[apps] Introduce new ad-hoc 569JNRXZghikmswarming for fconv,fmatmul,j…

…acobi
pulp-platform · Nov 28, 2023 · 0055270 · 0055270
1 parent 4f11b13
commit 0055270
Show file tree

Hide file tree

Showing 6 changed files with 533 additions and 1 deletion.
diff --git a/apps/benchmarks/benchmark/fconv3d.bmark b/apps/benchmarks/benchmark/fconv3d.bmark
@@ -40,8 +40,17 @@ extern int64_t CH;
 extern int64_t F;
 
 void warm_caches(uint64_t heat) {
+  volatile double buf;
+
   for (uint64_t k = 0; k < heat; ++k)
     fconv3d_CHx7x7(o, i, f, M, N, CH, F);
+// The following artificial warming ensures, with a larger cache,
+// not to experience any cache misses
+#ifdef AD_HOC_WARMING
+  for (uint64_t k = 0; k < F*F*CH; ++k)
+    buf = (volatile double) *(&(f[k]));
+  fconv3d_CHx7x7_warm(o, i, f, M, N, CH, F);
+#endif
 }
 
 int main() {

diff --git a/apps/benchmarks/benchmark/fmatmul.bmark b/apps/benchmarks/benchmark/fmatmul.bmark
@@ -42,12 +42,14 @@ void warm_caches(uint64_t heat) {
 
   for (uint64_t k = 0; k < heat; ++k)
     fmatmul(c, a, b, M, N, P);
+#ifdef AD_HOC_WARMING
     // Vector stores have invalidated the A mtx cache lines!
     // Fetch them again
     for (int m = 0; m < M; ++m) {
       buf = (volatile double) *a_;
       a_ += N;
     }
+#endif
 }
 
 int main() {

diff --git a/apps/benchmarks/benchmark/jacobi2d.bmark b/apps/benchmarks/benchmark/jacobi2d.bmark
@@ -102,8 +102,15 @@ extern DATA_TYPE A_s[] __attribute__((aligned(4 * NR_LANES), section(".l2")));
 extern DATA_TYPE B_s[] __attribute__((aligned(4 * NR_LANES), section(".l2")));
 
 void warm_caches(uint64_t heat, DATA_TYPE* A_fixed_v, DATA_TYPE* B_fixed_v) {
+
+  volatile double buf;
+
   for (uint64_t k = 0; k < heat; ++k)
     j2d_v(R, C, A_fixed_v, B_fixed_v, TSTEPS);
+#ifdef AD_HOC_WARMING
+  for (uint64_t k = 0; k < R*C; ++k)
+    buf = (volatile double)* &(A_fixed_v[k]);
+#endif
 }
 
 int main() {
@@ -117,7 +124,7 @@ int main() {
 
 #ifndef SPIKE
   // Warm-up caches
-  warm_caches(WARM_CACHES_ITER, A_fixed_s, B_fixed_s);
+  warm_caches(WARM_CACHES_ITER, A_fixed_v, B_fixed_s);
 #endif
 
   // Measure vector kernel execution

diff --git a/apps/fconv3d/fconv3d.h b/apps/fconv3d/fconv3d.h
@@ -28,6 +28,12 @@ void fconv3d_CHx7x7(double *o, double *i, double *f, int64_t M, int64_t N,
 void fconv3d_CHx7x7_block(double *o, double *i, double *f, int64_t M, int64_t N,
                           int64_t n_, int64_t C, int64_t F);
 
+void fconv3d_CHx7x7_warm(double *o, double *i, double *f, int64_t M, int64_t N,
+                    int64_t C, int64_t F);
+
+void fconv3d_warm(double *o, double *i, double *f, int64_t M, int64_t N,
+                          int64_t n_, int64_t C, int64_t F);
+
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 
 // Threshold for FP numbers comparison during the final check