RFR: 8318158: RISC-V: implement roundD/roundF intrinsics [v2]

[Olga Mikhaltsova]

On Mon, 13 Nov 2023 09:33:02 GMT, Andrew Haley <aph at openjdk.org> wrote:

>> I have consulted with our h/w team and they told me next:
>> multi-issue FP Unit can process few (data-independent) fp instructions at a time, even if they have different rounding mode.
>> The only issue is when the rounding mode is set to dynamic rounding (aka get it from csr), but it's not our case here
>
>> I have consulted with our h/w team and they told me next: multi-issue FP Unit can process few (data-independent) fp instructions at a time, even if they have different rounding mode. The only issue is when the rounding mode is set to dynamic rounding (aka get it from csr), but it's not our case here
> 
> OK. Do what you will, but you're coding for an architecture not an implementation, and this code may stand for many years.

First I've tried to implement rounding on riscv similar to aarch64, separated negative and positive numbers but branching is too expensive and I got not much performance improvement against the current java implementation.

I like Vladimir's idea, the algorithm gives significant performance improvement and why not to take this advantage. But the above mentioned fix, related to fadd_s(), should be made. Not all numbers are processed correctly without it. Having changed rounding mode of fadd_s() we've got two sequential instructions with the same rounding mode  - rdn. 

The performance improvement with the last fix on the T-Head RVB-ICE board is as follow:

Benchmark                              (TESTSIZE)   Mode  Cnt    Score   Error   Units
FpRoundingBenchmark.test_round_double        2048  thrpt   15  111.278  0.349  ops/ms
FpRoundingBenchmark.test_round_float         2048  thrpt   15  115.776  0.323  ops/ms

@theRealAph As you suggested I've also tested this algorithm on the full 32-bit range [0; 0xFFFFFFFF] using Float.intBitsToFloat(x) and all the numbers were processed correctly. The output of this algorithm is equal to the current java Math.round() implementation output.

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/16382#discussion_r1391794054