1 files changed, 46 insertions, 1 deletions
diff --git a/src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp b/src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp
index 2d61b72078..505a37174e 100644
--- a/src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp
+++ b/src/cpu/kernels/gemm_matrix_add/generic/neon/fp16.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2022 Arm Limited.
+ * Copyright (c) 2022-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -25,10 +25,55 @@
 
 #include "src/cpu/kernels/gemm_matrix_add/generic/neon/impl.h"
 
+#include <arm_neon.h>
+
 namespace arm_compute
 {
 namespace cpu
 {
+namespace
+{
+void matrix_addition_f16(const ITensor *src, ITensor *dst, const Window &window, float beta)
+{
+    ARM_COMPUTE_ERROR_ON_NULLPTR(src, dst);
+    const float16x8_t beta_f16 = vdupq_n_f16(beta);
+
+    constexpr int window_step_x  = 16;
+    const auto    window_start_x = static_cast<int>(window.x().start());
+    const auto    window_end_x   = static_cast<int>(window.x().end());
+
+    Window win = window.collapse_if_possible(window, Window::DimZ);
+    win.set(Window::DimX, Window::Dimension(0, 1, 1));
+
+    Iterator in(src, win);
+    Iterator out(dst, win);
+
+    execute_window_loop(win, [&](const Coordinates &)
+    {
+        const auto in_ptr  = reinterpret_cast<const float16_t *>(in.ptr());
+        const auto out_ptr = reinterpret_cast<float16_t *>(out.ptr());
+
+        int x = window_start_x;
+        for(; x < (window_end_x - window_step_x); x += window_step_x)
+        {
+            float16x8x2_t       alpha_ab = vld2q_f16(out_ptr + x);
+            const float16x8x2_t c        = vld2q_f16(in_ptr + x);
+            // Multiply matrix C by its weight and accumulate
+            alpha_ab.val[0] = vaddq_f16(alpha_ab.val[0], vmulq_f16(c.val[0], beta_f16));
+            alpha_ab.val[1] = vaddq_f16(alpha_ab.val[1], vmulq_f16(c.val[1], beta_f16));
+
+            vst2q_f16(out_ptr + x, alpha_ab);
+        }
+
+        // Left-over loop
+        for(; x < window_end_x; ++x)
+        {
+            *(out_ptr + x) += *(in_ptr + x) * static_cast<float16_t>(beta);
+        }
+    },
+    in, out);
+}
+} // namespace
 void neon_fp16_gemm_matrix_add(const ITensor *src, ITensor *dst, const Window &window, float beta)
 {
     return matrix_addition_f16(src, dst, window, beta);