void terminate(void)
{
fprintf(stderr, "unpartition !!\n");
starpu_data_unpartition(C_handle, 0);
starpu_data_unregister(C_handle);
gettimeofday(&end, NULL);
double timing = (double)((end.tv_sec - start.tv_sec)*1000000 + (end.tv_usec - start.tv_usec));
display_stats(timing);
#ifdef CHECK_OUTPUT
/* check results */
/* compute C = C - AB */
SGEMM("N", "N", ydim, xdim, zdim, -1.0f, A, ydim, B, zdim, 1.0f, C, ydim);
/* make sure C = 0 */
float err;
err = SASUM(xdim*ydim, C, 1);
if (err < xdim*ydim*0.001) {
fprintf(stderr, "Results are OK\n");
}
else {
fprintf(stderr, "There were errors ... err = %f\n", err);
}
#endif // CHECK_OUTPUT
}
int
slacon_(int *n, float *v, float *x, int *isgn, float *est, int *kase)
{
/* Table of constant values */
int c__1 = 1;
float zero = 0.0;
float one = 1.0;
/* Local variables */
static int iter;
static int jump, jlast;
static float altsgn, estold;
static int i, j;
float temp;
#ifdef _CRAY
extern int ISAMAX(int *, float *, int *);
extern float SASUM(int *, float *, int *);
extern int SCOPY(int *, float *, int *, float *, int *);
#else
extern int isamax_(int *, float *, int *);
extern float sasum_(int *, float *, int *);
extern int scopy_(int *, float *, int *, float *, int *);
#endif
#define d_sign(a, b) (b >= 0 ? fabs(a) : -fabs(a)) /* Copy sign */
#define i_dnnt(a) \
( a>=0 ? floor(a+.5) : -floor(.5-a) ) /* Round to nearest integer */
if ( *kase == 0 ) {
for (i = 0; i < *n; ++i) {
x[i] = 1. / (float) (*n);
}
*kase = 1;
jump = 1;
return 0;
}
switch (jump) {
case 1: goto L20;
case 2: goto L40;
case 3: goto L70;
case 4: goto L110;
case 5: goto L140;
}
/* ................ ENTRY (JUMP = 1)
FIRST ITERATION. X HAS BEEN OVERWRITTEN BY A*X. */
L20:
if (*n == 1) {
v[0] = x[0];
*est = fabs(v[0]);
/* ... QUIT */
goto L150;
}
#ifdef _CRAY
*est = SASUM(n, x, &c__1);
#else
*est = sasum_(n, x, &c__1);
#endif
for (i = 0; i < *n; ++i) {
x[i] = d_sign(one, x[i]);
isgn[i] = i_dnnt(x[i]);
}
*kase = 2;
jump = 2;
return 0;
/* ................ ENTRY (JUMP = 2)
FIRST ITERATION. X HAS BEEN OVERWRITTEN BY TRANSPOSE(A)*X. */
L40:
#ifdef _CRAY
j = ISAMAX(n, &x[0], &c__1);
#else
j = isamax_(n, &x[0], &c__1);
#endif
--j;
iter = 2;
/* MAIN LOOP - ITERATIONS 2,3,...,ITMAX. */
L50:
for (i = 0; i < *n; ++i) x[i] = zero;
x[j] = one;
*kase = 1;
jump = 3;
return 0;
/* ................ ENTRY (JUMP = 3)
X HAS BEEN OVERWRITTEN BY A*X. */
L70:
#ifdef _CRAY
SCOPY(n, x, &c__1, v, &c__1);
#else
scopy_(n, x, &c__1, v, &c__1);
#endif
estold = *est;
#ifdef _CRAY
*est = SASUM(n, v, &c__1);
#else
*est = sasum_(n, v, &c__1);
#endif
for (i = 0; i < *n; ++i)
if (i_dnnt(d_sign(one, x[i])) != isgn[i])
goto L90;
/* REPEATED SIGN VECTOR DETECTED, HENCE ALGORITHM HAS CONVERGED. */
goto L120;
L90:
/* TEST FOR CYCLING. */
if (*est <= estold) goto L120;
for (i = 0; i < *n; ++i) {
x[i] = d_sign(one, x[i]);
isgn[i] = i_dnnt(x[i]);
}
*kase = 2;
jump = 4;
return 0;
/* ................ ENTRY (JUMP = 4)
X HAS BEEN OVERWRITTEN BY TRANDPOSE(A)*X. */
L110:
