Optimize Neon™ Logistic Activation

- Use a 1d execution window to improve memory access pattern.

Resolves: [COMPMID-5465]
Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: Ida30669ffa06eb002ca43a6edf15e25a6eaad2f6
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8344
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 61 insertions, 6 deletions

diff --git a/src/core/helpers/WindowHelpers.cpp b/src/core/helpers/WindowHelpers.cpp
index fa152c9f58..a4d46db352 100644
--- a/src/core/helpers/WindowHelpers.cpp
+++ b/src/core/helpers/WindowHelpers.cpp
@@ -234,15 +234,15 @@ Window calculate_max_window_horizontal(const ValidRegion &valid_region, const St
 
 std::pair<Window, size_t> calculate_squashed_or_max_window(const ITensorInfo &src0, const ITensorInfo &src1)
 {
-    const auto &shape0 = src0.tensor_shape();
-    const auto &shape1 = src1.tensor_shape();
-    const auto &strides0 = src0.strides_in_bytes();
-    const auto &strides1 = src1.strides_in_bytes();
-    const auto num_dimensions = std::max(src0.num_dimensions(), src1.num_dimensions());
+    const auto &shape0         = src0.tensor_shape();
+    const auto &shape1         = src1.tensor_shape();
+    const auto &strides0       = src0.strides_in_bytes();
+    const auto &strides1       = src1.strides_in_bytes();
+    const auto  num_dimensions = std::max(src0.num_dimensions(), src1.num_dimensions());
 
     Window win;
     size_t split_dimension = Window::DimY;
-    size_t dim = 0;
+    size_t dim             = 0;
 
     size_t squashed_bytes = src0.element_size();
 
@@ -282,4 +282,47 @@ std::pair<Window, size_t> calculate_squashed_or_max_window(const ITensorInfo &sr
 
     return std::make_pair(win, split_dimension);
 }
+
+std::pair<Window, size_t> calculate_squashed_or_max_window(const ITensorInfo &src)
+{
+    const auto &shape          = src.tensor_shape();
+    const auto &strides        = src.strides_in_bytes();
+    const auto  num_dimensions = src.num_dimensions();
+
+    Window win;
+    size_t split_dimension = Window::DimY;
+    size_t dim             = 0;
+    size_t squashed_bytes  = src.element_size();
+
+    // Try to squash the low dimensions together.
+    for(; dim < num_dimensions; ++dim)
+    {
+        if(strides[dim] != squashed_bytes)
+        {
+            break;
+        }
+        squashed_bytes *= shape[dim];
+    }
+    if(dim == num_dimensions)
+    {
+        const auto squashed_elements = squashed_bytes / src.element_size();
+        split_dimension              = Window::DimX;
+        // The input tensor can be interpreted as 1D array.
+        win.set(0, Window::Dimension(0, squashed_elements, 1));
+        for(dim = 1; dim < Coordinates::num_max_dimensions; ++dim)
+        {
+            win.set(dim, Window::Dimension(0, 1, 1));
+        }
+    }
+    else
+    {
+        // Generate the max window.
+        for(dim = 0; dim < Coordinates::num_max_dimensions; ++dim)
+        {
+            win.set(dim, Window::Dimension(0, shape[dim], 1));
+        }
+    }
+    return std::make_pair(win, split_dimension);
+}
+
 } // namespace arm_compute
diff --git a/src/core/helpers/WindowHelpers.h b/src/core/helpers/WindowHelpers.h
index c9e5a135c0..eccf7f2d18 100644
--- a/src/core/helpers/WindowHelpers.h
+++ b/src/core/helpers/WindowHelpers.h
@@ -176,6 +176,18 @@ inline Window calculate_max_enlarged_window(const ITensorInfo &info, const Steps
     return calculate_max_enlarged_window(info.valid_region(), steps, border_size);
 }
 
+/** Calculate the squashed or maximum window for the given tensor shape.
+ *
+ * If the tensor data resides continuously in the memory, the tensor can be interpreted
+ * as 1D array and all the dimensions can be squashed together into the x-dimension.
+ * Otherwise, generate the max window for the given tensor shape.
+ *
+ * @param[in] src Tensor info object defining the shape of the input tensor.
+ *
+ * @return The maximum window the kernel can be executed on and the preferred split dimension.
+ */
+std::pair<Window, size_t> calculate_squashed_or_max_window(const ITensorInfo &src);
+
 /** Calculate the squashed or maximum window for the given tensor shapes.
  *
  * If the tensor data resides continuously in the memory, the tensor can be interpreted