//https://github.com/xianyi/OpenBLAS/issues/695
CTEST(potrf, bug_695){
openblas_complex_float A1[100] = {5.8525753+0.0*I, -0.79540455-0.7066077*I, 0.98274714-1.3824869*I, 2.619998-1.8532984*I, -1.8306153+1.2336911*I, 0.32275113-0.015575029*I, 2.1968813-1.0640624*I, 0.27894387-0.97911835*I, 3.0476584-0.18548489*I, 0.3842994-0.7050991*I,
-0.79540455+0.7066077*I, 8.313246+0.0*I, -1.8076122+0.8882447*I, 0.47806996-0.48494184*I, 0.5096429+0.5395974*I, -0.7285097+0.10360408*I, -1.1760061+2.7146957*I, -0.4271084-0.042899966*I, -1.7228563-2.8335886*I, 1.8942566-0.6389735*I,
0.98274714+1.3824869*I, -1.8076122-0.8882447*I, 9.367975+0.0*I, -0.1838578-0.6468568*I, -1.8338387-0.7064959*I, 0.041852742+0.6556877*I, 2.5673025-1.9732997*I, -1.1148382+0.15693812*I, 2.4704504+1.0389464*I, 1.0858271+1.298006*I,
2.619998+1.8532984*I, 0.47806996+0.48494184*I, -0.1838578+0.6468568*I, 3.1117508+0.0*I, -1.956626-0.22825956*I, 0.07081801+0.31801307*I, 0.3698375+0.5400855*I, 0.80686307-1.5315914*I, 1.5649154+1.6229297*I, -0.112077385-1.2014246*I,
-1.8306153-1.2336911*I, 0.5096429-0.5395974*I, -1.8338387+0.7064959*I, -1.956626+0.22825956*I, 3.6439795+0.0*I, -0.2594722-0.48786148*I, -0.47636223+0.27821827*I, -0.61608654+2.01858*I, -2.7767487-1.7693765*I, 0.048102796+0.9741874*I,
0.32275113+0.015575029*I, -0.7285097-0.10360408*I, 0.041852742-0.6556877*I, 0.07081801-0.31801307*I, -0.2594722+0.48786148*I, 3.624376+0.0*I, -1.6697118-0.4017511*I, -1.4397877+0.7550918*I, -0.31456697+1.0403451*I, -0.31978557-0.13701046*I,
2.1968813+1.0640624*I, -1.1760061-2.7146957*I, 2.5673025+1.9732997*I, 0.3698375-0.5400855*I, -0.47636223-0.27821827*I, -1.6697118+0.4017511*I, 6.8273163+0.0*I, -0.10051322-0.24303961*I, 1.4415971-0.29750675*I, 1.221786+0.85654986*I,
0.27894387+0.97911835*I, -0.4271084+0.042899966*I, -1.1148382-0.15693812*I, 0.80686307+1.5315914*I, -0.61608654-2.01858*I, -1.4397877-0.7550918*I, -0.10051322+0.24303961*I, 3.4057708+0.0*I, -0.5856801+1.0203559*I, 0.7103452-0.8422135*I,
3.0476584+0.18548489*I, -1.7228563+2.8335886*I, 2.4704504-1.0389464*I, 1.5649154-1.6229297*I, -2.7767487+1.7693765*I, -0.31456697-1.0403451*I, 1.4415971+0.29750675*I, -0.5856801-1.0203559*I, 7.005772+0.0*I, -0.9617417+1.2486815*I,
0.3842994+0.7050991*I, 1.8942566+0.6389735*I, 1.0858271-1.298006*I, -0.112077385+1.2014246*I, 0.048102796-0.9741874*I, -0.31978557+0.13701046*I, 1.221786-0.85654986*I, 0.7103452+0.8422135*I, -0.9617417-1.2486815*I, 3.4629636+0.0*I};
char up = 'U';
blasint n=10;
blasint info[1];
BLASFUNC(cpotrf)(&up, &n, (float*)(A1), &n, info);
//printf("%g+%g*I\n", creal(A1[91]), cimag(A1[91]));
openblas_complex_double A2[100] = {3.0607147216796875+0.0*I, -0.5905849933624268-0.29020825028419495*I, 0.321084201335907+0.45168760418891907*I, 0.8387917876243591-0.644718587398529*I, -0.3642411530017853+0.051274992525577545*I, 0.8071482181549072+0.33944568037986755*I, 0.013674172572791576+0.21422699093818665*I, 0.35476258397102356+0.42408594489097595*I, -0.5991537570953369-0.23082709312438965*I, -0.0600702166557312-0.2113417387008667*I,
-0.7954045534133911+0.7066076993942261*I, 2.807175397872925+0.0*I, -0.1691000759601593+0.313548743724823*I, -0.30911174416542053+0.7447023987770081*I, -0.22347848117351532+0.03316075727343559*I, -0.4088296890258789-1.0214389562606812*I, -0.2344931811094284+0.08056317269802094*I, 0.793269693851471-0.17507623136043549*I, 0.03163455054163933+0.20559945702552795*I, 0.13581633567810059-0.2110036462545395*I,
0.9827471375465393+1.3824869394302368*I, -1.8076121807098389-0.8882446885108948*I, 2.3277781009674072+0.0*I, 0.830405056476593-0.19296252727508545*I, 0.1394239068031311-0.5260677933692932*I, 1.239942193031311-0.09915469586849213*I, 0.06731037050485611-0.059320636093616486*I, 0.11507681757211685-0.1984301060438156*I, -0.6843825578689575+0.4647614359855652*I, 1.213119387626648-0.7757048010826111*I,
2.619997978210449+1.8532984256744385*I, 0.4780699610710144+0.48494184017181396*I, -0.18385779857635498+0.6468567848205566*I, 2.0811400413513184+0.0*I, -0.035075582563877106+0.09732913225889206*I, 0.27337002754211426-0.9032229781150818*I, -0.8374675512313843+0.0479498989880085*I, 0.6916252374649048+0.45711082220077515*I, 0.1883818507194519+0.06482727080583572*I, -0.32384994626045227+0.05857187137007713*I,
-1.8306152820587158-1.2336910963058472*I, 0.5096428990364075-0.5395973920822144*I, -1.833838701248169+0.7064958810806274*I, -1.956626057624817+0.22825956344604492*I, 1.706615924835205+0.0*I, -0.2895336151123047+0.17579378187656403*I, -0.923172116279602-0.4530014097690582*I, 0.5040621757507324-0.37026339769363403*I, -0.2824432849884033-1.0374568700790405*I, 0.1399831622838974+0.4977008104324341*I,
0.32275113463401794+0.015575028955936432*I, -0.7285097241401672-0.10360407829284668*I, 0.041852742433547974-0.655687689781189*I, 0.07081800699234009-0.318013072013855*I, -0.25947219133377075+0.4878614842891693*I, 1.5735365152359009+0.0*I, -0.2647853195667267-0.26654252409935*I, -0.6190430521965027-0.24699924886226654*I, -0.6288471221923828+0.48154571652412415*I, 0.02446540631353855-0.2611822783946991*I,
2.1968812942504883+1.0640623569488525*I, -1.1760060787200928-2.714695692062378*I, 2.5673024654388428+1.9732997417449951*I, 0.3698374927043915-0.54008549451828*I, -0.4763622283935547-0.27821826934814453*I, -1.6697118282318115+0.4017511010169983*I, 1.2674795389175415+0.0*I, 0.3079095482826233-0.07258892804384232*I, -0.5929520130157471-0.038360968232154846*I, 0.04388086497783661-0.025549031794071198*I,
0.27894386649131775+0.9791183471679688*I, -0.42710840702056885+0.0428999662399292*I, -1.1148382425308228-0.1569381207227707*I, 0.8068630695343018+1.5315914154052734*I, -0.6160865426063538-2.0185799598693848*I, -1.439787745475769-0.7550917863845825*I, -0.10051321983337402+0.24303960800170898*I, 0.9066106081008911+0.0*I, 0.05315789580345154-0.06136537343263626*I, -0.21304509043693542+0.6494344472885132*I,
3.0476584434509277+0.1854848861694336*I, -1.7228562831878662+2.8335886001586914*I, 2.4704504013061523-1.0389463901519775*I, 1.564915418624878-1.6229296922683716*I, -2.7767486572265625+1.769376516342163*I, -0.314566969871521-1.0403450727462769*I, 1.4415971040725708+0.29750674962997437*I, -0.5856801271438599-1.0203559398651123*I, 0.5668219923973083+0.0*I, 0.033351436257362366-0.07832501083612442*I,
0.3842993974685669+0.7050991058349609*I, 1.894256591796875+0.6389734745025635*I, 1.085827112197876-1.2980060577392578*I, -0.11207738518714905+1.2014245986938477*I, 0.04810279607772827-0.9741873741149902*I, -0.31978556513786316+0.13701045513153076*I, 1.2217860221862793-0.856549859046936*I, 0.7103452086448669+0.84221351146698*I, -0.9617416858673096-1.2486815452575684*I, 0.0756804421544075+0.0*I};
openblas_complex_double B[20] = {-0.21782716937787788-0.9222220085490986*I, -0.7620356655676837+0.15533508334193666*I, -0.905011814118756+0.2847570854574069*I, -0.3451346708401685+1.076948486041297*I, 0.25336108035924787+0.975317836492159*I, 0.11192755545114-0.1603741874112385*I, -0.20604111555491242+0.10570814584017311*I, -1.0568488936791578-0.06025820467086475*I, -0.6650468984506477-0.5000967284800251*I, -1.0509472322215125+0.5022165705328413*I,
-0.727775859267237+0.50638268521728*I, 0.39947219167701153-0.4576746001199889*I, -0.7122162951294634-0.630289556702497*I, 0.9870834574024372-0.2825689605519449*I, 0.0628393808469436-0.1253397353973715*I, 0.8439562576196216+1.0850814110398734*I, 0.562377322638969-0.2578030745663871*I, 0.12696236014017806-0.09853584666755086*I, -0.023682508769195098+0.18093440285319276*I, -0.7264975746431271+0.31670415674097235*I};
char lo = 'L';
blasint nrhs = 2;
BLASFUNC(zpotrs)(&lo, &n, &nrhs, (double*)(A2), &n, (double*)(B), &n, info);
// note that this is exactly equal to A1
openblas_complex_float A3[100] = {5.8525753+0.0*I, -0.79540455-0.7066077*I, 0.98274714-1.3824869*I, 2.619998-1.8532984*I, -1.8306153+1.2336911*I, 0.32275113-0.015575029*I, 2.1968813-1.0640624*I, 0.27894387-0.97911835*I, 3.0476584-0.18548489*I, 0.3842994-0.7050991*I,
-0.79540455+0.7066077*I, 8.313246+0.0*I, -1.8076122+0.8882447*I, 0.47806996-0.48494184*I, 0.5096429+0.5395974*I, -0.7285097+0.10360408*I, -1.1760061+2.7146957*I, -0.4271084-0.042899966*I, -1.7228563-2.8335886*I, 1.8942566-0.6389735*I,
0.98274714+1.3824869*I, -1.8076122-0.8882447*I, 9.367975+0.0*I, -0.1838578-0.6468568*I, -1.8338387-0.7064959*I, 0.041852742+0.6556877*I, 2.5673025-1.9732997*I, -1.1148382+0.15693812*I, 2.4704504+1.0389464*I, 1.0858271+1.298006*I,
2.619998+1.8532984*I, 0.47806996+0.48494184*I, -0.1838578+0.6468568*I, 3.1117508+0.0*I, -1.956626-0.22825956*I, 0.07081801+0.31801307*I, 0.3698375+0.5400855*I, 0.80686307-1.5315914*I, 1.5649154+1.6229297*I, -0.112077385-1.2014246*I,
