[FORK][FEATURE] jit_uni_reorder: relaxed isa condition to enable FP16 precision support on AVX2 platforms

The feature is required to enable fast FP16->BF16/FP32 precision conversion. Most commonly used for fast weights decompression which improves model compilation time.

Author: antonvor

src/cpu/x64/jit_uni_reorder.cpp

@@ -214,7 +214,7 @@ struct jit_uni_reorder_kernel_f32_t : public kernel_t, public jit_generator {
                 && IMPLICATION(utils::one_of(bf16, p.itype, p.otype),
                         mayiuse(avx512_core) || mayiuse(avx2_vnni_2))
                 && IMPLICATION(utils::one_of(f16, p.itype, p.otype),
-                        mayiuse(avx512_core_fp16) || mayiuse(avx2_vnni_2))
+                        mayiuse(avx512_core_fp16) || mayiuse(avx2))
                 && IMPLICATION(utils::one_of(f8_e5m2, p.itype, p.otype)
                                 || utils::one_of(f8_e4m3, p.itype, p.otype),
                         mayiuse(avx512_core_amx))
@@ -313,7 +313,7 @@ struct jit_uni_reorder_kernel_f32_t : public kernel_t, public jit_generator {
                             vcvtph2psx(dst, src);
                         else
                             vcvtph2psx(dst, Xmm(src.getIdx()));
-                    } else if (is_superset(isa_, avx2_vnni_2)) {
+                    } else if (is_superset(isa_, avx2)) {
                         if (src.isMEM())
                             vcvtph2ps(dst, src);
                         else