Added benchmark to compare tapenade and clad performance during disk offloading and the tapenade is fetched only at the build time

Author: Vedant2005goyal

benchmark/CMakeLists.txt

@@ -7,23 +7,37 @@ CB_ADD_GBENCHMARK(ArrayExpressionTemplates ArrayExpressionTemplates.cpp)
 if (CLAD_ENABLE_ENZYME_BACKEND)
   CB_ADD_GBENCHMARK(EnzymeCladComparison EnzymeCladComparison.cpp)
 endif(CLAD_ENABLE_ENZYME_BACKEND)
+
+include(FetchContent)
+FetchContent_Declare(
+  tapenade_kit
+  GIT_REPOSITORY https://gitlab.inria.fr/tapenade/tapenade.git
+  
+  # We use a commit hash from the 'develop' branch because
+  # the 'disk offloading' feature is not yet available in the stable release of tapenade.
+  GIT_TAG        6f8cb04a4168cac5f6242cdeb65f6816c413f4ee 
+  
+  GIT_SHALLOW    TRUE
+  GIT_PROGRESS   TRUE      
+)
+FetchContent_MakeAvailable(tapenade_kit)
+
+# Build the support library for the stack from the fetched source
+add_library(tapenade_support STATIC ${tapenade_kit_SOURCE_DIR}/ADFirstAidKit/adStack.c)
+target_include_directories(tapenade_support PUBLIC ${tapenade_kit_SOURCE_DIR}/ADFirstAidKit)
+
+target_compile_definitions(tapenade_support PRIVATE 
+      ADSTACK_MAX_SPACES=1        
+      ADSTACK_BLOCK_SIZE=4194304   
+)
+
 CB_ADD_GBENCHMARK(VectorModeComparison VectorModeComparison.cpp)
-CB_ADD_GBENCHMARK(MemoryComplexity MemoryComplexity.cpp)
+CB_ADD_GBENCHMARK(MemoryComplexity_tapenade MemoryComplexity.cpp)
 CB_ADD_GBENCHMARK(Multithreading Multithreading.cpp)
-CB_ADD_GBENCHMARK(Hessians Hessians.cpp)
-
-set (CLAD_BENCHMARK_DEPS clad)
-get_property(_benchmark_names DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR} PROPERTY TESTS)
 
-foreach (name ${_benchmark_names})
-  get_test_property(${name} LABELS _labels)
-  if (_labels MATCHES ".*benchmark.*")
-    get_test_property(${name} DEPENDS _deps)
-    list(APPEND CLAD_BENCHMARK_DEPS ${_deps})
-  endif()
-endforeach()
+target_link_libraries(MemoryComplexity_tapenade PRIVATE tapenade_support)
 
 add_custom_target(benchmark-clad COMMAND ${CMAKE_CTEST_COMMAND} -V
   DEPENDS ${CLAD_BENCHMARK_DEPS} WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
 
-set_target_properties(benchmark-clad PROPERTIES FOLDER "Clad benchmarks")
+set_target_properties(benchmark-clad PROPERTIES FOLDER "Clad benchmarks")

benchmark/MemoryComplexity.cpp

@@ -4,6 +4,11 @@
 #include "clad/Differentiator/Tape.h"
 #include <cstddef>
 #include <cstdint>
+
+extern "C" {
+void pushReal8(double x);
+void popReal8(double* x);
+}
 namespace {
   struct MemoryManager : public benchmark::MemoryManager {
     size_t cur_num_allocs = 0;
@@ -185,4 +190,44 @@ static void BM_CrashTest_Clad_Offload(benchmark::State& state) {
 }
 BENCHMARK(BM_CrashTest_Clad_Offload)->Iterations(1);
 
+// Tapenade Comparison Benchmarks
+struct Tapenade {
+  void push(double val) { pushReal8(val); }
+  void pop(double* val) { popReal8(val); }
+};
+struct CladDisk {
+  clad::tape_impl<double, 64, 1024, /*is_multithread=*/false,
+                  /*DiskOffload=*/true>
+      tape;
+
+  void push(double val) { tape.emplace_back(val); }
+  void pop(double* val) {
+    *val = tape.back();
+    tape.pop_back();
+  }
+};
+template <typename Strategy> static void BM_PushPop(benchmark::State& state) {
+  int64_t n = state.range(0);
+  Strategy s; // Instantiate the specific strategy (Clad or Tapenade)
+
+  for (auto _ : state) {
+    for (int64_t i = 0; i < n; ++i)
+      s.push(1.0);
+    for (int64_t i = 0; i < n; ++i) {
+      double val;
+      s.pop(&val);
+      benchmark::DoNotOptimize(val);
+    }
+  }
+}
+BENCHMARK_TEMPLATE(BM_PushPop, Tapenade)
+    ->RangeMultiplier(4)
+    ->Range(1024, 262144)
+    ->Name("BM_PushPop/TapenadeStack");
+
+BENCHMARK_TEMPLATE(BM_PushPop, CladDisk)
+    ->RangeMultiplier(4)
+    ->Range(1024, 262144)
+    ->Name("BM_PushPop/CladDiskStack");
+
 BENCHMARK_MAIN();