remove ARB_HAVE_GPU from header file (#574)

Move implementation of `gpu_context` from header to `cpp` file, so that `ARB_WITH_CUDA` doesn't leak from library implementation.

arbor/backends/gpu/stack.hpp

 #pragma once
 
 #include <algorithm>
+#include <memory>
 
 #include <arbor/assert.hpp>
 

arbor/gpu_context.cpp

 #include <memory>
 
+#ifdef ARB_HAVE_GPU
+#include <cuda.h>
+#include <cuda_runtime.h>
+#endif
+
 #include "gpu_context.hpp"
 
 namespace arb {
 
+bool gpu_context::has_concurrent_managed_access() const {
+    return attributes_ & gpu_flags::has_concurrent_managed_access;
+}
+
+bool gpu_context::has_atomic_double() const {
+    return attributes_ & gpu_flags::has_atomic_double;
+}
+
+#ifndef ARB_HAVE_GPU
+
+gpu_context::gpu_context(): has_gpu_(false), attributes_(0) {}
+void gpu_context::synchronize_for_managed_access() const {}
+
+#else
+
+gpu_context::gpu_context(): has_gpu_(true), attributes_(0) {
+    cudaDeviceProp prop;
+    cudaGetDeviceProperties(&prop, 0);
+    if (prop.concurrentManagedAccess) {
+        attributes_ |= gpu_flags::has_concurrent_managed_access;
+    }
+    if (prop.major*100 + prop.minor >= 600) {
+        attributes_ |= gpu_flags::has_atomic_double;
+    }
+};
+
+void gpu_context::synchronize_for_managed_access() const {
+    if(!has_concurrent_managed_access()) {
+        cudaDeviceSynchronize();
+    }
+}
+
+#endif
+
 std::shared_ptr<gpu_context> make_gpu_context() {
     return std::make_shared<gpu_context>();
 }

arbor/gpu_context.hpp

-#include <memory>
-
-#ifdef ARB_HAVE_GPU
-#include <cuda.h>
-#include <cuda_runtime.h>
-#endif
+#pragma once
 
 namespace arb {
 
-#ifndef ARB_HAVE_GPU
-struct gpu_context {
-    bool has_gpu_;
-    size_t attributes_;
-
-    gpu_context(): has_gpu_(false), attributes_(0) {}
-};
-
-#else
-
 enum gpu_flags {
     has_concurrent_managed_access = 1,
     has_atomic_double = 2
@@ -26,32 +11,11 @@ struct gpu_context {
     bool has_gpu_;
     size_t attributes_;
 
-    gpu_context() : has_gpu_(true) {
-        cudaDeviceProp prop;
-        cudaGetDeviceProperties(&prop, 0);
-        attributes_ = 0;
-        if (prop.concurrentManagedAccess) {
-            attributes_ |= gpu_flags::has_concurrent_managed_access;
-        }
-        if (prop.major*100 + prop.minor >= 600) {
-            attributes_ |= gpu_flags::has_atomic_double;
-        }
-    };
-
-    bool has_concurrent_managed_access() {
-        return attributes_ & gpu_flags::has_concurrent_managed_access;
-    }
-
-    bool has_atomic_double() {
-        return attributes_ & gpu_flags::has_atomic_double;
-    }
+    gpu_context();
 
-    void synchronize_for_managed_access() {
-        if(!has_concurrent_managed_access()) {
-            cudaDeviceSynchronize();
-        }
-    }
+    bool has_concurrent_managed_access() const;
+    bool has_atomic_double() const;
+    void synchronize_for_managed_access() const;
 };
 
-#endif
 }

include/CMakeLists.txt

@@ -27,15 +27,19 @@ add_custom_command(
 
 set(arb_features)
 if(ARB_WITH_ASSERTIONS)
+    # define ARB_ASSERT_ENABLED in version.hpp
     list(APPEND arb_features ASSERT)
 endif()
 if(ARB_WITH_MPI)
+    # define ARB_MPI_ENABLED in version.hpp
     list(APPEND arb_features MPI)
 endif()
 if(ARB_WITH_CUDA)
+    # define ARB_GPU_ENABLED in version.hpp
     list(APPEND arb_features GPU)
 endif()
 if(ARB_WITH_PROFILING)
+    # define ARB_PROFILE_ENABLED in version.hpp
     list(APPEND arb_features PROFILE)
 endif()
 

test/unit/CMakeLists.txt

@@ -140,6 +140,5 @@ add_executable(unit ${unit_sources} ${test_mech_sources})
 add_dependencies(unit build_test_mods)
 target_compile_options(unit PRIVATE ${ARB_CXXOPT_ARCH})
 target_compile_definitions(unit PRIVATE "-DDATADIR=\"${CMAKE_CURRENT_SOURCE_DIR}/swc\"")
-target_compile_definitions(unit PRIVATE ARB_HAVE_GPU)
 target_include_directories(unit PRIVATE "${CMAKE_CURRENT_BINARY_DIR}")
 target_link_libraries(unit PRIVATE gtest arbor arbor-private-headers arbor-aux)