[关闭]
@CrazyHenry 2018-04-19T12:27:53.000000Z 字数 6896 阅读 3704

faiss教程跟进--Makefile

hhhhfaiss

官方教程：https://github.com/facebookresearch/faiss/wiki

编译所有教程

  1. 在目录：
  2. /home/users/yingmin.li/proj/faiss_proj/faiss/tutorial/cpp
  3. 下执行：
  4. make all #即5个都编译
  6. make cpu #编译1、2、3
  7. make gpu #编译4、5
  8. make clean #删除二进制文件
  10. 只编译其中一个，比如编译4-GPU：
  12. make 4-GPU #注意没有.cpp

看懂了80%的Makefile：

  1. MAKEFILE_INC=../../makefile.inc
  3. -include $(MAKEFILE_INC)
  5. NVCCLDFLAGS =   -Xcompiler \"-Wl,-rpath=../../:../../gpu/\" \
  6.         -L../.. -L../../gpu -lfaiss -lgpufaiss
  8. LDFLAGS = -L../.. -Wl,-rpath=../.. -lfaiss
  11. #在此添加新的程序
  12. all: cpu gpu  #make all 指令所执行的范围
  14. cpu: 1-Flat 2-IVFFlat 3-IVFPQ #观察all: cpu gpu
  16. gpu: 4-GPU 5-Multiple-GPUs #观察all: cpu gpu
  18. 1-Flat: 1-Flat.cpp ../../libfaiss.$(SHAREDEXT)
  19. 	$(CXX) -o $@ $(CXXFLAGS) $< -I../../../ $(LDFLAGS)
  21. 2-IVFFlat: 2-IVFFlat.cpp ../../libfaiss.$(SHAREDEXT)
  22. 	$(CXX) -o $@ $(CXXFLAGS) $< -I../../../ $(LDFLAGS)
  24. 3-IVFPQ: 3-IVFPQ.cpp ../../libfaiss.$(SHAREDEXT)
  25. 	$(CXX) -o $@ $(CXXFLAGS) $< -I../../../ $(LDFLAGS)
  27. #GPU版本的写法同样可以编译CPU的程序
  28. 4-GPU: 4-GPU.cpp ../../libfaiss.$(SHAREDEXT) ../../gpu/libgpufaiss.$(SHAREDEXT)
  29.     $(NVCC) $(NVCCFLAGS) -o $@ $< $(NVCCLDFLAGS) -I../../../
  31. 5-Multiple-GPUs: 5-Multiple-GPUs.cpp ../../libfaiss.$(SHAREDEXT) \
  32. 	../../gpu/libgpufaiss.$(SHAREDEXT)
  33.     $(NVCC) $(NVCCFLAGS) -o $@ $< $(NVCCLDFLAGS) -I../../../
  35. #在此添加新的程序
  38. ../../libfaiss.$(SHAREDEXT):
  39. 	cd ../../ && make libfaiss.$(SHAREDEXT)
  41. ../../gpu/libgpufaiss.$(SHAREDEXT):
  42. 	cd ../../gpu/ && make libgpufaiss.$(SHAREDEXT)
  44. clean:
  45.     rm -f 1-Flat 2-IVFFlat 3-IVFPQ 4-GPU 5-Multiple-GPUs  #删除所有可执行程序

