  2099. // then the centroids have converged
  2100. if( changeCnt <= deltaStopCnt )
  2101. break;
  2102. if( maxIterCnt!=-1 && iterCnt>=maxIterCnt )
  2103. break;
  2104. // track the number of interations required to converge
  2105. ++iterCnt;
  2106. fprintf(stderr,"%i:%i (", iterCnt,changeCnt );
  2107. for(i=0; i<K; ++i)
  2108. fprintf(stderr,"%i ",nV[i]);
  2109. fprintf(stderr,") ");
  2110. fflush(stderr);
  2111. // update the centroids
  2112. for(ki=0; ki<K; ++ki)
  2113. if( roV[ki]==0 )
  2114. {
  2115. unsigned n = 0;
  2116. VECT_OP_FUNC(Zero)(centroidM + (ki*D), D );
  2117. // sum the all datapoints belonging to class ki
  2118. for(i=0; i<N; ++i)
  2119. if( classIdxV[i] == ki && ((sp=srcFunc(userSrcPtr,i))!=NULL))
  2120. {
  2121. VECT_OP_FUNC(AddVV)(centroidM + (ki*D), D, sp );
  2122. ++n;
  2123. }
  2124. // convert the sum to a mean to form the centroid
  2125. if( n > 0 )
  2126. VECT_OP_FUNC(DivVS)(centroidM + (ki*D), D, n );
  2127. }
  2128. }
  2129. cmMemPtrFree(&nV);
  2130. return iterCnt;
  2131. }
  2132. VECT_OP_TYPE* VECT_OP_FUNC(GaussPDF)( VECT_OP_TYPE* dbp, unsigned dn, const VECT_OP_TYPE* sbp, VECT_OP_TYPE mean, VECT_OP_TYPE stdDev )
  2133. {
  2134. VECT_OP_TYPE* rp = dbp;
  2135. const VECT_OP_TYPE* dep = dbp + dn;
  2136. VECT_OP_TYPE var = stdDev * stdDev;
  2137. VECT_OP_TYPE fact0 = 1.0/sqrt(2*M_PI*var);
  2138. VECT_OP_TYPE fact1 = 2.0 * var;
  2139. for(; dbp < dep; ++sbp )
  2140. *dbp++ = fact0 * exp( -((*sbp-mean)*(*sbp-mean))/ fact1 );
  2141. return rp;
  2142. }
  2143. /// Evaluate a multivariate normal distribution defined by meanV[D] and covarM[D,D]
  2144. /// at the data points held in the columns of xM[D,N]. Return the evaluation
  2145. /// results in the vector yV[N].
  2146. bool VECT_OP_FUNC(MultVarGaussPDF)( VECT_OP_TYPE* yV, const VECT_OP_TYPE* xM, const VECT_OP_TYPE* meanV, const VECT_OP_TYPE* covarM, unsigned D, unsigned N, bool diagFl )
  2147. {
  2148. VECT_OP_TYPE det0;
  2149. // calc the determinant of the covariance matrix
  2150. if( diagFl )
  2151. // kpl 1/16/11 det0 = VECT_OP_FUNC(LogDetDiagM)(covarM,D);
  2152. det0 = VECT_OP_FUNC(DetDiagM)(covarM,D);
  2153. else
  2154. // kpl 1/16/11 det0 = VECT_OP_FUNC(LogDetM)(covarM,D);
  2155. det0 = VECT_OP_FUNC(DetM)(covarM,D);
  2156. assert(det0 != 0 );
  2157. if( det0 == 0 )
  2158. return false;
  2159. // calc the inverse of the covariance matrix
