README_linalg

Pan Deng / Zora edited this page Aug 28, 2017 · 8 revisions

Internal linear algebra library

Table of Contents

Motivation

Linear algebra operations form the backbone for most of the computation components in any Machine Learning library. However, writing all of the required linear algebra operations from scratch is rather redundant and undesired, especially when we have some excellent open source alternatives. In Shogun, we prefer

  • Eigen3 for its speed and simplicity at the usage level,
  • ViennaCL version 1.5 for GPU powered linear algebra operations.

For Shogun maintainers, however, the usage of different external libraries for different operations can lead to a painful task.

  • For example, consider some part of an algorithm originally written using Eigen3 API. But a Shogun user wishes to use ViennaCL for that algorithm instead, hoping to obtain boosted performance utilizing a GPU powered platform. There is no way of doing that without having the algorithm rewritten by the developers using ViennaCL, which leads to duplication of code and effort.
  • Also, there is no way to do a performance comparison for the developers while using different external linear algebra libraries for the same algorithm in Shogun code.
  • It is also somewhat frustrating for a new developer who has to invest significant amount of time and effort to learn each of these external APIs just to add a new algorithm in Shogun.

Features of internal linear algebra library

Shogun's internal linear algebra library (will be referred as linalg hereinafter) is a work-in-progress attempt to overcome these issues. We designed linalg as a modularized internal library in order to

  • provide a uniform API for Shogun developers to choose any supported backend without having to worry about the syntactical differences in the external libraries' operations,
  • have the backend set for each operations at compile-time (for lesser runtime overhead) and therefore intended to be used internally by Shogun developers,
  • allow Shogun developers to add new linear algebra backend plug-ins easily.

For Shogun developers

Setting linalg backend

Users can switch between linalg backends via global variable sg_linalg.

  • Shogun uses Eigen3 backend as default linear algebra backend.
  • Enabling of GPU backend allows the data transfer between CPU and GPU, as well as the operations on GPU. ViennaCL(GPU) backend can be enabled by assigning new ViennaCL backend class to sg_linalg or canceled by:
    sg_linalg->set_gpu_backend(new LinalgBackendViennaCL());
    sg_linalg->set_gpu_backend(nullptr);
  • Though backends can be extended, only one CPU backend and one GPU backend are allowed to be registered each time.

Using linalg operations

linalg library works for both SGVectors and SGMatrices. The operations can be called by:

#include <shogun/mathematics/linalg/LinalgNamespace.h>
shogun::linalg::operation(args)
  • To use linalg operations on GPU data (vectors or matrices) and transfer data between GPU, one can call to_gpu and from_gpu methods. Users should pre-allocate memory to complete the transfer.
    // Pre-allocate SGVectors a, b, and c
    SGVector<int32_t> a(size), b, c;

    // Initialize SGVector a
    // SGVectors and SGMatrices are initialized on CPU by default
    a.range_fill(0);

    // Copy values from CPU SGVector(a) to GPU SGVector(b) by pre-assigned GPU backend
    // If SGVector a is already on GPU: no operation will be done
    // If there is no GPU backend available: shallow copy a to b
    to_gpu(a, b);

    // Copy values from GPU SGVector(b) to CPU SGVector(c) by pre-assigned GPU backend
    // If SGVector a is already on CPU: no operation will be done
    // If there is no GPU backend available: raise error
    from_gpu(b, c);

    // Transfer values from CPU to GPU
    to_gpu(a);
  • to_gpu method and from_gpu methods are atomic.

  • The operations will be carried out on GPU only if the data passed to the operations are on GPU and GPU backend is registered: sg_linalg->get_gpu_backend() == true. The linalg will be conducted on CPU if the data is on CPU.

  • The operations will be carried out on GPU only if the data passed to the operations are on GPU and GPU backend is registered: sg_linalg->get_gpu_backend() == true. The linalg will be conducted on CPU if the data is on CPU.

