sw: Make tests agnostic of core num

pulp-platform · Jul 28, 2024 · 1b2177c · 1b2177c
1 parent 54aeece
commit 1b2177c
Show file tree

Hide file tree

Showing 9 changed files with 25 additions and 36 deletions.
diff --git a/sw/dnn/softmax/src/softmax.h b/sw/dnn/softmax/src/softmax.h
@@ -117,7 +117,7 @@ static inline void softmax_layer(softmax_layer_t const l) {
 
         // printf("row_offset: %d, ldI: %d\n", row_offset, ldI);
         softmax_fp32(&ifmap[row_offset], &ofmap[row_offset], ldI, batch_offset,
-                     l.batch_size, l.seq_len / 8, l.input_samples);
+                     l.batch_size, l.seq_len / compute_num, l.input_samples);
 
     } else {
         snrt_cluster_hw_barrier();

diff --git a/sw/snRuntime/src/omp/omp.h b/sw/snRuntime/src/omp/omp.h
@@ -157,6 +157,10 @@ static inline unsigned omp_get_thread_num(void) {
     return snrt_cluster_core_idx();
 }
 
+static inline unsigned omp_get_num_threads(void) {
+    return snrt_cluster_compute_core_num();
+}
+
 static inline void parallelRegion(int32_t argc, void *data,
                                   void (*fn)(void *, uint32_t),
                                   int num_threads) {

diff --git a/sw/tests/caq.c b/sw/tests/caq.c
@@ -7,7 +7,7 @@
 
 #include <snrt.h>
 
-#define NUM_WORKERS 8
+#define NUM_WORKERS 4
 
 // To prevent X reads on non-CAQ-proofed systems, we need a sync
 inline void fp_sync() {

diff --git a/sw/tests/caq_frep.c b/sw/tests/caq_frep.c
@@ -7,7 +7,7 @@
 
 #include <snrt.h>
 
-#define NUM_WORKERS 8
+#define NUM_WORKERS 4
 
 // To prevent X reads on non-CAQ-proofed systems, we need a sync
 inline void fp_sync() {

diff --git a/sw/tests/non_null_exitcode.c b/sw/tests/non_null_exitcode.c
@@ -8,4 +8,9 @@
 // Should be used as a test to check that the simulator or whoever
 // is running the program actually captures an error when it occurs.
 
-int main() { return 14; }
+// Minimum number of cores the hardware must have for this test to pass
+#define MIN_CORES 4
+
+int main() {
+    if (snrt_cluster_core_idx() < MIN_CORES) return 14;
+}
diff --git a/sw/tests/openmp_for_static_schedule.c b/sw/tests/openmp_for_static_schedule.c
@@ -5,7 +5,6 @@
 #include "snrt.h"
 
 #define AXPY_N 64
-#define NTHREADS 8
 
 unsigned __attribute__((noinline)) static_schedule(void) {
     static double *data_x, *data_y, data_a;
@@ -24,13 +23,17 @@ unsigned __attribute__((noinline)) static_schedule(void) {
     // Compute AXPY
 #pragma omp parallel firstprivate(data_a, data_x, data_y)
     {
+        int nthreads = omp_get_num_threads();
         // DM, rep, bound, stride, data
-        __builtin_ssr_setup_1d_r(0, 0, AXPY_N / NTHREADS - 1, sizeof(double),
-                                 &data_x[AXPY_N / 8 * omp_get_thread_num()]);
-        __builtin_ssr_setup_1d_r(1, 0, AXPY_N / NTHREADS - 1, sizeof(double),
-                                 &data_y[AXPY_N / 8 * omp_get_thread_num()]);
-        __builtin_ssr_setup_1d_w(2, 0, AXPY_N / NTHREADS - 1, sizeof(double),
-                                 &data_y[AXPY_N / 8 * omp_get_thread_num()]);
+        __builtin_ssr_setup_1d_r(
+            0, 0, AXPY_N / nthreads - 1, sizeof(double),
+            &data_x[AXPY_N / nthreads * omp_get_thread_num()]);
+        __builtin_ssr_setup_1d_r(
+            1, 0, AXPY_N / nthreads - 1, sizeof(double),
+            &data_y[AXPY_N / nthreads * omp_get_thread_num()]);
+        __builtin_ssr_setup_1d_w(
+            2, 0, AXPY_N / nthreads - 1, sizeof(double),
+            &data_y[AXPY_N / nthreads * omp_get_thread_num()]);
         __builtin_ssr_enable();
 #pragma omp for schedule(static)
         for (unsigned i = 0; i < AXPY_N; i++) {

diff --git a/sw/tests/openmp_parallel.c b/sw/tests/openmp_parallel.c
@@ -16,7 +16,7 @@ unsigned __attribute__((noinline)) parallel_section(void) {
         tx = read_csr(minstret) - tx;
         __atomic_add_fetch(&sum, 10, __ATOMIC_RELAXED);
     }
-    return sum != 8 * 10;
+    return sum != snrt_cluster_compute_core_num() * 10;
 }
 
 int main() {

diff --git a/sw/tests/team_global.c b/sw/tests/team_global.c
diff --git a/target/snitch_cluster/sw/run.yaml b/target/snitch_cluster/sw/run.yaml
@@ -70,7 +70,7 @@ runs:
   - elf: tests/build/varargs_2.elf
   - elf: tests/build/zero_mem.elf
   - elf: tests/build/non_null_exitcode.elf
-    retcode: 126
+    retcode: 56
   - elf: tests/build/caq.elf
   - elf: tests/build/caq_frep.elf
     simulators: [vsim, vcs, verilator] # banshee does not model FREP timing