int main(int argc, char* argv[])
{
int result = EXIT_SUCCESS;
try {
const libxsmm_blasint benchmark = 1 < argc ? std::atoi(argv[1]) : 0;
LIBXSMM_GEMM_CONST libxsmm_blasint m = (2 < argc ? std::atoi(argv[2]) : 23);
LIBXSMM_GEMM_CONST libxsmm_blasint k = (4 < argc ? std::atoi(argv[4]) : m);
LIBXSMM_GEMM_CONST libxsmm_blasint n = (3 < argc ? std::atoi(argv[3]) : k);
const libxsmm_blasint q = (5 < argc ? std::atoi(argv[5]) : 0/*auto*/);
const libxsmm_blasint nrepeat = (6 < argc ? std::atoi(argv[6]) : (0 >= q ? 13 : 1));
LIBXSMM_GEMM_CONST libxsmm_blasint lda = m, ldb = k, ldc = m;
LIBXSMM_GEMM_CONST char transa = 'N', transb = 'N';
LIBXSMM_GEMM_CONST OTYPE alpha = 1, beta = 1;
const libxsmm_blasint asize = PAD(ITYPE, lda * k), bsize = PAD(ITYPE, ldb * n), csize = PAD(OTYPE, ldc * n);
const libxsmm_blasint max_size = ((2ULL << 30/*2 GB*/) / ((asize + bsize) * sizeof(ITYPE) + csize * sizeof(OTYPE)));
const libxsmm_blasint s = LIBXSMM_MIN(0 < q ? q : max_size, max_size);
const libxsmm_blasint aspace = LIBXSMM_ALIGNMENT / sizeof(ITYPE);
const size_t bwsize = static_cast<size_t>((asize/*load*/ + bsize/*load*/) * sizeof(ITYPE) + 2/*RFO*/ * csize * sizeof(OTYPE));
const double gflops = 2E-9 * s * m * n * k;
#if LIBXSMM_TYPEINFO(ITYPE, FP)
const char *const ops = "FLOPS";
const double scale = 1.0 / s;
#else
const char *const ops = "OPS";
const double scale = 1;
#endif
#if !defined(_DEBUG)
const char *const env_check = getenv("CHECK");
const int check = (0 == env_check ? 0 : atoi(env_check));
#else
/*const*/ int check = 1;
#endif
#if defined(LIBXSMM_OFFLOAD_TARGET)
# pragma offload target(LIBXSMM_OFFLOAD_TARGET)
#endif
{
#if defined(_OPENMP)
const libxsmm_blasint chunksize = s / omp_get_max_threads();
#endif
struct raii { // avoid std::vector (first-touch init. causes NUMA issue)
ITYPE *a, *b;
OTYPE *c, *d;
libxsmm_blasint *m_shuffle;
raii(libxsmm_blasint asize_, libxsmm_blasint bsize_, libxsmm_blasint csize_, libxsmm_blasint size_)
: a(new ITYPE[static_cast<size_t>(asize_)]), b(new ITYPE[static_cast<size_t>(bsize_)])
, c(new OTYPE[static_cast<size_t>(csize_)]), d(new OTYPE[static_cast<size_t>(csize_)])
, m_shuffle(new libxsmm_blasint[size_])
{
# if defined(_OPENMP)
# pragma omp parallel for schedule(static)
# endif
for (libxsmm_blasint i = 0; i < size_; ++i) m_shuffle[i] = libxsmm_rand_u32(size_);
}
~raii() { delete[] a; delete[] b; delete[] c; delete[] d; delete[] m_shuffle; }
#if defined(RANDOMIZED)
libxsmm_blasint shuffle(libxsmm_blasint i) const { return m_shuffle[i]; }
#else
libxsmm_blasint shuffle(libxsmm_blasint i) const { return i; }
#endif
} helper(s * asize + aspace - 1, s * bsize + aspace - 1, s * csize + aspace - 1, s);
ITYPE *const a = LIBXSMM_ALIGN(helper.a, LIBXSMM_ALIGNMENT);
ITYPE *const b = LIBXSMM_ALIGN(helper.b, LIBXSMM_ALIGNMENT);
OTYPE *const c = LIBXSMM_ALIGN(helper.c, LIBXSMM_ALIGNMENT);
OTYPE *const d = LIBXSMM_ALIGN(helper.d, LIBXSMM_ALIGNMENT);
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
LIBXSMM_MATINIT(ITYPE, 42 + helper.shuffle(i), a + helper.shuffle(i) * asize, m, k, lda, scale);
LIBXSMM_MATINIT(ITYPE, 24 + helper.shuffle(i), b + helper.shuffle(i) * bsize, k, n, ldb, scale);
LIBXSMM_MATINIT(OTYPE, 22 + i, c + i * csize, m, n, ldc, scale);
LIBXSMM_MATINIT(OTYPE, 22 + i, d + i * csize, m, n, ldc, scale);
}
#if defined(MKL_ENABLE_AVX512)
mkl_enable_instructions(MKL_ENABLE_AVX512);
#endif
// initialize LIBXSMM
libxsmm_init();
fprintf(stdout, "m=%lli n=%lli k=%lli size=%lli memory=%.1f MB (input=%s output=%s)\n\n",
