Tool Chain Overview

[And/or should this be part of the top-level intro?]

The procedure for turning a Halide app into a program running on a CGRA accelerator is shown below. This example uses the "pointwise multiply-by-two' Halide example from Jeff Setter's repository. The app simply reads an input image, multiplies every pixel in the image by two, and emits that as the output image.

1. CGRA generator builds a target CGRA

  git clone https://github.com/StanfordAHA/CGRAGenerator
  CGRAGenerator/hardware/generator_z/top/build_cgra.sh

CGRA generation should happen first as it dictates the design information that will be used by all succeeding steps (except the Halide front end, which compiles to design-independent intermediate form).

2. Halide compiler reads your Halide source app code and produces design-independent CoreIR intermediate code

  git clone https://github.com/jeffsetter/Halide_CoreIR
  make -C Halide_CoreIR/apps/coreir_examples/pointwise/ pointwise_design_top.json

See directory https://github.com/jeffsetter/Halide_CoreIR/apps/coreir_examples/pointwise/ to see how this Halide app was set up.

3. CoreIR mapper reads design-independent CoreIR and emits target-specific CoreIR

  git clone https://github.com/StanfordAHA/CGRAMapper
  CGRAMapper/bin/cgra-mapper pointwise_design_top.json pointwise_mapped.json

[Note the mapper does not appear to use CGRA-dependent info from the cgra_info.xml file...but it should...right?]

4. Place-and-route (PNR) turns target-specific CoreIR into CGRA assembly language

  git clone https://github.com/Kuree/cgra_pnr
  cgra_pnr/coreir_fix/fix_all.sh pointwise_mapped.json pointwise_mapped.json
  cgra_pnr/scripts/pnr_flow.sh cgra_info.xml pointwise_mapped.json pointwise_annotated.bsb

5. Assembler ("bsbuilder") turns assembly code into a CGRA configuration bitstream

  git clone https://github.com/StanfordAHA/CGRAGenerator
  CGRAGenerator/bitstream/bsbuilder/bsbuilder.py < pointwise_annotated.bsb > pointwise.bsa

6. Simulator ("run_tbg.csh") runs the app and emits a result for comparison

  git clone https://github.com/StanfordAHA/CGRAGenerator
  cd CGRAGenerator/verilator/generator_z_tb; ./run_tbg.csh \
       -config    <fullpath>/pointwise_pnr_bitstream \
       -io_config <fullpath>/pointwise_io.json \
       -input     <fullpath>/pointwise_input.png \
       -output    <fullpath>/pointwise_CGRA_out.raw \
       -out1      <fullpath>/pointwise_CGRA_out1.raw \
       -delay     0,0 \
       -nclocks    5M`

Notes

For more information you can look at the Travis test that runs after every CGRA build. The test performs several end-to-end tests, including compilation from Halide to both x86 and CGRA targets, and comparing the x86 vs. CGRA results to verify that both got the same answer.