Use EigenBlasGemm improve convolution computing performance in ARMv7 environment. #3549
Conversation
hedaoyuan
changed the title
| @@ -55,6 +55,7 @@ option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF) | |||
| option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) | |||
| option(GLIDE_INSTALL "Download and install go dependencies " ON) | |||
| option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) | |||
| option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) | |||
There was a problem hiding this comment.
建议将USE_EIGEN_FOR_BLAS的默认值改成ON，跑一遍单测看看有没有问题。确认没有问题之后，再将默认值改回OFF。
There was a problem hiding this comment.
在本地ON和OFF都是测试过（x86, armv7, armv8），CI只是编译，修改过来实际也跑不了单侧。
There was a problem hiding this comment.
若设置了USE_EIGEN_FOR_BLAS，teamcity上也会编译使用Eigen计算gemm的版本，是可以跑单测的。线下我只跑了mobilenet这一个模型，不确定测试的是否全面。
paddle/function/CMakeLists.txt
Outdated
| @@ -4,6 +4,8 @@ file(GLOB cpp_files . *Op.cpp) | |||
| list(APPEND h_files Function.h) | |||
| list(APPEND cpp_files Function.cpp) | |||
| list(APPEND cpp_files BufferArg.cpp) | |||
| list(APPEND cpp_files GemmFunctor.cpp) | |||
| list(APPEND cpp_files EigenGemm.cpp) | |||
There was a problem hiding this comment.
USE_EIGEN_FOR_BLAS为OFF时也要编译EigenGemm.cpp？
paddle/function/EigenGemm.cpp
Outdated
| c.device(device) += a.contract(b, dims); | ||
| } else { | ||
| c.device(device) = | ||
| c.constant(alpha) * a.contract(b, dims) + c.constant(beta) * c; |
There was a problem hiding this comment.
设置了transpose的size，在执行contract()时，内部就会自动transpose？
c.device(device) = c.constant(alpha) * a.contract(b, dims) + c.constant(beta) * c;
不能直接用c.device(device) = alpha * a.contract(b, dims) + beta * c;？
There was a problem hiding this comment.
内部就会自动transpose？
是的。
c.device(device) = alpha * a.contract(b, dims) + beta * c;
也是可以的。
| T* C, | ||
| const int ldc); | ||
| struct BlasGemm { | ||
| static void compute(const bool transA, |
There was a problem hiding this comment.
调用的时候省去一次构造，直接BlasGemm<Device, real>::compute。
| @@ -55,6 +55,7 @@ option(WITH_C_API "Compile PaddlePaddle with C-API(Prediction)" OFF) | |||
| option(WITH_GOLANG "Compile PaddlePaddle with GOLANG" OFF) | |||
| option(GLIDE_INSTALL "Download and install go dependencies " ON) | |||
| option(USE_NNPACK "Compile PaddlePaddle with NNPACK library" OFF) | |||
| option(USE_EIGEN_FOR_BLAS "Use matrix multiplication in Eigen" OFF) | |||
There was a problem hiding this comment.
若设置了USE_EIGEN_FOR_BLAS，teamcity上也会编译使用Eigen计算gemm的版本，是可以跑单测的。线下我只跑了mobilenet这一个模型，不确定测试的是否全面。
In some environment, the performance of Eigen's matrix multiplication is higher than OpenBlas (like ARMv7). So add an EigenBlasGemm.