看一下makefile.inc

  2. # Copyright (c) 2015-present, Facebook, Inc.
  3. # All rights reserved.
  4. #
  5. # This source code is licensed under the BSD+Patents license found in the
  6. # LICENSE file in the root directory of this source tree.
  8. # -*- makefile -*-
  9. # tested on CentOS 7, Ubuntu 16 and Ubuntu 14, see below to adjust flags to distribution.
  12. CC=gcc
  13. CXX=g++  #CXX原来是g++
  15. CFLAGS=-fPIC -m64 -Wall -g -O3 -mavx -msse4 -mpopcnt -fopenmp -Wno-sign-compare -fopenmp
  16. CXXFLAGS=$(CFLAGS) -std=c++11  #使用c++11
  17. LDFLAGS=-g -fPIC  -fopenmp
  19. # common linux flags
  20. SHAREDEXT=so   #这里就是一个so
  21. SHAREDFLAGS=-shared
  22. FAISSSHAREDFLAGS=-shared
  25. ##########################################################################
  26. # Uncomment one of the 4 BLAS/Lapack implementation options
  27. # below. They are sorted # from fastest to slowest (in our
  28. # experiments).
  29. ##########################################################################
  31. #
  32. # 1. Intel MKL
  33. #
  34. # This is the fastest BLAS implementation we tested. Unfortunately it
  35. # is not open-source and determining the correct linking flags is a
  36. # nightmare. See
  37. #
  38. #   https://software.intel.com/en-us/articles/intel-mkl-link-line-advisor
  39. #
  40. # The latest tested version is MLK 2017.0.098 (2017 Initial Release) and can
  41. # be downloaded here:
  42. #
  43. #   https://registrationcenter.intel.com/en/forms/?productid=2558&licensetype=2
  44. #
  45. # The following settings are working if MLK is installed on its default folder:
  46. #
  47. # MKLROOT=/opt/intel/compilers_and_libraries/linux/mkl/
  48. #
  49. # BLASLDFLAGS=-Wl,--no-as-needed -L$(MKLROOT)/lib/intel64   -lmkl_intel_ilp64 \
  50. # -lmkl_core -lmkl_gnu_thread -ldl -lpthread
  51. #
  52. # BLASCFLAGS=-DFINTEGER=long
  53. #
  54. # you may have to set the LD_LIBRARY_PATH=$MKLROOT/lib/intel64 at runtime.
  55. # If at runtime you get the error:
  56. #   Intel MKL FATAL ERROR: Cannot load libmkl_avx2.so or libmkl_def.so.
  57. # You may add set
  58. #   LD_PRELOAD=$MKLROOT/lib/intel64/libmkl_core.so:$MKLROOT/lib/intel64/libmkl_sequential.so
  59. # at runtime as well.
  61. # 默认使用OpenBLAS，比MKL慢30%
  62. # 2. Openblas
  63. #
  64. # The library contains both BLAS and Lapack. About 30% slower than MKL. Please see 
  65. #   https://github.com/facebookresearch/faiss/wiki/Troubleshooting#slow-brute-force-search-with-openblas
  66. # to fix performance problemes with OpenBLAS 
  68. BLASCFLAGS=-DFINTEGER=int
  70. # This is for Centos: #默认使用centos
  71. BLASLDFLAGS?=/usr/lib64/libopenblas.so.0
  73. # for Ubuntu 16:
  74. # sudo apt-get install libopenblas-dev python-numpy python-dev
  75. # BLASLDFLAGS?=/usr/lib/libopenblas.so.0
  77. # for Ubuntu 14:
  78. # sudo apt-get install libopenblas-dev liblapack3 python-numpy python-dev
  79. # BLASLDFLAGS?=/usr/lib/libopenblas.so.0 /usr/lib/lapack/liblapack.so.3.0
  80. #
  81. # 3. Atlas
  82. #
  83. # Automatically tuned linear algebra package. As the name indicates,
  84. # it is tuned automatically for a give architecture, and in Linux
  85. # distributions, it the architecture is typically indicated by the
  86. # directory name, eg. atlas-sse3 = optimized for SSE3 architecture.
  87. #
  88. # BLASCFLAGS=-DFINTEGER=int
  89. # BLASLDFLAGS=/usr/lib64/atlas-sse3/libptf77blas.so.3 /usr/lib64/atlas-sse3/liblapack.so
  90. #
  91. # 4. reference implementation
  92. #
  93. # This is just a compiled version of the reference BLAS
  94. # implementation, that is not optimized at all.