-1.8306153-1.2336911*I, 0.5096429-0.5395974*I, -1.8338387+0.7064959*I, -1.956626+0.22825956*I, 3.6439795+0.0*I, -0.2594722-0.48786148*I, -0.47636223+0.27821827*I, -0.61608654+2.01858*I, -2.7767487-1.7693765*I, 0.048102796+0.9741874*I,
0.32275113+0.015575029*I, -0.7285097-0.10360408*I, 0.041852742-0.6556877*I, 0.07081801-0.31801307*I, -0.2594722+0.48786148*I, 3.624376+0.0*I, -1.6697118-0.4017511*I, -1.4397877+0.7550918*I, -0.31456697+1.0403451*I, -0.31978557-0.13701046*I,
2.1968813+1.0640624*I, -1.1760061-2.7146957*I, 2.5673025+1.9732997*I, 0.3698375-0.5400855*I, -0.47636223-0.27821827*I, -1.6697118+0.4017511*I, 6.8273163+0.0*I, -0.10051322-0.24303961*I, 1.4415971-0.29750675*I, 1.221786+0.85654986*I,
0.27894387+0.97911835*I, -0.4271084+0.042899966*I, -1.1148382-0.15693812*I, 0.80686307+1.5315914*I, -0.61608654-2.01858*I, -1.4397877-0.7550918*I, -0.10051322+0.24303961*I, 3.4057708+0.0*I, -0.5856801+1.0203559*I, 0.7103452-0.8422135*I,
3.0476584+0.18548489*I, -1.7228563+2.8335886*I, 2.4704504-1.0389464*I, 1.5649154-1.6229297*I, -2.7767487+1.7693765*I, -0.31456697-1.0403451*I, 1.4415971+0.29750675*I, -0.5856801-1.0203559*I, 7.005772+0.0*I, -0.9617417+1.2486815*I,
0.3842994+0.7050991*I, 1.8942566+0.6389735*I, 1.0858271-1.298006*I, -0.112077385+1.2014246*I, 0.048102796-0.9741874*I, -0.31978557+0.13701046*I, 1.221786-0.85654986*I, 0.7103452+0.8422135*I, -0.9617417-1.2486815*I, 3.4629636+0.0*I};
BLASFUNC(cpotrf)(&up, &n, (float*)(A3), &n, info);
// printf("%g+%g*I\n", creal(A3[91]), cimag(A3[91]));
if(isnan(CREAL(A3[91])) || isnan(CIMAG(A3[91]))) {
CTEST_ERR("%s:%d got NaN", __FILE__, __LINE__);
}
}
void blas_shutdown(void){
int pos;
#ifdef SMP
BLASFUNC(blas_thread_shutdown)();
#endif
LOCK_COMMAND(&alloc_lock);
for (pos = 0; pos < release_pos; pos ++) {
release_info[pos].func(&release_info[pos]);
}
#ifdef SEEK_ADDRESS
base_address = 0UL;
#else
base_address = BASE_ADDRESS;
#endif
for (pos = 0; pos < NUM_BUFFERS; pos ++){
memory[pos].addr = (void *)0;
memory[pos].used = 0;
#if defined(WHEREAMI) && !defined(USE_OPENMP)
memory[pos].pos = -1;
#endif
memory[pos].lock = 0;
}
UNLOCK_COMMAND(&alloc_lock);
return;
}
CTEST(amax, samax){
blasint N=3, inc=1;
float te_max=0.0, tr_max=0.0;
float x[]={-1.1, 2.2, -3.3};
te_max=BLASFUNC(samax)(&N, x, &inc);
tr_max=3.3;
ASSERT_DBL_NEAR_TOL((double)(tr_max), (double)(te_max), SINGLE_EPS);
}
void test_samax()
{
int N=3, inc=1;
float te_max=0.0, tr_max=0.0;
float x[]={-1.1, 2.2, -3.3};
te_max=BLASFUNC(samax)(&N, x, &inc);
tr_max=BLASFUNC_REF(samax)(&N, x, &inc);
CU_ASSERT_DOUBLE_EQUAL(te_max, tr_max, CHECK_EPS);
}
void openblas_fork_handler()
{
// This handler shuts down the OpenBLAS-managed PTHREAD pool when OpenBLAS is
// built with "make USE_OPENMP=0".
// Hanging can still happen when OpenBLAS is built against the libgomp
// implementation of OpenMP. The problem is tracked at:
// http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60035
// In the mean time build with USE_OPENMP=0 or link against another
// implementation of OpenMP.
#if !(defined(OS_WINDOWS) || defined(OS_ANDROID)) && defined(SMP_SERVER)
int err;
err = pthread_atfork ((void (*)(void)) BLASFUNC(blas_thread_shutdown), NULL, NULL);
if(err != 0)
openblas_warning(0, "OpenBLAS Warning ... cannot install fork handler. You may meet hang after fork.\n");
#endif
}
void test_zdotu_offset_1(void)
{
int N=1,incX=1,incY=1;
double x1[]={1.0,2.0,3.0,4.0};
double y1[]={5.0,6.0,7.0,8.0};
double x2[]={1.0,2.0,3.0,4.0};
double y2[]={5.0,6.0,7.0,8.0};
double _Complex result1=0.0;
double _Complex result2=0.0;
//OpenBLAS
result1=BLASFUNC(zdotu)(&N,x1+1,&incX,y1+1,&incY);
//reference
result2=BLASFUNC_REF(zdotu)(&N,x2+1,&incX,y2+1,&incY);
CU_ASSERT_DOUBLE_EQUAL(creal(result1), creal(result2), CHECK_EPS);
CU_ASSERT_DOUBLE_EQUAL(cimag(result1), cimag(result2), CHECK_EPS);
// printf("\%lf,%lf\n",creal(result1),cimag(result1));
}
void test_csrot_inc_0(void)
{
int i=0;
int N=4,incX=0,incY=0;
float c=0.25,s=0.5;
float x1[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
float y1[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
float x2[]={1.0,3.0,5.0,7.0,1.0,3.0,5.0,7.0};
float y2[]={2.0,4.0,6.0,8.0,2.0,4.0,6.0,8.0};
//OpenBLAS
BLASFUNC(csrot)(&N,x1,&incX,y1,&incY,&c,&s);
//reference
BLASFUNC_REF(csrot)(&N,x2,&incX,y2,&incY,&c,&s);
for(i=0; i<2*N; i++){
CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
}
}
void test_drot_inc_0(void)
{
int i=0;
int N=4,incX=0,incY=0;
double c=0.25,s=0.5;
double x1[]={1.0,3.0,5.0,7.0};
double y1[]={2.0,4.0,6.0,8.0};
double x2[]={1.0,3.0,5.0,7.0};
double y2[]={2.0,4.0,6.0,8.0};
//OpenBLAS
BLASFUNC(drot)(&N,x1,&incX,y1,&incY,&c,&s);
//reference
BLASFUNC_REF(drot)(&N,x2,&incX,y2,&incY,&c,&s);
for(i=0; i<N; i++){
CU_ASSERT_DOUBLE_EQUAL(x1[i], x2[i], CHECK_EPS);
CU_ASSERT_DOUBLE_EQUAL(y1[i], y2[i], CHECK_EPS);
}
}
void NAME(blasint *M, blasint *N, FLOAT *Alpha,
FLOAT *x, blasint *INCX,
FLOAT *y, blasint *INCY,
FLOAT *a, blasint *LDA){
blasint m = *M;
blasint n = *N;
FLOAT alpha = *Alpha;
blasint incx = *INCX;
blasint incy = *INCY;
blasint lda = *LDA;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
blasint info;
PRINT_DEBUG_NAME;
info = 0;
if (lda < MAX(1,m)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (m < 0) info = 1;
if (info){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
blasint m, blasint n,
FLOAT alpha,
FLOAT *x, blasint incx,
FLOAT *y, blasint incy,
FLOAT *a, blasint lda) {
FLOAT *buffer;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
info = 0;
if (order == CblasColMajor) {
info = -1;
if (lda < MAX(1,m)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (m < 0) info = 1;
}
if (order == CblasRowMajor) {
info = -1;
t = n;
n = m;
m = t;
t = incx;
incx = incy;
incy = t;
buffer = x;
x = y;
y = buffer;
if (lda < MAX(1,m)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (m < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
/* Quick return if possible. */
if (m == 0 || n == 0) return;
if (alpha == 0.) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incy < 0) y -= (n - 1) * incy;
if (incx < 0) x -= (m - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
GER(m, n, 0, alpha, x, incx, y, incy, a, lda, buffer);
#ifdef SMP
} else {
GER_THREAD(m, n, alpha, x, incx, y, incy, a, lda, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
void NAME( char* ORDER, char* TRANS, blasint *rows, blasint *cols, FLOAT *alpha, FLOAT *a, blasint *lda, blasint *ldb)
{
char Order, Trans;
int order=-1,trans=-1;
blasint info = -1;
FLOAT *b;
size_t msize;
Order = *ORDER;
Trans = *TRANS;
TOUPPER(Order);
TOUPPER(Trans);
if ( Order == 'C' ) order = BlasColMajor;
if ( Order == 'R' ) order = BlasRowMajor;
if ( Trans == 'N' ) trans = BlasNoTrans;
if ( Trans == 'R' ) trans = BlasNoTrans;
if ( Trans == 'T' ) trans = BlasTrans;
if ( Trans == 'C' ) trans = BlasTrans;
#else
void CNAME( enum CBLAS_ORDER CORDER, enum CBLAS_TRANSPOSE CTRANS, blasint crows, blasint ccols, FLOAT calpha, FLOAT *a, blasint clda, blasint cldb)
{
char Order, Trans;
int order=-1,trans=-1;
blasint info = -1;
FLOAT *b;
size_t msize;
blasint *lda, *ldb, *rows, *cols;
FLOAT *alpha;
if ( CORDER == CblasColMajor) order = BlasColMajor;
if ( CORDER == CblasRowMajor) order = BlasRowMajor;
if ( CTRANS == CblasNoTrans || CTRANS == CblasConjNoTrans) trans = BlasNoTrans;
if ( CTRANS == CblasTrans || CTRANS == CblasConjTrans ) trans = BlasTrans;
rows = &crows;
cols = &ccols;
alpha = &calpha;
lda = &clda;
ldb = &cldb;
#endif
if ( order == BlasColMajor)
{
if ( trans == BlasNoTrans && *ldb < *rows ) info = 9;
if ( trans == BlasTrans && *ldb < *cols ) info = 9;
}
if ( order == BlasRowMajor)
{
if ( trans == BlasNoTrans && *ldb < *cols ) info = 9;
if ( trans == BlasTrans && *ldb < *rows ) info = 9;
}
if ( order == BlasColMajor && *lda < *rows ) info = 7;
if ( order == BlasRowMajor && *lda < *cols ) info = 7;
if ( *cols <= 0 ) info = 4;
if ( *rows <= 0 ) info = 3;
if ( trans < 0 ) info = 2;
if ( order < 0 ) info = 1;
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
if ( *lda > *ldb )
msize = (*lda) * (*ldb) * sizeof(FLOAT);
else
msize = (*ldb) * (*ldb) * sizeof(FLOAT);
b = malloc(msize);
if ( b == NULL )
{
printf("Memory alloc failed\n");
exit(1);
}
if ( order == BlasColMajor )
{
if ( trans == BlasNoTrans )
{
OMATCOPY_K_CN(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0 , b, *ldb, a, *ldb );
}
else
{
OMATCOPY_K_CT(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_CN(*rows, *cols, (FLOAT) 1.0, b, *ldb, a, *ldb );
}
}
else
{
if ( trans == BlasNoTrans )
{
OMATCOPY_K_RN(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, b, *ldb, a, *ldb );
}
else
{
OMATCOPY_K_RT(*rows, *cols, *alpha, a, *lda, b, *ldb );
