Frequently Asked Questions
This page contains the most frequently asked questions about the Intel® DPC++ Compatibility Tool.
General Information
How do I migrate source files that are using C++11 or newer C++ standard features on Linux* and Windows*?
On Linux, the default C++ standard for the Intel® DPC++ Compatibility Tool’s parser is C++98, with some C++11 features accepted. If you want to enable other C++11 or newer standard features in the Intel® DPC++ Compatibility Tool, you need to add the --extra-arg="-std=<value>" option to the command line. The supported values are:
c++11
c++14
c++17
On Windows, the default C++ standard for the Intel® DPC++ Compatibility Tool’s parser is C++14. If you want to enable C++17 features in the Intel® DPC++ Compatibility Tool, you need to add the option --extra-arg="-std=c++17" to the command line.
How do I migrate files on Windows when using a CMake project?
For a CMake project on a Windows OS, you can use CMake to generate Microsoft Visual Studio* project files (vcxproj files). Then choose one of the following options:
Migrate the source files on the command line by using the --vcxprojfile option of the Intel® DPC++ Compatibility Tool.
Migrate the entire project in Microsoft Visual Studio with an Intel® DPC++ Compatibility Tool Microsoft Visual Studio plugin.
How is the migrated code formatted?
The Intel® DPC++ Compatibility Tool provides two options to control the format of migrated code: --format-range and --format-style .
If input source code is well formatted, the Intel® DPC++ Compatibility Tool will use default options settings --format-range and --format-style to format the resulting code.
If input source code is not well formatted (for example, the tool detects mixed use of tabs and spaces or mixed indents) you can do one of the following:
The Intel® DPC++ Compatibility Tool will try to detect the indent size of the original code and apply it to the resulting code. You can guide the tool by setting TabWidth and UseTab in the .clang-format file. Because the input source code is not well formatted, the indents in the resulting code may still be inconsistent.
Run the Intel® DPC++ Compatibility Tool with the --format-range=all option to format the entire resulting file. The change between input source code and resulting source code may be large and make it more difficult to compare the code.
Format your input source code, then use the Intel® DPC++ Compatibility Tool with the same .clang-format file for migration.
Why does the compilation database (compile_commands.json) not contain all source file(s) in the project?
In the project build folder, the command intercept-build make [target] is used to generate the compilation database. The content of the compilation database depends on the optional [target] parameter. If you need to get the list of files corresponding to default build target, do not specify the [target] parameter.
How do I use the migrated module file in the new project?
.cu module files are compiled with the -ptx or -cubin options in the original project and dynamically loaded into other *.cu files with cuModuleLoad() or cuModuleLoadData().
The Intel® DPC++ Compatibility Tool migrates module file code in the same way as other *.cu files. In addition, it adds a wrapper function for each function in the module file that has the _global_ attribute.
You can compile the migrated module file into a dynamic library and load the library with a dynamic library API appropriate to your platform. For example:
In Linux, load a dynamic library (.so) using dlopen()
In Windows, load a dynamic library (.dll) using LoadLibraryA()
Is the memory space allocated by sycl::malloc_device, sycl::malloc_host, and dpct::dpct_malloc initialized?
The memory allocated by sycl::malloc_device, sycl::malloc_host, and dpct::dpct_malloc is not initialized. If your program explicitly or implicitly relies on the initial value of newly allocated memory, the program may fail at runtime. Adjust your code to avoid such failures.
For example, the following original code:
// original code int *device_mem = nullptr;device_mem = sycl::malloc_device<int>(size, dpct::get_default_queue()); device_mem[0] += somevalue;
is adjusted to initialize the newly allocated memory to 0 before use:
// fixed SYCL code int *device_mem = nullptr;device_mem = sycl::malloc_device<int>(size, dpct::get_default_queue()); dpct::get_default_queue().memset(0, size).wait(); device_mem[0] += somevalue;
How do I migrate CUDA* source code that contains CUB library implementation source code?
If you migrate the CUB library implementation code directly, you may not get the expected results. Instead, exclude CUB library implementation source code from your migration by adding --in-root-exclude=<path to CUB library source code> to your migration command.
Troubleshooting
How do I fix an error such as “error: unknown type name” when I migrate files with “dpct –in-root=srcdir –out-root=dstdir *.cu”?
The problem may be caused by files in the *.cu list, which can be used as header files (included with an #include statement) and are not supposed to be parsed as a standalone file. In this case, the Intel® DPC++ Compatibility Tool reports an error if it cannot parse the file because the file depends on the definitions/declarations in other files. Use one of the methods below to migrate your content:
Rely on the Intel® DPC++ Compatibility Tool to decide which files to migrate with: compile_commands.json: “dpct -p=compile_commands.json --in-root=srcdir --out-root=dstdir”
Manually pass specific files to migrate, but do not pass the files that are included in other files and not supposed to be compiled as a standalone file in the original application. The header files are migrated automatically when they are included by the files provided as the input to the tool and are located within the in-root folder: dpct --in-root= srcdir --out-root=dstdir sample.cu
How do I fix a parsing error such as “no member named ‘max’ in namespace ‘std’” or “no member named ‘min’ in namespace ‘std’” when migrating code on Windows?
