COMPMID-1494 Optimise NEON im2col and weights reshape for NHWC

Change-Id: I99ebae61024a7bce9d17292a02c28626ae6c29d5
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/144872
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Gian Marco Iodice <gianmarco.iodice@arm.com>

1 files changed, 5 insertions, 61 deletions

diff --git a/tests/validation/reference/Im2Col.cpp b/tests/validation/reference/Im2Col.cpp
index 0c41d88f3e..076b2aba07 100644
--- a/tests/validation/reference/Im2Col.cpp
+++ b/tests/validation/reference/Im2Col.cpp
@@ -93,52 +93,6 @@ template <typename T>
 void im2col_nhwc(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
 {
     ARM_COMPUTE_ERROR_ON(src.data_layout() != DataLayout::NHWC);
-    const int pad_x         = conv_info.pad().first;
-    const int pad_y         = conv_info.pad().second;
-    const int stride_x      = conv_info.stride().first;
-    const int stride_y      = conv_info.stride().second;
-    const int kernel_width  = kernel_dims.width;
-    const int kernel_height = kernel_dims.height;
-    const int src_width     = src.shape().y();
-    const int src_height    = src.shape().z();
-    const int src_depth     = src.shape().x();
-    const int batches       = src.shape().total_size_upper(3);
-    const int pad_val       = is_data_type_quantized_asymmetric(src.data_type()) ? src.quantization_info().offset : 0;
-    int       dst_idx       = 0;
-
-    const int lasty = src_height + (kernel_height > 1 ? pad_y : 0) - kernel_height;
-    const int lastx = src_width + (kernel_width > 1 ? pad_x : 0) - kernel_width;
-
-    for(int b = 0; b < batches; ++b)
-    {
-        for(int y = -pad_y; y <= lasty; y += stride_y)
-        {
-            for(int x = -pad_x; x <= lastx; x += stride_x)
-            {
-                for(int z = 0; z < src_depth; ++z)
-                {
-                    for(int patch_y = y; patch_y < (y + kernel_height); ++patch_y)
-                    {
-                        for(int patch_x = x; patch_x < (x + kernel_width); ++patch_x)
-                        {
-                            dst[dst_idx++] = tensor_elem_at(src, Coordinates(z, patch_x, patch_y, b), BorderMode::CONSTANT, static_cast<T>(pad_val));
-                        }
-                    }
-                }
-
-                if(has_bias)
-                {
-                    dst[dst_idx++] = static_cast<T>(1);
-                }
-            }
-        }
-    }
-}
-
-template <typename T>
-void im2col_nhwc_channel_first(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias)
-{
-    ARM_COMPUTE_ERROR_ON(src.data_layout() != DataLayout::NHWC);
     const int stride_x      = conv_info.stride().first;
     const int stride_y      = conv_info.stride().second;
     const int kernel_width  = kernel_dims.width;
@@ -185,7 +139,7 @@ void im2col_nhwc_channel_first(const SimpleTensor<T> &src, SimpleTensor<T> &dst,
 }
 
 template <typename T>
-void im2col(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const unsigned int num_groups, bool channels_first_output_nhwc)
+void im2col(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, const unsigned int num_groups)
 {
     switch(src.data_layout())
     {
@@ -196,14 +150,7 @@ void im2col(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kern
         }
         case DataLayout::NHWC:
         {
-            if(channels_first_output_nhwc)
-            {
-                im2col_nhwc_channel_first(src, dst, kernel_dims, conv_info, has_bias);
-            }
-            else
-            {
-                im2col_nhwc(src, dst, kernel_dims, conv_info, has_bias);
-            }
+            im2col_nhwc(src, dst, kernel_dims, conv_info, has_bias);
             break;
         }
         default:
@@ -214,12 +161,9 @@ void im2col(const SimpleTensor<T> &src, SimpleTensor<T> &dst, const Size2D &kern
     }
 }
 
-template void im2col(const SimpleTensor<uint8_t> &src, SimpleTensor<uint8_t> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups,
-                     bool channels_first_output_nhwc);
-template void im2col(const SimpleTensor<half> &src, SimpleTensor<half> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups,
-                     bool channels_first_output_nhwc);
-template void im2col(const SimpleTensor<float> &src, SimpleTensor<float> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups,
-                     bool channels_first_output_nhwc);
+template void im2col(const SimpleTensor<uint8_t> &src, SimpleTensor<uint8_t> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups);
+template void im2col(const SimpleTensor<half> &src, SimpleTensor<half> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups);
+template void im2col(const SimpleTensor<float> &src, SimpleTensor<float> &dst, const Size2D &kernel_dims, const PadStrideInfo &conv_info, bool has_bias, unsigned int num_groups);
 } // namespace reference
 } // namespace validation
 } // namespace test