  2160. VECT_OP_TYPE icM[D*D];
  2161. VECT_OP_FUNC(Copy)(icM,D*D,covarM);
  2162. VECT_OP_TYPE* r;
  2163. if( diagFl )
  2164. r = VECT_OP_FUNC(InvDiagM)(icM,D);
  2165. else
  2166. r = VECT_OP_FUNC(InvM)(icM,D);
  2167. if( r == NULL )
  2168. return false;
  2169. VECT_OP_FUNC(MultVarGaussPDF2)( yV, xM, meanV, icM, det0, D, N, diagFl );
  2170. return true;
  2171. }
  2172. VECT_OP_TYPE* VECT_OP_FUNC(MultVarGaussPDF2)( VECT_OP_TYPE* yV, const VECT_OP_TYPE* xM, const VECT_OP_TYPE* meanV, const VECT_OP_TYPE* icM, VECT_OP_TYPE logDet, unsigned D, unsigned N, bool diagFl )
  2173. {
  2174. unsigned i;
  2175. double fact = (-(cmReal_t)D/2) * log(2.0*M_PI) - 0.5*logDet;
  2176. for(i=0; i<N; ++i)
  2177. {
  2178. VECT_OP_TYPE dx[D];
  2179. VECT_OP_TYPE t[D];
  2180. // dx[] difference between mean and ith data point
  2181. VECT_OP_FUNC(SubVVV)(dx,D, xM + (i*D), meanV);
  2182. // t[] = dx[] * inv(covarM);
  2183. if( diagFl )
  2184. VECT_OP_FUNC(MultDiagVMV)(t,D,icM,D,dx);
  2185. else
  2186. VECT_OP_FUNC(MultVMV)(t,D,icM,D,dx);
  2187. // dist = sum(dx[] * t[])
  2188. cmReal_t dist = VECT_OP_FUNC(MultSumVV)(t,dx,D);
  2189. yV[i] = exp( fact - (0.5*dist) );
  2190. }
  2191. return yV;
  2192. }
  2193. VECT_OP_TYPE* VECT_OP_FUNC(MultVarGaussPDF3)(
  2194. VECT_OP_TYPE* yV,
  2195. const VECT_OP_TYPE* (*srcFunc)(void* funcDataPtr, unsigned frmIdx ),
  2196. void* funcDataPtr,
  2197. const VECT_OP_TYPE* meanV,
  2198. const VECT_OP_TYPE* icM,
  2199. VECT_OP_TYPE logDet,
  2200. unsigned D,
  2201. unsigned N,
  2202. bool diagFl )
  2203. {
  2204. unsigned i;
  2205. double fact = (-(cmReal_t)D/2) * log(2.0*M_PI) - 0.5*logDet;
  2206. for(i=0; i<N; ++i)
  2207. {
  2208. VECT_OP_TYPE dx[D];
  2209. VECT_OP_TYPE t[D];
  2210. const VECT_OP_TYPE* xV = srcFunc( funcDataPtr, i );
  2211. if( xV == NULL )
  2212. yV[i] = 0;
  2213. else
  2214. {
  2215. // dx[] difference between mean and ith data point
  2216. VECT_OP_FUNC(SubVVV)(dx, D, xV, meanV);
  2217. // t[] = dx[] * inv(covarM);
  2218. if( diagFl )
  2219. VECT_OP_FUNC(MultDiagVMV)(t,D,icM,D,dx);
  2220. else
  2221. VECT_OP_FUNC(MultVMV)(t,D,icM,D,dx);
  2222. // dist = sum(dx[] * t[])
  2223. cmReal_t dist = VECT_OP_FUNC(MultSumVV)(t,dx,D);
  2224. yV[i] = exp( fact - (0.5*dist) );
  2225. }
  2226. }
  2227. return yV;
  2228. }
  2229. /// stateV[timeN]
  2230. /// a[stateN,stateN],
  2231. /// b[stateN,timeN]
  2232. /// phi[stateN].
