Lego_v0

ONYX Docker Setup

Use docker_create.sh to create the docker container and within the docker container:

-- Within sam:

git checkout mapping_to_cgra
pip install -r requirements.txt
pip install -e .

-- Within Lego_v0

./requirements.sh

Your bashrc should have (you might need to mkdir these as well)

export SUITESPARSE_PATH=/aha/datasets/sparse-datasets/suitesparse/
export TACO_TENSOR_PATH=/aha/datasets/sparse-datasets/other
export FROSTT_PATH=/aha/datasets/sparse-datasets/frostt/
export SUITESPARSE_FORMATTED_PATH=/aha/datasets/sam/SUITESPARSE_FORMATTED
export FROSTT_FORMATTED_TACO_PATH=/aha/datasets/sam/FROST_FORMATTED_TACO
export FROSTT_FORMATTED_PATH=/aha/datasets/sam/FROST_FORMATTED

Input Language

-- program.txt

app_name: matmul_ijk_football                   \\ Name of the app
stmt: X(i, j) = B(i, k) * C(k, j)               \\ Tensor expression  
schedule_ap:   [ikj]                            \\ Schedule to pair the tiles
schedule_cp:   [ikj]                            \\ Schedule to pair the subtiles
schedule_cgra: [ikj]                            \\ Schedule on the CGRA 
i:split:10000:10000:30                          \\ <tensor-dim>:<tile-dim>:<subtile-dim> 
j:split:10000:10000:30
k:split:10000:10000:30

• So, if we need (30, 40) subtiles for B and (40, 30) subtiles for C, the schedule would be:

i:split:10000:10000:30                          
j:split:10000:10000:30
k:split:10000:10000:40

-- tensor.txt

tensor_name:dataset:dataset_name:gcn_flag sparse/dense:get_other_tensor_flag:random_matrix_density(%):data_type
B:ss:quilp:s:0:60:int 
C:ss:quilp:s:shift_transpose:60:int

• dataset:

  • ss: SUITESPARSE
  • frostt: FROSTT
  • gen: random tensor generation
  • sparse_ml: Sparse ML datasets
  • ex: extensor
  • New datasets could be updated in pre_process.py

• dataset_name: The name of the actual .mtx or .tns file within the dataset

• GCN Flag <TODO: Bo Wun>

• get_other_tensor_flag:

  • Process the tensor to generate a new tensor, Ex: shift_transpose, shifts the last mode and transposes the tensor. C = B.T
  • More operations can be found in pre_process,py

• random_matrix_density:

  • Percentage density of the random matrix being generated

• data_type: int - int16, f32 - float32, bf16 - bfloat16

-- bitstream.bs

-- design_meta.json

-- reg_write.h

All the above files are to be placed within: input/

Organization

• main.py generates the main.cpp code and the tiled_CSF formats for the tensors stored in lego_scratch/
• Executing main.cpp performs the subtile pairing and packing of data into the ONYX format
• Execution: python3 main.py --mode onyx // Check other arguments that might be useful
• ./lego_onyx_codegen.sh // For all apps
• ./lego_onyx_matmul_codegen.sh // For matmul data generation, with different schedules for tiling and pairing
• General output scripts are available at lego_scratch/
• Paired sub-tiles, extents and gold data are available at lego_scratch/app_name/