static_cast<long long>(m), static_cast<long long>(n), static_cast<long long>(k), static_cast<long long>(s),
1.0 * (s * ((asize + bsize) * sizeof(ITYPE) + csize * sizeof(OTYPE))) / (1 << 20),
LIBXSMM_TYPENAME(ITYPE), LIBXSMM_TYPENAME(OTYPE));
// LAPACK/BLAS3 (warm-up BLAS Library)
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k,
&alpha, a + helper.shuffle(i) * asize, &lda, b + helper.shuffle(i) * bsize, &ldb,
&beta, c + i * csize, &ldc);
}
#if (defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)) && (LIBXSMM_VERSION3(11, 3, 0) <= INTEL_MKL_VERSION)
std::vector<const ITYPE*> va_array(static_cast<size_t>(s)), vb_array(static_cast<size_t>(s));
std::vector<OTYPE*> vc_array(static_cast<size_t>(s));
const ITYPE* *const a_array = &va_array[0];
const ITYPE* *const b_array = &vb_array[0];
OTYPE* *const c_array = &vc_array[0];
const libxsmm_blasint group_count = 1;
for (libxsmm_blasint i = 0; i < s; ++i) { // setup batched (A,B,C)
a_array[i] = a + helper.shuffle(i) * asize; b_array[i] = b + helper.shuffle(i) * bsize; c_array[i] = d + i * csize;
}
// additional warm-up (also to eventually match the Gold result)
LIBXSMM_TPREFIX(ITYPE,gemm_batch)(&transa, &transb, &m, &n, &k,
&alpha, &a_array[0], &lda, &b_array[0], &ldb,
&beta, &c_array[0], &ldc, &group_count, &s);
#endif
switch (benchmark) {
case 0: { // batched
fprintf(stdout, "Batched (A,B,C)...\n");
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k,
&alpha, a + helper.shuffle(i) * asize, &lda, b + helper.shuffle(i) * bsize, &ldb,
&beta, c + i * csize, &ldc);
}
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * bwsize / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
} /* fallthrough */
#if (defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)) && (LIBXSMM_VERSION3(11, 3, 0) <= INTEL_MKL_VERSION)
case 1: { // batched indirect
fprintf(stdout, "Indirect (A,B,C)...\n");
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
LIBXSMM_TPREFIX(ITYPE,gemm_batch)(&transa, &transb, &m, &n, &k,
&alpha, &a_array[0], &lda, &b_array[0], &ldb,
&beta, &c_array[0], &ldc, &group_count, &s);
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * bwsize / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
if (0 == benchmark) { /* Gold result is available */
libxsmm_matdiff_info diff;
memset(&diff, 0, sizeof(diff));
for (libxsmm_blasint h = 0; h < s; ++h) {
const OTYPE *const u = c + h * csize, *const v = c_array[h];
libxsmm_matdiff_info dv;
if (EXIT_SUCCESS == libxsmm_matdiff(LIBXSMM_DATATYPE(OTYPE), m, n, u, v, &ldc, &ldc, &dv)) {
libxsmm_matdiff_reduce(&diff, &dv);
}
}
if (0 < diff.normf_rel) fprintf(stdout, "\tdiff: %.0f%%\n", 100.0 * diff.normf_rel);
}
}
#endif
break;
case 2: { // streaming A and C
fprintf(stdout, "Streamed (A,C)...\n");
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k,
&alpha, a + helper.shuffle(i) * asize, &lda, b, &ldb,
&beta, c + i * csize, &ldc);
}
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * (bwsize - bsize * sizeof(ITYPE)) / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
} /* fallthrough */
#if (defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)) && (LIBXSMM_VERSION3(11, 3, 0) <= INTEL_MKL_VERSION)
case 3: { // indirect A and C
fprintf(stdout, "Indirect (A,C)...\n");
for (libxsmm_blasint i = 0; i < s; ++i) { a_array[i] = a + helper.shuffle(i) * asize; b_array[i] = b; c_array[i] = d + i * csize; }
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
LIBXSMM_TPREFIX(ITYPE, gemm_batch)(&transa, &transb, &m, &n, &k,