  2233. void VECT_OP_FUNC(DiscreteViterbi)(unsigned* stateV, unsigned tN, unsigned sN, const VECT_OP_TYPE* phi, const VECT_OP_TYPE* a, const VECT_OP_TYPE* b )
  2234. {
  2235. unsigned* psiM = cmMemAlloc( unsigned, sN*tN ); // psi[sN,tN]
  2236. VECT_OP_TYPE* dV = cmMemAlloc( VECT_OP_TYPE, 2*sN );
  2237. VECT_OP_TYPE* d0V = dV;
  2238. VECT_OP_TYPE* d1V = dV + sN;
  2239. int t,i,j;
  2240. // calc the prob of starting in each state given the observations
  2241. VECT_OP_FUNC(MultVVV)( d0V, sN, phi, b );
  2242. VECT_OP_FUNC(NormalizeProbability)( d0V, sN ); // scale to prevent underflow
  2243. // for each time step
  2244. for(t=1; t<tN; ++t)
  2245. {
  2246. // for each possible next state
  2247. for(j=0; j<sN; ++j)
  2248. {
  2249. VECT_OP_TYPE mv = 0;
  2250. unsigned mi = 0;
  2251. // The following loop could be replaced with these vector op's:
  2252. // VECT_OP_TYPE tV[ sN ];
  2253. // VECT_OP_TYPE(MultVVV)(tV,sN,d0V,a + (j*sN));
  2254. // mi = VECT_OP_TYPE(MaxIndex)(tV,sN);
  2255. // mv = tV[mi];
  2256. // for each possible prev state
  2257. for(i=0; i<sN; ++i)
  2258. {
  2259. // calc prob of having ended in state i and transitioning to state j
  2260. VECT_OP_TYPE v = d0V[i] * a[ i + (j*sN) ];
  2261. // track the most likely transition ending in state j
  2262. if( v > mv )
  2263. {
  2264. mv = v;
  2265. mi = i;
  2266. }
  2267. }
  2268. // scale the prob of the most likely state by the prob of the obs given that state
  2269. d1V[j] = mv * b[ (t*sN) + j ];
  2270. // store the most likely previous state given that the current state is j
  2271. // (this is the key to understanding the backtracking step below)
  2272. psiM[ (t*sN) + j ] = mi;
  2273. }
  2274. VECT_OP_FUNC(NormalizeProbability)( d1V, sN ); // scale to prevent underflow
  2275. // swap d0V and d1V
  2276. VECT_OP_TYPE* tmp = d0V;
  2277. d0V = d1V;
  2278. d1V = tmp;
  2279. }
  2280. // store the most likely ending state
  2281. stateV[tN-1] = VECT_OP_FUNC(MaxIndex)( d0V, sN, 1 );
  2282. // given the most likely next step select the most likely previous step
  2283. for(t=tN-2; t>=0; --t)
  2284. stateV[t] = psiM[ ((t+1)*sN) + stateV[t+1] ];
  2285. cmMemPtrFree( &psiM );
  2286. cmMemPtrFree( &dV );
  2287. }
  2288. bool VECT_OP_FUNC(ClipLine2)( VECT_OP_TYPE x0, VECT_OP_TYPE y0, VECT_OP_TYPE x1, VECT_OP_TYPE y1, VECT_OP_TYPE xMin, VECT_OP_TYPE yMin, VECT_OP_TYPE xMax, VECT_OP_TYPE yMax, VECT_OP_TYPE* t0, VECT_OP_TYPE* t1 )
  2289. {
  2290. VECT_OP_TYPE dx = x1 - x0;
  2291. VECT_OP_TYPE dy = y1 - y0;
  2292. VECT_OP_TYPE p=0,q=0,r=0;
  2293. *t0 = 0.0;
  2294. *t1 = 1.0;
  2295. unsigned i;
  2296. for(i=0; i<4; ++i)
  2297. {
  2298. switch(i)
  2299. {
  2300. case 0: p=-dx; q=-(xMin - x0); break; // left
  2301. case 1: p= dx; q= (xMax - x0); break; // right
  2302. case 2: p=-dy; q=-(yMin - y0); break; // bottom
  2303. case 3: p= dy; q= (yMax - y0); break; // top
  2304. }
  2305. // if parallel to edge i
  2306. if( p == 0 )