Use one of the following methods to resolve the error:
Add #include <algorithm> to the source file before using std::min and std::max
Define the NOMINMAX macro by inserting #define NOMINMAX before including WinDef.h
How do I fix a compilation error such as “error: dlopen not declared” when I compile code on a Windows machine, that was originally migrated on Linux?
When the Intel® DPC++ Compatibility Tool generates the source code, it uses dynamic loading APIs specific to the OS on which the Intel® DPC++ Compatibility Tool is running.
For example, dlopen, dlclose, and dlsym are used on Linux and LoadLibraryA, FreeLibrary, and GetProcAddress are used on Windows.
If your code was migrated on a OS that is different from the OS you need to compile the generated code on, migrate the project again with the Intel® DPC++ Compatibility Tool on the target OS or fix the code manually.
Why didn’t the “atomic*” APIs get migrated?
The Intel® DPC++ Compatibility Tool may assume that the “atomic*” APIs are user-defined APIs, in which case they are not migrated.
This can occur in the following scenarios:
The CUDA include path is specified by both --cuda-include-path and -I*, but the paths are different
The CUDA include path is specified by -I*, but there are other CUDA include files located on the default CUDA install path
To make sure “atomic*” APIs are migrated, don’t use -I* to specify the CUDA include path with the dpct migration command. Instead, use only --cuda-include-path to specify the CUDA include path.
Why did my migration fail with “error: restrict requires a pointer or reference”?
The C++ standard does not support the restrict qualifier and the C standard supports the restrict qualifier only on pointers to an object type.
Based on these language standards the Intel® DPC++ Compatibility Tool emits the parsing error.
You may need to adjust the source code.
How do I resolve incorrect runtime behavior for dpct::dev_mgr and dpct:mem_mgr in a library project that is loaded more than once in another application?
dpct::dev_mgr and dpct::mem_mgr are singleton classes in the Intel® DPC++ Compatibility Tool helper functions. When the helper function headers are used to build an executable project, both dpct::dev_mgr and dpct::mem_mgr will have only one instance in the executable. However, when the helper function headers are used to build a library project and the library project is loaded more than once with dlopen() (or LoadLibraryA() for Windows) in an application, more than two instances of dpct::dev_mgr and dpct::mem_mgr will be created and result in incorrect runtime behavior.
For example, both files libA.cpp and libB.cpp include the Intel® DPC++ Compatibility Tool helper function header dpct.hpp, and they are built into dynamic libraries libA.so and libB.so respectively. If an application main.cpp imports the libraries with dlopen(), there will be two instances of dpct::dev_mgr and dpct::mem_mgr in the runtime of the application.
To resolve this issue, separate the implementation and the declaration of dpct::dev_mgr and dpct::mem_mgr in the Intel® DPC++ Compatibility Tool helper function:
Create a new C++ file dpct_helper.cpp.
Move the implementation of instance() in class dev_mgr from dpct/device.hpp to dpct_helper.cpp.
For example, the original dpct/device.hpp:
class dev_mgr { public: static dev_mgr &instance() { // the implementation and the declaration of dev_mgr::instance static dev_mgr d_m; return d_m; } ... }is updated to:
class dev_mgr { public: static dev_mgr &instance();//the declaration of dev_mgr::instance ... }and the new dpct_helper.cpp now contains the implementation of dev_mgr::instance():
#include <dpct/device.hpp> dpct::dev_mgr &dev_mgr::instance(){ // the implementation of dev_mgr::instance static dev_mgr d_m; return d_m; }Similar to step two, move the implementation of instance() in the class mem_mgr from dpct/memory.hpp to dpct_helper.cpp.
Build dpct_helper.cpp into a dynamic library libdpct_helper.
In Linux:
dpcpp -g -shared -o libdpct_helper.so -fPIC ./dpct_helper.cpp
In Windows:
cl.exe /LD dpct_helper.cpp
Add library libdpct_helper to the environment variables.
In Linux: Add the location of libdpct_helper.so into LD_LIBRARY_PATH.
In Windows: Add the location of libdpct_helper.dll into PATH.
Dynamically link libdpct_helper when building libraries and applications.
After performing the update steps, all the libraries and applications will share the same instance of the device manager dpct::dev_mgr and the memory manager dpct::mem_mgr in Intel® DPC++ Compatibility Tool helper functions.
Why do I get “warning: shift count >= width of type” when I compile migrated code with the Intel® oneAPI DPC++/C++ Compiler?
Shifting bits where the shift is greater than the type length is undefined behavior for the Intel® oneAPI DPC++/C++ Compiler and may result in different behavior on different devices. Adjust your code to avoid this type of shift.
For example, the migrated SYCL code:
// migrated SYCL code
void foo() {
...
unsigned int bit = index[tid] % 32;
unsigned int val = in[tid] << 32 - bit;
...
}
is adjusted to avoid a bit shift that is greater than the type length:
// fixed SYCL code
void foo() {
...
unsigned int bit = index[tid] % 32;
unsigned int val;
if(32 - bit == 32)
val = 0;
else
val = in[tid] << 32 - bit;
...
}