&alpha, &a_array[0], &lda, &b_array[0], &ldb,
&beta, &c_array[0], &ldc, &group_count, &s);
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * (bwsize - bsize * sizeof(ITYPE)) / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
}
#endif
break;
case 4: { // streaming B and C
fprintf(stdout, "Streamed (B,C)...\n");
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k,
&alpha, a, &lda, b + helper.shuffle(i) * bsize, &ldb,
&beta, c + i * csize, &ldc);
}
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * (bwsize - asize * sizeof(ITYPE)) / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
} /* fallthrough */
#if (defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)) && (LIBXSMM_VERSION3(11, 3, 0) <= INTEL_MKL_VERSION)
case 5: { // indirect B and C
fprintf(stdout, "Indirect (B,C)...\n");
for (libxsmm_blasint i = 0; i < s; ++i) { a_array[i] = a; b_array[i] = b + helper.shuffle(i) * bsize; c_array[i] = d + i * csize; }
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
LIBXSMM_TPREFIX(ITYPE, gemm_batch)(&transa, &transb, &m, &n, &k,
&alpha, &a_array[0], &lda, &b_array[0], &ldb,
&beta, &c_array[0], &ldc, &group_count, &s);
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * (bwsize - asize * sizeof(ITYPE)) / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
}
#endif
break;
case 6: { // streaming A and B
fprintf(stdout, "Streamed (A,B)...\n");
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
#if defined(_OPENMP) /* attempt to write to disjunct cachelines */
const libxsmm_blasint j = omp_get_thread_num() * chunksize * csize;
#else
const libxsmm_blasint j = 0;
#endif
LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k,
&alpha, a + helper.shuffle(i) * asize, &lda, b + helper.shuffle(i) * bsize, &ldb,
&beta, c + j, &ldc);
}
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * (bwsize - 2 * csize * sizeof(OTYPE)) / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
} /* fallthrough */
#if (defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)) && (LIBXSMM_VERSION3(11, 3, 0) <= INTEL_MKL_VERSION)
case 7: { // indirect A and B
fprintf(stdout, "Indirect (A,B)...\n");
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
a_array[i] = a + helper.shuffle(i) * asize; b_array[i] = b + helper.shuffle(i) * bsize;
#if defined(_OPENMP) /* attempt to write to disjunct cachelines */
c_array[i] = d + omp_get_thread_num() * chunksize * csize;
#else
c_array[i] = d;
#endif
}
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
LIBXSMM_TPREFIX(ITYPE, gemm_batch)(&transa, &transb, &m, &n, &k,
&alpha, &a_array[0], &lda, &b_array[0], &ldb,
&beta, &c_array[0], &ldc, &group_count, &s);
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", s * (bwsize - 2 * csize * sizeof(OTYPE)) / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
}
#endif
break;
case 8: { // cached
fprintf(stdout, "Cached...\n");
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
#if defined(_OPENMP) /* attempt to write to disjunct cachelines */
const libxsmm_blasint j = omp_get_thread_num() * chunksize * csize;
#else
const libxsmm_blasint j = 0;
#endif
LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k,
&alpha, a, &lda, b, &ldb, &beta, c + j, &ldc);
}
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
} /* fallthrough */
#if (defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)) && (LIBXSMM_VERSION3(11, 3, 0) <= INTEL_MKL_VERSION)
case 9: { // indirect cached
fprintf(stdout, "Indirect cached...\n");
#if defined(_OPENMP)
# pragma omp parallel for schedule(static)
#endif
