On Thu, 19 Oct 2023 12:14:52 GMT, Ilya Gavrilin <igavri...@openjdk.org> wrote:
>> Hi all, please review this changes into risc-v floating point copysign and
>> signum intrinsics.
>> CopySign - returns first argument with the sign of second. On risc-v we have
>> `fsgnj.x` instruction, which can implement this intrinsic.
>> Signum - returns input value if it is +/- 0.0 or NaN, otherwise 1.0 with the
>> sign of input value returned. On risc-v we can use `fclass.x` to specify
>> type of input value and return appropriate value.
>>
>> Tests:
>> Performance tests on t-head board:
>> With intrinsics:
>>
>> Benchmark (seed) Mode Cnt Score Error Units
>> MathBench.copySignDouble 0 thrpt 8 34156.580 ± 76.272 ops/ms
>> MathBench.copySignFloat 0 thrpt 8 34181.731 ± 38.182 ops/ms
>> MathBench.signumDouble 0 thrpt 8 31977.258 ± 1122.327 ops/ms
>> MathBench.signumFloat 0 thrpt 8 31836.852 ± 56.013 ops/ms
>>
>> Intrinsics turned off (`-XX:+UnlockDiagnosticVMOptions
>> -XX:-UseCopySignIntrinsic -XX:-UseSignumIntrinsic`):
>>
>> Benchmark (seed) Mode Cnt Score Error Units
>> MathBench.copySignDouble 0 thrpt 8 31000.996 ± 943.094 ops/ms
>> MathBench.copySignFloat 0 thrpt 8 30678.016 ± 28.087 ops/ms
>> MathBench.signumDouble 0 thrpt 8 25435.010 ± 2047.085 ops/ms
>> MathBench.signumFloat 0 thrpt 8 25257.058 ± 79.175 ops/ms
>>
>> Regression tests: tier1, hotspot:tier2 on risc-v board.
>>
>> Also, changed name of one micro test: before we had: `sigNumDouble` and
>> `signumFloat` tests, they does not matches to `signum` or `sigNum`. Now we
>> have similar part: `signum`.
>> Performance tests has been changed a bit, to check intrinsics result better,
>> diff to modify tests:
>>
>> diff --git a/test/micro/org/openjdk/bench/java/lang/MathBench.java
>> b/test/micro/org/openjdk/bench/java/lang/MathBench.java
>> index 6cd1353907e..0bee25366bf 100644
>> --- a/test/micro/org/openjdk/bench/java/lang/MathBench.java
>> +++ b/test/micro/org/openjdk/bench/java/lang/MathBench.java
>> @@ -143,12 +143,12 @@ public double ceilDouble() {
>>
>> @Benchmark
>> public double copySignDouble() {
>> - return Math.copySign(double81, doubleNegative12);
>> + return Math.copySign(double81, doubleNegative12) +
>> Math.copySign(double81, double2) + Math.copySign(double4Dot1,
>> doubleNegative12);
>> }
>>
>> @Benchmark
>> public float copySignFloat() {
>> - return Math.copySign(floatNegative99, float1);
>> + return ...
>
> Ilya Gavrilin has updated the pull request incrementally with one additional
> commit since the last revision:
>
> Changed branch inside signum implementation
Change branch inside `signum`, now `dst` always moves to `src`, so we can
remove `j(done); `
Performance results:
Without intrinsic:
Benchmark (seed) Mode Cnt Score Error Units
MathBench.signumDouble 0 thrpt 8 35666.674 ± 6.317 ops/ms
MathBench.signumFloat 0 thrpt 8 34040.220 ± 13.783 ops/ms
With old version:
Benchmark (seed) Mode Cnt Score Error Units
MathBench.signumDouble 0 thrpt 8 41601.513 ± 16.570 ops/ms
MathBench.signumFloat 0 thrpt 8 39414.511 ± 28.290 ops/ms
With new version:
Benchmark (seed) Mode Cnt Score Error Units
MathBench.signumDouble 0 thrpt 8 44060.456 ± 12.483 ops/ms
MathBench.signumFloat 0 thrpt 8 40481.776 ± 28.512 ops/ms
-------------
PR Comment: https://git.openjdk.org/jdk/pull/16186#issuecomment-1770835814