  • linalg will report errors if the data is on GPU but no GPU backend is available anymore. Errors will also occur when an operation requires multiple inputs but the inputs are not on the same backend.

  • The status of data can be checked by: data.on_gpu(). True means the data is on GPU and false means the data is on CPU.

Examples

Here we show how to do vector dot with linalg library operations on CPU and GPU.

// CPU dot operation

#include <shogun/lib/SGVector.h>
#include <shogun/mathematics/linalg/LinalgNamespace.h>

using namespace shogun;

// Create SGVectors
const index_t size = 3;
SGVector<int32_t> a(size), b(size);
a.range_fill(0);
b.range_fill(0);

auto result = linalg::dot(a, b);
// GPU dot operation

#include <shogun/lib/SGVector.h>
#include <shogun/mathematics/linalg/LinalgNamespace.h>
#include <shogun/mathematics/linalg/LinalgBackendViennaCL.h>

using namesapce shogun;

// Set gpu backend
sg_linalg->set_gpu_backend(new LinalgBackendViennaCL());

// Create SGVectors
const index_t size = 3;
SGVector<int32_t> a(size), b(size), a_gpu, b_gpu;
a.range_fill(0);
b.range_fill(0);

// Transfer vectors to GPU
linalg::to_gpu(a, a_gpu);
linalg::to_gpu(b, b_gpu);

// run dot operation
auto result = linalg::dot(a_gpu, b_gpu);

If the result is a vector or matrix, it needs to be transferred back

#include <shogun/lib/SGVector.h>
#include <shogun/mathematics/linalg/LinalgNamespace.h>
#include <shogun/mathematics/linalg/LinalgBackendViennaCL.h>

using namesapce shogun;

// set gpu backend
sg_linalg->set_gpu_backend(new LinalgBackendViennaCL());

// Create a SGVector
SGVector<float32_t> a(5), a_gpu;
a.range_fill(0);

// Transfer the vector to gpu
linalg::to_gpu(a, a_gpu);

// Run scale operation and transfer the result back to CPU
auto result_gpu = linalg::scale(a_gpu, 0.3);
SGVector<float32_t> result;
from_gpu(result_gpu, result);

For linalg developers

linalg consists of three groups of components:

  • The interface that decides which backend to use for each operation (LinalgNameSpace.h)
  • The structure serves as interface of GPU backend libraries (GPUMemory*.h)
  • The operation implementations in each backend (LinalgBackend*.h).

Understanding operation interface LinalgNamespace.h

  • LinalgNamespace.h defines multiple linalg operation interfaces in namespace linalg. All operation methods will call infer_backend() method on the inputs, and decide the backend to call.

Understanding backend interfaces

  • GPUMemoryBase class is a generic base class serving as GPU memory library interface. The GPU data is referred as GPUMemoryBase pointer once it is generated by to_GPU() method, and is cast back to specific GPU memory type during operations.

  • GPUMemoryViennaCL is ViennaCL specific GPU memory library interface, which defines the operations to access and manipulate data on GPU with ViennaCL operations.

Understanding operation implementations of different backends

  • LinalgBackendBase is the base class for operations on all different backends. The macros in LinalgBackendBase class defined the linalg operations and data transfer operations available in at least one backend.

  • LinalgBackendGPUBase has two pure virtual methods: to_gpu() and from_gpu(). LinalgBackendViennaCL and other user-defined GPU backend classes are required to be derived from LinalgBackendGPUBase class, and thus GPU transfer methods are required to be implemented.

  • LinalgBackendEigen and LinalgBackendViennaCL* classes provide the specific implementations of linear algebra operations with Eigen3 library and ViennaCL library.

Extend external libraries

Current linalg framework allows easy addition of external linear algebra libraries. To add CPU-based algebra libraries, users just need to derive from LinalgBackendBase and re-implement the methods with new library. For GPU-based libraries, users need to add new class derived from LinalgBackendGPUBase, as well as the GPU memory library interface class derived from GPUMemoryBase class.

Clone this wiki locally
You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.
Press h to open a hovercard with more details.