multi function invoke (#22)

* move concurrent queue to cmake fetch

* add tensor movement test

* fix asan, tsan bugs

Author: jeffreyqdd

CMakeLists.txt

@@ -1,23 +1,71 @@
-cmake_minimum_required(VERSION 3.14)
+cmake_minimum_required(VERSION 3.27)
 cmake_policy(SET CMP0148 NEW)
 project(pyscheduler VERSION 0.1.0 LANGUAGES CXX)
 
 set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+# fetch concurrent queue
+include(FetchContent)
+FetchContent_Declare(
+    concurrentqueue
+    GIT_REPOSITORY https://github.com/cameron314/concurrentqueue.git
+    GIT_TAG master
+)
+FetchContent_MakeAvailable(concurrentqueue)
 
 find_package(pybind11 CONFIG REQUIRED)
-find_package(concurrentqueue CONFIG REQUIRED)
+find_package(Threads REQUIRED)
 
 # create library target
 add_library(${PROJECT_NAME} INTERFACE)
 target_link_libraries(${PROJECT_NAME} INTERFACE
     pybind11::embed
-    concurrentqueue::concurrentqueue
+    concurrentqueue
+    Threads::Threads
 )
 target_include_directories(${PROJECT_NAME} INTERFACE
     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
     $<INSTALL_INTERFACE:include>
 )
 
+# Diagnostic flags
+option(ENABLE_GPROF "Enable gprof profiling flags (-pg)" OFF)
+option(ENABLE_FP "Enable frame pointer for flamegraph generation" OFF)
+option(ENABLE_ASAN  "Enable AddressSanitizer" OFF)
+option(ENABLE_TSAN  "Enable ThreadSanitizer" OFF)
+if (ENABLE_ASAN AND ENABLE_TSAN)
+    message(FATAL_ERROR "ASan and TSan cannot be enabled simultaneously.")
+endif()
+
+if (CMAKE_CXX_COMPILER_ID MATCHES "GNU|Clang")
+    if (ENABLE_GPROF)
+        message(STATUS "Enabling gprof (-pg)")
+        add_compile_options(-pg)
+        add_link_options(-pg)
+    endif()
+
+    if (ENABLE_FP)
+        message(STATUS "Enabling frame pointer (-fno-omit-frame-pointer)")
+        add_compile_options(-fno-omit-frame-pointer -g)
+        if (CMAKE_BUILD_TYPE MATCHES "Debug|RelWithDebInfo")
+            add_compile_options(-g)
+        endif()
+    endif()
+
+    if (ENABLE_ASAN)
+        message(STATUS "Using AddressSanitizer")
+        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+        add_link_options(-fsanitize=address)
+    endif()
+
+    if (ENABLE_TSAN)
+        message(STATUS "Using ThreadSanitizer")
+        add_compile_options(-fsanitize=thread -fno-omit-frame-pointer)
+        add_link_options(-fsanitize=thread)
+    endif()
+endif()
+
 # Build example programs
 option(BUILD_EXAMPLES "Build example programs" OFF)
 if(BUILD_EXAMPLES)

configure.sh

@@ -9,6 +9,11 @@ BUILD_EXAMPLES=OFF
 BUILD_TESTS=OFF
 PYTHON_UDL_INTERFACE_PREFIX="${PYTHON_UDL_INTERFACE_PREFIX:-/usr/local}"
 
