Mentor Catapult C HLS的C++库源码/文档和AI参考设计

appleleaf

Algorighm C 数据类型库，包括定点，浮点，和复杂数据类型，用于描述bit accurate的算法以综合成硬件电路



Algorighm C 数学库，包含synthesizable C++ function，实现reciprocal, log, exponent, square-root, sin/cos/tan等功能，


Algorighm C DSP库，包括synthesizable C++实现的滤波器， FIR filters, IIR filters, Moving Average Filters and Polyphase Interpolation/Decimation， 以及傅里叶变换Decimation-in-Time and Decimation-in-Frequency等，

appleleaf

Catapult C 较为复杂的设计实例：

EdgeDetect_Algorithm.h - Original algorithm
EdgeDetect_BitAccurate.h - Adding bit-accurate data types
EdgeDetect_Synthesizable.h - Prepared for HLS
EdgeDetect_MemoryArch.h - Refinining memory architecture for 1PPC
EdgeDetect_Hierarchy.h - Modularized into hierarchy for performance
EdgeDetect_SinglePort.h - Recode to use single-port memories
EdgeDetect_SinglePort_Programable.h - Recode to make image size programable
EdgeDetect_CircularBuf.h - Recode to have line buffers operate in a circular fasion for power reduction

appleleaf

添加一个机器学习参考设计，
Algorithmic C Machine Learning Reference Designs
实例来自 HLS SEMINAR 2021 和 DVCON_2021

这个需要用到英伟达的 SystemC  库 MatchLib
MatchLib Connections is a SystemC library implementing latency-insensitive channels for use by High-Level synthesis tools.

RTL仿真比C++仿真慢了30倍，参见8楼补充图片，

tracy6969

thanks

watertree

thanks . study

liqunxiong

感谢分享

appleleaf

一个数学库里开平方的例子：

//=========================================================================
// Function: ac_sqrt (for ac_int)
//
// Description:
//    Calculation of square root of real inputs, passed as ac_int
//    variables.
//    + input: unsigned ac_int argument
//    + output: ac_int square root
//
//-------------------------------------------------------------------------

  template< int XW, int OW, bool OS >
  void ac_sqrt(
    ac_int<XW,false> x,
    ac_int<OW,OS> &sqrt
  )
  {
    const int RW = (XW+1)/2;
    // masks used only to hint synthesis on precision
    ac_int<RW+2, false> mask_d = 0;

    ac_int<RW+2, false> d = 0;
    ac_int<RW, false> r = 0;
    ac_int<2*RW, false> z = x;

    // needs to pick 2 bits of z for each iteration starting from
    // the 2 MSB bits. Inside loop, z will be shifted left by 2 each
    // iteration. The bits of interest are always on the same
    // position (z_shift+1 downto z_shift)
    unsigned int z_shift = (RW-1)*2;

    for (int i = RW-1; i >= 0; i--) {
      r <<= 1;

      mask_d = (mask_d << 2) | 0x3;
      d = (mask_d & (d << 2)) | ((z >> z_shift) & 0x3 );

      ac_int<RW+2, false> t = (ac_int<RW+2, false>)(d - (( ((ac_int<RW+1, false>)r) << 1) | 0x1));
      if ( !t[RW+1] ) { // since t is unsigned, look at MSB
        r |= 0x1;
        d = mask_d & t;
      }
      z <<= 2;
    }
    sqrt = r;
  }

appleleaf

机器学习参考设计里面的一个例子：

RTL仿真比C++仿真慢了30倍

zoe0615

Thank you a lot.

minidust

十分感谢分享