---
tags: CSIE,Research
---
# CCF Manual

## 1. CCF
- 全名叫做**C**GRA **C**ompilation **F**ramework
> CGRA:Coarse-Grained Reconfigurable Array

- 主要simulate的架構為heterogeneous processor core+CGRAs architecture
>heterogeneous:使用不同的ISA與不同的architecture做運算
### 1.1. Overview of CCF

- 可以在C/C++ code增加`#pragma CGRA`即可使CGRA做運算
> CCF-Compiler會自動extracts(提取)

### 1.2. Organization of Source Code Directories

- `RAMP`:負責mapping loop到CGRA上(CONFIG在這!!)
- `InstructionGenerator`:傳遞需要的 machine instructions
- `scripts`:包含下載的資料名稱，主要有CGRA Library function
- `work/benchmarks/Mibench`:範例的程式(有用到CCF)
- `basicmath`:基本的運算及迴圈
- `bitcount`:二位元的運算及迴圈(待確認)
---
## 2. Code Generation and Validation Using CCF
### 2.1. Loop Annotation

想要CGRA跑的話，在需要的loop註記`#pragma CGRA`

### 2.2. Make

原本是用`gcc`這個compiler，取代成`cgracc`
>cgracc全名為CCF's CGRA Compiler Collection

CCF-compiler會說明有沒有用到CGRA
>如果在loop裡有system call，卻沒有用到CGRA
>或是compiler對code vectorlized(向量化?)
>就不會產生給CGRA的code
>同時也會說明為何不產生

>如果compiler有產生給CGRA的code
>則會有一個`CGRAExec`的資料夾
>裡面會包含給CGRA的所有loops的information

### 2.3. Simulate Heterogeneous Execution

- simulate platform是建立在`gem5`上的，可以執行`se_hetro.py`來取代`se.py`
>SE - System call emulation mode
>hetro - 使用多於一種的processor執行
- 在CCF的資料夾中可以輸入來執行模擬用了 CGRA+CPU 的gem5 model
>n=2 cores, 1 core is specified as a CGRA, and another n-1=1 is the processor core
>`-c`:Run Executable File

gem5/build/ARM/gem5.opt gem5/configs/example/se_hetro.py -n 2 --cpu-type atomic -c ./work/benchmarks/MiBench/basicmath-CCF/basicmath12/basicmath_small > output_small.txt

- 如果想要直接在外面嘗試的話要改成這樣

#!/bin/sh
#qemu-arm ./bitcnts 75000 > output_small.txt

gem5/build/ARM/gem5.opt gem5/configs/example/se_hetro.py -n 2 --cpu-type atomic -c ./work/benchmarks/MiBench/bitcount-CCF/bitcount9/bitcnts --options="75000" > output_small.txt

> 在work中裡的資料夾都可以去改
- 如果對debug mode有興趣,可以增加一些參數，可以看到CGRA的micro-architecture conponents

--debug-flags=CGRA,CGRA_Detailed

---
## 3. CCF’s Code Generation Steps

說明如何產生CGRA的code

### 3.1. Extraction of the Annotated Loop(s)

- CCF-compiler從C/C++的程式碼去找`#pragma CGRA`
> CCF-compiler可以透過clang去修改
- 之後會找到intermediate representation (IR)
> `temporary.ll`
> 在這裡若有符合annotated loop 的metadata則CCF-compiler會來分析及轉換到CGRA上
- CCF target擁有最高優化的執行權
> 舉例來說auto-vectorization enabled
- 若IR符合CCF-compiler的要求，則會產生DDG
> DDG: data dependency graph
### 3.2. Generation of DDG and Communication of the Live Data

- LLVM pass(DDGGen)
> DDG可以被視覺化 [Graphviz tool (online)]( http://www.graphviz.org/)

### 3.3. Mapping of DDG on the Target CGRA

- DDG map到CGRA上時會順便產生prologue, kernel, and the epilogue of the mapping

- **The source code of our framework is parameterized and can model different CGRA configurations**
>XML based input for target configuration

- 他的target就是16個不同的pe然後是一個4x4的2-D Network
> each PE has 4 local registers.

