Deep Neural Network Library (DNNL)  1.2.0
Performance library for Deep Learning
 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Groups Pages
Intel MKL-DNN to DNNL Transition Guide

To simplify library naming and differentiate it from Intel MKL, starting with version 1.1 the library name is changed to Deep Neural Network Library (DNNL).

The library will maintain backward compatibility with respect to API, environment variables and build options until the next major release. However, there are some incompatibilities that are described in Broken compatibility with Intel MKL-DNN section below.

1. Summary of Changes

In short, the migration can be as simple as just replacing all MKLDNN/mkldnn substrings with DNNL/dnnl.

1.1. Source code changes

All headers, functions, types, and namespaces renamed by replacing mkldnn with dnnl. The macros with MKLDNN are replaced with DNNL counterparts.

An example of code with Intel MKL-DNN v1.0:

#include "mkldnn.hpp"
using namespace mkldnn;
mkldnn_memory_desc_t md;
if (md.format_kind == mkldnn_blocked) {}
conv.exec(stream, {{MKLDNN_ARGS_SRC, src}, ...});

The updated example with DNNL v1.1:

#include "dnnl.hpp"
using namespace dnnl;
if (md.format_kind == dnnl_blocked) {}
conv.exec(stream, {{DNNL_ARGS_SRC, src}, ...});

To API compatibility with Intel MKL-DNN is based on include/mkldnn_dnnl_mangling.h header file that maps all Intel MKL-DNN symbols to DNNL ones using C preprocessor:

// ...
#define mkldnn_memory_desc_t dnnl_memory_desc_t
#define mkldnn_memory_desc_init_by_tag dnnl_memory_desc_init_by_tag
// ...

This file is included to every former Intel MKL-DNN header files (for instance, see mkldnn.h) along with the DNNL counterpart.

1.2. Build process

The changes to the build options are similar to the ones in the source code. All the options and namespace with MKLDNN are replaced with DNNL:

Intel MKL-DNN DNNL
MKLDNN (namespace) DNNL (namespace)
MKLDNN_ARCH_OPT_FLAGS DNNL_ARCH_OPT_FLAGS
MKLDNN_BUILD_EXAMPLES DNNL_BUILD_EXAMPLES
MKLDNN_BUILD_FOR_CI DNNL_BUILD_FOR_CI
MKLDNN_BUILD_TESTS DNNL_BUILD_TESTS
MKLDNN_CPU_RUNTIME DNNL_CPU_RUNTIME
MKLDNN_ENABLE_CONCURRENT_EXEC DNNL_ENABLE_CONCURRENT_EXEC
MKLDNN_ENABLE_JIT_PROFILING DNNL_ENABLE_JIT_PROFILING
MKLDNN_GPU_BACKEND DNNL_GPU_BACKEND
MKLDNN_GPU_RUNTIME DNNL_GPU_RUNTIME
MKLDNN_INSTALL_MODE DNNL_INSTALL_MODE
MKLDNN_LIBRARY_TYPE DNNL_LIBRARY_TYPE
MKLDNN_THREADING DNNL_THREADING
MKLDNN_USE_CLANG_SANITIZER DNNL_USE_CLANG_SANITIZER
MKLDNN_VERBOSE DNNL_VERBOSE
MKLDNN_WERROR DNNL_WERROR

Similarly to the source code, DNNL preserves compatibility for build process as well. It should be possible to continue using:

# Through find package
find_package(mkldnn MKLDNN CONFIG REQUIRED)
target_link_libraries(project_app MKLDNN::mkldnn)
# Or direct sub-project inclusion
add_subdirectory(${MKLDNN_DIR} MKLDNN)
include_directories(${MKLDNN_DIR}/include)
target_link_libraries(project_app mkldnn)

Though it is preferable to switch to:

# Through find package
find_package(dnnl DNNL CONFIG REQUIRED)
target_link_libraries(project_app DNNL::dnnl)
# Or direct sub-project inclusion
add_subdirectory(${DNNL_DIR} DNNL)
include_directories(${DNNL_DIR}/include)
target_link_libraries(project_app dnnl)

In case both style options are set, the DNNL one takes precedence.

1.3. Runtime parameters

DNNL supports both old Intel MKL-DNN and new DNNL environment variable controls. DNNL value takes precedence over Intel MKL-DNN ones.

Intel MKL-DNN DNNL
Verbose MKLDNN_VERBOSE DNNL_VERBOSE
Dumping jit kernels MKLDNN_JIT_DUMP DNNL_JIT_DUMP

2. Broken compatibility with Intel MKL-DNN

Unfortunately, the full compatibility after renaming is not implemented. DNNL is not compatible with Intel MKL-DNN in the following things:

3. Information for developers

The implementation of renaming (several patches and scripts that rename the library) can be found here.

Also it is worth mentioning that DNNL team finally switched to the mandatory code formatting based on _clang-format file in the root of the repository. The corresponding changes were done by this and neighbor commits.