[vector] RFR (L): Generalized intrinsics for vector operations (first batch)

[Vladimir Ivanov]

Thanks, Razvan!

Best regards,
Vladimir Ivanov

On 2/13/18 12:58 AM, Lupusoru, Razvan A wrote:
> Hi Vladimir,
> 
> The patch looks great and you have my +1 as well.
> 
> --Razvan
> 
> -----Original Message-----
> From: panama-dev [mailto:panama-dev-bounces at openjdk.java.net] On Behalf Of Vladimir Ivanov
> Sent: Friday, February 09, 2018 2:55 PM
> To: 'panama-dev at openjdk.java.net' <panama-dev at openjdk.java.net>
> Subject: Re: [vector] RFR (L): Generalized intrinsics for vector operations (first batch)
> 
> One thing I forgot to mention: original memory vector operations
> (fromArray/intoArray) miss bounds checks.
> 
> The patch put them in place, but current shape of the checks has significant performance overhead (up to 2x on simple benchmarks).
> 
> I proposed some ideas how to improve them [1] and believe it's possible to optimize them as ordinary array OOB checks, but it's a task for future exploration.
> 
> Meanwhile, there's a temporary flag to turn OOB checks off (by default, the checks are turned on):
> 
>     -Djdk.incubator.vector.VECTOR_ACCESS_OOB_CHECK=false
> 
> Best regards,
> Vladimir Ivanov
> 
> [1]
> http://mail.openjdk.java.net/pipermail/panama-dev/2017-December/000889.html
> 
> On 2/10/18 12:38 AM, Vladimir Ivanov wrote:
>> http://cr.openjdk.java.net/~vlivanov/panama/vector.generalized_intrins
>> ics/webrev.06/
>>
>>
>> Here's the first batch of rewritten intrinsics for vector operations.
>>
>> Main motivation for new implementation is:
>>     * reduce number of intrinsic needed;
>>     * improve intrinsification robustness;
>>     * minimize changes needed in shared C2 code.
>>
>> The idea of generalized intrinsics is to parameterize them with
>> additional information (passed as constant arguments), so JIT-compiler
>> has enough information to dispatch to proper implementation during
>> intrinsification.
>>
>> For example, binary vector operation can be represented as:
>>
>>     // (V,V) -> V
>>     @HotSpotIntrinsicCandidate
>>     static <V> V binaryOp(int oprId, Class<V> vectorClass,
>>                           Class<?> elementType, int vlen,
>>                           V v1, V v2,
>>                           BiFunction<V,V,V> defaultImpl) {
>>         return defaultImpl.apply(v1, v2);
>>     }
>>
>> and used as:
>>
>>     // (Int256Vector,Int256Vector) -> Int256Vector
>>
>>     public Int256Vector add(Vector<Integer,Shapes.S256Bit> v) {
>>       return (Int256Vector) VectorIntrinsics.binaryOp(
>>               VECTOR_OP_ADD,
>>               Int256Vector.class, int.class, 8,
>>               this, (Int256Vector)v,
>>               (v1, v2) -> ((Int256Vector)v1).bOp(v2,
>>                                      (i, a, b) -> (int)(a+b)));
>>     }
>>
>> where:
>>     oprId encodes actual operation (VECTOR_OP_ADD);
>>
>>     vectorClass, elementType, vlen describe concrete vector class
>> (Int256Vector);
>>
>>     v1, v2 are actual arguments of vector binary operations;
>>
>>     defaultImpl - scalar implementation which is used when
>> intrinsification fails.
>>
>>
>> Generalized intrinsics are declared on
>> jdk.incubator.vector.VectorIntrinsics and the patch contains 6 of them:
>>     * broadcastCoerced
>>     * reductionCoerced
>>     * binaryOp
>>     * load/store
>>     * test
>>
>> The following vector operations were ported to new mechanism:
>>     * broadcast: zero, broadcast, trueMask, falseMask
>>     * reduction: addAll, mulAll
>>     * binaryOp: add, sub, mul, div, and, or, xor
>>     * load/store: intoArray, fromArray
>>     * test: anyTrue, allTrue
>>
>> There's new flag added to turn new intrinsics on/off:
>>
>> +  product(bool, UseVectorApiGeneralizedIntrinsics, true,
>>
>> Previous discussions [1] [2].
>>
>> The patch adds alternative implementations, but doesn't remove
>> existing ones, since there are some code dependencies on them. Once
>> the dependencies are broken, the code will go away.
>>
>> Thanks!
>>
>> Best regards,
>> Vladimir Ivanov
>>
>> [1]
>> http://mail.openjdk.java.net/pipermail/panama-dev/2017-November/000748
>> .html
>>
>> [2]
>> http://mail.openjdk.java.net/pipermail/panama-dev/2017-December/000884
>> .html