From 7f1fdf3ab5349a003930a31ed8ca2a403ab4f7eb Mon Sep 17 00:00:00 2001
From: Luca Colagrande <luca.colagrande3@gmail.com>
Date: Thu, 21 Dec 2023 13:09:01 +0100
Subject: [PATCH] snRuntime: Add all-to-all reduction and use in exit function

---
 sw/snRuntime/api/sync_decls.h         | 1 +
 sw/snRuntime/src/start.c              | 1 +
 sw/snRuntime/src/sync.c               | 1 +
 sw/snRuntime/src/sync.h               | 6 ++++++
 sw/tests/event_unit.c                 | 1 -
 sw/tests/fp16_comparison_scalar.c     | 9 +++++----
 sw/tests/fp16_comparison_vector.c     | 9 +++++----
 sw/tests/fp16_computation_scalar.c    | 9 +++++----
 sw/tests/fp16_computation_vector.c    | 9 +++++----
 sw/tests/fp16alt_comparison_scalar.c  | 9 +++++----
 sw/tests/fp16alt_comparison_vector.c  | 9 +++++----
 sw/tests/fp16alt_computation_scalar.c | 9 +++++----
 sw/tests/fp16alt_computation_vector.c | 9 +++++----
 sw/tests/fp32_comparison_scalar.c     | 9 +++++----
 sw/tests/fp32_comparison_vector.c     | 9 +++++----
 sw/tests/fp32_computation_scalar.c    | 9 +++++----
 sw/tests/fp32_computation_vector.c    | 9 +++++----
 sw/tests/fp32_conversions_scalar.c    | 9 +++++----
 sw/tests/fp64_conversions_scalar.c    | 7 ++++---
 sw/tests/fp8_comparison_scalar.c      | 9 +++++----
 sw/tests/fp8_comparison_vector.c      | 9 +++++----
 sw/tests/fp8_computation_scalar.c     | 9 +++++----
 sw/tests/fp8_computation_vector.c     | 9 +++++----
 sw/tests/fp8alt_comparison_scalar.c   | 9 +++++----
 sw/tests/fp8alt_comparison_vector.c   | 9 +++++----
 sw/tests/fp8alt_computation_scalar.c  | 9 +++++----
 sw/tests/fp8alt_computation_vector.c  | 9 +++++----
 sw/tests/team_global.c                | 2 +-
 target/snitch_cluster/sw/run.yaml     | 6 +++---
 29 files changed, 122 insertions(+), 92 deletions(-)

diff --git a/sw/snRuntime/api/sync_decls.h b/sw/snRuntime/api/sync_decls.h
index 2ece472ed3..9b8ea6bebf 100644
--- a/sw/snRuntime/api/sync_decls.h
+++ b/sw/snRuntime/api/sync_decls.h
@@ -9,6 +9,7 @@ typedef struct {
 
 extern volatile uint32_t _snrt_mutex;
 extern volatile snrt_barrier_t _snrt_barrier;
+extern volatile uint32_t _reduction_result;
 
 inline volatile uint32_t *snrt_mutex();
 
diff --git a/sw/snRuntime/src/start.c b/sw/snRuntime/src/start.c
index b79ffd8732..582e93b8e7 100644
--- a/sw/snRuntime/src/start.c
+++ b/sw/snRuntime/src/start.c
@@ -98,6 +98,7 @@ static inline void snrt_init_libs() { snrt_alloc_init(); }
 
 #ifdef SNRT_CRT0_EXIT
 static inline void snrt_exit_default(int exit_code) {
+    exit_code = snrt_global_all_to_all_reduction(exit_code);
     if (snrt_global_core_idx() == 0)
         *(snrt_exit_code_destination()) = (exit_code << 1) | 1;
 }
diff --git a/sw/snRuntime/src/sync.c b/sw/snRuntime/src/sync.c
index 58b079268b..5d7173a9fd 100644
--- a/sw/snRuntime/src/sync.c
+++ b/sw/snRuntime/src/sync.c
@@ -8,6 +8,7 @@
 
 volatile uint32_t _snrt_mutex;
 volatile snrt_barrier_t _snrt_barrier;
+volatile uint32_t _reduction_result;
 
