int main(int argc, char* argv[]) { LIBXSMM_GEMM_CONST libxsmm_blasint m = (1 < argc ? atoi(argv[1]) : 1024); LIBXSMM_GEMM_CONST libxsmm_blasint k = (3 < argc ? atoi(argv[3]) : m); LIBXSMM_GEMM_CONST libxsmm_blasint n = (2 < argc ? atoi(argv[2]) : k); const libxsmm_blasint bm = (4 < argc ? atoi(argv[4]) : 32); const libxsmm_blasint bk = (6 < argc ? atoi(argv[6]) : bm); const libxsmm_blasint bn = (5 < argc ? atoi(argv[5]) : bk); const libxsmm_bgemm_order order = (libxsmm_bgemm_order)(7 < argc ? atoi(argv[7]) : 0); const int nrepeat = (8 < argc ? atoi(argv[8]) : 100); const libxsmm_blasint b_m1 = (9 < argc ? atoi(argv[9]) : 1); const libxsmm_blasint b_n1 = (10 < argc ? atoi(argv[10]) : 1); const libxsmm_blasint b_k1 = (11 < argc ? atoi(argv[11]) : 1); const libxsmm_blasint b_k2 = (12 < argc ? atoi(argv[12]) : 1); const int ab = (13 < argc ? atoi(argv[13]) : 0); LIBXSMM_GEMM_CONST libxsmm_blasint lda = (14 < argc ? atoi(argv[13]) : m); LIBXSMM_GEMM_CONST libxsmm_blasint ldb = (15 < argc ? atoi(argv[14]) : k); LIBXSMM_GEMM_CONST libxsmm_blasint ldc = (16 < argc ? atoi(argv[15]) : m); LIBXSMM_GEMM_CONST char transa = 'N', transb = 'N'; /* no transposes */ LIBXSMM_GEMM_CONST ITYPE alpha = 1, beta = 1; const int gemm_flags = LIBXSMM_GEMM_FLAGS(transa, transb); const double gflops = 2.0 * m * n * k * 1E-9; int result = EXIT_SUCCESS; #if defined(CHECK) const char *const env_check = getenv("CHECK"); const double check = LIBXSMM_ABS(0 == env_check ? 0 : atof(env_check)); #endif if (argc > 1 && !strncmp(argv[1], "-h", 3)) { /* check command line */ printf("\nUsage: ./bgemm [M] [N] [K] [bm] [bn] [bk] [order] [reps] [b_m1] [b_n1] [b_k1] [b_k2] [verbose]\n\n"); return result; } MYASSERT(m % b_m1 == 0); MYASSERT(n % b_n1 == 0); MYASSERT(k % b_k1 == 0); MYASSERT(m/b_m1 % bm == 0); MYASSERT(n/b_n1 % bn == 0); MYASSERT(k/b_k1/b_k2 % bk == 0); #if defined(LIBXSMM_OFFLOAD_TARGET) # pragma offload target(LIBXSMM_OFFLOAD_TARGET) #endif { ITYPE* agold = (ITYPE*)libxsmm_malloc((size_t)(lda * k * sizeof(ITYPE))); ITYPE* bgold = (ITYPE*)libxsmm_malloc((size_t)(ldb * n * sizeof(ITYPE))); ITYPE* cgold = (ITYPE*)libxsmm_malloc((size_t)(ldc * n * sizeof(ITYPE))); ITYPE* a = (ITYPE*)libxsmm_malloc((size_t)(m * k * sizeof(ITYPE))); ITYPE* b = (ITYPE*)libxsmm_malloc((size_t)(k * n * sizeof(ITYPE))); ITYPE* c = (ITYPE*)libxsmm_malloc((size_t)(m * n * sizeof(ITYPE))); libxsmm_bgemm_handle* handle = 0; unsigned long long start; double duration; handle = libxsmm_bgemm_handle_create( LIBXSMM_GEMM_PRECISION(ITYPE), LIBXSMM_GEMM_PRECISION(ITYPE), m, n, k, &bm, &bn, &bk, &b_m1, &b_n1, &b_k1, &b_k2, &alpha, &beta, &gemm_flags, NULL/*auto-prefetch*/, &order); if (0 != handle) { LIBXSMM_MATINIT(ITYPE, 42, agold, m, k, lda, 1.0); LIBXSMM_MATINIT(ITYPE, 24, bgold, k, n, ldb, 1.0); LIBXSMM_MATINIT(ITYPE, 0, cgold, m, n, ldc, 1.0); libxsmm_bgemm_copyin_a(handle, agold, &lda, a); libxsmm_bgemm_copyin_b(handle, bgold, &ldb, b); libxsmm_bgemm_copyin_c(handle, cgold, &ldc, c); #if defined(MKL_ENABLE_AVX512) mkl_enable_instructions(MKL_ENABLE_AVX512); #endif /* warm-up OpenMP (populate thread pool) */ libxsmm_bgemm_omp(handle, a, b, c, 1); #if defined(CHECK) if (!LIBXSMM_FEQ(0, check)) { LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k, &alpha, agold, &lda, bgold, &ldb, &beta, cgold, &ldc); } #endif if (!ab) { libxsmm_gemm_print(stdout, LIBXSMM_GEMM_PRECISION(ITYPE), &transa, &transb, &m, &n, &k, &alpha, a, &lda, b, &ldb, &beta, c, &ldc); fprintf(stdout, "\n\n"); } start = libxsmm_timer_tick(); libxsmm_bgemm_omp(handle, a, b, c, nrepeat); duration = libxsmm_timer_duration(start, libxsmm_timer_tick()); if (0 < duration) { if (ab) { fprintf(stdout, "\tLIBXSMM: %.1f GFLOPS/s | %lli,%lli,%lli,%lli,%lli,%lli,%i,%lli,%lli,%lli,%lli\n", gflops * nrepeat / duration, (long long)m, (long