  2307. {
  2308. // if entirely outside of window
  2309. if( q < 0 )
  2310. return false;
  2311. continue;
  2312. }
  2313. r = p/q;
  2314. // if travelling right/up
  2315. if( p < 0 )
  2316. {
  2317. // travelling away from x1,y1
  2318. if( r > *t1 )
  2319. return false;
  2320. // update distance on line to point of intersection
  2321. if( r > *t0 )
  2322. *t0 = r;
  2323. }
  2324. else // if travelling left/down
  2325. {
  2326. // travelling away from x1,y1
  2327. if( r < *t0 )
  2328. return false;
  2329. // update distance on line to point of intersection
  2330. if( r < *t1 )
  2331. *t1 = r;
  2332. }
  2333. }
  2334. return true;
  2335. }
  2336. /// (Uses the Laing-Barsky clipping algorithm)
  2337. /// From: http://www.skytopia.com/project/articles/compsci/clipping.html
  2338. bool VECT_OP_FUNC(ClipLine)( VECT_OP_TYPE* x0, VECT_OP_TYPE* y0, VECT_OP_TYPE* x1, VECT_OP_TYPE* y1, VECT_OP_TYPE xMin, VECT_OP_TYPE yMin, VECT_OP_TYPE xMax, VECT_OP_TYPE yMax )
  2339. {
  2340. VECT_OP_TYPE t0;
  2341. VECT_OP_TYPE t1;
  2342. if( VECT_OP_FUNC(ClipLine2)(*x0,*y0,*x1,*y1,xMin,yMin,xMax,yMax,&t0,&t1) )
  2343. {
  2344. VECT_OP_TYPE dx = *x1 - *x0;
  2345. VECT_OP_TYPE dy = *y1 - *y0;
  2346. *x0 = *x0 + t0*dx;
  2347. *x1 = *x0 + t1*dx;
  2348. *y0 = *y0 + t0*dy;
  2349. *y1 = *y0 + t1*dy;
  2350. return true;
  2351. }
  2352. return false;
  2353. }
  2354. bool VECT_OP_FUNC(IsLineInRect)( VECT_OP_TYPE x0, VECT_OP_TYPE y0, VECT_OP_TYPE x1, VECT_OP_TYPE y1, VECT_OP_TYPE xMin, VECT_OP_TYPE yMin, VECT_OP_TYPE xMax, VECT_OP_TYPE yMax )
  2355. {
  2356. VECT_OP_TYPE t0;
  2357. VECT_OP_TYPE t1;
  2358. return VECT_OP_FUNC(ClipLine2)(x0,y0,x1,y1,xMin,yMin,xMax,yMax,&t0,&t1);
  2359. }
  2360. VECT_OP_TYPE VECT_OP_FUNC(PtToLineDistance)( VECT_OP_TYPE x0, VECT_OP_TYPE y0, VECT_OP_TYPE x1, VECT_OP_TYPE y1, VECT_OP_TYPE px, VECT_OP_TYPE py)
  2361. {
  2362. // from:http://en.wikipedia.org/wiki/Distance_from_a_point_to_a_line
  2363. double normalLength = sqrt((x1 - x0) * (x1 - x0) + (y1 - y0) * (y1 - y0));
  2364. if( normalLength <= 0 )
  2365. return 0;
  2366. return (VECT_OP_TYPE)fabs((px - x0) * (y1 - y0) - (py - y0) * (x1 - x0)) / normalLength;
  2367. }
  2368. void VECT_OP_FUNC(Lsq1)(const VECT_OP_TYPE* x, const VECT_OP_TYPE* y, unsigned n, VECT_OP_TYPE* b0, VECT_OP_TYPE* b1 )
  2369. {
  2370. VECT_OP_TYPE sx = 0;
  2371. VECT_OP_TYPE sy = 0;
  2372. VECT_OP_TYPE sx_2 = 0;
  2373. VECT_OP_TYPE sxy = 0;
  2374. unsigned i;
  2375. if( x == NULL )
  2376. {
  2377. for(i=0; i<n; ++i)
  2378. {
  2379. VECT_OP_TYPE xx = i;
  2380. sx += xx;
  2381. sx_2 += xx * xx;
  2382. sxy += xx * y[i];
  2383. sy += y[i];
  2384. }
  2385. }
  2386. else
  2387. {
  2388. for(i=0; i<n; ++i)
  2389. {
  2390. sx += x[i];
  2391. sx_2 += x[i] * x[i];
  2392. sxy += x[i] * y[i];
  2393. sy += y[i];
  2394. }
  2395. }
  2396. *b1 = (sxy * n - sx * sy) / (sx_2 * n - sx*sx);
  2397. *b0 = (sy - (*b1) * sx) / n;
  2398. }
  2399. #endif