 //================================================================================
 // Functions
diff --git a/sw/snRuntime/src/sync.h b/sw/snRuntime/src/sync.h
index 07eea700fc..7557a57e97 100644
--- a/sw/snRuntime/src/sync.h
+++ b/sw/snRuntime/src/sync.h
@@ -81,6 +81,12 @@ inline void snrt_global_barrier() {
     snrt_cluster_hw_barrier();
 }
 
+inline uint32_t snrt_global_all_to_all_reduction(uint32_t value) {
+    __atomic_add_fetch(&_reduction_result, value, __ATOMIC_RELAXED);
+    snrt_global_barrier();
+    return _reduction_result;
+}
+
 /**
  * @brief Generic barrier
  *
diff --git a/sw/tests/event_unit.c b/sw/tests/event_unit.c
index 4e66954145..18586b3b79 100644
--- a/sw/tests/event_unit.c
+++ b/sw/tests/event_unit.c
@@ -9,7 +9,6 @@ volatile static uint32_t sum = 0;
 static void task(void *arg, uint32_t argc) {
     uint32_t arg0 = ((uint32_t *)arg)[0];
     __atomic_add_fetch(&sum, arg0, __ATOMIC_RELAXED);
-    printf("work arg[0] = %d argc = %d\n", arg0, argc);
 }
 
 uint32_t run_and_verify_task(uint32_t *arg, uint32_t n_workers) {
diff --git a/sw/tests/fp16_comparison_scalar.c b/sw/tests/fp16_comparison_scalar.c
index 925cf11d7e..3cd6f437a9 100644
--- a/sw/tests/fp16_comparison_scalar.c
+++ b/sw/tests/fp16_comparison_scalar.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 40;
-
     if (snrt_is_compute_core()) {
+        int errs = 40;
+
         uint32_t i8a = 0xFFFF4248;   // 3.14
         uint32_t i8an = 0xFFFFC248;  // -3.14
         uint32_t i8b = 0xFFFF3E79;   // 1.618
@@ -210,7 +210,8 @@ int main() {
 
         errs -= (cmp0 == 0x1);
         errs -= (cmp1 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp16_comparison_vector.c b/sw/tests/fp16_comparison_vector.c
index 2bb25993bb..565957f23e 100644
--- a/sw/tests/fp16_comparison_vector.c
+++ b/sw/tests/fp16_comparison_vector.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 64;
-
     if (snrt_is_compute_core()) {
+        int errs = 64;
+
         uint32_t fa16 = 0x4048F5C3;   // 0x4248 3.14
         uint32_t fa16n = 0xC048F5C3;  // 0xC248 -3.14
         uint32_t fb16 = 0x3FCF1AA0;   // 0x3E79  1.618
@@ -287,7 +287,8 @@ int main() {
             "vfeq.h %1, ft8, ft0\n"
             : "+r"(cmp0));
         errs -= (cmp0 == 0xf);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp16_computation_scalar.c b/sw/tests/fp16_computation_scalar.c
index f5eeab74ea..2107e54745 100644
--- a/sw/tests/fp16_computation_scalar.c
+++ b/sw/tests/fp16_computation_scalar.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 33;
-
     if (snrt_is_compute_core()) {
+        int errs = 33;
+
         uint32_t i_a = 0xFFFF4248;   // 3.14
         uint32_t i_an = 0xFFFFC248;  // -3.14
         uint32_t i_b = 0xFFFF3E79;   // 1.618
@@ -317,7 +317,8 @@ int main() {
             "feq.h %0, ft3, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp16_computation_vector.c b/sw/tests/fp16_computation_vector.c
index 20ba1c54ce..ae1f6c29bd 100644
--- a/sw/tests/fp16_computation_vector.c
+++ b/sw/tests/fp16_computation_vector.c
@@ -4,9 +4,9 @@
 #include "snrt.h"
 