OMATCOPY_K_RN(*rows, *cols, (FLOAT) 1.0, b, *ldb, a, *ldb );
}
}
free(b);
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX, FLOAT *BETA, FLOAT *y, blasint *INCY){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha = *ALPHA;
blasint lda = *LDA;
blasint incx = *INCX;
FLOAT beta = *BETA;
blasint incy = *INCY;
int (*symv[])(BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
SYMV_U, SYMV_L,
};
#ifdef SMP
int (*symv_thread[])(BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, int) = {
SYMV_THREAD_U, SYMV_THREAD_L,
};
#endif
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT alpha,
FLOAT *a, blasint lda, FLOAT *x, blasint incx, FLOAT beta, FLOAT *y, blasint incy) {
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMP
int nthreads;
#endif
int (*symv[])(BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
SYMV_U, SYMV_L,
};
#ifdef SMP
int (*symv_thread[])(BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, int) = {
SYMV_THREAD_U, SYMV_THREAD_L,
};
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
info = -1;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if (beta != ONE) SCAL_K(n, 0, 0, beta, y, abs(incy), NULL, 0, NULL, 0);
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
if (incy < 0 ) y -= (n - 1) * incy;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(symv[uplo])(n, n, alpha, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(symv_thread[uplo])(n, alpha, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * n / 2 + 2 * n, 2 * n * n);
IDEBUG_END;
return;
}
void NAME(char *TRANSA, char *TRANSB,
blasint *M, blasint *N, blasint *K,
FLOAT *alpha,
FLOAT *a, blasint *ldA,
FLOAT *b, blasint *ldB,
FLOAT *beta,
FLOAT *c, blasint *ldC){
blas_arg_t args;
int transa, transb, nrowa, nrowb;
blasint info;
char transA, transB;
FLOAT *buffer;
FLOAT *sa, *sb;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
#if defined(SMP) && !defined(NO_AFFINITY) && !defined(USE_SIMPLE_THREADED_LEVEL3)
int nodes;
#endif
PRINT_DEBUG_NAME;
args.m = *M;
args.n = *N;
args.k = *K;
args.a = (void *)a;
args.b = (void *)b;
args.c = (void *)c;
args.lda = *ldA;
args.ldb = *ldB;
args.ldc = *ldC;
args.alpha = (void *)alpha;
args.beta = (void *)beta;
transA = *TRANSA;
transB = *TRANSB;
TOUPPER(transA);
TOUPPER(transB);
transa = -1;
transb = -1;
if (transA == 'N') transa = 0;
if (transA == 'T') transa = 1;
#ifndef COMPLEX
if (transA == 'R') transa = 0;
if (transA == 'C') transa = 1;
#else
if (transA == 'R') transa = 2;
if (transA == 'C') transa = 3;
#endif
if (transB == 'N') transb = 0;
if (transB == 'T') transb = 1;
#ifndef COMPLEX
if (transB == 'R') transb = 0;
if (transB == 'C') transb = 1;
#else
if (transB == 'R') transb = 2;
if (transB == 'C') transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = 0;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
if (info){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB,
blasint m, blasint n, blasint k,
#ifndef COMPLEX
FLOAT alpha,
#else
FLOAT *alpha,
#endif
FLOAT *a, blasint lda,
FLOAT *b, blasint ldb,
#ifndef COMPLEX
FLOAT beta,
#else
FLOAT *beta,
#endif
FLOAT *c, blasint ldc) {
blas_arg_t args;
int transa, transb;
blasint nrowa, nrowb, info;
XFLOAT *buffer;
XFLOAT *sa, *sb;
#ifdef SMP
#ifndef COMPLEX
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_REAL;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_REAL;
#else
int mode = BLAS_SINGLE | BLAS_REAL;
#endif
#else
#ifdef XDOUBLE
int mode = BLAS_XDOUBLE | BLAS_COMPLEX;
#elif defined(DOUBLE)
int mode = BLAS_DOUBLE | BLAS_COMPLEX;
#else
int mode = BLAS_SINGLE | BLAS_COMPLEX;
#endif
#endif
#endif
#if defined(SMP) && !defined(NO_AFFINITY) && !defined(USE_SIMPLE_THREADED_LEVEL3)
int nodes;
#endif
PRINT_DEBUG_CNAME;
#ifndef COMPLEX
args.alpha = (void *)α
args.beta = (void *)β
#else
args.alpha = (void *)alpha;
args.beta = (void *)beta;
#endif
transa = -1;
transb = -1;
info = 0;
if (order == CblasColMajor) {
args.m = m;
args.n = n;
args.k = k;
args.a = (void *)a;
args.b = (void *)b;
args.c = (void *)c;
args.lda = lda;
args.ldb = ldb;
args.ldc = ldc;
if (TransA == CblasNoTrans) transa = 0;
if (TransA == CblasTrans) transa = 1;
#ifndef COMPLEX
if (TransA == CblasConjNoTrans) transa = 0;
if (TransA == CblasConjTrans) transa = 1;
#else
if (TransA == CblasConjNoTrans) transa = 2;
if (TransA == CblasConjTrans) transa = 3;
#endif
if (TransB == CblasNoTrans) transb = 0;
if (TransB == CblasTrans) transb = 1;
#ifndef COMPLEX
if (TransB == CblasConjNoTrans) transb = 0;
if (TransB == CblasConjTrans) transb = 1;
#else
if (TransB == CblasConjNoTrans) transb = 2;
if (TransB == CblasConjTrans) transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = -1;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
}
if (order == CblasRowMajor) {
args.m = n;
args.n = m;
args.k = k;
args.a = (void *)b;
args.b = (void *)a;
args.c = (void *)c;
args.lda = ldb;
args.ldb = lda;
args.ldc = ldc;
if (TransB == CblasNoTrans) transa = 0;
if (TransB == CblasTrans) transa = 1;
#ifndef COMPLEX
if (TransB == CblasConjNoTrans) transa = 0;
if (TransB == CblasConjTrans) transa = 1;
#else
if (TransB == CblasConjNoTrans) transa = 2;
if (TransB == CblasConjTrans) transa = 3;
#endif
if (TransA == CblasNoTrans) transb = 0;
if (TransA == CblasTrans) transb = 1;
#ifndef COMPLEX
if (TransA == CblasConjNoTrans) transb = 0;
if (TransA == CblasConjTrans) transb = 1;
#else
if (TransA == CblasConjNoTrans) transb = 2;
if (TransA == CblasConjTrans) transb = 3;
#endif
nrowa = args.m;
if (transa & 1) nrowa = args.k;
nrowb = args.k;
if (transb & 1) nrowb = args.n;
info = -1;
if (args.ldc < args.m) info = 13;
if (args.ldb < nrowb) info = 10;
if (args.lda < nrowa) info = 8;
if (args.k < 0) info = 5;
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (transb < 0) info = 2;
if (transa < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if ((args.m == 0) || (args.n == 0)) return;
#if 0
fprintf(stderr, "m = %4d n = %d k = %d lda = %4d ldb = %4d ldc = %4d\n",
args.m, args.n, args.k, args.lda, args.ldb, args.ldc);
#endif
IDEBUG_START;
FUNCTION_PROFILE_START();
buffer = (XFLOAT *)blas_memory_alloc(0);
sa = (XFLOAT *)((BLASLONG)buffer +GEMM_OFFSET_A);
sb = (XFLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#ifdef SMP
mode |= (transa << BLAS_TRANSA_SHIFT);
mode |= (transb << BLAS_TRANSB_SHIFT);
args.common = NULL;
args.nthreads = num_cpu_avail(3);
if (args.nthreads == 1) {
#endif
(gemm[(transb << 2) | transa])(&args, NULL, NULL, sa, sb, 0);
#ifdef SMP
} else {
#ifndef USE_SIMPLE_THREADED_LEVEL3
#ifndef NO_AFFINITY
nodes = get_num_nodes();
if ((nodes > 1) && get_node_equal()) {
args.nthreads /= nodes;
gemm_thread_mn(mode, &args, NULL, NULL, gemm[16 | (transb << 2) | transa], sa, sb, nodes);
} else {
#endif
(gemm[16 | (transb << 2) | transa])(&args, NULL, NULL, sa, sb, 0);
#else
GEMM_THREAD(mode, &args, NULL, NULL, gemm[(transb << 2) | transa], sa, sb, args.nthreads);
#endif
#ifndef USE_SIMPLE_THREADED_LEVEL3
#ifndef NO_AFFINITY
}
#endif
#endif
#endif
#ifdef SMP
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE, args.m * args.k + args.k * args.n + args.m * args.n, 2 * args.m * args.n * args.k);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA,
FLOAT *x, blasint *INCX, FLOAT *a){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint incx = *INCX;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
if (n == 0) return;
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(spr[uplo])(n, alpha_r, alpha_i, x, incx, a, buffer);
#ifdef SMP
} else {
(spr_thread[uplo])(n, ALPHA, x, incx, a, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
int NAME(char *UPLO, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){
blas_arg_t args;
blasint uplo_arg = *UPLO;
blasint uplo;
blasint info;
FLOAT *buffer;
#ifdef PPC440
extern
#endif
FLOAT *sa, *sb;
PRINT_DEBUG_NAME;
args.n = *N;
args.a = (void *)a;
args.lda = *ldA;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (args.lda < MAX(1,args.n)) info = 4;
if (args.n < 0) info = 2;
if (uplo < 0) info = 1;
if (info) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
*Info = - info;
return 0;
}
*Info = 0;
if (args.n == 0) return 0;
IDEBUG_START;
FUNCTION_PROFILE_START();
#ifndef PPC440
buffer = (FLOAT *)blas_memory_alloc(1);
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#endif
#ifdef SMP
args.common = NULL;
args.nthreads = num_cpu_avail(4);
if (args.nthreads == 1) {
#endif
*Info = (lauum_single[uplo])(&args, NULL, NULL, sa, sb, 0);
#ifdef SMP
} else {
*Info = (lauum_parallel[uplo])(&args, NULL, NULL, sa, sb, 0);
}
#endif
#ifndef PPC440
blas_memory_free(buffer);
#endif
FUNCTION_PROFILE_END(1, .5 * args.n * args.n,
2. * args.n * (1./3. + args.n * ( 1./2. + args.n * 1./6.))
