commit 41eb2d92c89274200d59ff97653e2bd66819b310
author Pablo Marquez Tello <pablo.tello@arm.com> Thu Jun 23 16:02:05 2022 +0100
committer Pablo Marquez Tello <pablo.tello@arm.com> Fri Jun 24 09:17:19 2022 +0000
tree 2cb2d9022e16ac8780bba3ab308f73aa15ce5811
parent 700b913ed9257c44147372378bc8a0dadcfd2ac2

Improve LUT Neon Hard-Swish

* Changed window_step from 16 to tensor_shape().x() when calling
  into the assembly byte substitution code.

* Resolve COMPMID-5211

Change-Id: I5c1f5273455999bb35f94c76a8afb4290e728858
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7843
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

src/cpu/kernels/activation/generic/neon/qasymm8.cpp

diff --git a/src/cpu/kernels/activation/generic/neon/qasymm8.cpp b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
index f35d0d2..29f5e6b 100644
--- a/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
+++ b/src/cpu/kernels/activation/generic/neon/qasymm8.cpp
@@ -421,33 +421,16 @@
 {
     ARM_COMPUTE_ERROR_ON(act_info.activation() != ActivationLayerInfo::ActivationFunction::HARD_SWISH);
 #ifdef __aarch64__
-    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());
-
+    const int window_step_x  = src->info()->tensor_shape().x();
     Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
     win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
-
     Iterator input(src, win_collapsed);
     Iterator output(dst, win_collapsed);
-
     execute_window_loop(win_collapsed, [&](const Coordinates &)
     {
-        // Compute S elements per iteration
-        int x = window_start_x;
-        for(; x <= (window_end_x - window_step_x); x += window_step_x)
-        {
-            const auto input_ptr  = reinterpret_cast<const uint8_t *>(input.ptr() + x);
-            auto       output_ptr = reinterpret_cast<uint8_t *>(output.ptr() + x);
-            substitute_bytes_neon(act_info.lut().data(), 1u, window_step_x, &input_ptr, &output_ptr);
-        }
-        // Compute left-over elements
-        for(; x < window_end_x; ++x)
-        {
-            const auto input_ptr  = reinterpret_cast<const uint8_t *>(input.ptr() + x);
-            auto       output_ptr = reinterpret_cast<uint8_t *>(output.ptr() + x);
-            substitute_bytes_neon(act_info.lut().data(), 1u, 1u, &input_ptr, &output_ptr);
-        }
+        const auto input_ptr  = reinterpret_cast<const uint8_t *>(input.ptr());
+        auto       output_ptr = reinterpret_cast<uint8_t *>(output.ptr());
+        substitute_bytes_neon(act_info.lut().data(), 1u, window_step_x, &input_ptr, &output_ptr);
     },
     input, output);
 #else  // #ifdef __aarch64__