jlast = j;
#ifdef _CRAY
j = ISAMAX(n, &x[0], &c__1);
#else
j = isamax_(n, &x[0], &c__1);
#endif
--j;
if (x[jlast] != fabs(x[j]) && iter < 5) {
++iter;
goto L50;
}
/* ITERATION COMPLETE. FINAL STAGE. */
L120:
altsgn = 1.;
for (i = 1; i <= *n; ++i) {
x[i-1] = altsgn * ((float)(i - 1) / (float)(*n - 1) + 1.);
altsgn = -altsgn;
}
*kase = 1;
jump = 5;
return 0;
/* ................ ENTRY (JUMP = 5)
X HAS BEEN OVERWRITTEN BY A*X. */
L140:
#ifdef _CRAY
temp = SASUM(n, x, &c__1) / (float)(*n * 3) * 2.;
#else
temp = sasum_(n, x, &c__1) / (float)(*n * 3) * 2.;
#endif
if (temp > *est) {
#ifdef _CRAY
SCOPY(n, &x[0], &c__1, &v[0], &c__1);
#else
scopy_(n, &x[0], &c__1, &v[0], &c__1);
#endif
*est = temp;
}
L150:
*kase = 0;
return 0;
} /* slacon_ */
int
dlacon_(int *n, double *v, double *x, int *isgn, double *est, int *kase)
{
/*
Purpose
=======
DLACON estimates the 1-norm of a square matrix A.
Reverse communication is used for evaluating matrix-vector products.
Arguments
=========
N (input) INT
The order of the matrix. N >= 1.
V (workspace) DOUBLE PRECISION array, dimension (N)
On the final return, V = A*W, where EST = norm(V)/norm(W)
(W is not returned).
X (input/output) DOUBLE PRECISION array, dimension (N)
On an intermediate return, X should be overwritten by
A * X, if KASE=1,
A' * X, if KASE=2,
and DLACON must be re-called with all the other parameters
unchanged.
ISGN (workspace) INT array, dimension (N)
EST (output) DOUBLE PRECISION
An estimate (a lower bound) for norm(A).
KASE (input/output) INT
On the initial call to DLACON, KASE should be 0.
On an intermediate return, KASE will be 1 or 2, indicating
whether X should be overwritten by A * X or A' * X.
On the final return from DLACON, KASE will again be 0.
Further Details
======= =======
Contributed by Nick Higham, University of Manchester.
Originally named CONEST, dated March 16, 1988.
Reference: N.J. Higham, "FORTRAN codes for estimating the one-norm of
a real or complex matrix, with applications to condition estimation",
ACM Trans. Math. Soft., vol. 14, no. 4, pp. 381-396, December 1988.
=====================================================================
*/
/* Table of constant values */
int c__1 = 1;
double zero = 0.0;
double one = 1.0;
/* Local variables */
static int iter;
static int jump, jlast;
static double altsgn, estold;
static int i, j;
double temp;
#ifdef _CRAY
extern int ISAMAX(int *, double *, int *);
extern double SASUM(int *, double *, int *);
extern int SCOPY(int *, double *, int *, double *, int *);
#else
extern int idamax_(int *, double *, int *);
extern double dasum_(int *, double *, int *);
extern int dcopy_(int *, double *, int *, double *, int *);
#endif
#define d_sign(a, b) (b >= 0 ? fabs(a) : -fabs(a)) /* Copy sign */
#define i_dnnt(a) \
( a>=0 ? floor(a+.5) : -floor(.5-a) ) /* Round to nearest integer */
if ( *kase == 0 ) {
for (i = 0; i < *n; ++i) {
x[i] = 1. / (double) (*n);
}
*kase = 1;
jump = 1;
return 0;
}
switch (jump) {
case 1: goto L20;