+ args.n * (args.n * args.n - 1));
IDEBUG_END;
return 0;
}
int NAME(char *UPLO, char *DIAG, blasint *N, FLOAT *a, blasint *ldA, blasint *Info){
blas_arg_t args;
blasint uplo_arg = *UPLO;
blasint diag_arg = *DIAG;
blasint uplo, diag;
blasint info;
FLOAT *buffer;
#ifdef PPC440
extern
#endif
FLOAT *sa, *sb;
PRINT_DEBUG_NAME;
args.n = *N;
args.a = (void *)a;
args.lda = *ldA;
TOUPPER(uplo_arg);
TOUPPER(diag_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
diag = -1;
if (diag_arg == 'U') diag = 0;
if (diag_arg == 'N') diag = 1;
info = 0;
if (args.lda < MAX(1,args.n)) info = 5;
if (args.n < 0) info = 3;
if (diag < 0) info = 2;
if (uplo < 0) info = 1;
if (info) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
*Info = - info;
return 0;
}
*Info = 0;
if (args.n <= 0) return 0;
IDEBUG_START;
FUNCTION_PROFILE_START();
#ifndef PPC440
buffer = (FLOAT *)blas_memory_alloc(1);
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#endif
info = (trti2[(uplo << 1) | diag])(&args, NULL, NULL, sa, sb, 0);
*Info = info;
#ifndef PPC440
blas_memory_free(buffer);
#endif
FUNCTION_PROFILE_END(1, .5 * args.n * args.n,
2. * args.n * (1./3. + args.n * ( 1./2. + args.n * 1./6.))
+ 6. * args.n * (1./3. + args.n * (-1./2. + args.n * 1./6.)));
IDEBUG_END;
return 0;
}
int main(int argc, char *argv[]) {
int i,j,k;
machineInformation currentMachine;
counterSessionInfo session;
initializeCUDA();
// Set machine information from CounterHomeBrew.h
currentMachine.cpu_model = CPU_MODEL;
currentMachine.num_sockets = NUM_SOCKETS;
currentMachine.num_phys_cores_per_socket = NUM_PHYS_CORES_PER_SOCKET;
currentMachine.num_cores_per_socket = NUM_CORES_PER_SOCKET;
currentMachine.num_cores = NUM_CORES;
currentMachine.num_cbos = NUM_PHYS_CORES_PER_SOCKET; // should multiply by NUM_SOCKETS???
currentMachine.core_gen_counter_num_max = CORE_GEN_COUNTER_MAX;
currentMachine.cbo_counter_num_max = CBO_COUNTER_NUM_MAX;
// Set session events, umasks and counters used
// int32 core_event_numbers[] = {FP_COMP_OPS_EXE_EVTNR,SIMD_FP_256_EVTNR,0x51,0xF1,0x80};
// int32 core_umasks[] = {FP_COMP_OPS_EXE_SCALAR_DOUBLE_UMASK,SIMD_FP_256_PACKED_DOUBLE_UMASK,0x01, 0x07,0x01};
session.core_gen_counter_num_used = 5;
int32 core_event_numbers[] = {0x10,0x10,0x11,0x51,0xF1};
int32 core_umasks[] = {0x20,0x40,0x01,0x01, 0x07};
session.cbo_counter_num_used = 1;
int32 cbo_event_numbers[] = {0x37};
int32 cbo_umasks[] = {0xf};
session.cbo_filter = 0x1f;
for (i = 0; i < session.core_gen_counter_num_used; i++) {
session.core_event_numbers[i] = core_event_numbers[i];
session.core_umasks[i] = core_umasks[i];
}
for (i = 0; i < session.cbo_counter_num_used; i++) {
session.cbo_event_numbers[i] = cbo_event_numbers[i];
session.cbo_umasks[i] = cbo_umasks[i];
}
int fd[NUM_CORES];
// Arrays to hold counter data...
counterData before;
counterData after;
// some data for doing a naive matmul to test flop counting...
// initloop(N);
// M,N,K are multiples of the block size....
int gpuOuter = atoi(argv[1]);
int gpuInner = atoi(argv[2]);
int cpuInner = atoi(argv[3]);
double minRuntime = atoi(argv[4]);
int Md = atoi(argv[5])*block_size;
int Nd = atoi(argv[6])*block_size;
int Kd = atoi(argv[7])*block_size;
int Mh = atoi(argv[8]);
int Nh = atoi(argv[9]);
int Kh = atoi(argv[10]);
char *ts1,*ts2,*ts3,*ts4;
char *ts5,*ts6,*ts7,*ts8;
double fineTimeStamps[8];
double gTime = 0.0;
double cTime = 0.0;
double seconds = 0.0;
int num_iters;
uint64 *coreSums;
coreSums = (uint64*)calloc(currentMachine.num_sockets*session.core_gen_counter_num_used,sizeof(uint64));
uint64 *sums;
sums = (uint64*)calloc(currentMachine.num_sockets*session.cbo_counter_num_used,sizeof(uint64));
float *Atmp = NULL;
float *Btmp = NULL;
float *Ctmp = NULL;
Atmp = (float*) malloc( Mh * Nh * sizeof(float) );
Btmp = (float*) malloc( Nh * sizeof(float) );
Ctmp = (float*) malloc( Mh * sizeof(float) );
randomInit(Atmp,Mh*Nh);
randomInit(Btmp,Nh);
for (num_iters = cpuInner; seconds < minRuntime; num_iters *=2) {
seconds = 0.0;
for (i =0; i < num_iters; i++)
BLASFUNC( CblasColMajor,CblasNoTrans,Mh,Nh, 1, Atmp,Mh, Btmp,1, 1, Ctmp,1 );
seconds = read_timer()-seconds;
}
// num_iters /= 2;
free(Atmp);
free(Btmp);
free(Ctmp);
int readyThreads = 0;
#pragma omp parallel
{
int threadNum = omp_get_thread_num();
int numThreads = omp_get_num_threads();
assert(numThreads==2);
if (threadNum == 0) {
cudaError_t error;
int memSizeA = sizeof(float)*Md*Nd;
int memSizeB = sizeof(float)*Nd;
int memSizeC = sizeof(float)*Md;
float *Ahost,*Bhost,*Chost;
// use pinned memory on the host for BW and asynch memory transfers..