long)n, (long long)k, (long long)bm, (long long)bn, (long long)bk, (int)order, (long long)b_m1, (long long)b_n1, (long long)b_k1, (long long)b_k2); } else { fprintf(stdout, "\tLIBXSMM: %.1f GFLOPS/s\n", gflops * nrepeat / duration); } } #if defined(CHECK) if (!LIBXSMM_FEQ(0, check)) { /* validate result against LAPACK/BLAS xGEMM */ ITYPE* ctest = 0; int i; start = libxsmm_timer_tick(); for (i = 0; i < nrepeat; ++i) { LIBXSMM_GEMM_SYMBOL(ITYPE)(&transa, &transb, &m, &n, &k, &alpha, agold, &lda, bgold, &ldb, &beta, cgold, &ldc); } duration = libxsmm_timer_duration(start, libxsmm_timer_tick()); if (0 < duration) { fprintf(stdout, "\tBLAS: %.1f GFLOPS/s\n", gflops * nrepeat / duration); } /* free memory not needed further; avoid double-free later on */ libxsmm_free(agold); agold = 0; libxsmm_free(bgold); bgold = 0; libxsmm_free(a); a = 0; libxsmm_free(b); b = 0; /* allocate C-matrix in regular format, and perform copy-out */ ctest = (ITYPE*)libxsmm_malloc((size_t)(ldc * n * sizeof(ITYPE))); if (0 != ctest) { libxsmm_matdiff_info diff; libxsmm_bgemm_copyout_c(handle, c, &ldc, ctest); if (EXIT_SUCCESS == libxsmm_matdiff(LIBXSMM_DATATYPE(ITYPE), m, n, cgold, ctest, &ldc, &ldc, &diff)) { fprintf(stdout, "\tdiff: L2abs=%f Linf=%f\n", diff.l2_abs, diff.linf_abs); if (check < 100.0 * diff.normf_rel) { fprintf(stderr, "FAILED with an error of %f%%!\n", 100.0 * diff.normf_rel); result = EXIT_FAILURE; } } libxsmm_free(ctest); } } #endif libxsmm_bgemm_handle_destroy(handle); } else { fprintf(stderr, "FAILED to create BGEMM-handle! For details retry with LIBXSMM_VERBOSE=1.\n"); result = EXIT_FAILURE; } libxsmm_free(agold); libxsmm_free(bgold); libxsmm_free(cgold); libxsmm_free(a); libxsmm_free(b); libxsmm_free(c); } if(!ab) { fprintf(stdout, "Finished\n"); } return result; }
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[]) { double rng_stddev = 0; float* rngs; float vrng[16]; libxsmm_matdiff_info info; libxsmm_blasint num_rngs; libxsmm_blasint i; unsigned long long start; if (2 < argc) { fprintf(stderr, "Usage:\n %s number_rngs\n", argv[0]); return EXIT_SUCCESS; } /* parse the command line and set up the test parameters */ num_rngs = (1 < argc ? atoi(argv[1]) : 1000); assert(num_rngs >= 1); rngs = (float*)malloc((size_t)(sizeof(float) * num_rngs)); if (NULL == rngs) num_rngs = 0; libxsmm_rng_set_seed( (uint32_t)(time(0))); /* fill array with random floats */ libxsmm_rng_f32_seq( rngs, num_rngs ); /* some quality measure; variance is based on discovered average rather than expected value */ if (EXIT_SUCCESS == libxsmm_matdiff(&info, LIBXSMM_DATATYPE_F32, 1/*m*/, num_rngs, NULL/*ref*/, rngs/*tst*/, NULL/*ldref*/, NULL/*ldtst*/)) { rng_stddev = sqrt(info.var_tst); } start = libxsmm_timer_tick(); for (i = 0; i < num_rngs; ++i) { libxsmm_rng_f32_seq( rngs, 1 ); } printf("\nlibxsmm_rng_float: %llu cycles per random number (scalar)\n", libxsmm_timer_cycles(start, libxsmm_timer_tick()) / num_rngs); start = libxsmm_timer_tick(); for (i = 0; i < num_rngs; ++i) { libxsmm_rng_f32_seq( vrng, 16 ); } printf("\nlibxsmm_rng_float: %llu cycles per random number (vlen=16)\n", libxsmm_timer_cycles(start, libxsmm_timer_tick()) / ((size_t)num_rngs*16)); /* let's compute some values of the random numbers */ printf("\nWe have generated %lli random numbers uniformly distributed in [0,1(\n", (long long)num_rngs); printf("We expect the following values E=0.5, Var=0.083333, Stddev=0.288675\n\n"); printf("minimum random number is: %f\n", info.min_tst); printf("maximum random number is: %f\n", info.max_tst); printf("sum of random numbers is: %f\n", info.l1_tst); printf("Expected Value of random numbers is: %f\n", info.avg_tst); printf("Variance of random numbers is: %f\n", info.var_tst); printf("StdDev of random numbers is: %f\n\n", rng_stddev); free( rngs ); return EXIT_SUCCESS; }