case 2: goto L40;
case 3: goto L70;
case 4: goto L110;
case 5: goto L140;
}
/* ................ ENTRY (JUMP = 1)
FIRST ITERATION. X HAS BEEN OVERWRITTEN BY A*X. */
L20:
if (*n == 1) {
v[0] = x[0];
*est = fabs(v[0]);
/* ... QUIT */
goto L150;
}
#ifdef _CRAY
*est = SASUM(n, x, &c__1);
#else
*est = dasum_(n, x, &c__1);
#endif
for (i = 0; i < *n; ++i) {
x[i] = d_sign(one, x[i]);
isgn[i] = i_dnnt(x[i]);
}
*kase = 2;
jump = 2;
return 0;
/* ................ ENTRY (JUMP = 2)
FIRST ITERATION. X HAS BEEN OVERWRITTEN BY TRANSPOSE(A)*X. */
L40:
#ifdef _CRAY
j = ISAMAX(n, &x[0], &c__1);
#else
j = idamax_(n, &x[0], &c__1);
#endif
--j;
iter = 2;
/* MAIN LOOP - ITERATIONS 2,3,...,ITMAX. */
L50:
for (i = 0; i < *n; ++i) x[i] = zero;
x[j] = one;
*kase = 1;
jump = 3;
return 0;
/* ................ ENTRY (JUMP = 3)
X HAS BEEN OVERWRITTEN BY A*X. */
L70:
#ifdef _CRAY
SCOPY(n, x, &c__1, v, &c__1);
#else
dcopy_(n, x, &c__1, v, &c__1);
#endif
estold = *est;
#ifdef _CRAY
*est = SASUM(n, v, &c__1);
#else
*est = dasum_(n, v, &c__1);
#endif
for (i = 0; i < *n; ++i)
if (i_dnnt(d_sign(one, x[i])) != isgn[i])
goto L90;
/* REPEATED SIGN VECTOR DETECTED, HENCE ALGORITHM HAS CONVERGED. */
goto L120;
L90:
/* TEST FOR CYCLING. */
if (*est <= estold) goto L120;
for (i = 0; i < *n; ++i) {
x[i] = d_sign(one, x[i]);
isgn[i] = i_dnnt(x[i]);
}
*kase = 2;
jump = 4;
return 0;
/* ................ ENTRY (JUMP = 4)
X HAS BEEN OVERWRITTEN BY TRANDPOSE(A)*X. */
L110:
jlast = j;
#ifdef _CRAY
j = ISAMAX(n, &x[0], &c__1);
#else
j = idamax_(n, &x[0], &c__1);
#endif
--j;
if (x[jlast] != fabs(x[j]) && iter < 5) {
++iter;
goto L50;
}
/* ITERATION COMPLETE. FINAL STAGE. */
L120:
altsgn = 1.;
for (i = 1; i <= *n; ++i) {
x[i-1] = altsgn * ((double)(i - 1) / (double)(*n - 1) + 1.);
altsgn = -altsgn;
}
*kase = 1;
jump = 5;
return 0;
/* ................ ENTRY (JUMP = 5)
X HAS BEEN OVERWRITTEN BY A*X. */
L140:
#ifdef _CRAY
temp = SASUM(n, x, &c__1) / (double)(*n * 3) * 2.;
#else
temp = dasum_(n, x, &c__1) / (double)(*n * 3) * 2.;
#endif
if (temp > *est) {
#ifdef _CRAY
SCOPY(n, &x[0], &c__1, &v[0], &c__1);
#else
dcopy_(n, &x[0], &c__1, &v[0], &c__1);
#endif
*est = temp;
}
L150:
*kase = 0;
return 0;
} /* dlacon_ */
int main( int argc, char *argv[] )
{
unsigned int numspes = 1, i;
unsigned int numblocks = 8, blocksize = 4;
unsigned int type = 1;
// There are arguments
if ( argc > 1 )
{
// The first argument is present
if ( argc > 1 )
{
numblocks = atoi( argv[1] );
}
if ( argc > 2 )
{
blocksize = atoi( argv[2] );
}
if ( argc > 3 )
{
type = atoi( argv[3] );
}
if ( argc > 4 )
{
numspes = atoi( argv[4] );
}
}
else
{
printf( "Usage pputest <numblocks> <blocksize^2> <type> <num spes>\n" );
printf( "type: 1: sdot, 2: sdotv, 3: snrm2, 4: snrm2v\n" );
return -1;
}
init( numspes );
paddingx = 0;
paddingy = 1;
unsigned int size, numelements = numblocks*blocksize*blocksize;
printf( "Testing BLAS()\n" );
printf( "Num SPEs:\t%u\n", speThreads );
printf( "------------------\n" );