int flags = cudaHostAllocDefault;
ts5 = getTimeStamp();
fineTimeStamps[0] = read_timer();
error = cudaHostAlloc((void**)&Ahost,memSizeA,flags);if (error != cudaSuccess){printf("cudaHostMalloc Ahost returned error code %d, line(%d)\n", error, __LINE__);exit(EXIT_FAILURE);}
error = cudaHostAlloc((void**)&Bhost,memSizeB,flags);if (error != cudaSuccess){printf("cudaHostMalloc Bhost returned error code %d, line(%d)\n", error, __LINE__);exit(EXIT_FAILURE);}
error = cudaHostAlloc((void**)&Chost,memSizeC,flags);if (error != cudaSuccess){printf("cudaHostMalloc Chost returned error code %d, line(%d)\n", error, __LINE__);exit(EXIT_FAILURE);}
// set local arrays
randomInit(Ahost,Md*Nd);
randomInit(Bhost,Nd);
// allocate device memory
float *Adevice,*Bdevice,*Cdevice;
error = cudaMalloc((void**)&Adevice,memSizeA); if (error != cudaSuccess){printf("cudaMalloc Adevice returned error code %d, line(%d)\n", error, __LINE__);exit(EXIT_FAILURE);}
error = cudaMalloc((void**)&Bdevice,memSizeB); if (error != cudaSuccess){printf("cudaMalloc Bdevice returned error code %d, line(%d)\n", error, __LINE__);exit(EXIT_FAILURE);}
error = cudaMalloc((void**)&Cdevice,memSizeC); if (error != cudaSuccess){printf("cudaMalloc Cdevice returned error code %d, line(%d)\n", error, __LINE__);exit(EXIT_FAILURE);}
fineTimeStamps[1] = read_timer();
ts6 = getTimeStamp();
#pragma omp critical
{
readyThreads += 1;
}
// fprintf(stderr,"Incremented ready GPU\n");
while (readyThreads < 2){sleep(1);fprintf(stderr,"Thread 0: %d\n",readyThreads);};
//#pragma omp single
//{
cudaStream_t stream1;
cudaStreamCreate ( &stream1) ;
ts3 = getTimeStamp();
fineTimeStamps[2] = read_timer();
gTime = read_timer();
for (int i = 0; i < gpuOuter; i++)
GPUsgemv(gpuInner,Md,Nd,Kd,Adevice,Bdevice,Cdevice,Ahost,Bhost,Chost,&stream1);
cudaStreamSynchronize(stream1);
gTime = read_timer() - gTime;
fineTimeStamps[3] = read_timer();
ts4 = getTimeStamp();
cudaFreeHost(Ahost);
cudaFreeHost(Bhost);
cudaFreeHost(Chost);
} else {
// uint64 min_iters = strtoull(argv[4],NULL,0);
float *A = NULL;
float *B = NULL;
float *C = NULL;
ts7 = getTimeStamp();
fineTimeStamps[4] = read_timer();
A = (float*) malloc( Mh * Nh * sizeof(float) );
B = (float*) malloc( Nh * sizeof(float) );
C = (float*) malloc( Mh * sizeof(float) );
randomInit(A,Mh*Nh);
randomInit(B,Nh);
fineTimeStamps[5] = read_timer();
ts8 = getTimeStamp();
#pragma omp critical
{
readyThreads += 1;
}
// fprintf(stderr,"Incremented ready CPU\n");
while (readyThreads < 2){sleep(1);fprintf(stderr,"Thread 1: %d\n",readyThreads);};
// open the msr files for each core on the machine
for (i = 0; i < currentMachine.num_cores; i++)
open_msr_file(i,&fd[i]);
int socketsProgrammed = 0;
for (i = 0; i < currentMachine.num_cores; i++) {
int currentCoreFD = fd[i];
stopCounters(i, currentCoreFD, ¤tMachine, &session);
programCoreFixedCounters(currentCoreFD);
programGeneralPurposeRegisters(currentCoreFD, ¤tMachine, &session);
/* Program the Uncore as desired...*/
// Only program the first physical core on each socket.
// NOTE: Some assumptions about topology here...check /proc/cpuinfo to confirm.
if (i % currentMachine.num_phys_cores_per_socket == 0 && socketsProgrammed < currentMachine.num_sockets) {
programUncoreCounters( currentCoreFD, ¤tMachine, &session);
socketsProgrammed++;
}
}
seconds = 0.0;
// start the programmed counters...
for (i = 0; i < currentMachine.num_cores; i++)
startCounters( i, fd[i], ¤tMachine, &session);
/* READ COUNTERS BEFORE STUFF */
readCounters(fd,¤tMachine,&session, &before);
ts1 = getTimeStamp();
fineTimeStamps[6] = read_timer();
seconds = read_timer();
/* DO STUFF */
for (i =0; i < num_iters; i++)
BLASFUNC( CblasColMajor,CblasNoTrans,Mh,Nh, 1, A,Mh, B,1, 1, C,1 );
/* END DOING STUFF */
seconds = read_timer()-seconds;
fineTimeStamps[7] = read_timer();
ts2 = getTimeStamp();
/* READ COUNTERS AFTER STUFF */
for (i = 0; i < currentMachine.num_cores; i++)
stopCounters(i,fd[i],¤tMachine, &session);
// printf("num_iters = %"PRIu64", runtime is %g\n",num_iters,seconds);
readCounters(fd,¤tMachine,&session,&after);
diffCounterData(¤tMachine, &session, &after, &before, &after);
for (i = 0; i < currentMachine.num_sockets; i++) {
// printf("Socket %d\n",i);
for (j = 0; j < currentMachine.num_cores_per_socket; j++) {
// printf("%d,",j);
for (k = 0; k < session.core_gen_counter_num_used; k++){
// printf("%"PRIu64",",after.generalCore[i*currentMachine.num_cores_per_socket + j][k]);
// bug in the indexing of the core sums???
// coreSums[i*session.core_gen_counter_num_used + k] += after.generalCore[i*currentMachine.num_cores_per_socket + j][k];
coreSums[k] += after.generalCore[i*currentMachine.num_cores_per_socket + j][k];
}
// printf("\n");
}
}
for (i = 0; i < currentMachine.num_sockets; i++) {
// printf("%d,",i);
for (j = 0; j < currentMachine.num_cbos; j++) {
// printf("%d,",j);
for (k = 0; k < session.cbo_counter_num_used; k++) {
// printf("%llu,",after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k]);
// bug in the indexing of the core sums???
// sums[i*session.cbo_counter_num_used + k] += after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k];
sums[k] += after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k];
}
}
}
// printf("\n");
// Stop counters, reset PMU, close msr files
cleanup(fd,¤tMachine,&session);
free(A);
free(B);
free(C);
}
} // end parallel region
printf("%s,%s,%s,%s,%s,%s,%s,%s,%d,%d,%d,%d,%d,%d,%d,%d,%d,%f,%f,%f,",ts7,ts8,ts1,ts2,ts5,ts6,ts3,ts4,Mh,Nh,Kh,Md/block_size,Nd/block_size,Kd/block_size,num_iters,gpuOuter,gpuInner,seconds,gTime,(float)(gpuOuter*(Md*Kd+Nd+Md))/16.0);
for (int i = 0; i < 8; i++)
printf("%f,",fineTimeStamps[i]);
for (j = 0; j < session.core_gen_counter_num_used; j++)
printf("%llu,",coreSums[j]);
for (j = 0; j < session.cbo_counter_num_used; j++)
if (j == session.cbo_counter_num_used-1)
printf("%llu",sums[j]);
else
printf("%llu,",sums[j]);
printf("\n");
free(sums);
free(coreSums);
return 0;
}
void NAME(char *TRANS, blasint *M, blasint *N,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT alpha = *ALPHA;
FLOAT beta = *BETA;
FLOAT *buffer;
#ifdef SMP
int nthreads;
int nthreads_max;
int nthreads_avail;
double MNK;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 0;
if (trans == 'C') i = 1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
FLOAT alpha,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT beta,
FLOAT *y, blasint incy){
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
int nthreads_max;
int nthreads_avail;
double MNK;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
PRINT_DEBUG_CNAME;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
info = -1;
t = n;
n = m;
m = t;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
//printf("m=%d, n=%d, trans=%d, incx=%d, incy=%d, alpha=%f, beta=%f\n", m, n, trans, incx, incy, alpha, beta);
if ((m==0) || (n==0)) return;
lenx = n;
leny = m;
if (trans) lenx = m;
if (trans) leny = n;
if (beta != ONE) SCAL_K(leny, 0, 0, beta, y, abs(incy), NULL, 0, NULL, 0);
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx;
if (incy < 0) y -= (leny - 1) * incy;
#ifdef MAX_STACK_ALLOC
// make it volatile because some gemv implementation (ex: dgemv_n.S)
// do not restore all register
volatile int stack_alloc_size = 0;
//for gemv_n and gemv_t, try to allocate on stack
stack_alloc_size = m + n;
#ifdef ALIGNED_ACCESS
stack_alloc_size += 3;
#endif
// if(stack_alloc_size < 128)
//dgemv_n.S require a 128 bytes buffer
// increasing instead of capping 128
// ABI STACK for windows 288 bytes
stack_alloc_size += 288 / sizeof(FLOAT) ;
if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(FLOAT))
stack_alloc_size = 0;
// stack overflow check
volatile double stack_check = 3.14159265358979323846;
FLOAT stack_buffer[stack_alloc_size];
buffer = stack_alloc_size ? stack_buffer : (FLOAT *)blas_memory_alloc(1);
// printf("stack_alloc_size=%d\n", stack_alloc_size);
#else
//Original OpenBLAS/GotoBLAS codes.
