Optimize CL reduction operation

* Batch dimension is added to reduction operation.
  - All the dimensions higher than the batch dimension are collapsed
    so that the input and output tensors are always 3-4D.
  - CL kernel is called once instead of being repeatedly called
    to process each sliding window.

Resolves: COMPMID-6443
Signed-off-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Change-Id: Icd99939d52d3bb648f08537e5f52ef27e894061b
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/10456
Reviewed-by: Jakub Sujak <jakub.sujak@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

2 files changed, 25 insertions, 9 deletions

diff --git a/arm_compute/core/Window.h b/arm_compute/core/Window.h
index 4863b95045..e93d2863c9 100644
--- a/arm_compute/core/Window.h
+++ b/arm_compute/core/Window.h
@@ -21,8 +21,8 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
-#ifndef ARM_COMPUTE_WINDOW_H
-#define ARM_COMPUTE_WINDOW_H
+#ifndef ACL_ARM_COMPUTE_CORE_WINDOW_H
+#define ACL_ARM_COMPUTE_CORE_WINDOW_H
 
 #include "arm_compute/core/Coordinates.h"
 #include "arm_compute/core/Error.h"
@@ -213,15 +213,17 @@ public:
      */
     void shift(size_t dimension, int shift_value);
 
-    /** Shift down all the dimensions of a window
+    /** Shift down all the dimensions of a window starting from the specified dimension.
      *
-     * i.e new_dims[n] = old_dims[n+shift_value].
+     * new_dims[i] = old_dims[i]             for all i < start_dim.
+     * new_dims[i] = old_dims[i+shift_value] for all i >= start_dim.
      *
      * @param[in] shift_value Number of dimensions to shift the window by.
+     * @param[in] start_dim   The dimension from which the dimensions start to shift.
      *
      * @return The window with the shifted dimensions.
      */
-    Window shift_dimensions(unsigned int shift_value) const;
+    Window shift_dimensions(unsigned int shift_value, unsigned int start_dim = 0) const;
 
     /** Adjust the start or end of a given dimension by the given value
      *
@@ -460,4 +462,4 @@ private:
 };
 } // namespace arm_compute
 #include "Window.inl"
-#endif /*ARM_COMPUTE_WINDOW_H */
+#endif // ACL_ARM_COMPUTE_CORE_WINDOW_H
diff --git a/arm_compute/core/Window.inl b/arm_compute/core/Window.inl
index d935507b1d..0f7c4fbdd7 100644
--- a/arm_compute/core/Window.inl
+++ b/arm_compute/core/Window.inl
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016-2020, 2022 Arm Limited.
+ * Copyright (c) 2016-2020, 2022-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -21,6 +21,10 @@
  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  * SOFTWARE.
  */
+
+#ifndef ACL_ARM_COMPUTE_CORE_WINDOW_INL
+#define ACL_ARM_COMPUTE_CORE_WINDOW_INL
+
 namespace arm_compute
 {
 inline Window::Window(const Window &src)
@@ -100,13 +104,21 @@ inline Window Window::collapse_if_possible(const Window &full_window,
     return collapsed;
 }
 
-inline Window Window::shift_dimensions(unsigned int shift_value) const
+inline Window Window::shift_dimensions(unsigned int shift_value, unsigned int start_dim) const
 {
     Window shifted_window;
-    for (size_t n = 0; n < (Coordinates::num_max_dimensions - shift_value); n++)
+    size_t n = 0;
+
+    for (; n < start_dim; ++n)
+    {
+        shifted_window.set(n, _dims[n]);
+    }
+
+    for (; n < (Coordinates::num_max_dimensions - shift_value); n++)
     {
         shifted_window.set(n, _dims[n + shift_value]);
     }
+
     return shifted_window;
 }
 
@@ -313,3 +325,5 @@ inline bool operator==(const Window &lhs, const Window &rhs)
     return (lhs._dims == rhs._dims) && (lhs._is_broadcasted == rhs._is_broadcasted);
 }
 } // namespace arm_compute
+
+#endif // ACL_ARM_COMPUTE_CORE_WINDOW_INL