/* **************************************************************************** * * ======= * AUTHOR: * ======= * * Michael Holst, Postdoc Email: holst@ama.caltech.edu * Applied Math 217-50 Phone: (818) 395-4549 * Caltech, Pasadena, CA 91125 FAX: (818) 683-3549 * * ======== * PURPOSE: * ======== * * A simple benchmark code to compute raw floating point performance of * common computational kernels occuring in numerical simulation codes. * * Our main interest is in solving partial differential equations, so in * this benchmark we stress vector-vector and SPARSE matrix-vector * operations, as opposed to DENSE matrix-vector operations often used for * quoting benchmarks by others. Sparse matrix (perhaps with structures * such as diagonal bands, but still quite sparse) operations are much more * common than dense matrix operations in any application involving the * numerical solution of ordinary or partial differential equations, and * therefore the floating point performance on sparse matrix operations is * a much more important indicator of simulation speed than dense matrix * performance. However, the latter is often quoted as the floating point * speed of the architecture, using LINPACK-type dense matrix benchmarks, * which is unfortunate since dense matrix operations allow for much higher * levels of cache reuse, and artifically pump up the floating point * performance in the sense that one will never see such performance for * a typical simulation code. This benchmark code is intended to time some * of the kernels that really occur in simulation codes. * * ==== * USE: * ==== * * The only parameter that may need to be set for different machines is the * parameter "irepeat", which determines how many times the loops are * executed for the timing. The larger the number, the longer the test * takes, but the more accurate the timings are. For slower machines you * will need to set this to be a smaller number (such as irepeat=2 for an * Intel 80386) unless you are willing to wait forever. For faster * machines, if you don't set this number large enough (such as at least * irepeat=100 for an Intel P5-90) then you will get inaccurate timings, * and in fact may get floating point exceptions when the code tries to * divide the number of operations by the time taken (time=0 if fast the * machine is fast and irepeat small, or if you timing routine does not * provide high resolution.) * * ================ * SAMPLE MAKEFILE: * ================ * * ######################################################################### * # purpose: makefile for benchmark code. * # author: michael holst * ######################################################################### * ARCH = -DLINUX * TOPLEVC = /usr * CC = gcc * CCFLAGS = -c * COPT = -O2 * CDEBUG = * CLIBPATHS = -L$(TOPLEVC)/lib * CLIBRARIES = * CINCLUDES = -I$(TOPLEVC)/include * CLIBS = $(CLIBPATHS) $(CLIBRARIES) * CFLAGS = $(CINCLUDES) $(COPT) $(CDEBUG) $(ARCH) * bench : bench.o * $(CC) -o bench bench.o $(CLIBS) * .c.o : * $(CC) $(CCFLAGS) $(CFLAGS) $*.c * clean : * rm -f *.o bench * * ================== * CONDITIONS OF USE: * ================== * * Total unrestricted use of the code is granted, provided that the user * be kind enough to forward interesting benchmark results on interesting * machines to the author. In addition, if you must make modifications in * order to run the code on a particular machine, the author would also * like to know exactly what was required so that the code can continue to * become more general. Any modifications should (hopefully) be isolated * to the last routine in the file, "tsecnd()", and can be "#ifdef"-ed such * as was need for SOLARIS and AIX. At the bottom of this comment block * is a list of benchmark results on various architectures, which will be * updated as results on new machines are obtained. * * =========================================== * BENCHMARK RESULTS ON VARIOUS ARCHITECTURES: * =========================================== * * All performance numbers for "gcc -O2" unless otherwise indicated. * Kernels timed are the following. Timings