RFR: 8352675: Support Intel AVX10 converged vector ISA feature detection [v10]
Vladimir Ivanov
vlivanov at openjdk.org
Sat May 3 07:44:48 UTC 2025
On Sat, 3 May 2025 07:26:29 GMT, Jatin Bhateja <jbhateja at openjdk.org> wrote:
>> - Intel AVX10[1] extends and enhances the capabilities of Intel AVX-512 to benefit all Intel® products and will be the vector ISA of choice moving into the future.
>> - It supports a new ISA versioning scheme which simplifies the existing AVX512 feature enumeration scheme. Feature set supported by an AVX10 ISA version will be supported by all the versions above it.
>> - The initial, fully-featured version of Intel® AVX10 will be enumerated as Version 2 (denoted as Intel® AVX10.2). This will include the new ISA extension over the existing AVX512 instructions.
>> - An early version of Intel® AVX10 (Version 1, or Intel® AVX10.1) that only enumerates the Intel® AVX-512 instruction set at 128, 256, and 512 bits will be enabled on the Granite Rapids Server for software pre-enabling.
>>
>> This patch adds the necessary CPUID feature detection for AVX10 ISA version 1 and 2. In terms of architectural state save restoration, AVX10 is isomorphic to AVX512 support up till Granite Rapids. State components affected by AVX10 extension include SSE, AVX, Opmask, ZMM_Hi256, and Hi16_ZMM registers.
>>
>> The patch has been regressed through tier1 and jvmci tests
>>
>> Please review and share your feedback.
>>
>> Best Regards,
>> Jatin
>>
>> [1] https://www.intel.com/content/www/us/en/content-details/844829/intel-advanced-vector-extensions-10-2-intel-avx10-2-architecture-specification.html
>
> Jatin Bhateja has updated the pull request incrementally with one additional commit since the last revision:
>
> Review comments resolution
Ok, thanks! I wasn't sure you finished the pass.
I'm still seeing dynamic memory allocation which IMO unnecessarily complicates the implementation. Bitmap size is fixed and well-known at compile time. It enables `VM_Feature` class to embed the array of proper size inline. And it eliminates all the problems related to undesired sharing of backed array. (Also, `pre_initialize()` is not needed as well.)
-------------
PR Comment: https://git.openjdk.org/jdk/pull/24329#issuecomment-2848488960
More information about the graal-dev
mailing list