buffer = (FLOAT *)blas_memory_alloc(1);
#endif
#ifdef SMP
nthreads_max = num_cpu_avail(2);
nthreads_avail = nthreads_max;
MNK = (double) m * (double) n;
if ( MNK <= (24.0 * 24.0 * (double) (GEMM_MULTITHREAD_THRESHOLD*GEMM_MULTITHREAD_THRESHOLD) ) )
nthreads_max = 1;
if ( nthreads_max > nthreads_avail )
nthreads = nthreads_avail;
else
nthreads = nthreads_max;
if (nthreads == 1) {
#endif
(gemv[(int)trans])(m, n, 0, alpha, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gemv_thread[(int)trans])(m, n, alpha, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
// stack overflow check
assert(stack_check==3.14159265358979323846);
#ifdef MAX_STACK_ALLOC
if(!stack_alloc_size){
blas_memory_free(buffer);
}
#else
blas_memory_free(buffer);
#endif
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *b, blasint *INCX, FLOAT *BETA, FLOAT *c, blasint *INCY){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint lda = *LDA;
blasint incx = *INCX;
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
blasint incy = *INCY;
int (*symv[])(BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
SYMV_U, SYMV_L,
};
#ifdef SMP
int (*symv_thread[])(BLASLONG, FLOAT *, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, BLASLONG, FLOAT *, int) = {
SYMV_THREAD_U, SYMV_THREAD_L,
};
#endif
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incy == 0) info = 10;
if (incx == 0) info = 7;
if (lda < MAX(1, n)) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
if (n == 0) return;
if ((beta_r != ONE) || (beta_i != ZERO)) SCAL_K(n, 0, 0, beta_r, beta_i, c, abs(incy), NULL, 0, NULL, 0);
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) b -= (n - 1) * incx * COMPSIZE;
if (incy < 0 ) c -= (n - 1) * incy * COMPSIZE;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(symv[uplo])(n, n, alpha_r, alpha_i, a, lda, b, incx, c, incy, buffer);
#ifdef SMP
} else {
(symv_thread[uplo])(n, ALPHA, a, lda, b, incx, c, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + 2 * n, 2 * n * n);
IDEBUG_END;
return;
}
int main(int argc, char *argv[]) {
int i,j,k;
char *ts1,*ts2;
machineInformation currentMachine;
counterSessionInfo session;
double seconds = 0.0;
// Set machine information from CounterHomeBrew.h
currentMachine.cpu_model = CPU_MODEL;
currentMachine.num_sockets = NUM_SOCKETS;
currentMachine.num_phys_cores_per_socket = NUM_PHYS_CORES_PER_SOCKET;
currentMachine.num_cores_per_socket = NUM_CORES_PER_SOCKET;
currentMachine.num_cores = NUM_CORES;
currentMachine.num_cbos = NUM_PHYS_CORES_PER_SOCKET;
currentMachine.core_gen_counter_num_max = CORE_GEN_COUNTER_MAX;
currentMachine.cbo_counter_num_max = CBO_COUNTER_NUM_MAX;
// Set session events, umasks and counters used
// NHM-EX
session.core_gen_counter_num_used = 0;
int32 core_event_numbers[] = {};
int32 core_umasks[] = {};
session.cbo_counter_num_used = 1;
int32 cbo_event_numbers[] = {0x14};
int32 cbo_umasks[] = {0x7};
// JKT
/*
session.core_gen_counter_num_used = 5;
int32 core_event_numbers[] = {0x10,0x10,0x11,0x51,0xF1};
int32 core_umasks[] = {0x80,0x10,0x02,0x01, 0x07};
session.cbo_counter_num_used = 1;
int32 cbo_event_numbers[] = {0x37};
int32 cbo_umasks[] = {0xf};
session.cbo_filter = 0x1f;
*/
for (i = 0; i < session.core_gen_counter_num_used; i++) {
session.core_event_numbers[i] = core_event_numbers[i];
session.core_umasks[i] = core_umasks[i];
}
for (i = 0; i < session.cbo_counter_num_used; i++) {
session.cbo_event_numbers[i] = cbo_event_numbers[i];
session.cbo_umasks[i] = cbo_umasks[i];
}
int fd[NUM_CORES];
// Arrays to hold counter data...
counterData before;
counterData after;
// some data for doing a naive matmul to test flop counting...
// initloop(N);
uint64 min_iters = 2;
double minRuntime = 10.0;
int M = atoi(argv[1]);
int K = atoi(argv[2]);
double *A = NULL;
double *b = NULL;
double *c = NULL;
// posix_memalign((void**)A,64,M*K*sizeof(double));
// posix_memalign((void**)B,64,K*N*sizeof(double));
// posix_memalign((void**)C,64,M*N*sizeof(double));
A = (double*) malloc( M * K * sizeof(double) );
b = (double*) malloc( K * sizeof(double) );
c = (double*) malloc( M * sizeof(double) );
fill( A, M * K );
fill( b, K );
fill( c, M );
// open the msr files for each core on the machine
for (i = 0; i < currentMachine.num_cores; i++)
open_msr_file(i,&fd[i]);
// warm up da caches...
BLASFUNC( CblasColMajor,CblasNoTrans,M,K, 1, A,M, b,1, 1, c,1 );
// Program the counters!!!
int socketsProgrammed = 0;
for (i = 0; i < currentMachine.num_cores; i++) {
int currentCoreFD = fd[i];
/* clear global control register before programming */
stopCounters(i, currentCoreFD, ¤tMachine, &session);
/* set up the fixed counters on each core */
programCoreFixedCounters(currentCoreFD);
/* set up the general purpose registers for each core */
programGeneralPurposeRegisters(currentCoreFD, ¤tMachine, &session);
/* Program the Uncore as desired...*/
// Only program the first physical core on each socket.
// NOTE: Some assumptions about topology here...check /proc/cpuinfo to confirm.
#if CPU_MODEL == JAKETOWN
if (i % currentMachine.num_phys_cores_per_socket == 0 && socketsProgrammed < currentMachine.num_sockets)
#elif CPU_MODEL == NEHALEM_EX
if (i < currentMachine.num_sockets && socketsProgrammed < currentMachine.num_sockets)
#elif CPU_MODEL == IVY_BRIDGE
if (i < currentMachine.num_sockets && socketsProgrammed < currentMachine.num_sockets)
#endif
{
programUncoreCounters( currentCoreFD, ¤tMachine, &session);
socketsProgrammed++;
}
/* set global control register to active counters */
// startCounters( i, currentCoreFD, ¤tMachine, &session);
}
uint64 num_iters;
for (num_iters = min_iters; seconds < minRuntime; num_iters *=2) {
if (num_iters != min_iters) {
free(ts1);
free(ts2);
}
sleep(5);
seconds = 0.0;
// start the programmed counters...
for (i = 0; i < currentMachine.num_cores; i++)
startCounters( i, fd[i], ¤tMachine, &session);
/* READ COUNTERS BEFORE STUFF */
readCounters(fd,¤tMachine,&session, &before);
ts1 = getTimeStamp();
seconds = read_timer();
/* DO STUFF */
for (i =0; i < num_iters; i++)
BLASFUNC( CblasColMajor,CblasNoTrans,M,K, 1, A,M, b,1, 1, c,1 );
/* END DOING STUFF */
seconds = read_timer()-seconds;
ts2 = getTimeStamp();
/* READ COUNTERS AFTER STUFF */
for (i = 0; i < currentMachine.num_cores; i++)
stopCounters(i,fd[i],¤tMachine, &session);
}
num_iters /= 2;
readCounters(fd,¤tMachine,&session,&after);
diffCounterData(¤tMachine, &session, &after, &before, &after);
uint64 *coreSums;
coreSums = (uint64*)calloc(currentMachine.num_sockets*session.core_gen_counter_num_used,sizeof(uint64));
for (i = 0; i < currentMachine.num_sockets; i++) {
for (j = 0; j < currentMachine.num_cores_per_socket; j++) {
for (k = 0; k < session.core_gen_counter_num_used; k++)
// coreSums[i*session.core_gen_counter_num_used + k] += after.generalCore[i*currentMachine.num_cores_per_socket + j][k];
coreSums[k] += after.generalCore[i*currentMachine.num_cores_per_socket + j][k];
}
}
uint64 *sums;
sums = (uint64*)calloc(currentMachine.num_sockets*session.cbo_counter_num_used,sizeof(uint64));
for (i = 0; i < currentMachine.num_sockets; i++) {
// printf("Socket %d\n",i);
for (j = 0; j < currentMachine.num_cbos; j++) {
// printf("%d,",j);
for (k = 0; k < session.cbo_counter_num_used; k++) {
// printf("%"PRIu64",",after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k]);
// bug in the indexing of the core sums???
// sums[i*session.cbo_counter_num_used + k] += after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k];
sums[k] += after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k];
}
// printf("\n");
}
}
/*
sums = (uint64*)calloc(currentMachine.num_sockets*session.cbo_counter_num_used,sizeof(uint64));
for (i = 0; i < currentMachine.num_sockets; i++) {
for (j = 0; j < currentMachine.num_cbos; j++) {
for (k = 0; k < session.cbo_counter_num_used; k++)
// sums[i*session.cbo_counter_num_used + k] += after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k];
sums[k] += after.cboUncore[i*currentMachine.num_phys_cores_per_socket + j][k];
}
}
*/
// only print data from first socket and core
printf("%s,%s,%"PRIu64",%d,%d,%d,%f,",ts1,ts2,num_iters,M,K,K,seconds/(double)num_iters);
for (j = 0; j < session.core_gen_counter_num_used; j++)
// printf("%"PRIu64",",after.generalCore[0][j]);
printf("%f,",coreSums[j]/(double)num_iters);
for (j = 0; j < session.cbo_counter_num_used; j++)
printf("%f,",sums[j]/(double)num_iters);
printf("\n");
free(sums);
free(coreSums);
// Stop counters, reset PMU, close msr files
cleanup(fd,¤tMachine,&session);
free(A);
free(b);
free(c);
return 0;
}
void NAME(char *SIDE, char *UPLO,
blasint *M, blasint *N,
FLOAT *alpha, FLOAT *a, blasint *ldA,
FLOAT *b, blasint *ldB,
FLOAT *beta, FLOAT *c, blasint *ldC){
char side_arg = *SIDE;
char uplo_arg = *UPLO;
blas_arg_t args;
FLOAT *buffer;
FLOAT *sa, *sb;
#if defined(SMP) && !defined(NO_AFFINITY)
int nodes;
#endif
blasint info;
int side;
int uplo;
PRINT_DEBUG_NAME;
args.alpha = (void *)alpha;
args.beta = (void *)beta;
TOUPPER(side_arg);
TOUPPER(uplo_arg);
side = -1;
uplo = -1;
if (side_arg == 'L') side = 0;
if (side_arg == 'R') side = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
args.m = *M;
args.n = *N;
args.c = (void *)c;
args.ldc = *ldC;
info = 0;
if (args.ldc < MAX(1, args.m)) info = 12;
if (!side) {
args.a = (void *)a;
args.b = (void *)b;
args.lda = *ldA;
args.ldb = *ldB;
if (args.ldb < MAX(1, args.m)) info = 9;
if (args.lda < MAX(1, args.m)) info = 7;
} else {
args.a = (void *)b;
args.b = (void *)a;
args.lda = *ldB;
args.ldb = *ldA;
if (args.lda < MAX(1, args.m)) info = 9;
if (args.ldb < MAX(1, args.n)) info = 7;
