snRuntime: Add all-to-all reduction and use in exit function

pulp-platform · Jan 10, 2024 · 7f1fdf3 · 7f1fdf3
1 parent 7c2f936
commit 7f1fdf3
Show file tree

Hide file tree

Showing 29 changed files with 122 additions and 92 deletions.
diff --git 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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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;
 }