- 但CCF專注在support general-purpose application

### 3.4. Generation of Machine Instructions

#### 3.4.1 Instruction Formats(是32-bit)

- R-Type Instruction Format

| 31:28 | 27 | 26:24 | 23:21 | 20:19 | 18:17 | 16:15 | 14 | 13 | 12 | 11:0 |
|:------:|:---:|:-----:|:-----:|:-----:|:-----:|:-----:|:---:|:---:|:---:|:---------:|
| opcode | P | LMUX | RMUX | R1 | R2 | RW | WE | AB | DB | Immediate |

- P-Type Instruction Format

| 31:28 | 27 | 26:24 | 23:21 | 20:19 | 18:17 | 16:15 | 14:12 | 11:0 |
|:------:|:---:|:-----:|:-----:|:-----:|:-----:|:-----:|:---:|:---------:|
| opcode | P | LMUX | RMUX | R1 | R2 | RP | PMUX | Immediate |

### 3.5. Architectural Simulation

- 建立在gem5的platform，他把CGRA當作ARM Cortex的seperate cortex，並且當作heterogeneous platform

### 3.6. Techniques Implemented

>看起來像縮寫的解釋
- CCF: A CGRA Compilation Framework
- RAMP: Resource-Aware Mapping for CGRAs
- URECA: A Compiler Solution to Manage Unified Register File for CGRAs
- REGIMap: Register-aware Application Mapping on Coarse-grained Reconfigurable Architectures (CGRAs) (for implementation of clique-based mapping heuristic, routing through registers)
- Power-Efficient Predication Techniques for Acceleration of Control Flow Execution on CGRA (partial predication based execution of conditional operations)

## test

s108321042@lab304-ubuntu:~$ docker attach dd
root@479fceca14b2:/# cd ~
root@479fceca14b2:~# ls
ccf  ccf-origin  ccf2*2  cmake-3.11.1-Linux-x86_64  cmake-3.11.1-Linux-x86_64.tar.gz  gem5
root@479fceca14b2:~# cd ccf
root@479fceca14b2:~/ccf# ls
CCF Manual.pdf  InstructionGenerator  RAMP  README.md  gem5  install  llvm  plugin-llvm-gold  scripts  work
root@479fceca14b2:~/ccf# cd work/benchmarks/MiBench/basicmath-CCF/
basicmath1/  basicmath11/ basicmath13/ basicmath15/ basicmath2/  basicmath4/  basicmath6/  basicmath8/  org/
basicmath10/ basicmath12/ basicmath14/ basicmath16/ basicmath3/  basicmath5/  basicmath7/  basicmath9/
root@479fceca14b2:~/ccf# cd work/benchmarks/MiBench/basicmath-CCF/basicmath13
root@479fceca14b2:~/ccf/work/benchmarks/MiBench/basicmath-CCF/basicmath13# ls
CGRAExec  LICENSE       Makefile    basicmath_large.c  basicmath_small.c  cubic.c    isqrt.c  newIR.ll          pi.h       round.h         runme_small.sh  sniptype.h  temporary.ll
COMPILE   LoopInfo.txt  TCInfo.txt  basicmath_small    combinedIR.ll      debugfile  m5out    output_small.txt  rad2deg.c  runme_large.sh  snipmath.h      temp.ll     temporaryIR.ll
root@479fceca14b2:~/ccf/work/benchmarks/MiBench/basicmath-CCF/basicmath13# make clean
rm -rf basicmath_small basicmath_large output* *.ll CGRAExec m5out *.s
root@479fceca14b2:~/ccf/work/benchmarks/MiBench/basicmath-CCF/basicmath13# make
cgracc  -static -O3 basicmath_small.c   rad2deg.c  cubic.c   isqrt.c -o basicmath_small -lm
Loop has        Depth=1 Number of BBs=1 With one exiting branch
Loop has        Depth=2 Number of BBs=1 With one exiting branch
Total Loops Compiled for CGRA:  2
root@479fceca14b2:~/ccf/work/benchmarks/MiBench/basicmath-CCF/basicmath13#./runme_small.sh