nervanagpu 0.3.1

简介

nervanagpu 是一个用于深度学习的Python模块。它包括，

• 矩阵乘法（GEMM）、卷积和池化内核，使用自定义 汇编器 优化，

• 元素级和广播操作，这些操作会自动组合成高效的内核，

• 一个简单但功能强大的数组类，利用并部分借用自 pycuda1 的代码，

• 用于构建网络的层类，以进行基准测试，

• 完整的汇编源代码，以鼓励社区贡献。

使用方法

nervanagpu包括一个工厂类NervanaGPU和一个类似NumPy的数组类GPUTensor。张量和GEMM操作的内存布局为行顺序。以下是如何使用它们的示例。

import numpy as np
import pycuda.autoinit
from nervanagpu import NervanaGPU

# initialize factory class
ng = NervanaGPU(stochastic_round=False)

m, n, k  = 10, 20, 10
dtype = np.float16

# create matrices on host
cpuA = np.random.randn(k,m)
cpuB = np.random.randn(k,n)

# transfer to device
devA = ng.array(cpuA, dtype=dtype)
devB = ng.array(cpuB, dtype=dtype)
devC = ng.empty((m,n), dtype=dtype)

# do GEMM operation
ng.dot(devA.T, devB, devC, relu=False)

# get from device
cpuC = devC.get()

# import and initialize NervanaGPU, transfer matrix from cpu to dev as above

devC[:] = ng.mean(ng.square(devA - ng.mean(devA, axis=1)), axis=1)

# import and initialize NervanaGPU as above

eps  = .001
A    = ng.empty((128, 32), dtype=np.float16)
A[:] = ng.rand() # generate uniform random on device between 0 and 1

# Normalize batch data by batch mean and variance,
A[:] = ng.reciprocal(ng.sqrt(ng.var(A, axis=1) + eps)) * (A - ng.mean(A, axis=1))

nervanagpu包含内核的完整汇编代码。要构建内核，安装**maxas**，Nervana为NVIDIA Maxwell提供的汇编器。然后可以通过运行以下命令构建模块

make kernels      # build the kernels
make python       # build python bindings
make test         # run nosetests
make doc          # build sphinx docs

简单的make将执行前两个步骤。

目前文档和测试较少。请贡献力量。

性能

nervanagpu在nervanagpu/benchmarks下附带一组基准脚本。还包括验证内核结果与cuBLAS和cuDNN的脚本来验证。

以下是使用Soumith Chintala的基准页面上列出的网络在单个TitanX上以默认时钟和功率限制运行的benchmarks/convnet-benchmarks.py的示例运行

---------------------------------------------
Alexnet (dtype=float16, N=128)  Results:
---------------------------------------------
Avg(10) fprop:   29.498 msecs 6043.698 gflops
Avg(10) bprop:   66.562 msecs 5356.689 gflops
Avg(10) total:   96.059 msecs 5567.654 gflops
---------------------------------------------
Alexnet (dtype=float32, N=128)  Results:
---------------------------------------------
Avg(10) fprop:   31.251 msecs 5704.698 gflops
Avg(10) bprop:   77.567 msecs 4596.660 gflops
Avg(10) total:  108.818 msecs 4914.869 gflops

---------------------------------------------
Overfeat (dtype=float16, N=128)  Results:
---------------------------------------------
Avg(10) fprop:  116.723 msecs 6134.994 gflops
Avg(10) bprop:  242.084 msecs 5916.054 gflops
Avg(10) total:  358.807 msecs 5987.277 gflops
---------------------------------------------
Overfeat (dtype=float32, N=128)  Results:
---------------------------------------------
Avg(10) fprop:  124.569 msecs 5748.559 gflops
Avg(10) bprop:  278.408 msecs 5144.192 gflops
Avg(10) total:  402.977 msecs 5331.015 gflops

---------------------------------------------
VGG (dtype=float16, N=64)  Results:
---------------------------------------------
Avg(10) fprop:  162.186 msecs 5978.348 gflops
Avg(10) bprop:  357.850 msecs 5419.051 gflops
Avg(10) total:  520.036 msecs 5593.481 gflops
---------------------------------------------
VGG (dtype=float32, N=64)  Results:
---------------------------------------------
Avg(10) fprop:  170.822 msecs 5676.112 gflops
Avg(10) bprop:  438.031 msecs 4427.108 gflops
Avg(10) total:  608.853 msecs 4777.533 gflops