CNTK/Makefile_kaldi2.gpu

206 строки
9.1 KiB
Plaintext

# WORK IN PROGRESS, not currently complete nor usable
# NOTE: We are in the process of consolidating all Makefile's into a single one,
# with different build options. In the meantime, all active work will go into Makefile.gpu
# and other Makefile's, like this one, may fall behind.
# makefile for a Linux/GCC build of CNTK
# This needs ACML_PATH. E.g. in tcsh, say: setenv ACML_PATH C:/AMD/acml5.3.1/ifort64_mp
# This is work in progress and not at all complete or usable.
#
# The Linux and Windows versions are not different branches, but rather build off the same
# source files, using different makefiles. This current makefile has the purpose of enabling
# work to make all sources compile with GCC, and also to check for GCC-compat regressions due to
# modifications which are currently done under Windows.
#
# The planned steps are:
# - runnable non-GPU GCC-built version under Cygwin
# - get all CPU-only sources to compile with GCC/x64 under Cygwin --currently ongoing work
# - port the dynamic-loading mechanism
# - runnable non-GPU version on actual Linux
# - enable CUDA on Linux (=makefile code and figuring out the right compiler options)
#
# Any help is welcome, of course!
#
# This makefile will be extended/completed as we go.
#
# You may need to do the following or something similar for all this to work
# export LD_LIBRARY_PATH=/usr/local/acml5.3.0/gfortran64/lib:/usr/local/cuda/lib64:/usr/local/lib
# export PATH=$PATH:/usr/local/bin:/usr/local/cuda/bin
#
# WARNING:
# Since we now compile against Kaldi lattices, which depend on OpenFst, you
# will have to re-compile Kaldi with OpenFst 1.4.1 (lower version of OpenFst
# does not support c++11). You can do the following:
# 1. In kaldi-trunk/tools/Makefile, uncomment # OPENFST_VERSION = 1.4.1, and
# re-install OpenFst using the makefile.
# 2. In kaldi-trunk/src/, do ./configure --shared; make depend -j 8; make -j 8;
# and re-compile Kaldi.
#
# Besides, if you compile Kaldi with -O4 option instead of -g, it will be faster
# in sequence training.
CC = g++-4.8
NVCC = nvcc
ARCH = x86_64
DEVICE = gpu
#BUILDTYPE = debug
BUILDTYPE = release
# comment following and uncomment the next one to enable MKL library
#MATHLIB = acml
MATHLIB = mkl
# modify relevant path below for your system
MKL_PATH = /usr/users/yzhang87/tools/composer_xe_2015.2.164
ACML_PATH = /usr/users/yzhang87/code/acml/gfortran64
#######
BUILDFOR = $(ARCH).$(DEVICE).$(BUILDTYPE).$(MATHLIB)
OBJDIR = .build/$(BUILDFOR)
BINDIR = bin/$(BUILDFOR)
ifeq ($(BUILDTYPE),debug)
BUILDTYPE_OPT = -g
GPU_BUILDTYPE_OPT = -G
else
BUILDTYPE_OPT = -O4
GPU_BUILDTYPE_OPT =
endif
ifeq ($(MATHLIB),mkl)
MATHLIB_INCLUDE = $(MKL_PATH)/mkl/include
MATHLIB_LIB = -L$(MKL_PATH)/compiler/lib/intel64 -L$(MKL_PATH)/mkl/lib/intel64 -L$(MKL_PATH)/compiler/lib/mic -L$(MKL_PATH)/mkl/lib/mic -lmkl_intel_lp64 -lmkl_intel_thread -lmkl_core -lm -liomp5 -lpthread
MATHLIB_DEFINE = -DUSE_MKL
else
MATHLIB_INCLUDE = $(ACML_PATH)/include
MATHLIB_LIB = -L$(ACML_PATH)/lib -lacml -lm -lpthread
MATHLIB_DEFINE = -DUSE_ACML
endif