}
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (uplo < 0) info = 2;
if (side < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_SIDE Side, enum CBLAS_UPLO Uplo,
blasint m, blasint n,
#ifndef COMPLEX
FLOAT alpha,
#else
FLOAT *alpha,
#endif
FLOAT *a, blasint lda,
FLOAT *b, blasint ldb,
#ifndef COMPLEX
FLOAT beta,
#else
FLOAT *beta,
#endif
FLOAT *c, blasint ldc) {
blas_arg_t args;
int side, uplo;
blasint info;
FLOAT *buffer;
FLOAT *sa, *sb;
#if defined(SMP) && !defined(NO_AFFINITY)
int nodes;
#endif
PRINT_DEBUG_CNAME;
#ifndef COMPLEX
args.alpha = (void *)α
args.beta = (void *)β
#else
args.alpha = (void *)alpha;
args.beta = (void *)beta;
#endif
args.c = (void *)c;
args.ldc = ldc;
side = -1;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Side == CblasLeft) side = 0;
if (Side == CblasRight) side = 1;
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
args.m = m;
args.n = n;
if (args.ldc < MAX(1, args.m)) info = 12;
if (!side) {
args.a = (void *)a;
args.b = (void *)b;
args.lda = lda;
args.ldb = ldb;
if (args.ldb < MAX(1, args.m)) info = 9;
if (args.lda < MAX(1, args.m)) info = 7;
} else {
args.a = (void *)b;
args.b = (void *)a;
args.lda = ldb;
args.ldb = lda;
if (args.lda < MAX(1, args.m)) info = 9;
if (args.ldb < MAX(1, args.n)) info = 7;
}
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (uplo < 0) info = 2;
if (side < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Side == CblasLeft) side = 1;
if (Side == CblasRight) side = 0;
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
info = -1;
args.m = n;
args.n = m;
if (args.ldc < MAX(1, args.m)) info = 12;
if (!side) {
args.a = (void *)a;
args.b = (void *)b;
args.lda = lda;
args.ldb = ldb;
if (args.ldb < MAX(1, args.m)) info = 9;
if (args.lda < MAX(1, args.m)) info = 7;
} else {
args.a = (void *)b;
args.b = (void *)a;
args.lda = ldb;
args.ldb = lda;
if (args.lda < MAX(1, args.m)) info = 9;
if (args.ldb < MAX(1, args.n)) info = 7;
}
if (args.n < 0) info = 4;
if (args.m < 0) info = 3;
if (uplo < 0) info = 2;
if (side < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (args.m == 0 || args.n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
buffer = (FLOAT *)blas_memory_alloc(0);
sa = (FLOAT *)((BLASLONG)buffer + GEMM_OFFSET_A);
sb = (FLOAT *)(((BLASLONG)sa + ((GEMM_P * GEMM_Q * COMPSIZE * SIZE + GEMM_ALIGN) & ~GEMM_ALIGN)) + GEMM_OFFSET_B);
#ifdef SMP
args.common = NULL;
args.nthreads = num_cpu_avail(3);
if (args.nthreads == 1) {
#endif
(symm[(side << 1) | uplo ])(&args, NULL, NULL, sa, sb, 0);
#ifdef SMP
} else {
#ifndef NO_AFFINITY
nodes = get_num_nodes();
if (nodes > 1) {
args.nthreads /= nodes;
gemm_thread_mn(MODE, &args, NULL, NULL,
symm[4 | (side << 1) | uplo ], sa, sb, nodes);
} else {
#endif
#ifndef USE_SIMPLE_THREADED_LEVEL3
(symm[4 | (side << 1) | uplo ])(&args, NULL, NULL, sa, sb, 0);
#else
GEMM_THREAD(MODE, &args, NULL, NULL, symm[(side << 1) | uplo ], sa, sb, args.nthreads);
#endif
#ifndef NO_AFFINITY
}
#endif
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(COMPSIZE * COMPSIZE,
(!side)? args.m * (args.m / 2 + args.n) : args.n * (args.m + args.n / 2),
(!side)? 2 * args.m * args.m * args.n : 2 * args.m * args.n * args.n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, char *TRANS, char *DIAG,
blasint *N, FLOAT *a, FLOAT *x, blasint *INCX){
char uplo_arg = *UPLO;
char trans_arg = *TRANS;
char diag_arg = *DIAG;
blasint n = *N;
blasint incx = *INCX;
blasint info;
int uplo;
int unit;
int trans;
FLOAT *buffer;
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
TOUPPER(trans_arg);
TOUPPER(diag_arg);
trans = -1;
unit = -1;
uplo = -1;
if (trans_arg == 'N') trans = 0;
if (trans_arg == 'T') trans = 1;
if (trans_arg == 'R') trans = 0;
if (trans_arg == 'C') trans = 1;
if (diag_arg == 'U') unit = 0;
if (diag_arg == 'N') unit = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 7;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
blasint n, FLOAT *a, FLOAT *x, blasint incx) {
int trans, uplo, unit;
blasint info;
FLOAT *buffer;
PRINT_DEBUG_CNAME;
unit = -1;
uplo = -1;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 7;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 7;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
(tpsv[(trans<<2) | (uplo<<1) | unit])(n, a, x, incx, buffer);
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
void NAME(char *TRANS, blasint *M, blasint *N,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG,
FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T, GEMV_R, GEMV_C,
GEMV_O, GEMV_U, GEMV_S, GEMV_D,
};
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
FLOAT alpha_r = *(ALPHA + 0);
FLOAT alpha_i = *(ALPHA + 1);
FLOAT beta_r = *(BETA + 0);
FLOAT beta_i = *(BETA + 1);
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 2;
if (trans == 'C') i = 3;
if (trans == 'O') i = 4;
if (trans == 'U') i = 5;
if (trans == 'S') i = 6;
if (trans == 'D') i = 7;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1,m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
FLOAT *ALPHA,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT *BETA,
FLOAT *y, blasint incy){
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG,
FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T, GEMV_R, GEMV_C,
GEMV_O, GEMV_U, GEMV_S, GEMV_D,
};
PRINT_DEBUG_CNAME;
FLOAT alpha_r = *(ALPHA + 0);
FLOAT alpha_i = *(ALPHA + 1);
FLOAT beta_r = *(BETA + 0);
FLOAT beta_i = *(BETA + 1);
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 2;
if (TransA == CblasConjTrans) trans = 3;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 3;
if (TransA == CblasConjTrans) trans = 2;
info = -1;
t = n;
n = m;
m = t;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
/* Quick return if possible. */
if (m == 0 || n == 0) return;
lenx = n;
leny = m;
if (trans & 1) lenx = m;
if (trans & 1) leny = n;
if (beta_r != ONE || beta_i != ZERO) SCAL_K(leny, 0, 0, beta_r, beta_i, y, abs(incy), NULL, 0, NULL, 0);
if (alpha_r == ZERO && alpha_i == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx * 2;
if (incy < 0) y -= (leny - 1) * incy * 2;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(gemv[(int)trans])(m, n, 0, alpha_r, alpha_i, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gemv_thread[(int)trans])(m, n, ALPHA, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
float BLASFUNC(sdsdot)(int* n, float* alpha, float* x, int* incx, float* y, int* incy)
{ return *alpha + BLASFUNC(dsdot)(n, x, incx, y, incy); }
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA,
FLOAT *x, blasint *INCX, FLOAT *a){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha = *ALPHA;
blasint incx = *INCX;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_UPLO Uplo,
blasint n,
FLOAT alpha,
FLOAT *x, blasint incx,
FLOAT *a) {
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
info = -1;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(spr[uplo])(n, alpha, x, incx, a, buffer);
#ifdef SMP
} else {
(spr_thread[uplo])(n, alpha, x, incx, a, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, char *TRANS, char *DIAG,
blasint *N, blasint *K,
FLOAT *a, blasint *LDA, FLOAT *x, blasint *INCX){
char uplo_arg = *UPLO;
char trans_arg = *TRANS;
char diag_arg = *DIAG;
blasint n = *N;
blasint k = *K;
blasint lda = *LDA;
blasint incx = *INCX;
blasint info;
int uplo;
int unit;
int trans;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
TOUPPER(trans_arg);
TOUPPER(diag_arg);
trans = -1;
unit = -1;
uplo = -1;
if (trans_arg == 'N') trans = 0;
if (trans_arg == 'T') trans = 1;
if (trans_arg == 'R') trans = 0;
if (trans_arg == 'C') trans = 1;
if (diag_arg == 'U') unit = 0;
if (diag_arg == 'N') unit = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 9;
if (lda < k + 1) info = 7;
if (k < 0) info = 5;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
blasint n, blasint k, FLOAT *a, blasint lda, FLOAT *x, blasint incx) {
int trans, uplo, unit;
blasint info;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
unit = -1;
uplo = -1;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 9;
if (lda < k + 1) info = 7;
if (k < 0) info = 5;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 9;
if (lda < k + 1) info = 7;
if (k < 0) info = 5;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(tbmv[(trans<<2) | (uplo<<1) | unit])(n, k, a, lda, x, incx, buffer);
#ifdef SMP
} else {
(tbmv_thread[(trans<<2) | (uplo<<1) | unit])(n, k, a, lda, x, incx, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, n * k / 2 + n, n * k);
IDEBUG_END;
return;
}
void NAME(char *TRANS, blasint *M, blasint *N,
blasint *KU, blasint *KL,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint ku = *KU;
blasint kl = *KL;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 2;
if (trans == 'C') i = 3;
if (trans == 'O') i = 4;
if (trans == 'U') i = 5;
if (trans == 'S') i = 6;
if (trans == 'D') i = 7;
if (incy == 0) info = 13;
if (incx == 0) info = 10;
if (lda < kl + ku + 1) info = 8;
if (kl < 0) info = 5;
if (ku < 0) info = 4;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
blasint ku, blasint kl,
FLOAT *ALPHA,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT *BETA,
FLOAT *y, blasint incy){
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 2;
if (TransA == CblasConjTrans) trans = 3;
info = -1;
if (incy == 0) info = 13;
if (incx == 0) info = 10;
if (lda < kl + ku + 1) info = 8;
if (kl < 0) info = 5;
if (ku < 0) info = 4;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 3;
if (TransA == CblasConjTrans) trans = 2;
info = -1;
t = n;
n = m;
m = t;
t = ku;
ku = kl;
kl = t;
