GAMES 201 Lecture 1 Taichi Programing language

Lecture 1

目录->GAMES201学习笔记

Phase of a taichi program

1. Initialization: ti.init((debug=True),arch=ti....)
2. Tensor allocation:
   i) ti.var(现为 ti.field)
   ii) ti.Vector
3. Computation (launch kernels, access tensors in Python-scope)
4. Optional: restart the Taichi system
   (i.e. clear memory, destory all variables and kernels):ti.reset()

• Note: For now, after the first kernel launch or tensor access in Python-scope, no more tensor allocation is allowed.

Data in Taichi programing language

Taichi is a data-oriented programing language where tensors are first-class citzens

• Tensor

  • Tensors are essentially multi-dimensional arrays.
  • An element of a tensor can be either a scalar, a vector (ti.Vector) or a matrix (ti.Matrix)
  • Tensor elements are always accessed via the a[i, j, k] synatx (No pointer!)
  • Tensors can be spatially sparse
    ex: a=ti.field (dt=ti.f32, shape=(42,63))
    注：太极中的矩阵规模一般在2x2~3x3左右，更大规模的矩阵需要用tensor实现

Kernels

• Kernels must be decorated with @ti.kernel, kernel arguments and return values must be type-hinted.
  ex:

@ti.kernel
def calc() -> ti.i32
	... ...
	return

• Taichi function can be called by Taichi kernels and other Taichi functions.->(But function can not call kernels)

Functions

• Functions must be decorated with @ti.func
  ex:

@ti.func
def triple(x):
	return x*3


@ti.kernel
def triple_array:
	for i in range(128):
		a[i]=triple(a[i])

• Note:Taichi functions will be force-inlined, recursion is not allowed.contain at most one return statement.
• Most python math operators are supported in taichi.
  
• Taichi supports chain comparisions, e.g. a<b<=c!=d

Matrices and linear algebra

• ti.Matrix is for small matrices only.
  • NOTE:Differentiate element-wise product-> "*"
    Matrix product->"@"
    ex:A.transpose()->转置
    A.inverse()->逆
    A.trace()->迹
    A.determinant(type)->行列式
    A.normalized
    A.cast(type)->类型转换
    ti.sin(A)/cos(A)->element-wise的类型转换

For Loops in Taichi

Range-for loops

• no different from Python->for i in range(...): ...
• it will be paralleled when used at the outermost scope
• can be nested

Struct-for loops (稀疏数据结构常用)

• Iterates over（sparse）tensor elements
  Taichi Kernel -> for loops 最外层自动并行化
  Note: It is the loop at "the outermost scope" that gets parallized, not the outermost loop.

Atomic operations

• In taichi, augmented assignments (e.g. x[i]+=1) -> automatically atomic.
  ex: total[None] += x[i] -> atomic
  ti.atomic_add(total[None],x[i]) ->atomic
  toatal[None] = total[None] + x[i] ->not atomic
  

Scope

Taichi-Scope

• ->Everything decorated with ti.kernel,ti.func.
• Code in Taichi-Scope will be compiled by the Taichi compiler and run on parallel devices.

Python-Scope

->Code outside the Taichi-Scope

• Code in Python-Scope is simply Python code and will executed by the Python interpreter.
  补充：使用 yapf ( +PEP8 )实现代码的格式化
• ※：随着太极版本更新，更多更详细的内容请参照太极的文档->Taichi 文档
  finished.