+ENABLE_GPROF=OFF
+ENABLE_ASAN=OFF
+ENABLE_TSAN=OFF
+ENABLE_FP=OFF
+
 # Parse arguments
 while [[ "$#" -gt 0 ]]; do
     case $1 in
@@ -30,13 +35,29 @@ while [[ "$#" -gt 0 ]]; do
         PYTHON_UDL_INTERFACE_PREFIX="$2"
         shift
         ;;
+    --gprof)
+        ENABLE_GPROF=ON
+        ;;
+    --asan)
+        ENABLE_ASAN=ON
+        ;;
+    --tsan)
+        ENABLE_TSAN=ON
+        ;;
+    --flame)
+        ENABLE_FP=ON
+        ;;
     -h | --help)
         echo "Usage: ./configure.sh [options]"
-        echo "  -m | --mode      : Build type (default: Debug)"
+        echo "  -m | --mode      : Build type (Debug/Release/RelWithDebInfo)"
         echo "  -d | --dir       : Build directory (default: build)"
-        echo "  -e | --examples  : Enable building examples (default: OFF)"
-        echo "  -t | --tests     : Enable building tests (default: OFF)"
-        echo "  -p | --prefix    : Install prefix (default: /usr/local or \$PYTHON_UDL_INTERFACE_PREFIX)"
+        echo "  -e | --examples  : Enable building examples"
+        echo "  -t | --tests     : Enable building tests"
+        echo "  -p | --prefix    : Install prefix"
+        echo "  --gprof          : Enable gprof profiling"
+        echo "  --asan           : Enable AddressSanitizer"
+        echo "  --tsan           : Enable ThreadSanitizer"
+        echo "  --flame          : Enable frame pointer for flamegraph generation"
         exit 0
         ;;
     *)
@@ -47,12 +68,19 @@ while [[ "$#" -gt 0 ]]; do
     shift
 done
 
+# Summary
+BUILD_DIR="${BUILD_DIR}-${BUILD_TYPE}"
+
 echo "Configuring project..."
 echo "  Build type     : $BUILD_TYPE"
 echo "  Build dir      : $BUILD_DIR"
 echo "  Build examples : $BUILD_EXAMPLES"
 echo "  Build tests    : $BUILD_TESTS"
 echo "  Install prefix : $PYTHON_UDL_INTERFACE_PREFIX"
+echo "  ENABLE_GPROF   : $ENABLE_GPROF"
+echo "  ENABLE_ASAN    : $ENABLE_ASAN"
+echo "  ENABLE_TSAN    : $ENABLE_TSAN"
+echo "  ENABLE_FP      : $ENABLE_FP"
 
 # Create build directory
 mkdir -p "$BUILD_DIR"
@@ -63,6 +91,10 @@ cmake -B "$BUILD_DIR" \
     -DCMAKE_INSTALL_PREFIX="$PYTHON_UDL_INTERFACE_PREFIX" \
     -DBUILD_EXAMPLES="$BUILD_EXAMPLES" \
     -DBUILD_TESTS="$BUILD_TESTS" \
+    -DENABLE_GPROF="$ENABLE_GPROF" \
+    -DENABLE_ASAN="$ENABLE_ASAN" \
+    -DENABLE_TSAN="$ENABLE_TSAN" \
+    -DENABLE_FP="$ENABLE_FP" \
     -S .
 
 echo "Configuration complete. You can now run: cmake --build $BUILD_DIR"

examples/CMakeLists.txt

@@ -1,2 +1,5 @@
-add_executable(add_example add/main.cpp)
-target_link_libraries(add_example PRIVATE pyscheduler)
+add_executable(add add/main.cpp)
+target_link_libraries(add PRIVATE pyscheduler)
+
+add_executable(tensor_share tensor_share/main.cpp)
+target_link_libraries(tensor_share PRIVATE pyscheduler)

examples/add/main.cpp

@@ -6,6 +6,9 @@ int main() {
 	pyscheduler::PyManager manager;
 	pyscheduler::PyManager::InvokeHandler add =
 		manager.loadPythonModule("examples.add.python_modules.add", "invoke");
-	std::cout << add.invoke<int64_t>(3000, -1234) << std::endl;
+	for (int i = 0; i < 200; i++) {
+		std::cout << add.invoke<int64_t>(3000, -1234) << "\n";
+	}
+	std::cout.flush();
 	return 0;
 }