if (incy == 0) info = 13;
if (incx == 0) info = 10;
if (lda < kl + ku + 1) info = 8;
if (kl < 0) info = 5;
if (ku < 0) info = 4;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if ((m==0) || (n==0)) return;
lenx = n;
leny = m;
if (trans & 1) lenx = m;
if (trans & 1) leny = n;
if (beta_r != ONE || beta_i != ZERO) SCAL_K(leny, 0, 0, beta_r, beta_i, y, abs(incy), NULL, 0, NULL, 0);
if (alpha_r == ZERO && alpha_i == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx * 2;
if (incy < 0) y -= (leny - 1) * incy * 2;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(gbmv[(int)trans])(m, n, kl, ku, alpha_r, alpha_i, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gbmv_thread[(int)trans])(m, n, kl, ku, ALPHA, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, m * n / 2 + n, m * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, FLOAT *ALPHA,
FLOAT *x, blasint *INCX, FLOAT *y, blasint *INCY, FLOAT *a, blasint *LDA){
char uplo_arg = *UPLO;
blasint n = *N;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (lda < MAX(1, n)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo, blasint n, FLOAT *ALPHA, FLOAT *x, blasint incx, FLOAT *y, blasint incy, FLOAT *a, blasint lda) {
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (lda < MAX(1, n)) info = 9;
if (incy == 0) info = 7;
if (incx == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 3;
if (Uplo == CblasLower) uplo = 2;
info = -1;
if (lda < MAX(1, n)) info = 9;
if (incx == 0) info = 7;
if (incy == 0) info = 5;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx * 2;
if (incy < 0 ) y -= (n - 1) * incy * 2;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(her2[uplo])(n, alpha_r, alpha_i, x, incx, y, incy, a, lda, buffer);
#ifdef SMP
} else {
(her2_thread[uplo])(n, ALPHA, x, incx, y, incy, a, lda, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + 2 * n, 2 * n * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, char *TRANS, char *DIAG,
blasint *N, FLOAT *a, blasint *LDA, FLOAT *x, blasint *INCX){
char uplo_arg = *UPLO;
char trans_arg = *TRANS;
char diag_arg = *DIAG;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint info;
int uplo;
int unit;
int trans, buffer_size;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
TOUPPER(trans_arg);
TOUPPER(diag_arg);
trans = -1;
unit = -1;
uplo = -1;
if (trans_arg == 'N') trans = 0;
if (trans_arg == 'T') trans = 1;
if (trans_arg == 'R') trans = 2;
if (trans_arg == 'C') trans = 3;
if (diag_arg == 'U') unit = 0;
if (diag_arg == 'N') unit = 1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
info = 0;
if (incx == 0) info = 8;
if (lda < MAX(1, n)) info = 6;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
blasint n, FLOAT *a, blasint lda, FLOAT *x, blasint incx) {
int trans, uplo, unit, buffer_size;
blasint info;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
PRINT_DEBUG_CNAME;
unit = -1;
uplo = -1;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 2;
if (TransA == CblasConjTrans) trans = 3;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 8;
if (lda < MAX(1, n)) info = 6;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 1;
if (Uplo == CblasLower) uplo = 0;
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 3;
if (TransA == CblasConjTrans) trans = 2;
if (Diag == CblasUnit) unit = 0;
if (Diag == CblasNonUnit) unit = 1;
info = -1;
if (incx == 0) info = 8;
if (lda < MAX(1, n)) info = 6;
if (n < 0) info = 4;
if (unit < 0) info = 3;
if (trans < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx * 2;
#ifdef SMP
// Calibrated on a Xeon E5-2630
if(1L * n * n > 36L * sizeof(FLOAT) * sizeof(FLOAT) * GEMM_MULTITHREAD_THRESHOLD) {
nthreads = num_cpu_avail(2);
if(nthreads > 2 && 1L * n * n < 64L * sizeof(FLOAT) * sizeof(FLOAT) * GEMM_MULTITHREAD_THRESHOLD)
nthreads = 2;
} else
nthreads = 1;
if(nthreads > 1) {
buffer_size = n > 16 ? 0 : n * 4 + 40;
}
else
#endif
{
buffer_size = ((n - 1) / DTB_ENTRIES) * 2 * DTB_ENTRIES + 32 / sizeof(FLOAT);
if(incx != 1)
buffer_size += n * 2;
}
STACK_ALLOC(buffer_size, FLOAT, buffer);
#ifdef SMP
if (nthreads == 1) {
#endif
(trmv[(trans<<2) | (uplo<<1) | unit])(n, a, lda, x, incx, buffer);
#ifdef SMP
} else {
(trmv_thread[(trans<<2) | (uplo<<1) | unit])(n, a, lda, x, incx, buffer, nthreads);
}
#endif
STACK_FREE(buffer);
FUNCTION_PROFILE_END(4, n * n / 2 + n, n * n);
IDEBUG_END;
return;
}
void NAME(char *TRANS, blasint *M, blasint *N,
FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX,
FLOAT *BETA, FLOAT *y, blasint *INCY){
char trans = *TRANS;
blasint m = *M;
blasint n = *N;
blasint lda = *LDA;
blasint incx = *INCX;
blasint incy = *INCY;
FLOAT alpha = *ALPHA;
FLOAT beta = *BETA;
FLOAT *buffer;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
blasint info;
blasint lenx, leny;
blasint i;
PRINT_DEBUG_NAME;
TOUPPER(trans);
info = 0;
i = -1;
if (trans == 'N') i = 0;
if (trans == 'T') i = 1;
if (trans == 'R') i = 0;
if (trans == 'C') i = 1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (i < 0) info = 1;
trans = i;
if (info != 0){
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_TRANSPOSE TransA,
blasint m, blasint n,
FLOAT alpha,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT beta,
FLOAT *y, blasint incy){
FLOAT *buffer;
blasint lenx, leny;
int trans;
blasint info, t;
#ifdef SMP
int nthreads;
#endif
int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG, FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
GEMV_N, GEMV_T,
};
PRINT_DEBUG_CNAME;
trans = -1;
info = 0;
if (order == CblasColMajor) {
if (TransA == CblasNoTrans) trans = 0;
if (TransA == CblasTrans) trans = 1;
if (TransA == CblasConjNoTrans) trans = 0;
if (TransA == CblasConjTrans) trans = 1;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (order == CblasRowMajor) {
if (TransA == CblasNoTrans) trans = 1;
if (TransA == CblasTrans) trans = 0;
if (TransA == CblasConjNoTrans) trans = 1;
if (TransA == CblasConjTrans) trans = 0;
info = -1;
t = n;
n = m;
m = t;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < MAX(1, m)) info = 6;
if (n < 0) info = 3;
if (m < 0) info = 2;
if (trans < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if ((m==0) || (n==0)) return;
lenx = n;
leny = m;
if (trans) lenx = m;
if (trans) leny = n;
if (beta != ONE) SCAL_K(leny, 0, 0, beta, y, abs(incy), NULL, 0, NULL, 0);
if (alpha == ZERO) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0) x -= (lenx - 1) * incx;
if (incy < 0) y -= (leny - 1) * incy;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMP
int nthreads_max = num_cpu_avail(2);
int nthreads_avail = nthreads_max;
double MNK = (double) m * (double) n;
if ( MNK <= (24.0 * 24.0 * (double) (GEMM_MULTITHREAD_THRESHOLD*GEMM_MULTITHREAD_THRESHOLD) ) )
nthreads_max = 1;
if ( nthreads_max > nthreads_avail )
nthreads = nthreads_avail;
else
nthreads = nthreads_max;
if (nthreads == 1) {
#endif
(gemv[(int)trans])(m, n, 0, alpha, a, lda, x, incx, y, incy, buffer);
#ifdef SMP
} else {
(gemv_thread[(int)trans])(m, n, alpha, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
IDEBUG_END;
return;
}
void NAME(char *UPLO, blasint *N, blasint *K, FLOAT *ALPHA, FLOAT *a, blasint *LDA,
FLOAT *x, blasint *INCX, FLOAT *BETA, FLOAT *y, blasint *INCY){
char uplo_arg = *UPLO;
blasint n = *N;
blasint k = *K;
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
blasint lda = *LDA;
blasint incx = *INCX;
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
blasint incy = *INCY;
blasint info;
int uplo;
FLOAT *buffer;
#ifdef SMPBUG
int nthreads;
#endif
PRINT_DEBUG_NAME;
TOUPPER(uplo_arg);
uplo = -1;
if (uplo_arg == 'U') uplo = 0;
if (uplo_arg == 'L') uplo = 1;
if (uplo_arg == 'V') uplo = 2;
if (uplo_arg == 'M') uplo = 3;
info = 0;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < k + 1) info = 6;
if (k < 0) info = 3;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
if (info != 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#else
void CNAME(enum CBLAS_ORDER order,
enum CBLAS_UPLO Uplo,
blasint n, blasint k,
FLOAT *ALPHA,
FLOAT *a, blasint lda,
FLOAT *x, blasint incx,
FLOAT *BETA,
FLOAT *y, blasint incy){
FLOAT alpha_r = ALPHA[0];
FLOAT alpha_i = ALPHA[1];
FLOAT beta_r = BETA[0];
FLOAT beta_i = BETA[1];
FLOAT *buffer;
int uplo;
blasint info;
#ifdef SMPBUG
int nthreads;
#endif
PRINT_DEBUG_CNAME;
uplo = -1;
info = 0;
if (order == CblasColMajor) {
if (Uplo == CblasUpper) uplo = 0;
if (Uplo == CblasLower) uplo = 1;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < k + 1) info = 6;
if (k < 0) info = 3;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (order == CblasRowMajor) {
if (Uplo == CblasUpper) uplo = 3;
if (Uplo == CblasLower) uplo = 2;
info = -1;
if (incy == 0) info = 11;
if (incx == 0) info = 8;
if (lda < k + 1) info = 6;
if (k < 0) info = 3;
if (n < 0) info = 2;
if (uplo < 0) info = 1;
}
if (info >= 0) {
BLASFUNC(xerbla)(ERROR_NAME, &info, sizeof(ERROR_NAME));
return;
}
#endif
if (n == 0) return;
if ((beta_r != ONE) || (beta_i != ZERO)) SCAL_K(n, 0, 0, beta_r, beta_i, y, abs(incy), NULL, 0, NULL, 0);
if ((alpha_r == ZERO) && (alpha_i == ZERO)) return;
IDEBUG_START;
FUNCTION_PROFILE_START();
if (incx < 0 ) x -= (n - 1) * incx * COMPSIZE;
if (incy < 0 ) y -= (n - 1) * incy * COMPSIZE;
buffer = (FLOAT *)blas_memory_alloc(1);
#ifdef SMPBUG
nthreads = num_cpu_avail(2);
if (nthreads == 1) {
#endif
(hbmv[uplo])(n, k, alpha_r, alpha_i, a, lda, x, incx, y, incy, buffer);
#ifdef SMPBUG
} else {
(hbmv_thread[uplo])(n, k, ALPHA, a, lda, x, incx, y, incy, buffer, nthreads);
}
#endif
blas_memory_free(buffer);
FUNCTION_PROFILE_END(4, n * k / 2 + n, n * k);
IDEBUG_END;
return;
}