CUDA_PATH = /scratch/cuda-6.5
CUDA_INCLUDE = $(CUDA_PATH)/include
CUDA_LIB = -L$(CUDA_PATH)/lib64 -lcublas -lcudart -lcuda -lcurand -lcusparse -lnvidia-ml
NVML_INCLUDE = /scratch/usr/include/nvidia/gdk # install and include it @ https://developer.nvidia.com/gpu-deployment-kit
# Add KALDI (you need to add your Kaldi path into this file)
include kaldi_vars.mk
INCFLAGS = -I Common/Include -I Math/Math -I MachineLearning/CNTK -I $(CUDA_INCLUDE) -I $(MATHLIB_INCLUDE) $(KALDI_INCLUDES) -I $(NVML_INCLUDE)
CFLAGS = -msse3 -std=c++0x -std=c++11 -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K $(MATHLIB_DEFINE) -fopenmp -fpermissive $(KALDI_DEFINES)
NVCCFLAGS = -std=c++11 -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K -arch=compute_20 $(KALDI_DEFINES)
COMMON_SRC = Common/fileutil.cpp Common/DataWriter.cpp Common/ConfigFile.cpp Common/DataReader.cpp \
Common/Eval.cpp Common/File.cpp Common/BestGpu.cpp Common/TimerUtility.cpp
MATH_SRC = Math/Math/Matrix.cpp Math/Math/GPUMatrix.cu Math/Math/GPUMatrixCUDAKernels.cu Math/Math/GPUSparseMatrix.cu Math/Math/GPUWatcher.cu \
Math/Math/CPUMatrix.cpp Math/Math/CPUSparseMatrix.cpp #Math/Math/InstantiateTemplates.cu
CN_SRC = MachineLearning/CNTK/NetworkDescriptionLanguage.cpp MachineLearning/CNTK/CNTK.cpp MachineLearning/CNTK/ComputationNode.cpp \
MachineLearning/CNTK/ModelEditLanguage.cpp \
MachineLearning/CNTK/SimpleNetworkBuilder.cpp \
MachineLearning/CNTK/Profiler.cpp MachineLearning/CNTK/tests.cpp MachineLearning/CNTKEval/CNTKEval.cpp
BINARYREADER_SRC = #DataReader/BinaryReader/BinaryWriter.cpp DataReader/BinaryReader/BinaryReader.cpp DataReader/BinaryReader/BinaryFile.cpp
HTKMLFREADER_SRC = DataReader/HTKMLFReader_linux/HTKMLFWriter.cpp DataReader/HTKMLFReader_linux/DataWriter.cpp DataReader/HTKMLFReader_linux/DataReader.cpp DataReader/HTKMLFReader_linux/HTKMLFReader.cpp
KALDIREADER_SRC = DataReader/KaldiReader/HTKMLFWriter.cpp DataReader/KaldiReader/DataWriter.cpp DataReader/KaldiReader/DataReader.cpp DataReader/KaldiReader/HTKMLFReader.cpp
KALDI2READER_SRC = DataReader/Kaldi2Reader/HTKMLFWriter.cpp DataReader/Kaldi2Reader/DataWriter.cpp DataReader/Kaldi2Reader/DataReader.cpp DataReader/Kaldi2Reader/HTKMLFReader.cpp DataReader/Kaldi2Reader/UtteranceDerivativeBuffer.cpp DataReader/Kaldi2Reader/KaldiSequenceTrainingDerivative.cpp
SEQUENCEREADER_SRC = DataReader/LMSequenceReader/SequenceReader.cpp DataReader/LMSequenceReader/SequenceParser.cpp DataReader/LMSequenceReader/Exports.cpp
LUSEQUENCEREADER_SRC = DataReader/LUSequenceReader/LUSequenceReader.cpp DataReader/LUSequenceReader/LUSequenceParser.cpp DataReader/LUSequenceReader/Exports.cpp
UCIFASTREADER_SRC = DataReader/UCIFastReader/UCIParser.cpp DataReader/UCIFastReader/UCIFastReader.cpp DataReader/UCIFastReader/Exports.cpp
READER_SRC = $(UCIFASTREADER_SRC) $(LUSEQUENCEREADER_SRC) $(HTKMLFREADER_SRC) $(SEQUENCEREADER_SRC) $(BINARYREADER_SRC) $(KALDIREADER_SRC) $(KALDI2READER_SRC)
CORE_SRC = $(CN_SRC) $(MATH_SRC) $(COMMON_SRC)
SRC = $(READER_SRC) $(CORE_SRC)
VPATH := $(sort $(dir $(SRC)))
OBJ_TMP := $(patsubst %.cpp, $(OBJDIR)/%.o, $(SRC))