examples/add/python_modules/tensor_mul.py

@@ -0,0 +1,51 @@
+#include <chrono>
+#include <pyscheduler/pyscheduler.hpp>
+
+using namespace std::chrono;
+using namespace std;
+
+constexpr int const NUM_ITERATIONS = 600;
+constexpr int const DIM = 100;
+
+int main() {
+	auto module_load_start = chrono::high_resolution_clock::now();
+	pyscheduler::PyManager manager;
+	pyscheduler::PyManager::InvokeHandler generator = manager.loadPythonModule("examples.tensor_share.python_modules.tensor_juggler", "generate_tensor");
+	pyscheduler::PyManager::InvokeHandler fma = manager.loadPythonModule("examples.tensor_share.python_modules.tensor_juggler", "multiply_sum_tensors");
+	auto module_load_end = chrono::high_resolution_clock::now();
+
+	vector<std::pair<pybind11::object, pybind11::object>> cache;
+
+	for (int i = 0; i < NUM_ITERATIONS; i++) {
+		cache.emplace_back(
+			std::make_pair(
+				generator.invoke([&](pybind11::object x){return x; }, DIM, DIM),
+				generator.invoke([&](pybind11::object x){return x; }, DIM, DIM)
+			)
+		);
+	}
+
+	// solve problems
+	std::vector<std::future<int>> promises;
+	promises.reserve(NUM_ITERATIONS);
+
+
+	auto module_solve_start = chrono::high_resolution_clock::now();
+	for (int i = 0; i < NUM_ITERATIONS; i++) {
+		auto [a, b] = std::move(cache[i]);
+		promises.emplace_back(
+			fma.queue_invoke(
+				[](pybind11::object x){return x.cast<int>();}, 
+				std::move(a),
+				std::move(b)
+			)
+		);
+	}
+	for(auto &o : promises) {
+		o.wait();
+	}
+	auto module_solve_end = chrono::high_resolution_clock::now();
+	std::cout << duration_cast<microseconds>(module_solve_end - module_solve_start).count() << std::endl;
+
+	return 0;
+}

examples/tensor_share/main.cpp

@@ -0,0 +1,30 @@
+import torch.utils.dlpack as dlpack
+import torch
+import time
+def generate_tensor(length: int, width: int):
+    tensor = torch.rand(length, width, dtype=torch.float32)
+    return dlpack.to_dlpack(tensor) # type: ignore
+
+def multiply_sum_tensors(a, b) -> int:
+    a_tensor = dlpack.from_dlpack(a)
+    b_tensor = dlpack.from_dlpack(b)
+    c_tensor = a_tensor * b_tensor
+    d_tensor = torch.sum(c_tensor)
+    return int(d_tensor.item())
+
+
+def main():
+    DIM = 100
+    NUM_ITERATIONS = 600
+
+    x = time.perf_counter_ns()
+    a_tensors = [ generate_tensor(DIM, DIM) for _ in range(NUM_ITERATIONS) ]
+    b_tensors = [ generate_tensor(DIM, DIM) for _ in range(NUM_ITERATIONS) ]
+
+    for i in range(NUM_ITERATIONS):
+        result = multiply_sum_tensors(a_tensors[i], b_tensors[i])
+    y = time.perf_counter_ns()
+    print(f"Elapsed Time (us): {(y - x) / 1e3}")
+
+if __name__ == "__main__":
+    main()

examples/tensor_share/main.py

@@ -0,0 +1,12 @@
+import torch.utils.dlpack as dlpack
+import torch
+def generate_tensor(length: int, width: int):
+    tensor = torch.rand(length, width, dtype=torch.float32)
+    return dlpack.to_dlpack(tensor) # type: ignore
+
+def multiply_sum_tensors(a, b) -> int:
+    a_tensor = dlpack.from_dlpack(a)
+    b_tensor = dlpack.from_dlpack(b)
+    c_tensor = a_tensor * b_tensor
+    d_tensor = torch.sum(c_tensor)
+    return int(d_tensor.item())