on multiprocessor machines * are for one node only. * * ---------------------------------------------------------------------------- * Benchmarking your machine with... * ---------------------------------------------------------------------------- * Vector Copy D[i] = A[i] * Vector Add D[i] = A[i] + B[i] * Vector Multiply D[i] = A[i] * B[i] * Vector Divide D[i] = A[i] / B[i] * Vector Add-Multiply D[i] = A[i] + B[i] * C[i] * Vector Add-Divide D[i] = A[i] + B[i] / C[i] * Matrix-vector product (5-diagonal sparse matrix) * ---------------------------------------------------------------------------- * * ------------ -------- -------- -------- -------- -------- -------- -------- * Architecture Copy Add Multiply Divide Add-Mult Add-Divi SpMatvec * ------------ -------- -------- -------- -------- -------- -------- -------- * Intel 80386 3.20e-01 1.83e-01 1.28e-01 4.27e-01 5.12e-01 3.66e-01 9.85e-01 * Sun SPARC-2 1.85e+00 1.15e+00 1.10e+00 6.58e-01 1.52e+00 1.10e+00 3.50e+00 * Sun SPARC-5 4.17e+00 2.94e+00 2.70e+00 1.47e+00 4.08e+00 2.47e+00 8.93e+00 * Sun SPARC-10 4.00e+00 3.12e+00 2.70e+00 2.08e+00 6.67e+00 4.55e+00 9.17e+00 * Sun SPARC-20 4.17e+00 4.17e+00 4.17e+00 2.27e+00 5.41e+00 3.85e+00 9.90e+00 * Intel Paragon 4.00e+00 4.00e+00 3.57e+00 2.51e-01 6.06e+00 4.45e-01 9.52e+00 * Intel P5-90 2.43e+00 2.46e+00 2.72e+00 1.45e+00 3.88e+00 2.60e+00 9.57e+00 * IBM RS6000 3.70e+00 6.25e+00 6.25e+00 2.33e+00 9.52e+00 4.88e+00 1.47e+01 * RS6000 (xlc) 1.25e+01 1.25e+01 7.69e+00 2.50e+00 1.25e+01 4.35e+00 3.23e+01 * Your Machine ... ... ... ... ... ... ... * ------------ -------- -------- -------- -------- -------- -------- -------- * * ============== * LAST MODIFIED: * ============== * * 6-18-94 * * ************************************************************************* */ #define irepeat 1000 /* **************************************************************************** * Shouldn't have to change the code below this point, other than possibly * the details of the timing routine "tsecnd()" which appears at the very * end of the code, before the benchmark database. * ************************************************************************* */ #include #include #include #include #include extern double tsecnd(); extern void tstart(); extern void tstop (); #define II2(i,j) ( ( (j-J0) * (I1-I0+1) ) + (i-I0) ) #define nhalf 100 #define n (nhalf * nhalf) #define ndim ( (nhalf+1) * (nhalf+1) ) #define nloops 100 main() { /* vectors for vector-vector and matrix-vector operations */ double a0[ndim], a1[ndim], a2[ndim], a3[ndim], a4[ndim], a5[ndim]; double a6[ndim], a7[ndim], a8[ndim], a9[ndim]; double b0[ndim], b1[ndim], b2[ndim], b3[ndim], b4[ndim], b5[ndim]; double b6[ndim], b7[ndim], b8[ndim], b9[ndim]; double x[ndim], y[ndim]; double oC[ndim], oN[ndim], oE[ndim], oNW[ndim], oNE[ndim]; /* keep track of timings and operation counts */ double rtime[nloops],rscale[nloops]; /* various other business */ double bf, oh, bf_g, oh_g, ttime, scale, s, mflop; long int icount, index, i, j, I0, I1, J0, J1, ij, im1j, ip1j, ijm1, ijp1; long int my_loop; char* loop_name[78]; /* global timing */ tstart(&bf_g, &oh_g); /* Set loop names */ /* '1234567890123456789012345678901234567890' */ loop_name[0] = " Vector Copy D[i] = A[i] "; loop_name[1] = " Vector Add D[i] = A[i] + B[i] "; loop_name[2] = " Vector Multiply D[i] = A[i] * B[i] "; loop_name[3] = " Vector Divide D[i] = A[i] / B[i] "; loop_name[4] = " Vector Add-Multiply D[i] = A[i] + B[i] * C[i] "; loop_name[5] = " Vector Add-Divide D[i] = A[i] + B[i] / C[i] "; loop_name[6] = " Matrix-vector product (5-diagonal sparse matrix) "; /* some i/o */ printf(" * --------------------------------------"); printf("--------------------------------------\n"); printf(" * Benchmarking your machine with...\n"); printf(" * --------------------------------------"); printf("--------------------------------------\n"); for (j=0; j<7; j++) printf(" * %s\n",loop_name[j]); printf(" * --------------------------------------"); printf("--------------------------------------\n"); printf(" * \n"); /* Zero time vector */ for (i=0; i