Paddle Tensor Computation

[hedaoyuan]

About Paddle Tensor Computation
https://github.com/hedaoyuan/Paddle/tree/tensor_doc/paddle/math/doc

当前Paddle Tensor Compute主要Code在BaseMatrix中，BaseMatrix在函数定义，功能扩展和性能优化上都有不足，所以新增TensorExpression模块来解决并替换BaseMatrix。后续在Refactoring Computation API #977 的时候尽量使用TensorExpression中的实现。

1. 函数定义上
   TensorExpression能够简化代码实现，比如上面链接中的sgdUpdate例子。
2. 功能扩展上
   BaseMatrix受限于 hl_matrix_ops.cuh一个表达式中最多只能支持4个element，而TensorExpression没有限制。
3. 性能优化上
   TensorExpression可以通过Lazy Assignment将多条表达式在一个Kernel里面展开执行。

[hedaoyuan]

Paddle Function New Features

• support aggregate
• support add vector(row vector/col vector)
• support avx
• support cuda fp16

[hedaoyuan]

Some Simple Example
对比基于BaseMatrix和基于TensorExpression的实现
// Tensor can be BaseMatrix, CpuMatrix, GpuMatrix, CpuVector, GpuVector

template<typename Tensor>
void testTensorSubScalar(Tensor& A1, Tensor& A2) {
  real p = 2.5;
  A1.subScalar(p);  // a -= p
  A2 -= p;
  TensorCheckEqual(A1, A2);
}

template<typename Tensor>
void testTensorAdd(Tensor& A1, Tensor& A2, Tensor& B, Tensor& C) {
  A1.add2(B, C, p1, p2, p3);  // a = p1 * a + p2 * b + p3 * c
  A2 = A2 * p1 + B * p2 + C * p3;
  TensorCheckEqual(A1, A2);
}

[hedaoyuan]

CrossMapNormal
这是CrossMapNormal分别基于BaseMatrix API和基于TensorExpression的算法实现。两段代码的算法实现逻辑是一样的，但性能上第二段基于TensorExpression的实现要高一些，相比第一段代码表现在：

1. 不用增加额外内存（tmp）；
2. 更少的load/store；

{
  CpuMatrix tmp(channels, height * width);
  CpuMatrix tmp2(tmp.getData(), 1, channels * height * width);
  denoms.zero();
  const int start = -((int)sizeX - 1) / 2;
  const int end = (int)sizeX + start;
  for (size_t i = 0; i < numSample; i++) {
    input.subMatrix(i, 1)->square2(tmp2);
    CpuMatrix subDen(
        denoms.subMatrix(i, 1)->getData(), channels, height * width);
    for (int c = 0; c < (int)channels; c++) {
      for (int s = start; s < end; s++) {
        if (c + s >= 0 && c + s < (int)channels) {
          subDen.subMatrix(c, 1)->add(*tmp.subMatrix(c + s, 1));
        }
      }
    }
  }
  denoms.add(scale, (real)1);
  this->pow2(denoms, -pow);
  this->dotMul(input);
}

  denomsV = denomsV.constant(1.0);
  const int start = -((int)size - 1) / 2;
  const int end = (int)size + start;
  for (size_t i = 0; i < numSamples; i++) {
    real* oneDenom = denoms + i * oneSample;
    real* oneInput = const_cast<real*>(inputs) + i * oneSample;
    for (int c = 0; c < (int)channels; c++) {
      CpuVector denom(oneImage, oneDenom + c * oneImage);
      for (int s = start; s < end; s++) {
        if (c + s >= 0 && c + s < (int)channels) {
          CpuVector input(oneImage, oneInput + (c + s) * oneImage);
          denom += input.square() * scale;
        }
      }
    }
  }
  outputsV = inputsV * denomsV.pow(-pow);

[gongweibao]

Closing this issue due to inactivity, feel free to reopen it.