OBJ := $(patsubst %.cu, $(OBJDIR)/%.o, $(OBJ_TMP))
CORE_OBJ_TMP := $(patsubst %.cpp, $(OBJDIR)/%.o, $(CORE_SRC))
CORE_OBJ := $(patsubst %.cu, $(OBJDIR)/%.o, $(CORE_OBJ_TMP))
UCIFASTREADER_OBJ := $(patsubst %.cpp, $(OBJDIR)/%.o, $(UCIFASTREADER_SRC))
LUSEQUENCEREADER_OBJ := $(patsubst %.cpp, $(OBJDIR)/%.o, $(LUSEQUENCEREADER_SRC))
SEQUENCEREADER_OBJ := $(patsubst %.cpp, $(OBJDIR)/%.o, $(SEQUENCEREADER_SRC))
HTKMLFREADER_OBJ := $(patsubst %.cpp, $(OBJDIR)/%.o, $(HTKMLFREADER_SRC))
KALDIREADER_OBJ := $(patsubst %.cpp, $(OBJDIR)/%.o, $(KALDIREADER_SRC))
KALDI2READER_OBJ := $(patsubst %.cpp, $(OBJDIR)/%.o, $(KALDI2READER_SRC))
DEP := $(patsubst %.o, %.d, $(OBJ))
SEPARATOR = "=-----------------------------------------------------------="
all: $(BINDIR)/cntk $(BINDIR)/UCIFastReader.so $(BINDIR)/LMSequenceReader.so $(BINDIR)/LUSequenceReader.so $(BINDIR)/HTKMLFReader.so $(BINDIR)/Kaldi2Reader.so
ln -sf $(CURDIR)/$(BINDIR)/* bin
$(BINDIR)/UCIFastReader.so: $(UCIFASTREADER_OBJ) $(CORE_OBJ)
@echo $(SEPARATOR)
$(CC) $(BUILDTYPE_OPT) -fPIC -shared -o $@ $^
$(BINDIR)/LMSequenceReader.so: $(SEQUENCEREADER_OBJ) $(CORE_OBJ)
@echo $(SEPARATOR)
$(CC) $(BUILDTYPE_OPT) -fPIC -shared -o $@ $^
$(BINDIR)/LUSequenceReader.so: $(LUSEQUENCEREADER_OBJ) $(CORE_OBJ)
@echo $(SEPARATOR)
$(CC) $(BUILDTYPE_OPT) -fPIC -shared -o $@ $^
$(BINDIR)/HTKMLFReader.so: $(HTKMLFREADER_OBJ) $(CORE_OBJ)
@echo $(SEPARATOR)
$(CC) $(BUILDTYPE_OPT) -fPIC -shared -o $@ $^
$(BINDIR)/KaldiReader.so: $(KALDIREADER_OBJ) $(CORE_OBJ)
@echo $(SEPARATOR)
$(CC) $(BUILDTYPE_OPT) -fPIC -shared -o $@ $^ $(KALDI_LIBS)
$(BINDIR)/KaldiWriter.so: $(KALDIREADER_OBJ) $(CORE_OBJ)
@echo $(SEPARATOR)
$(CC) $(BUILDTYPE_OPT) -fPIC -shared -o $@ $^
$(BINDIR)/Kaldi2Reader.so: $(KALDI2READER_OBJ) $(CORE_OBJ)
@echo $(SEPARATOR)
$(CC) $(BUILDTYPE_OPT) -fPIC -shared -o $@ $^ $(KALDI_LIBS)
#$(BINDIR)/HTKMLFReader.so: ${HTKMLFREADER_SRC:.cpp=.o} ${COMMON_SRC:.cpp=.o} $(CORE_OBJ)
# @echo $(SEPARATOR)
# $(CC) -o $(addsuffix .so, $@) $^ -fPIC -shared
#BinaryReader: ${BINARYREADER_SRC:.cpp=.o} ${COMMON_SRC:.cpp=.o}
# $(CC) -o $(addsuffix .so, $@) $^ -fPIC -shared
$(BINDIR)/cntk: $(CORE_OBJ)
@echo $(SEPARATOR)
@mkdir -p $(dir $@)
@echo building output for $(ARCH) with build type $(BUILDTYPE)
$(CC) $(BUILDTYPE_OPT) -o $@ $^ $(CUDA_LIB) $(MATHLIB_LIB) -fopenmp -ldl -fPIC
-include ${DEP}
$(OBJDIR)/%.o : %.cu Makefile
@echo $(SEPARATOR)
@echo creating $@ for $(ARCH) with build type $(BUILDTYPE)
@mkdir -p $(dir $@)
$(NVCC) -c $< -o $@ $(BUILDTYPE_OPT) $(GPU_BUILDTYPE_OPT) $(NVCCFLAGS) $(INCFLAGS) -Xcompiler -fPIC
$(OBJDIR)/%.o : %.cpp Makefile
@echo $(SEPARATOR)
@echo creating $@ for $(ARCH) with build type $(BUILDTYPE)
@mkdir -p $(dir $@)
$(CC) -c $< -o $@ $(BUILDTYPE_OPT) $(CPPFLAGS) $(CFLAGS) $(INCFLAGS) -fPIC -MD -MP -MF ${@:.o=.d}
$(OBJDIR)/%.o : %.cc Makefile
@echo $(SEPARATOR)
@echo creating $@ for $(ARCH) with build type $(BUILDTYPE)
@mkdir -p $(dir $@)
$(CC) -c $< -o $@ $(BUILDTYPE_OPT) $(CPPFLAGS) $(CFLAGS) $(INCFLAGS) -fPIC -MD -MP -MF ${@:.o=.d}
.PHONY: clean
clean:
rm -rf $(OBJDIR)
rm -rf $(BINDIR)