chore(cuda): Add concrete-core as submodule in order to integrate concrete-cuda as a cmake subproject

.gitmodules

@@ -6,3 +6,7 @@
 	path = compiler/concrete-optimizer
 	url = git@github.com:zama-ai/concrete-optimizer.git
 	shallow = true
+[submodule "compiler/concrete-core"]
+	path = compiler/concrete-core
+	url = git@github.com:zama-ai/concrete-core.git
+	shallow = true

compiler/CMakeLists.txt

@@ -19,17 +19,17 @@ endif()
 
 # CMake library generation settings.
 set(BUILD_SHARED_LIBS
-    OFF
-    CACHE BOOL "Default to building a static mondo-lib")
+  OFF
+  CACHE BOOL "Default to building a static mondo-lib")
 set(CMAKE_PLATFORM_NO_VERSIONED_SONAME
-    ON
-    CACHE BOOL "Python soname linked libraries are bad")
+  ON
+  CACHE BOOL "Python soname linked libraries are bad")
 set(CMAKE_VISIBILITY_INLINES_HIDDEN
-    ON
-    CACHE BOOL "Hide inlines")
+  ON
+  CACHE BOOL "Hide inlines")
 
 # The -fvisibility=hidden option only works for static builds.
-if(BUILD_SHARED_LIBS AND (CMAKE_CXX_VISIBILITY_PRESET STREQUAL "hidden"))
+if(BUILD_SHARED_LIBS AND(CMAKE_CXX_VISIBILITY_PRESET STREQUAL "hidden"))
   message(FATAL_ERROR "CMAKE_CXX_VISIBILITY_PRESET=hidden is incompatible \
                         with BUILD_SHARED_LIBS.")
 endif()
@@ -75,10 +75,8 @@ if(CONCRETELANG_CUDA_SUPPORT)
   message(STATUS "Found CUDA version: ${CUDAToolkit_VERSION}")
   message(STATUS "Found CUDA library dir: ${CUDAToolkit_LIBRARY_DIR}")
   link_directories(${CUDAToolkit_LIBRARY_DIR})
+  add_subdirectory(${CONCRETE_CORE_PATH}/concrete-cuda/cuda)
   include_directories(${CONCRETE_CORE_PATH}/concrete-cuda/cuda/include)
-  add_library(ConcreteCUDA STATIC IMPORTED)
-  set_target_properties(ConcreteCUDA PROPERTIES IMPORTED_LOCATION
-                                                ${CONCRETE_CORE_PATH}/concrete-cuda/cuda/build/src/libconcrete_cuda.a)
   add_compile_options(-DCONCRETELANG_CUDA_SUPPORT)
 endif()
 
@@ -110,7 +108,7 @@ if(CONCRETELANG_DATAFLOW_EXECUTION_ENABLED)
   include_directories(SYSTEM ${HPX_INCLUDE_DIRS})
   list(APPEND CMAKE_MODULE_PATH "${HPX_CMAKE_DIR}")
   add_compile_options(-DCONCRETELANG_DATAFLOW_EXECUTION_ENABLED
-                      -DHPX_DEFAULT_CONFIG_FILE="${PROJECT_SOURCE_DIR}/hpx.ini")
+    -DHPX_DEFAULT_CONFIG_FILE="${PROJECT_SOURCE_DIR}/hpx.ini")
 
 else()
   message(STATUS "ConcreteLang dataflow execution disabled.")

compiler/Makefile

@@ -6,8 +6,8 @@ DATAFLOW_EXECUTION_ENABLED=OFF
 TIMING_ENABLED=OFF
 CC_COMPILER=
 CXX_COMPILER=
-CUDA_SUPPORT=OFF
-CONCRETE_CORE_PATH?=
+CUDA_SUPPORT?=OFF
+CONCRETE_CORE_PATH?= $(shell pwd)/concrete-core
 
 CONCRETE_OPTIMIZER_DIR ?= $(shell pwd)/concrete-optimizer
 
@@ -82,10 +82,6 @@ else
 		PYTHON_TESTS_MARKER="not parallel"
 endif
 
-ifneq ($(CONCRETE_CORE_PATH),)
-	CONCRETE_CORE_PATH_OPTION=-DCONCRETE_CORE_PATH=$(CONCRETE_CORE_PATH)
-endif
-
 all: concretecompiler python-bindings build-tests build-benchmarks build-mlbench doc
 
 # concrete-core-ffi #######################################
@@ -149,7 +145,7 @@ $(BUILD_DIR)/configured.stamp:
 	-DLLVM_EXTERNAL_CONCRETELANG_SOURCE_DIR=. \
 	-DPython3_EXECUTABLE=${Python3_EXECUTABLE} \
 	-DCONCRETE_OPTIMIZER_DIR=${CONCRETE_OPTIMIZER_DIR} \
-	$(CONCRETE_CORE_PATH_OPTION) \
+	-DCONCRETE_CORE_PATH=$(CONCRETE_CORE_PATH) \
 	-DCONCRETELANG_CUDA_SUPPORT=${CUDA_SUPPORT}
 	touch $@
 

compiler/lib/Runtime/CMakeLists.txt

@@ -6,7 +6,10 @@ if(CONCRETELANG_DATAFLOW_EXECUTION_ENABLED)
 endif()
 
 if(CONCRETELANG_CUDA_SUPPORT)
-  target_link_libraries(ConcretelangRuntime PRIVATE ConcreteCUDA cudart)
+  target_link_libraries(
+    ConcretelangRuntime
+    PRIVATE
+    concrete_cuda)
 endif()
 
 if(${CMAKE_SYSTEM_NAME} MATCHES "Darwin")

compiler/lib/Runtime/wrappers.cpp

@@ -180,8 +180,9 @@ void memref_bootstrap_lwe_cuda_u64(
   // run gpu bootstrap
   cuda_bootstrap_amortized_lwe_ciphertext_vector_64(
       stream, out_gpu, test_vector_gpu, test_vector_idxes_gpu, ct0_gpu,
-      fbsk_gpu, input_lwe_dim, poly_size, base_log, level, num_samples,
-      num_test_vectors, lwe_idx, cuda_get_max_shared_memory(gpu_idx));
+      fbsk_gpu, input_lwe_dim, glwe_dim, poly_size, base_log, level,
+      num_samples, num_test_vectors, lwe_idx,
+      cuda_get_max_shared_memory(gpu_idx));
   // copy output ciphertext back to cpu
   memcpy_async_ct_to_cpu(out_allocated, out_aligned, out_offset, out_size,
                          out_stride, out_gpu, out_size, gpu_idx, stream);