RFR: 8261542: X86 slice and unslice intrinsics for 256-bit byte/short vectors
Vladimir Kozlov
kvn at openjdk.java.net
Thu Feb 18 19:17:44 UTC 2021
On Thu, 11 Feb 2021 02:37:35 GMT, Sandhya Viswanathan <sviswanathan at openjdk.org> wrote:
> The slice and unslice intrinsics for 256-bit byte/short vectors can be implemented for x86 platforms supporting AVX2 using a sequence of instructions.
>
> JBS: https://bugs.openjdk.java.net/browse/JDK-8261542
>
> The PerfSliceOrigin.java jmh test attached to the JBS shows the following performance on AVX2 platform.
>
> Before:
> Benchmark (size) Mode Cnt Score Error Units
> PerfSliceOrigin.vectorSliceOrigin 1024 thrpt 5 18.887 ± 1.128 ops/ms
> PerfSliceOrigin.vectorSliceUnsliceOrigin 1024 thrpt 5 9.374 ± 0.370 ops/ms
>
> After:
> Benchmark (size) Mode Cnt Score Error Units
> PerfSliceOrigin.vectorSliceOrigin 1024 thrpt 5 13861.420 ± 19.071 ops/ms
> PerfSliceOrigin.vectorSliceUnsliceOrigin 1024 thrpt 5 7895.199 ± 142.580 ops/ms
Please, add a test which verifies correctness of results when this code is used. If we don't have it already.
src/hotspot/cpu/x86/x86.ad line 7506:
> 7504: instruct rearrangeB_avx(legVec dst, legVec src, vec shuffle, legVec vtmp1, legVec vtmp2, rRegP scratch) %{
> 7505: predicate(vector_element_basic_type(n) == T_BYTE &&
> 7506: vector_length(n) == 32 && !VM_Version::supports_avx512_vbmi());
Predicate matches bail-out condition in match_rule_supported_vector(). Does it mean this code never used before?
So you are implementing it now. Right?
src/hotspot/cpu/x86/x86.ad line 7550:
> 7548: // only byte shuffle instruction available on these platforms
> 7549: int vlen_in_bytes = vector_length_in_bytes(this);
> 7550: if (UseAVX == 0) {
This code will not be executed with vector length 16 because match_rule_supported_vector() bailout with (size_in_bits == 256 && UseAVX < 2).
-------------
Changes requested by kvn (Reviewer).
PR: https://git.openjdk.java.net/jdk/pull/2520
More information about the hotspot-compiler-dev
mailing list