Aarch64 pa ACL f16 enablement

Author: ashwins990

src/common/transformations/src/transformations/common_optimizations/convert_pagedattn_inputs.cpp

@@ -72,10 +72,16 @@ ov::pass::ConvertPagedAttnInputs::ConvertPagedAttnInputs(const KVCacheConfig& co
         const auto pa_op = m.get_match_root();
         auto key_cache = ov::as_type_ptr<ov::op::v0::Parameter>(pa_op->get_input_node_shared_ptr(3));
         auto value_cache = ov::as_type_ptr<ov::op::v0::Parameter>(pa_op->get_input_node_shared_ptr(4));
+#if defined(OPENVINO_ARCH_ARM64)
+        auto format_cache_precision = [](ov::element::Type cache_precision, ov::element::Type infer_precision) {
+            return ov::element::u8;
+        };
+#else
         auto format_cache_precision = [](ov::element::Type cache_precision, ov::element::Type infer_precision) {
             return cache_precision == ov::element::f16 && infer_precision == ov::element::bf16 ? infer_precision
                                                                                                : cache_precision;
         };
+#endif
         auto init_cache_shape = [&](const size_t head_nums,
                                     const size_t head_size,
                                     const size_t block_size,

src/plugins/intel_cpu/src/nodes/kernels/aarch64/pa_kernels.hpp

@@ -0,0 +1,68 @@
+// Copyright (C) 2024 FUJITSU LIMITED
+// SPDX-License-Identifier: Apache-2.0
+//
+#include <arm_sve.h>
+
+#include "openvino/core/type/float16.hpp"
+
+#define SIZE_IN_BITS(t_var) sizeof(t_var) * 8
+#define __ce(expr, bits, ...)     \
+    if constexpr (expr == bits) { \
+        __VA_ARGS__               \
+    }
+
+#define SVE_PREDICATE(var, t_var)                                                                         \
+    svbool_t var;                                                                                         \
+                                                                                                          \
+    __ce(SIZE_IN_BITS(t_var), 8, var = svptrue_b8();) __ce(SIZE_IN_BITS(t_var), 16, var = svptrue_b16();) \
+        __ce(SIZE_IN_BITS(t_var), 32, var = svptrue_b32();) __ce(SIZE_IN_BITS(t_var), 64, var = svptrue_b64();)
+
+#define SVE_VLEN(var, t_var)                                                                     \
+    size_t var;                                                                                  \
+                                                                                                 \
+    __ce(SIZE_IN_BITS(t_var), 8, var = svcntb();) __ce(SIZE_IN_BITS(t_var), 16, var = svcnth();) \
+        __ce(SIZE_IN_BITS(t_var), 32, var = svcntw();) __ce(SIZE_IN_BITS(t_var), 64, var = svcntd();)
+
+#define SVE_PREDICATE_WHILELT(var, t_var, arg1, arg2)                       \
+    svbool_t var;                                                           \
+                                                                            \
+    __ce(SIZE_IN_BITS(t_var), 8, var = svwhilelt_b8(arg1, arg2);)           \
+        __ce(SIZE_IN_BITS(t_var), 16, var = svwhilelt_b16(arg1, arg2);)     \
+            __ce(SIZE_IN_BITS(t_var), 32, var = svwhilelt_b32(arg1, arg2);) \
+                __ce(SIZE_IN_BITS(t_var), 64, var = svwhilelt_b64(arg1, arg2);)
+
+namespace ov::Extensions::Cpu::XARCH {
+static void cvt_copy(float* dst, ov::float16* src, size_t n) {
+    auto src_ptr = reinterpret_cast<float16_t*>(src);
+    auto pg_vl2 = svwhilelt_b16(svcnth() / 2, svcnth());
+    auto vlen = svcnth() / 2;
+    auto pg_dst = svptrue_b32();
+    size_t i = 0;
+    for (; i + vlen <= n; i += vlen) {
+        auto load_src = svld1_f16(pg_vl2, src_ptr + i);
+        auto src_interleave = svzip1_f16(load_src, load_src);
+        auto cvt_dst = svcvt_f32_f16_z(pg_dst, src_interleave);
+        svst1(pg_dst, dst + i, cvt_dst);
+    }
+    for (; i < n; i++) {
+        dst[i] = src[i];
+    }
+}
+
+static void cvt_copy(ov::float16* dst, float* src, size_t n) {
+    auto dst_ptr = reinterpret_cast<float16_t*>(dst);
+    auto pg_src = svptrue_b32();
+    auto pg_dst = svwhilelt_b16(svcnth() / 2, svcnth());
+    auto vlen = svcntw();
+    size_t i = 0;
+    for (; i + vlen < n; i += vlen) {
+        auto load_src = svld1_f32(pg_src, src + i);
+        auto cvt_dst = svcvt_f16_f32_z(pg_src, load_src);
+        auto str_dst = svuzp1(cvt_dst, cvt_dst);
+        svst1_f16(pg_dst, dst_ptr + i, str_dst);
+    }
+    for (; i < n; i++) {
+        dst[i] = src[i];
+    }
+}
+}  // namespace ov::Extensions::Cpu::XARCH