The Sparse Abstract Machine (SAM) IR, Compiler, and Simulator

SAM Front-end Compiler

Overview: tensor expression + format language + schedule --> SAM Graph --> dot file and png of dot file -->
RTL Graph or Simulator Graph

Compiling SAM graphs

Init the taco/ repo as a submodule

make submodules

Setup the compilation for the taco/ repo

make taco/build

Run the script to generate a handful of example sam graphs

make sam

The example sam graphs should now be located in compiler/sam-outputs/ in both the dot/ and png/ folers.

Naming convention

Naming rules

• all (block) types are lower case: repeat, repeat_gen, fiber_lookup, fiber_write, reduce, intersect, union, sparse_accum
• network signal types are: crd, ref, val, repsig, and bv
• Tensor casing: Matrices and higher order tensors are upper case, scalars and vectors are lower case
• Index variables are going to be i, j, k, ..., etc.
• Tensor ranks are going to correspond to 0, 1, 2, ..., etc. (no longer using rows and columns)
• For a given expression result is always 'x' (or 'X') and the inputs start from 'b, c, ..., etc.' of equivalently 'B, C, ..., etc.'

Metadata Naming Metadata naming convention for other blocks: - Metadata naming convention for fiber (lookup and write) blocks: fiber<lookup|write>-<glb?> Examples:

1. fiber_lookup_Bi_B0_compressed
2. repeat_Ci

SAM Simulator

Installing SAM Simulator as a Package

 pip install -e .

Running Tests

The simulator uses pytest to run tests

To run all tests type

cd sam/sim/
pytest

Use the following pytest optional arguments below

--debug-sim                 Turn on debug mode for sim
--count=<n>                 Repeat each test for n iterations 
-k <testname>[<paramlist>]  Run only tests with testname and paramlist
-vv                         Double verbose
-s                          Forward printouts to stdout
--full-trace                Print full trace to stdout

Test Naming Convention

Full kernel tests follow the naming convention test_<kernel>_<rand|direct>_<outformat>_<in1format>_<in2format>_...<innformat> where:

1. <kernel> is the name of the tensor algebra kernel being tested (e.g. mat_elemmul, mat_mul, vec_elemmul, etc.)
2. *format takes on u | c | s for formats uncompressed, compressed, singleton respectively
3. `<rand|direct> specifies if the test is randomly generated or a directed (handwritten) test

Primitive unit tests follow the naming convention test_<primitive>_<feature>_<order> where:

1. <primitive> is the name of the primitive being tested (e.g. array, intersect, union, etc.)
2. <feature> is the name of the feature being tested (e.g. for an array we can test both loads and stores)
3. <order> is the name of the order of stream being tested (1d for vectors, 2d for matrices, ..., and nd for all dimensions/tensor orders, etc.)

Directory Structure

sim
│   
│       
│
└───src
│   │   base.py
│   │   joiner.py
│   │   ...         # All primitive block classes
│   
└───test
│   │   test.py
│   │   file022.txt
│   │
│   └───apps
│       │   test_mat_elemmul.py
│       │   ...     # Full kernel/expression tests
│
└───────primitives
        │   test_joiner.py
        │   ...     # Primitive unit tests

SAM Binding to Onyx

See the README in sam/sam/onyx

License

All files in this project (code, scripts, documentaiton) are released under the MIT License