numpy 可能表现不佳的原因是什么？

如何解决numpy 可能表现不佳的原因是什么？

我一直遇到 numpy 的性能问题。我测试了计算点积的不同方法，但其他 numpy 函数也出现了类似的性能问题。

import timeit
import numpy as np
import tensorflow as tf

def np_dot(U,X):
    tic = timeit.default_timer()
    X_bar = np.dot(U.T,X)
    toc = timeit.default_timer()
    print("np.dot took {:.4f} s".format(toc-tic))
    return X_bar

def np_at(U,X):
    tic = timeit.default_timer()
    X_bar = U.T @ X
    toc = timeit.default_timer()
    print("np @ took {:.4f} s".format(toc-tic))
    return X_bar

def np_matmul(U,X):
    tic = timeit.default_timer()
    X_bar = np.matmul(U.T,X)
    toc = timeit.default_timer()
    print("np.matmul took {:.4f} s".format(toc-tic))
    return X_bar

def np_einsum(U,X):
    tic = timeit.default_timer()
    X_bar = np.einsum('ij,jk',U.T,X)
    toc = timeit.default_timer()
    print("np.einsum took {:.4f} s".format(toc-tic))
    return X_bar

def tf_matmul(U,X):
    tic = timeit.default_timer()
    X_bar = tf.matmul(U.T,X)
    toc = timeit.default_timer()
    print("tf.matmul took {:.4f} s".format(toc-tic))
    return X_bar


if __name__ == "__main__":
    print("is_gpu_available?",tf.test.is_gpu_available(),tf.test.is_built_with_cuda())
    print("numpy version:",np.__version__)
    print("tensorflow version:",tf.__version__)
    
    M = N = 1000
    X = np.random.rand(M,N)
    U = np.random.rand(M,N)
    X_bar = np_dot(U,X)
    X_bar0 = np_at(U,X)
    X_bar1 = np_matmul(U,X)
    X_bar2 = np_einsum(U,X)
    X_bar3 = tf_matmul(U,X)
    print(np.allclose(X_bar,X_bar0))
    print(np.allclose(X_bar,X_bar1))
    print(np.allclose(X_bar,X_bar2))
    print(np.allclose(X_bar,X_bar3))

我的笔记本电脑（使用 Windows 10 和 python 3.8）上的输出是：

is_gpu_available? False False
numpy version: 1.19.2
tensorflow version: 2.2.0
np.dot took 12.5368 s
np @ took 14.3303 s
np.matmul took 14.0634 s
np.einsum took 0.5937 s
tf.matmul took 0.0263 s
True
True
True
True

意味着标准的 @ 比实际速度慢 545 倍！在朋友的笔记本电脑上，结果看起来更像预期的（使用 Ubuntu 和 python 3.6.9）： >

is_gpu_available? False False
numpy version: 1.19.5
tensorflow version: 1.14.0
np.dot took 0.0192 s
np @ took 0.0196 s
np.matmul took 0.0208 s
np.einsum took 0.3893 s
tf.matmul took 0.0280 s
True
True
True
True

有趣的是，numpys einsum 似乎不受我笔记本电脑出现问题的影响，tensorflow 也不受此影响。

这里发生了什么？ 造成如此巨大性能差异的潜在原因是什么？

编辑：

始终确保 numpy 正在使用某些 BLAS 库。在此处阅读更多相关信息：Boosting numpy: Why BLAS Matters。

就我而言，numpy.show_config() 返回以下内容：

blas_mkl_info:
  NOT AVAILABLE
blis_info:
  NOT AVAILABLE
openblas_info:
  NOT AVAILABLE
atlas_3_10_blas_threads_info:
  NOT AVAILABLE
atlas_3_10_blas_info:
  NOT AVAILABLE
atlas_blas_threads_info:
  NOT AVAILABLE
atlas_blas_info:
  NOT AVAILABLE
accelerate_info:
  NOT AVAILABLE
blas_info:
  NOT AVAILABLE
blas_src_info:
  NOT AVAILABLE
blas_opt_info:
  NOT AVAILABLE
lapack_mkl_info:
  NOT AVAILABLE
openblas_lapack_info:
  NOT AVAILABLE
openblas_clapack_info:
  NOT AVAILABLE
flame_info:
  NOT AVAILABLE
atlas_3_10_threads_info:
  NOT AVAILABLE
atlas_3_10_info:
  NOT AVAILABLE
atlas_threads_info:
  NOT AVAILABLE
atlas_info:
  NOT AVAILABLE
lapack_info:
  NOT AVAILABLE
lapack_src_info:
  NOT AVAILABLE
lapack_opt_info:
  NOT AVAILABLE
numpy_linalg_lapack_lite:
    language = c
    define_macros = [('HAVE_BLAS_ILP64',None),('BLAS_SYMBOL_SUFFIX','64_')]