working on experiments for benchmarking.

tarunprabhu · Feb 6, 2024 · ab61a3c · ab61a3c
1 parent f510df1
commit ab61a3c
Show file tree

Hide file tree

Showing 8 changed files with 16 additions and 9 deletions.
diff --git a/kitsune/experiments/euler3d/euler3d-forall.cpp b/kitsune/experiments/euler3d/euler3d-forall.cpp
@@ -401,7 +401,7 @@ int main(int argc, char** argv)
     return 0;
   }
 
-  int iterations = 2000;  
+  int iterations = 4000;  
   if (argc > 2)
     iterations = atoi(argv[2]);
 

diff --git a/kitsune/experiments/euler3d/euler3d-kokkos.cpp b/kitsune/experiments/euler3d/euler3d-kokkos.cpp
@@ -443,7 +443,7 @@ int main(int argc, char** argv)
     return 0;
   }
 
-  int iterations = 2000;
+  int iterations = 4000;
   if (argc > 2)
     iterations = atoi(argv[2]);
 

diff --git a/kitsune/experiments/launch-attr/makefile b/kitsune/experiments/launch-attr/makefile
@@ -13,7 +13,7 @@ all: ${targets}
 srad-forall.cuda.${host_arch}: srad-forall.cpp 
 	@echo $@
 	@$(TIME_CMD) $(KIT_CXX) $(TAPIR_CUDA_FLAGS)  -DTHREADS_PER_BLOCK=$(THREADS_PER_BLOCK) \
-		-mllvm --cuabi-occupancy-launches=false -o $@ $< -Xlinker -rpath=$(KITSUNE_PREFIX)/lib
+		-o $@ $< -Xlinker -rpath=$(KITSUNE_PREFIX)/lib
 	@$(FILE_SIZE)
 
 srad-forall-exprs.cuda.${host_arch}: srad-forall-exprs.cpp 

diff --git a/kitsune/experiments/launch-attr/srad-forall.cpp b/kitsune/experiments/launch-attr/srad-forall.cpp
@@ -111,11 +111,13 @@ int main(int argc, char* argv[])
 
   cout << "  Starting benchmark...\n" << std::flush;  
   auto start_time = chrono::steady_clock::now();
+  [[kitsune::launch(8)]]
   forall(int i = 0; i < rows; i++) {
     iN[i] = i-1;
     iS[i] = i+1;
   }
 
+  [[kitsune::launch(8)]]
   forall(int j=0; j < cols; j++) {
     jW[j] = j-1;
     jE[j] = j+1;
@@ -126,6 +128,7 @@ int main(int argc, char* argv[])
   jW[0] = 0;
   jE[cols-1] = cols-1;
 
+  [[kitsune::launch(64)]]
   forall(int k = 0;  k < size_I; k++)
     J[k] = (float)exp(I[k]) ;
 
@@ -149,7 +152,7 @@ int main(int argc, char* argv[])
     q0sqr   = varROI / (meanROI*meanROI);
 
     auto loop1_start_time = chrono::steady_clock::now();
-    [[kitsune::launch(THREADS_PER_BLOCK)]]
+    [[kitsune::launch(16)]]
     forall(int i = 0 ; i < rows; i++) {
 
       for(int j = 0; j < cols; j++) {
@@ -191,7 +194,7 @@ int main(int argc, char* argv[])
       loop1_min_time = etime;
 
     auto loop2_start_time = chrono::steady_clock::now();
-    [[kitsune::launch(THREADS_PER_BLOCK)]]
+    [[kitsune::launch(8)]]
     forall(int i = 0; i < rows; i++) {
       for(int j = 0; j < cols; j++) {
         // current index

diff --git a/kitsune/experiments/launch-bench.sh b/kitsune/experiments/launch-bench.sh
@@ -12,7 +12,7 @@ echo "gpu: $gpuname" >> $outfile
 echo "time stamp: $tstamp" >> $outfile
 echo "ThreadsPerBlock,Time" >> $outfile
 
-for tpb in {16..1024..16}
+for tpb in {8..512..16}
 do 
   export KITCUDA_THREADS_PER_BLOCK=$tpb
   echo "  running '$*' with $tpb threads-per-block"

diff --git a/kitsune/experiments/makefile b/kitsune/experiments/makefile
@@ -21,5 +21,6 @@ $(subdirs):
 run: $(subdirs)
 	$(MAKE) -C $< $(MAKECMDGOALS)
 
-
+launch_bench: $(subdirs)
+	$(MAKE) -C $< $(MAKECMDGOALS)
 
diff --git a/kitsune/experiments/noforall/makefile b/kitsune/experiments/noforall/makefile
@@ -31,6 +31,9 @@ run: $(targets)
 
 .PHONY: all clean run
 
+launch_bench: 
+	@echo "launch benchmark invalid... skipping..."
+
 clean:
 	-rm -f *.${host_arch} *.ll *.o
 	-rm -f *~ core *.log 

diff --git a/kitsune/experiments/srad/makefile b/kitsune/experiments/srad/makefile
@@ -29,8 +29,8 @@ srad-forall.hip.${host_arch}: srad-forall.cpp
 	@$(FILE_SIZE)
 
 # kokkos-based tests (w/out views)
-#srad-kokkos.cuda.kitsune.${host_arch}: srad-kokkos-no-view.cpp 
-#	/usr/bin/time $(KIT_CXX) $(TAPIR_CUDA_FLAGS) $(KITSUNE_KOKKOS_FLAGS) -o $@ $< -Xlinker -rpath=$(KITSUNE_PREFIX)/lib
+srad-kokkos.cuda.kitsune.${host_arch}: srad-kokkos-no-view.cpp 
+	/usr/bin/time $(KIT_CXX) $(TAPIR_CUDA_FLAGS) $(KITSUNE_KOKKOS_FLAGS) -o $@ $< -Xlinker -rpath=$(KITSUNE_PREFIX)/lib
 #srad-kokkos.hip.kitsune.${host_arch}: srad-kokkos-no-view.cpp 
 #	/usr/bin/time $(KIT_CXX) $(TAPIR_HIP_FLAGS) -o $@ $< -Xlinker -rpath=$(KITSUNE_PREFIX)/lib