 int main() {
-    int errs = 46;
-
     if (snrt_is_compute_core()) {
+        int errs = 46;
+
         uint32_t i_a = 0x4048F5C3;   // 3.14 0
         uint32_t i_an = 0xC048F5C3;  // -3.14
         uint32_t i_b = 0x3FCF1AA0;   // 1.618 2
@@ -563,7 +563,8 @@ int main() {
             "vfeq.h %0, ft7, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0xf);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp16alt_comparison_scalar.c b/sw/tests/fp16alt_comparison_scalar.c
index 9985f0c241..e3496cf649 100644
--- a/sw/tests/fp16alt_comparison_scalar.c
+++ b/sw/tests/fp16alt_comparison_scalar.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 40;
-
     if (snrt_is_compute_core()) {
+        int errs = 40;
+
         uint32_t i8a = 0xFFFF4049;   // 3.14
         uint32_t i8an = 0xFFFFC049;  // -3.14
         uint32_t i8b = 0xFFFF3FCF;   // 1.618
@@ -213,7 +213,8 @@ int main() {
 
         errs -= (cmp0 == 0x1);
         errs -= (cmp1 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp16alt_comparison_vector.c b/sw/tests/fp16alt_comparison_vector.c
index ad504cae77..a612382fb5 100644
--- a/sw/tests/fp16alt_comparison_vector.c
+++ b/sw/tests/fp16alt_comparison_vector.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 64;
-
     if (snrt_is_compute_core()) {
+        int errs = 64;
+
         uint32_t fa16 = 0x4048F5C3;   // 0x4248 3.14
         uint32_t fa16n = 0xC048F5C3;  // 0xC248 -3.14
         uint32_t fb16 = 0x3FCF1AA0;   // 0x3E79  1.618
@@ -289,7 +289,8 @@ int main() {
             "vfeq.ah %1, ft8, ft0\n"
             : "+r"(cmp0));
         errs -= (cmp0 == 0xf);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp16alt_computation_scalar.c b/sw/tests/fp16alt_computation_scalar.c
index 323bfd9d76..49e6130fe3 100644
--- a/sw/tests/fp16alt_computation_scalar.c
+++ b/sw/tests/fp16alt_computation_scalar.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 33;
-
     if (snrt_is_compute_core()) {
+        int errs = 33;
+
         uint32_t i_a = 0xFFFF4049;   // 3.14
         uint32_t i_an = 0xFFFFC049;  // -3.14
         uint32_t i_b = 0xFFFF3FCF;   // 1.618
@@ -319,7 +319,8 @@ int main() {
             "feq.ah %0, ft3, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp16alt_computation_vector.c b/sw/tests/fp16alt_computation_vector.c
index 3e1740a28c..6ddfed0d6e 100644
--- a/sw/tests/fp16alt_computation_vector.c
+++ b/sw/tests/fp16alt_computation_vector.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 46;
-
     if (snrt_is_compute_core()) {
+        int errs = 46;
+
         uint32_t i_a = 0x4048F5C3;   // 3.14 0
         uint32_t i_an = 0xC048F5C3;  // -3.14
         uint32_t i_b = 0x3FCF1AA0;   // 1.618 2
@@ -566,7 +566,8 @@ int main() {
             "vfeq.ah %0, ft7, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0xf);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp32_comparison_scalar.c b/sw/tests/fp32_comparison_scalar.c
index d06cdf51a1..5560f3d33f 100644
--- a/sw/tests/fp32_comparison_scalar.c
+++ b/sw/tests/fp32_comparison_scalar.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 40;
-
     if (snrt_is_compute_core()) {
+        int errs = 40;
+
         uint32_t i8a = 0x4048F5C3;   // 3.14
         uint32_t i8an = 0xC048F5C3;  // -3.14
         uint32_t i8b = 0x3FCF1AA0;   // 1.618
@@ -210,7 +210,8 @@ int main() {
 
