BMLang C++ 用户开发文档¶

声明¶

声明

介绍¶

编程模型¶

BMLang程序的基本元素¶

Tensor
- 张量(Tensor)
- 标量(Scalar)
Operator
- NN/Matrix Operator
  - conv
  - deconv
  - conv3d
  - matrix_mul
  - batch_norm
  - softmax
- NN/Matrix Int Operator
  - conv_v2
  - conv3d_v2
  - deconv_v2
  - deconv3d_v2
  - matrix_mul_v2
  - fc_int
  - softmax_int
- Base Element-wise Operator
  - mul
  - div
  - add
  - sub
  - max
  - min
  - adds
  - subs
  - muls
  - divs
  - rsubs
  - rdivs
  - axpy
  - maxs
  - mins
  - active
  - active_param
  - ceil
  - floor
  - round
  - clamp
  - cast
  - mul_shift
  - add_shift
  - sub_shift
  - muls_shift
  - adds_shift
  - subs_shift
  - rsubs_shift
  - copy
- Element-wise Compare Operator
  - gt
  - ge
  - lt
  - le
  - eq
  - ne
  - eqs
  - nes
  - ges
  - gts
  - les
  - lts
- Element-Wise Bit Logic Operator
  - arith_lshift
  - arith_lshifts
  - arith_rshift
  - arith_rshifts
  - round_rshift
  - round_rshifts
  - bit_and
  - bit_ands
  - bit_or
  - bit_ors
  - bit_xor
  - bit_xors
- Triangle Operator
  - sin
  - cos
- Stride Element-Wise Operator
- Data Arrange Operator
  - permute
  - tile
  - tile_v2(TODO)
  - broadcast
  - extract
  - extract_v2
  - interleave
  - concat
  - split
  - pad
  - pad_v2
  - space_to_batch
  - space_to_batch_v2(TODO)
  - batch_to_space
  - batch_to_space_v2(TODO)
  - repeat
- Select Operator
  - condition_select
  - select
  - masked_select
  - index_select
  - where
  - lut
- Up/Down Scaling Operator
  - pool
  - pool_int
  - pool3d_int
  - upsample
  - reduce
  - hw_avg
  - hw_mul
- Sort Operator
  - arg
  - topk
  - sort
  - argsort
- Shape About Operator
- Quant Operator
- Data Genenrate Operator
- Vision Operator
- Custom Operator
  - customed_ops
- Shift Operator
Control Functions

对用户DL框架和模型支持¶

基于BMNet编译器开发自定义算子¶

对非DL算法的加速¶

实现高性能的BMLang程序¶