kaushikcfd/feinsum

Issues

• feinsum.query should accept a cl.Device instead of cl.Context

  #38 opened a year ago
  0

• Validate the prescription of the transform space by fuzz testing

  #32 opened 2 years ago
  0

• Pretty print a Batched einsum

  #31 opened 2 years ago
  0

• [docs] clarify that we use the same namespace for argument names and index names

  #30 opened 2 years ago
  0

• Rename FusedEinsum -> BatchedEinsum

  #29 opened 2 years ago
  0

• Support matching `max` and `min` reductions

  #27 opened 2 years ago
  0

• Use vertex coloring while canonicalizing the FusedEinsum

  #26 opened a year ago
  1

• Add parallel tuning capability

  #22 opened 2 years ago
  5

• Remove dependency on OpenTuner.

  #21 opened 2 years ago
  4

• Add a CI to do some trivial checks on the database entries

  #20 opened 2 years ago
  0

• Add helpers to compare the best result in the database with other tools

  #19 opened 2 years ago
  1

• Add empirical maximum FLOP rate measurement

  #18 opened 2 years ago
  2

• Add empirical global memory bandwidth measurement

  #17 opened 2 years ago
  0

• Add .rtd-env-py3.yml

  #15 opened 2 years ago
  0

• Fusion algorithm for composition of batched einsums

  #12 opened 2 years ago
  0

• Avoid `lp.fix_parameters` while timing/empirically verifying the transformation

  #11 opened 2 years ago
  0

• Add a utility to check whether two Fused-einsums are performance-isomorphic.

  #10 opened a year ago
  1

• `f.loopy_utils` should provide a transformation to make i-th einsum operand as substitution

  #9 opened 2 years ago
  0

• Database should not store the kernel string rather a filepath that can be located

  #8 opened 2 years ago
  1

• Check statement dependencies in `f.match_t_unit_to_einsum`

  #6 opened 2 years ago
  0

• Check dtypes in `fnsm.match_t_unit_to_einsum`

  #5 opened a year ago
  1

• Verify the output before reporting the performance

  #4 opened 2 years ago
  0

• Attach metadata to inames, kernel arguments

  #3 opened a year ago
  1

• Choose long_dim_length based after computing the minimum invariance dim size

  #2 opened 2 years ago
  0

• `feinsum.pprint_comparison_vs_roofline` should use number of flops from `opt_einsum`.

  #1 opened 2 years ago
  0