From e31f0f1dedb09449d386b43dd978a637b6943c36 Mon Sep 17 00:00:00 2001 From: Rodrigo Arias Date: Thu, 20 Jun 2024 17:39:42 +0200 Subject: [PATCH] Add fibonacci and heat RT tests for mark API --- test/rt/nodes/CMakeLists.txt | 3 + test/rt/nodes/fibonacci-mark.c | 47 ++++++++++ test/rt/nodes/heat-mark.c | 161 +++++++++++++++++++++++++++++++++ 3 files changed, 211 insertions(+) create mode 100644 test/rt/nodes/fibonacci-mark.c create mode 100644 test/rt/nodes/heat-mark.c diff --git a/test/rt/nodes/CMakeLists.txt b/test/rt/nodes/CMakeLists.txt index 5b6d4ab..4ba2a0e 100644 --- a/test/rt/nodes/CMakeLists.txt +++ b/test/rt/nodes/CMakeLists.txt @@ -82,3 +82,6 @@ elseif(ENABLE_ALL_TESTS) else() message(STATUS "Disabling perf paranoid tests for NODES") endif() + +nodes_rt_test(fibonacci-mark.c SORT) +nodes_rt_test(heat-mark.c SORT) diff --git a/test/rt/nodes/fibonacci-mark.c b/test/rt/nodes/fibonacci-mark.c new file mode 100644 index 0000000..5e9a62d --- /dev/null +++ b/test/rt/nodes/fibonacci-mark.c @@ -0,0 +1,47 @@ +/* Copyright (c) 2024 Barcelona Supercomputing Center (BSC) + * SPDX-License-Identifier: GPL-3.0-or-later */ + +#include "common.h" +#include "compat.h" +#include "ovni.h" +#include +#include + +enum { INDEX = 0, RUN = 1 }; + +#pragma oss task label("fibonacci") +static void +fib2(long index, long *res, int run) +{ + if (index <= 1) { + *res = index; + } else { + ovni_mark_push(RUN, run); + ovni_mark_push(INDEX, index); + + long a, b; + fib2(index-1, &a, run); + fib2(index-2, &b, run); + #pragma oss taskwait + *res = a + b; + + ovni_mark_pop(INDEX, index); + ovni_mark_pop(RUN, run); + } +} + +int +main(void) +{ + ovni_mark_type(INDEX, OVNI_MARK_STACK, "Fibonacci index"); + ovni_mark_type(RUN, OVNI_MARK_STACK, "Fibonacci run"); + + long res = 0; + for (int i = 1; i <= 30; i++) { + fib2(15, &res, i); + #pragma oss taskwait + } + + #pragma oss taskwait + return 0; +} diff --git a/test/rt/nodes/heat-mark.c b/test/rt/nodes/heat-mark.c new file mode 100644 index 0000000..ce93df0 --- /dev/null +++ b/test/rt/nodes/heat-mark.c @@ -0,0 +1,161 @@ +/* Copyright (c) 2024 Barcelona Supercomputing Center (BSC) + * SPDX-License-Identifier: GPL-3.0-or-later */ + +#include "common.h" +#include "compat.h" +#include "ovni.h" +#include +#include +#include + +enum { ITER = 0, ROW, COL, PRIO, TIME }; + +long rows; +long cols; + +long rbs; +long cbs; +long timesteps; + +long nrb; +long ncb; + +int use_prio = 0; +int prio_types = 4; + +static double +get_time(void) +{ + struct timespec ts; + clock_gettime(CLOCK_MONOTONIC, &ts); + return (double) ts.tv_sec + (double) ts.tv_nsec * 1.0e-9; +} + +static void +compute_block(long rstart, long rend, long cstart, long cend, double M[rows][cols]) +{ + for (long r = rstart; r <= rend; ++r) { + for (long c = cstart; c <= cend; ++c) { + M[r][c] = 0.25*(M[r-1][c] + M[r+1][c] + M[r][c-1] + M[r][c+1]); + } + } +} + +static void +step(double M[rows][cols], char reps[nrb][ncb], long iter) +{ + for (long R = 1; R < nrb-1; ++R) { + for (long C = 1; C < ncb-1; ++C) { + long prio = 0; + /* Higher number = runs before */ + if (use_prio == 0) + prio = 1; + else if (use_prio == 1) + prio = 1 + nrb * ncb - (R * ncb + C); + else if (use_prio == 2) + prio = 1 + nrb + ncb - (R + C); + else if (use_prio == 3) + prio = (timesteps + 1) * nrb * ncb - (iter * ncb * nrb + R * ncb + C); + + #pragma oss task label("block computation") \ + in(reps[R-1][C]) in(reps[R+1][C]) \ + in(reps[R][C-1]) in(reps[R][C+1]) \ + priority(prio) \ + inout(reps[R][C]) + { + ovni_mark_push(ITER, iter); + ovni_mark_push(ROW, R); + ovni_mark_push(COL, C); + ovni_mark_push(PRIO, prio); + compute_block((R-1)*rbs+1, R*rbs, (C-1)*cbs+1, C*cbs, M); + ovni_mark_pop(PRIO, prio); + ovni_mark_pop(COL, C); + ovni_mark_pop(ROW, R); + ovni_mark_pop(ITER, iter); + } + } + } +} + + +static void +solve(double *raw_matrix, long n) +{ + /* Ignore halos */ + double (*matrix)[cols] = (double (*)[cols]) raw_matrix; + char representatives[nrb][ncb]; + for (long t = 1; t <= n; t++) + step(matrix, representatives, t); + #pragma oss taskwait +} + +static void +reset(double *matrix) +{ + /* Set all elements to zero */ + memset(matrix, 0, rows * cols * sizeof(double)); + + /* Set the left side to 1.0 */ + for (long i = 0; i < rows; i++) + matrix[i * cols] = 1.0; +} + +int +main(void) +{ + ovni_mark_type(ITER, OVNI_MARK_STACK, "Heat iteration"); + ovni_mark_type(ROW, OVNI_MARK_STACK, "Heat row"); + ovni_mark_type(COL, OVNI_MARK_STACK, "Heat col"); + ovni_mark_type(PRIO, OVNI_MARK_STACK, "Heat priority"); + ovni_mark_type(TIME, OVNI_MARK_STACK, "Heat time"); + + /* The real number of allocated rows and columns will be bigger */ + long _rows = 8L * 1024L; + long _cols = _rows; + + rows = _rows + 2; + cols = _cols + 2; + + rbs = 128L; + cbs = rbs; + timesteps = 10L; + + nrb = (rows - 2) / rbs + 2; + ncb = (cols - 2) / cbs + 2; + + long pagesize = sysconf(_SC_PAGESIZE); + if (pagesize <= 0) + die("cannot read pagesize"); + + double *matrix; + int err = posix_memalign((void **) &matrix, pagesize, rows * cols * sizeof(double)); + if (err || matrix == NULL) + die("posix_memalign failed:"); + + /* Do a warmup iteration */ + reset(matrix); + solve(matrix, 1); + + sleep_us(100000); + + char buf[1024]; + for (int i = 0; i < prio_types; i++) { + use_prio = i; + reset(matrix); + double t0 = get_time(); + solve(matrix, timesteps); + double t1 = get_time(); + double dt = t1 - t0; + + /* Use label to write a custom string */ + sprintf(buf, "prio=%d took %e s", i, dt); + ovni_mark_label(TIME, i+1, buf); + ovni_mark_push(TIME, i+1); + sleep_us(100000); + ovni_mark_pop(TIME, i+1); + } + + free(matrix); + + return 0; +}