  95. #
  96. # BLASCFLAGS=-DFINTEGER=int
  97. # BLASLDFLAGS=/usr/lib64/libblas.so.3 /usr/lib64/liblapack.so.3.2
  98. #
  101. ##########################################################################
  102. # SWIG and Python flags #对于我没啥用
  103. ##########################################################################
  105. # SWIG executable. This should be at least version 3.x
  106. SWIGEXEC=swig
  108. # The Python include directories for a given python executable can
  109. # typically be found with
  110. #
  111. # python -c "import distutils.sysconfig; print distutils.sysconfig.get_python_inc()"
  112. # python -c "import numpy ; print numpy.get_include()"
  113. #
  114. # or, for Python 3, with
  115. #
  116. # python3 -c "import distutils.sysconfig; print(distutils.sysconfig.get_python_inc())"
  117. # python3 -c "import numpy ; print(numpy.get_include())"
  118. #
  120. PYTHONCFLAGS=-I/usr/include/python2.7/ -I/usr/lib64/python2.7/site-packages/numpy/core/include/
  123. ###########################################################################
  124. # Cuda GPU flags 
  125. ###########################################################################
  129. # root of the cuda 8 installation
  130. CUDAROOT=/usr/local/cuda-8.0/  #使用cuda-8.0
  132. CUDACFLAGS=-I$(CUDAROOT)/include
  134. NVCC=$(CUDAROOT)/bin/nvcc
  136. NVCCFLAGS= $(CUDAFLAGS) \
  137.    -I $(CUDAROOT)/targets/x86_64-linux/include/ \
  138.    -Xcompiler -fPIC \
  139.    -Xcudafe --diag_suppress=unrecognized_attribute \
  140.    -gencode arch=compute_35,code="compute_35" \  #可以删除此行，因为我们的显卡很强
  141.    -gencode arch=compute_52,code="compute_52" \
  142.    -gencode arch=compute_60,code="compute_60" \
  143.    --std c++11 -lineinfo \
  144.    -ccbin $(CXX) -DFAISS_USE_FLOAT16
  146.   # GeForce GTX TITAN X 5.2是公司显卡的型号
  148. # 在Makefile里写$(NVCC) $(NVCCFLAGS) -o $@ $< $(NVCCLDFLAGS) -I../../../
  149. # 等价于 $($(CUDAROOT)/bin/nvcc) $($(CUDAFLAGS) \
  150. #   -I $(CUDAROOT)/targets/x86_64-linux/include/ \
  151. #   -Xcompiler -fPIC \
  152. #   -Xcudafe --diag_suppress=unrecognized_attribute \
  153. #   -gencode arch=compute_35,code="compute_35" \
  154. #   -gencode arch=compute_52,code="compute_52" \
  155. #   -gencode arch=compute_60,code="compute_60" \
  156. #   --std c++11 -lineinfo \
  157. #   -ccbin $(CXX) -DFAISS_USE_FLOAT16) -o $@ $< $(NVCCLDFLAGS) -I../../../
  158. #    NVCCLDFLAGS这个参数是存在于Makefile文件中的
  159. #   NVCCLDFLAGS=   -Xcompiler \"-Wl,-rpath=../../:../../gpu/\" \
  160. #       -L../.. -L../../gpu -lfaiss -lgpufaiss
  162. # BLAS LD flags for nvcc (used to generate an executable)
  163. # if BLASLDFLAGS contains several flags, each one may
  164. # need to be prepended with -Xlinker
  165. BLASLDFLAGSNVCC=-Xlinker $(BLASLDFLAGS)
  167. # Same, but to generate a .so
  168. BLASLDFLAGSSONVCC=-Xlinker  $(BLASLDFLAGS)

结论：

显卡型号：
image_1cbeti0381av7140e1qaecj7qqm.png-219.7kB

查阅官网得：
image_1cbetk5gu1r3pej6f72isr1rqp13.png-10kB

1.改成 -gencode arch=compute_52,code="compute_52"之后，计算能力进一步提升！！！！

2.GPU的完整编译指令是这样的：

  1. /usr/local/cuda-8.0//bin/nvcc  -I /usr/local/cuda-8.0//targets/x86_64-linux/include/ -Xcompiler -fPIC -Xcudafe --diag_suppress=unrecognized_attribute -gencode arch=compute_52,code="compute_52" -gencode arch=compute_60,code="compute_60" --std c++11 -lineinfo -ccbin g++ -DFAISS_USE_FLOAT16 -o 5-Multiple-GPUs 5-Multiple-GPUs.cpp -Xcompiler \"-Wl,-rpath=../../:../../gpu/\" -L../.. -L../../gpu -lfaiss -lgpufaiss -I../../../

image_1cbetalls12djj1dbpo1bo417p39.png-141.4kB

添加新批注
在作者公开此批注前，只有你和作者可见。
回复批注