Skip to content

Latest commit

 

History

History
107 lines (70 loc) · 4.76 KB

README.md

File metadata and controls

107 lines (70 loc) · 4.76 KB
Raw

DAEDAL

DAEDAL provides LLVM tools for Decoupled Access Execute. It is currently being developed at Uppsala University.

Getting Started

  • Clone this repository to your local machine
  • Compile DAEDAL by running make

How to Generate DAEDAL Code

There are several pre-defined levels of indirection and granularity, as well as compilation target.

  • Indirection indicates the maximum numver of indirections that should be considered for data prefetch. All loads with an indirection lower or equal to this number will be turned into prefetches.
  • Granularity indicates the number of iterations that should be prefetched and consumed at a time
  • Target can be choosen from: DAE, CAE and ORIGINAL, where DAE is Decoupled Access-Execute applied with both indirections and granularities, CAE is coupled Access-Execute applied with only granularities and ORIGINAL is default compiled without any transformations.

Changes of settings can be done under:

$(path to daedal)/sources/common/DAE/Makefile.targets

Variables INDIR_COUNT and GRAN_COUNT can be changed in #line.8 and #line.9 respectively.

Targets can be modified at #line.24.

Small Benchmark Example

To illustrate the usage of DAEDAL, we provide one example benchmark which can be found at:

$(path to daedal)/sources/myBenchmark
  • In order to build binaries, change $(path to daedal)/sources/common/Makefile.environment to contain the correct path to DAEDAL:
COMPILER_LIB=$(path to daedal)/daedal/compiler/build/projects-build/lib
LLVM_BIN=$(path to daedal)/daedal/compiler/build/llvm-build/bin/
  • Run make in $(path to daedal)/sources (parallel make is allowed):
make

File Generation

The flow of generated output files from TARGET=DAE goes as follows (which can also be observed from Makefile.defaults):

  1. all .c/.cpp/.cc files get compiled to generate their respective .ll files

  2. all .ll files are used to generate .stats.ll files which are used for purpose of observations on executable runtime statistics (access/execute times)

  3. all .stats.ll files are used to mark hot loops for transformation and output .marked.ll files

  4. all .marked.ll files are used to chunck marked loops and output .gran.ll files

  5. two global files DAE-header.ll and Globals.ll are genarated (once only). They contain the value that should be used for granulairty. For each granularity DAEDAL will create a copy and replace the granularity with the current one. This copy will then be linked into the final benchmark.

  6. based on INDIR_COUNT setting, for each indirection Y,

6.1) based on GRAN_COUNT settings, for each granularity X .gran.ll files are used for loop extraction and output .granX**.extract.ll** files

6.1.a) remove redundant prefetches from .granX**.extract.ll** file and output .granX**.indirY.dae.ll** file

6.1.b) Apply pretches on __kernel__ marked functions from .granX**.indirY.dae.ll** file and output .granX**.indirY.dae.O3.ll** file

6.1.c) Global information from DAE-header.ll and Globals.ll are attached together and a new file .granX**.indirY.dae.GV_DAE.ll** is created

6.1.d) with files from 6.1.b) and 6.1.c), an executable file .granX**.indirY.dae** then gets generated

The flow of generated output files for TARGET=CAE is analogous but simpler by skipping steps 6.1.a) and 6.1.b) from above, in step 6.1.d) files are from 6.1) and 6.1.c).

All compilation process is recorded and can be found in:

$(path to daedal)/sources/myBenchmark/bin/log.txt

Adapt DAEDAL to run your own benchmarks

Feel free to try the example benchmark and change/extend to your own applications.

  • Copy the myBenchmark directory and replace the sources with your application's source.
  • Change the Makefile in your new benchmark directory to use the correct sources, compiler flags, and name:
$(path to daedal)/sources/myBenchmark/src/Makefile
  • Change the source Makefile in order to build your benchmark. Add the name of the directory of your benchmark to the list in here:
$(path to daedal)/sources/Makefile
  • Mark the loop to transform (using a vectorize pragma). See an example here:
$(path to daedal)/sources/myBenchmark/src/small_benchmark.cpp
  • You should be ready to compile your own benchmark now!

Others

More detailed information on DAEDAL can be found in Multiversioned decoupled access-execute: the key to energy-efficient compilation of general-purpose programs. Compiler Construction, March 17-18, 2016: 121-131.

For more information please visit our website.

Please contact us at contact@etascale.com