printf( "Vector size: \t\t%u\n", blocksize*blocksize*numblocks-paddingx );
printf( "Num blocks:\t\t%u\n", numblocks );
printf( "Num elements pr block: \t%u\n", blocksize*blocksize );
printf( "1 block in bytes: \t%u\n", blocksize*blocksize*4 );
printf( "Num elements:\t\t%u\n", numblocks*blocksize*blocksize );
printf( "Total size in MB: \t%f\n", (double)(numblocks*blocksize*blocksize*4)/(1024*1024) );
printf( "Total size in MB: \t%f\n", (double)(numblocks*blocksize*blocksize*4)/(1024*1024)*2 );
printf( "------------------\n" );
// PyArrayObject
PyArrayObject pyobj1;
PyArrayObject pyobj2;
PyArrayObject pyscalar1;
sMakeMatrix( numblocks, 1, blocksize, 1.0f, &pyobj1 );
sMakeMatrix( numblocks, 1, blocksize, 1.0f, &pyobj2 );
sMakeMatrix( 1, 1, 4, 2.334f, &pyscalar1 );
// printf( "First block address=%#x\n", pyobj1.blockData[0] );
// printf( "Sum is %f\n", SumBlock( pyobj1.blockData[0], blocksize ) );
unsigned int *shader;
double time;
switch( type )
{
case 1:
shader = blas_1_sdot;
printf( "Calling SDOT();\n" );
size = (numblocks*blocksize*blocksize*4) * 2;
time = SDOT( &pyobj1, &pyobj2, blas_1_sdot_size, shader );
break;
case 2:
shader = blas_1_sdotv;
printf( "Calling SDOTvvvv();\n" );
size = (numblocks*blocksize*blocksize*4) * 2;
time = SDOT( &pyobj1, &pyobj2, blas_1_sdotv_size, shader );
break;
case 3:
shader = snrm2;
printf( "Calling SNRM2();\n" );
size = (numblocks*blocksize*blocksize*4);
time = SNRM2( &pyobj1, snrm2_size, shader );
break;
case 4:
shader = snrm2v;
printf( "Calling SNRM2vvvvv();\n" );
size = (numblocks*blocksize*blocksize*4);
time = SNRM2( &pyobj1, snrm2v_size, shader );
break;
case 5:
shader = blas_1_sscal;
printf( "Calling SSCAL();\n" );
size = (numblocks*blocksize*blocksize*4);
time = SSCAL( &pyobj1, &pyscalar1, blas_1_sscal_size, shader );
break;
case 7:
shader = blas_1_sasum;
printf( "Calling SASUM();\n" );
size = (numblocks*blocksize*blocksize*4);
time = SASUM( &pyobj1, blas_1_sasum_size, shader );
break;
case 8:
shader = blas_1_sasumv;
printf( "Calling SASUMv();\n" );
size = (numblocks*blocksize*blocksize*4);
time = SASUM( &pyobj1, blas_1_sasumv_size, shader );
break;
case 9:
shader = blas_1_isamaxv;
printf( "Calling ISAMAXv();\n" );
size = (numblocks*blocksize*blocksize*4);
//ISAMAX( &pyobj1,blas_1_isamaxv_size, shader );
break;
case 102:
shader = blas_1_sdotv;
printf( "Calling SDOT2vvvv();\n" );
size = (numblocks*blocksize*blocksize*4) * 2;
time = SDOT2( &pyobj1, &pyobj2, blas_1_sdotv_size, shader );
break;
}
unsigned int state = 0;
for ( i = 0 ; i < speThreads ; i++ )
{
spe_in_mbox_write ( speData[i].spe_ctx, &state, 1, SPE_MBOX_ALL_BLOCKING );
}
// Wait for all the SPE threads to complete.
for ( i = 0 ; i < speThreads ; i++ )
{
CompleteSPEThreads( &speData[i] );
}
double GB = (double)size / (1024*1024*1024);
printf( "Time: %f\n", time );
printf( "Size: %u\n", size );
printf( "GigaBytes: %f\n", GB );
printf( "GB/s: %f\n", GB/time );
printf( "GFlops: %f\n", ( numelements/time ) / ( 1024*1024*1024 ) );
return 1;
}