Flop counting#623

Open

inducer wants to merge 31 commits intomainfrom

flop-counting

Commits on Apr 24, 2026

add MapAsIndexLambdaMixin
majosm
committed
reimplement AxesTagsEquationCollector using MapAsIndexLambdaMixin
majosm
committed
add flop counting functions
majosm
committed
Update baseline

inducer
authored and
majosm
committed
add note to docs about assumptions when handling conditional expressions
majosm
committed
change 'pass' -> '...' in ReductionOperation abstract methods
majosm
committed
move op_name_to_num_flops type declaration to FlopCounter class body
majosm
committed
add some details about how flop counts are computed
majosm
committed
clarify how flop counting functions behave w.r.t. DAG functions
majosm
committed
don't try to count flops for function calls
majosm
committed
move own flop counting out of rec and into its own method in _PerEntryFlopCounter
majosm
committed
change is_materializable -> is_materialized / has_taggable_materialization
majosm
committed
use explicit isinstance() check instead of try/except around to_index_lambda()
majosm
committed
remove a couple of FIXMEs
majosm
committed
add placeholder class for operator flop counts that aren't specified
majosm
committed
fix swapped and/or op names

majosm
and
Copilot
committed
fix flop counting for negative power

majosm
and
Copilot
committed
make UndefinedOpFlopCountError not a dataclass to avoid issues with __init__
majosm
committed
update note about function traversal for MaterializedNodeFlopCounter
majosm
committed
add InputUseCounter in scalar_expr
majosm
committed
abandon is_materialized/has_taggable_materialization approach in favor of collecting materialized arrays ahead of time
majosm
committed
fix flop counting for reductions
majosm
committed
handle CSR matmul in flop counting
majosm
committed
handle unused bindings

majosm
and
Copilot
committed
enumerate all undefined op flop counts in exception message

majosm
and
Copilot
committed
forbid loopy calls in flop counting
majosm
committed
add type annotation for scalar_op_name overrides
majosm
committed
make power flop count more precise
majosm
committed
fix remainder flop counting
majosm
committed
test recursion in test_scalar_flop_count
majosm
committed
add note about flop counting for subscript indices
majosm
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Flop counting#623

Flop counting#623
inducer wants to merge 31 commits intomainfrom
flop-counting

Commits on Apr 24, 2026

add MapAsIndexLambdaMixin

reimplement AxesTagsEquationCollector using MapAsIndexLambdaMixin

add flop counting functions

Update baseline

add note to docs about assumptions when handling conditional expressions

change 'pass' -> '...' in ReductionOperation abstract methods

move op_name_to_num_flops type declaration to FlopCounter class body

add some details about how flop counts are computed

clarify how flop counting functions behave w.r.t. DAG functions

don't try to count flops for function calls

move own flop counting out of rec and into its own method in _PerEntryFlopCounter

change is_materializable -> is_materialized / has_taggable_materialization

use explicit isinstance() check instead of try/except around to_index_lambda()

remove a couple of FIXMEs

add placeholder class for operator flop counts that aren't specified

fix swapped and/or op names

fix flop counting for negative power

make UndefinedOpFlopCountError not a dataclass to avoid issues with init

update note about function traversal for MaterializedNodeFlopCounter

add InputUseCounter in scalar_expr

abandon is_materialized/has_taggable_materialization approach in favor of collecting materialized arrays ahead of time

fix flop counting for reductions

handle CSR matmul in flop counting

handle unused bindings

enumerate all undefined op flop counts in exception message

forbid loopy calls in flop counting

add type annotation for scalar_op_name overrides

make power flop count more precise

fix remainder flop counting

test recursion in test_scalar_flop_count

add note about flop counting for subscript indices