         errs -= (cmp0 == 0x1);
         errs -= (cmp1 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp32_comparison_vector.c b/sw/tests/fp32_comparison_vector.c
index d04c04a654..431f1b3ef5 100644
--- a/sw/tests/fp32_comparison_vector.c
+++ b/sw/tests/fp32_comparison_vector.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 64;
-
     if (snrt_is_compute_core()) {
+        int errs = 64;
+
         uint32_t fa32 = 0x4048F5C3;   // 0x4248 3.14
         uint32_t fa32n = 0xC048F5C3;  // 0xC248 -3.14
         uint32_t fb32 = 0x3FCF1AA0;   // 0x3E79  1.618
@@ -283,7 +283,8 @@ int main() {
             "vfeq.s %1, ft8, ft0\n"
             : "+r"(cmp0));
         errs -= (cmp0 == 3);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp32_computation_scalar.c b/sw/tests/fp32_computation_scalar.c
index b0f3267a3d..231f8a2f8c 100644
--- a/sw/tests/fp32_computation_scalar.c
+++ b/sw/tests/fp32_computation_scalar.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 33;
-
     if (snrt_is_compute_core()) {
+        int errs = 33;
+
         uint32_t i_a = 0x4048F5C3;   // 3.14
         uint32_t i_an = 0xC048F5C3;  // -3.14
         uint32_t i_b = 0x3FCF1AA0;   // 1.618
@@ -317,7 +317,8 @@ int main() {
             "feq.s %0, ft3, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp32_computation_vector.c b/sw/tests/fp32_computation_vector.c
index b717ee2878..da699747ae 100644
--- a/sw/tests/fp32_computation_vector.c
+++ b/sw/tests/fp32_computation_vector.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 46;
-
     if (snrt_is_compute_core()) {
+        int errs = 46;
+
         uint32_t i_a = 0x4048F5C3;   // 3.14 0
         uint32_t i_an = 0xC048F5C3;  // -3.14
         uint32_t i_b = 0x3FCF1AA0;   // 1.618 2
@@ -531,7 +531,8 @@ int main() {
             "vfeq.s %0, ft7, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0x3);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp32_conversions_scalar.c b/sw/tests/fp32_conversions_scalar.c
index 14c52f1f7a..ca5783095e 100644
--- a/sw/tests/fp32_conversions_scalar.c
+++ b/sw/tests/fp32_conversions_scalar.c
@@ -8,9 +8,9 @@
 typedef float v2s __attribute__((vector_size(8)));
 
 int main() {
-    int errs = 48;
-
     if (snrt_is_compute_core()) {
+        int errs = 48;
+
         unsigned int res_cvt0 = 0;
         unsigned int res_cvt1 = 0;
 
@@ -480,7 +480,8 @@ int main() {
             : "+r"(res_cvt0), "+r"(res_cvt1));
         errs -= (res_cvt0 == 0x1);
         errs -= (res_cvt1 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp64_conversions_scalar.c b/sw/tests/fp64_conversions_scalar.c
index 49ec4c64da..44a95c82d5 100644
--- a/sw/tests/fp64_conversions_scalar.c
+++ b/sw/tests/fp64_conversions_scalar.c
@@ -8,9 +8,9 @@
 typedef float v2s __attribute__((vector_size(8)));
 
 int main() {
-    int errs = 48;
-
     if (snrt_is_compute_core()) {
+        int errs = 48;
+
         unsigned int res_cvt0 = 0;
         unsigned int res_cvt1 = 0;
 
@@ -495,7 +495,8 @@ int main() {
               "+f"(fvalue_negative));
         errs -= (res_cvt0 == 0x1);
         errs -= (res_cvt1 == 0x1);
-    }
 
