mbutrovich opened a new issue, #7697: URL: https://github.com/apache/datafusion-comet/issues/7697
### What is the problem the feature request solves? Several of Comet's own loops divide every row by a divisor that is fixed for the whole batch or the whole shuffle. Each row pays for a hardware divide, or for a `__divti3` or `i256` long-division call. apache/arrow-rs#11421 adds precomputed divisors to `arrow-buffer`, which turn each division into a high multiply and a shift. Its prototype measured `pmod`-style `u32 % 13` at 4.0x and Comet's `i256` multiply at 2.7x to 4.2x. These are the loops that divide by a fixed divisor: - [`pmod`](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/shuffle/src/comet_partitioning.rs#L123-L129) divides each row's `i32` hash by the partition count to pick its shuffle partition. - [`div_round_half_up`](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/spark-expr/src/math_funcs/wide_decimal_binary_expr.rs#L121-L146) divides each `i256` result of `WideDecimalBinaryExpr` by `10^k` ([add and subtract](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/spark-expr/src/math_funcs/wide_decimal_binary_expr.rs#L255), [multiply](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/spark-expr/src/math_funcs/wide_decimal_binary_expr.rs#L293)). - The decimal to integer casts divide each `i128` by `10^scale` ([`cast_decimal_to_int16_down`](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/spark-expr/src/conversion_funcs/numeric.rs#L545), [`cast_decimal_to_int32_up`](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/spark-expr/src/conversion_funcs/numeric.rs#L600)). - `ceil` and `floor` on decimals divide each `i128` by `10^scale` when the scale isn't a compile-time constant ([`decimal_round_f`](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/spark-expr/src/math_funcs/ceil_floor.rs#L214-L218)). - `round` on decimals divides each `i128` by `10^k` ([`decimal_round_f`](https://github.com/apache/datafusion-comet/blob/d187e6823f9052e7331e8a810b08a1202debb91c/native/spark-expr/src/math_funcs/round.rs#L218-L233)). ### Describe the potential solution Once Comet upgrades to an arrow release that includes them, build a divisor once per batch (once per shuffle for `pmod`) and divide each row with it: - [ ] `pmod` with `DivisorI32`. - [ ] `div_round_half_up` with the `i256` divisor, and with `DivisorI128` for batches whose values fit in `i128`. - [ ] The decimal to integer casts with `DivisorI128`, and with `DivisorI64` for batches whose values fit in `i64`. - [ ] Decimal `ceil` and `floor` the same way. - [ ] Decimal `round` the same way. apache/arrow-rs#11421 lists what a kernel change must do, and the same applies here. Each change should report benchmark bot numbers with `-Ctarget-cpu=x86-64-v3` and `-Ctarget-cpu=neoverse-n1`, keep today's behavior for nulls and errors in ANSI mode, and include cases with more than half of the values null. Arrow kernels that adopt divisors under apache/arrow-rs#11421, such as integer `/` and `%` by a scalar (apache/arrow-rs#12076), need no change in Comet. ### Additional context The `i32` and `i64` divisors are in apache/arrow-rs#11467, and the `i128` and `i256` divisors are in apache/arrow-rs#12232. Comet won't port them ahead of the arrow release, so that Comet and arrow-rs share one implementation. ### Willingness to contribute I can contribute a fix for this bug independently -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