for (libxsmm_blasint i = 0; i < s; ++i) {
a_array[i] = a; b_array[i] = b;
#if defined(_OPENMP) /* attempt to write to disjunct cachelines */
c_array[i] = d + omp_get_thread_num() * chunksize * csize;
#else
c_array[i] = d;
#endif
}
const unsigned long long start = libxsmm_timer_tick();
for (libxsmm_blasint r = 0; r < nrepeat; ++r) {
LIBXSMM_TPREFIX(ITYPE, gemm_batch)(&transa, &transb, &m, &n, &k,
&alpha, &a_array[0], &lda, &b_array[0], &ldb,
&beta, &c_array[0], &ldc, &group_count, &s);
}
const unsigned long long ncycles = libxsmm_timer_diff(start, libxsmm_timer_tick());
const double duration = libxsmm_timer_duration(0, ncycles) / nrepeat;
if (0 < duration && 0 != ncycles) {
fprintf(stdout, "\tpseudo-perf.: %.1f %s/cycle\n", (2 * k - 1) * (double)(s * m * n) / ncycles, ops);
fprintf(stdout, "\tperformance: %.1f G%s/s\n", gflops / duration, ops);
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
}
#endif
break;
default: throw "invalid case selected!";
} /*switch*/
if (0 != check) {
libxsmm_matdiff_info diff;
if (EXIT_SUCCESS == libxsmm_matdiff(LIBXSMM_DATATYPE(OTYPE), m, n, 0 == (benchmark & 1) ? c : d, NULL, &ldc, &ldc, &diff)) {
fprintf(stdout, "\tcheck: %f\n", diff.l1_ref);
}
}
// finalize LIBXSMM
libxsmm_finalize();
fprintf(stdout, "Finished\n");
}
}
catch(const std::exception& e) {
fprintf(stderr, "Error: %s\n", e.what());
result = EXIT_FAILURE;
}
catch(const char* message) {
fprintf(stderr, "Error: %s\n", message);
result = EXIT_FAILURE;
}
catch(...) {
fprintf(stderr, "Error: unknown exception caught!\n");
result = EXIT_FAILURE;
}
return result;
}
int main(int argc, char* argv[])
{
const libxsmm_blasint m = 1 < argc ? atoi(argv[1]) : 4096;
const libxsmm_blasint n = 2 < argc ? atoi(argv[2]) : m;
const libxsmm_blasint lda = LIBXSMM_MAX(3 < argc ? atoi(argv[3]) : 0, m);
const libxsmm_blasint ldb = LIBXSMM_MAX(4 < argc ? atoi(argv[4]) : 0, n);
REAL_TYPE *const a = (REAL_TYPE*)malloc(lda * n * sizeof(REAL_TYPE));
REAL_TYPE *const b = (REAL_TYPE*)malloc(ldb * m * sizeof(REAL_TYPE));
const unsigned int size = m * n * sizeof(REAL_TYPE);
unsigned long long start;
libxsmm_blasint i, j;
double duration;
fprintf(stdout, "m=%i n=%i lda=%i ldb=%i size=%.fMB (%s)\n", m, n, lda, ldb,
1.0 * size / (1 << 20), 8 == sizeof(REAL_TYPE) ? "DP" : "SP");
for (i = 0; i < n; ++i) {
for (j = 0; j < m; ++j) {
a[i*lda+j] = initial_value(i, j, lda);
}
}
start = libxsmm_timer_tick();
libxsmm_transpose_oop(b, a, sizeof(REAL_TYPE), m, n, lda, ldb);
libxsmm_transpose_oop(a, b, sizeof(REAL_TYPE), n, m, ldb, lda);
duration = libxsmm_timer_duration(start, libxsmm_timer_tick());
for (i = 0; i < n; ++i) {
for (j = 0; j < m; ++j) {
if (0 < fabs(a[i*lda+j] - initial_value(i, j, lda))) {
i = n + 1;
break;
}
}
}
if (i <= n) {
if (0 < duration) {
fprintf(stdout, "\tbandwidth: %.1f GB/s\n", size / (duration * (1 << 30)));
}
fprintf(stdout, "\tduration: %.0f ms\n", 1000.0 * duration);
}
else {
fprintf(stderr, "Validation failed!\n");
}
#if defined(__MKL) || defined(MKL_DIRECT_CALL_SEQ) || defined(MKL_DIRECT_CALL)
{
double mkl_duration;
start = libxsmm_timer_tick();
LIBXSMM_CONCATENATE(mkl_, LIBXSMM_TPREFIX(REAL_TYPE, omatcopy))('C', 'T', m, n, 1, a, lda, b, ldb);
LIBXSMM_CONCATENATE(mkl_, LIBXSMM_TPREFIX(REAL_TYPE, omatcopy))('C', 'T', n, m, 1, b, ldb, a, lda);
mkl_duration = libxsmm_timer_duration(start, libxsmm_timer_tick());
if (0 < mkl_duration) {
fprintf(stdout, "\tMKL: %.1fx\n", duration / mkl_duration);
}
}
#endif
free(a);
free(b);
return EXIT_SUCCESS;
}