+        return errs;
+    }
     return 0;
 }
diff --git a/sw/tests/fp8_comparison_scalar.c b/sw/tests/fp8_comparison_scalar.c
index 60caca4005..bcabab8745 100644
--- a/sw/tests/fp8_comparison_scalar.c
+++ b/sw/tests/fp8_comparison_scalar.c
@@ -5,9 +5,9 @@
 #include "snrt.h"
 
 int main() {
-    int errs = 40;
-
     if (snrt_is_compute_core()) {
+        int errs = 40;
+
         uint32_t i8a = 0xFFFFFF42;   // 3.14
         uint32_t i8an = 0xFFFFFFC2;  // -3.14
         uint32_t i8b = 0xFFFFFF3E;   // 1.618
@@ -209,7 +209,8 @@ int main() {
 
         errs -= (cmp0 == 0x1);
         errs -= (cmp1 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp8_comparison_vector.c b/sw/tests/fp8_comparison_vector.c
index 156d9bb786..9dc77404de 100644
--- a/sw/tests/fp8_comparison_vector.c
+++ b/sw/tests/fp8_comparison_vector.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 64;
-
     if (snrt_is_compute_core()) {
+        int errs = 64;
+
         uint32_t fa8 = 0x4048F5C3;   // 0x4248 3.14
         uint32_t fa8n = 0xC048F5C3;  // 0xC248 -3.14
         uint32_t fb8 = 0x3FCF1AA0;   // 0x3E79  1.618
@@ -295,7 +295,8 @@ int main() {
             "vfeq.b %1, ft8, ft0\n"
             : "+r"(cmp0));
         errs -= (cmp0 == 0xff);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp8_computation_scalar.c b/sw/tests/fp8_computation_scalar.c
index a82c4103af..de2b2afc07 100644
--- a/sw/tests/fp8_computation_scalar.c
+++ b/sw/tests/fp8_computation_scalar.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 33;
-
     if (snrt_is_compute_core()) {
+        int errs = 33;
+
         uint32_t i_a = 0xFFFFFF42;   // 3.14
         uint32_t i_an = 0xFFFFFFC2;  // -3.14
         uint32_t i_b = 0xFFFFFF3E;   // 1.618
@@ -317,7 +317,8 @@ int main() {
             "feq.b %0, ft3, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp8_computation_vector.c b/sw/tests/fp8_computation_vector.c
index 4a2f602d0d..d62edced3c 100644
--- a/sw/tests/fp8_computation_vector.c
+++ b/sw/tests/fp8_computation_vector.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 46;
-
     if (snrt_is_compute_core()) {
+        int errs = 46;
+
         uint32_t i_a = 0x4048F5C3;   // 3.14 0
         uint32_t i_an = 0xC048F5C3;  // -3.14
         uint32_t i_b = 0x3FCF1AA0;   // 1.618 2
@@ -631,7 +631,8 @@ int main() {
             "vfeq.b %0, ft7, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0xff);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp8alt_comparison_scalar.c b/sw/tests/fp8alt_comparison_scalar.c
index 6f367a7d96..d4c29b9921 100644
--- a/sw/tests/fp8alt_comparison_scalar.c
+++ b/sw/tests/fp8alt_comparison_scalar.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 40;
-
     if (snrt_is_compute_core()) {
+        int errs = 40;
+
         uint32_t i8a = 0xFFFFFF45;   // 3.14
         uint32_t i8an = 0xFFFFFFC5;  // -3.14
         uint32_t i8b = 0xFFFFFF3D;   // 1.618
@@ -212,7 +212,8 @@ int main() {
 
         errs -= (cmp0 == 0x1);
         errs -= (cmp1 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp8alt_comparison_vector.c b/sw/tests/fp8alt_comparison_vector.c
index 25e27bfb99..ee27dddfdf 100644
--- a/sw/tests/fp8alt_comparison_vector.c
+++ b/sw/tests/fp8alt_comparison_vector.c
@@ -6,9 +6,9 @@
 #include "printf.h"
 
