The build process (how Kaldi is compiled)

This page describes in general terms how the Kaldi build process works.

See also External matrix libraries for an explanation of how the matrix code uses external libraries and the linking errors that can arise from this; Downloading and installing Kaldi may also be of interest.

Build process on Windows

The build process for Windows is separate from the build process for UNIX-like systems, and is described in windows/INSTALL (tested some time ago with Windows 7 and Microsoft Visual Studio 2013). We use scripts to create the Visual Studio 10.0 solution file. There are two options for the math library on Windows: either Intel MKL, or use Cygwin to compile ATLAS. Detailed instructions are provided. However, note that the Windows setup is becoming out of date and is not regularly tested, and not all the may compile.

How our configure script works (for UNIX variants)

The "configure" script, located in src/, should be run by typing ./configure. This script takes various options. For instance, you can run

   ./configure --shared

if you want to build the setup with shared libraries instead of static libraries (this can make the binaries smaller), and it also takes various options to enable you to specify different mat libraries, for instance to use OpenBlas. Look at the top of the script to see some example command lines. The configure script currently supports are Cygwin, Darwin (which is Apple's version of BSD UNIX), and Linux. The configure script produces a file called kaldi.mk. This file will be textually included by the Makefiles in the subdirectores (see below).

Editing kaldi.mk

Changes that you might want to make to kaldi.mk after running "configure" are the following:

  • Changing the debug level:
    • The default is "-O1"
    • Easy to debug binaries can be enabled by uncommenting the options "-O0 -DKALDI_PARANOID".
    • For maximum speed and no checking, you can replace the "-O0 -DKALDI_PARANOID" options with "-O2 -DNDEBUG" or "-O3 -DNDEBUG"
  • Changing the default precision
    • To test algorithms in double precision (e.g. if you suspect that roundoff is affecting your results), you can change 0 to 1 in the option -DKALDI_DOUBLEPRECISION=0.
  • Suppressing warnings
    • To suppress warnings caused by signed-unsigned comparisons in OpenFst code, you can add -Wno-sign-compare to CXXFLAGS

It is also possible to modify kaldi.mk to use different math libraries (e.g. to use CLAPACK instead of ATLAS versions of LAPACK functions) and to link to math libraries dynamically instead of statically, but this is quite complicated and we can't give any generic instructions that will enable you to do this (see External matrix libraries to understand the compilation process for our math code). It will probably be easier to work with the options to configure script instead.

Targets defined by the Makefiles

The targets defined by the Makefiles are:

  • "make depend" will rebuild the dependencies. It's a good idea to run this before building the toolkit. If the .depend files gets out of date (because you haven't run "make depend"), you may get errors that look like this:
    make[1]: *** No rule to make target `/usr/include/foo/bar', needed by `baz.o'.  Stop.
    
  • "make all" (or just "make") will compile all the code, including testing code.
  • "make test" will run the testing code (useful for making sure the build worked on your system, and that you have not introduced bugs)
  • "make clean" will remove all compiled binaries, .o (object) files and .a (archive) files.
  • "make valgrind" will run the test programs under valgrind to check for memory leaks.
  • "make cudavalgrind" will run the test program (in cudamatrix) to check for memory leaks for machine with GPU card which support NVIDIA CUDA and OS with CUDA installation.

Where do the compiled binaries go?

Currently, the Makefiles do not put the compiled binaries in a special place; they just leave them in the directories where the corresponding code is. Currently, binaries exist in the directories "bin/", "gmmbin/", featbin/", "fstbin/", and "lm/", all of which are subdirectories of "src/". In the future we may designate a single place to put all the binaries.

How our Makefiles work

Currently the file src/Makefile just invokes the Makefiles in all the source sub-directories (src/base, src/matrix and so on). These directories have their own Makefiles, all of which share a common structure. They all include the line:

include ../kaldi.mk

This is like an #include line in C (it includes the text of kaldi.mk). When reading kaldi.mk, bear in mind that it is to be invoked from one directory below where it is actually located (it is located in src/). An example of what the kaldi.mk file looks like is as follows. This is for a Linux system; we have removed some rules relating to valgrind that are not very important.

ATLASLIBS = /usr/local/lib/liblapack.a /usr/local/lib/libcblas.a \
          /usr/local/lib/libatlas.a /usr/local/lib/libf77blas.a
CXXFLAGS = -msse -Wall -I.. \
      -DKALDI_DOUBLEPRECISION=0 -msse2 -DHAVE_POSIX_MEMALIGN \
     -DHAVE_EXECINFO_H=1 -rdynamic -DHAVE_CXXABI_H \
      -DHAVE_ATLAS -I ../../tools/ATLAS/include \
       -I ../../tools/openfst/include \
      -g -O0 -DKALDI_PARANOID
LDFLAGS = -rdynamic
LDLIBS = ../../tools/openfst/lib/libfst.a -ldl $(ATLASLIBS) -lm
CC = g++
CXX = g++
AR = ar
AS = as
RANLIB = ranlib

So kaldi.mk is responsible for setting up include paths, defining preprocessor variables, setting compile options, linking with libraries, and so on.

Which platforms has Kaldi been compiled on?

We have compiled Kaldi on Windows, Cygwin, various flavors of Linux (including Ubuntu, CentOS, Debian, Red Hat and SUSE), and Darwin. We recommend you use g++ version 4.7 or above, although other compilers such as llvm and Intel's icc are also known to work.