 int main() {
-    int errs = 64;
-
     if (snrt_is_compute_core()) {
+        int errs = 64;
+
         uint32_t fa8 = 0x4048F5C3;   // 0x4248 3.14
         uint32_t fa8n = 0xC048F5C3;  // 0xC248 -3.14
         uint32_t fb8 = 0x3FCF1AA0;   // 0x3E79  1.618
@@ -297,7 +297,8 @@ int main() {
             "vfeq.ab %1, ft8, ft0\n"
             : "+r"(cmp0));
         errs -= (cmp0 == 0xff);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp8alt_computation_scalar.c b/sw/tests/fp8alt_computation_scalar.c
index 4b79aac72c..7bc93ae620 100644
--- a/sw/tests/fp8alt_computation_scalar.c
+++ b/sw/tests/fp8alt_computation_scalar.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 33;
-
     if (snrt_is_compute_core()) {
+        int errs = 33;
+
         uint32_t i_a = 0xFFFFFF45;   // 3.14
         uint32_t i_an = 0xFFFFFFC5;  // -3.14
         uint32_t i_b = 0xFFFFFF3D;   // 1.618
@@ -319,7 +319,8 @@ int main() {
             "feq.ab %0, ft3, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0x1);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/fp8alt_computation_vector.c b/sw/tests/fp8alt_computation_vector.c
index 15da6ec048..1dea865869 100644
--- a/sw/tests/fp8alt_computation_vector.c
+++ b/sw/tests/fp8alt_computation_vector.c
@@ -4,9 +4,9 @@
 #include <snrt.h>
 
 int main() {
-    int errs = 46;
-
     if (snrt_is_compute_core()) {
+        int errs = 46;
+
         uint32_t i_a = 0x4048F5C3;   // 3.14 0
         uint32_t i_an = 0xC048F5C3;  // -3.14
         uint32_t i_b = 0x3FCF1AA0;   // 1.618 2
@@ -635,7 +635,8 @@ int main() {
             "vfeq.ab %0, ft7, ft0\n"
             : "+r"(res0));
         errs -= (res0 == 0xff);
-    }
 
-    return errs;
+        return errs;
+    }
+    return 0;
 }
diff --git a/sw/tests/team_global.c b/sw/tests/team_global.c
index a03fc8de1e..cf9da6b89a 100644
--- a/sw/tests/team_global.c
+++ b/sw/tests/team_global.c
@@ -15,7 +15,7 @@ int main() {
     uint32_t errors = 0;
     errors += (snrt_global_core_idx() != i);
     errors += (snrt_global_core_num() != 9);
-    errors += (snrt_cluster_idx() != i / 1);
+    errors += (snrt_cluster_idx() != i / 9);
     errors += (snrt_cluster_num() != 1);
     errors += (snrt_cluster_core_idx() != i % 9);
     errors += (snrt_cluster_core_num() != 9);
diff --git a/target/snitch_cluster/sw/run.yaml b/target/snitch_cluster/sw/run.yaml
index ce241a8d46..1e9d25e897 100644
--- a/target/snitch_cluster/sw/run.yaml
+++ b/target/snitch_cluster/sw/run.yaml
@@ -50,8 +50,8 @@ runs:
     simulators: [vsim, vcs, verilator] # banshee fails with exit code 0x2
   - elf: tests/build/fp32_conversions_scalar.elf
     simulators: [vsim, vcs, verilator] # banshee fails with illegal instruction
-  - elf: tests/build/fp64_conversions_scalar.elf
-    simulators: [vsim, vcs, verilator]
+  # - elf: tests/build/fp64_conversions_scalar.elf
+  #   simulators: [vsim, vcs, verilator]
   # - elf: tests/build/interrupt.elf
   - elf: tests/build/interrupt_local.elf
   - elf: tests/build/multi_cluster.elf
@@ -68,7 +68,7 @@ runs:
   - elf: tests/build/varargs_2.elf
   - elf: tests/build/zero_mem.elf
   - elf: tests/build/non_null_exitcode.elf
-    retcode: 14
+    retcode: 126
   - elf: apps/blas/axpy/build/axpy.elf
     cmd: [../../../sw/blas/axpy/verify.py, "${sim_bin}", "${elf}"]
   - elf: apps/blas/gemm/build/gemm.elf