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>

5 files changed, 52 insertions, 127 deletions

diff --git a/tests/validation/CL/Im2Col.cpp b/tests/validation/CL/Im2Col.cpp
index cf7c79ad72..ebf2331e5e 100644
--- a/tests/validation/CL/Im2Col.cpp
+++ b/tests/validation/CL/Im2Col.cpp
@@ -94,17 +94,15 @@ template <typename T>
 using CLIm2ColFixture = Im2ColValidationFixture<CLTensor, CLAccessor, CLIm2Col, T, true>;
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)),
-                                                                                                      conv_args),
-                                                                                              framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)),
+                                                                                              conv_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)),
-                                                                                                          conv_args),
-                                                                                                  framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)),
+                                                                                                  conv_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -112,16 +110,14 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<float>, framework::DatasetMode:
 TEST_SUITE_END()
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)),
-                                                                                                     conv_args),
-                                                                                             framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)),
+                                                                                             conv_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)),
-                                                                                                         conv_args),
-                                                                                                 framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)),
+                                                                                                 conv_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -130,16 +126,14 @@ TEST_SUITE_END()
 TEST_SUITE_END()
 
 TEST_SUITE(QASYMM8)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
-                                                                                                        conv_args),
-                                                                                                framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                conv_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
-                                                                                                            conv_args),
-                                                                                                    framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                    conv_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -148,19 +142,17 @@ TEST_SUITE_END()
 
 TEST_SUITE(Grouped)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::GroupedIm2ColSmallShapes(), framework::dataset::make("DataType",
-                                                                                                              DataType::F32)),
-                                                                                                      grouped_args),
-                                                                                              framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::GroupedIm2ColSmallShapes(), framework::dataset::make("DataType",
+                                                                                                      DataType::F32)),
+                                                                                              grouped_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::GroupedIm2ColLargeShapes(), framework::dataset::make("DataType",
-                                                                                                                  DataType::F32)),
-                                                                                                          grouped_args),
-                                                                                                  framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::GroupedIm2ColLargeShapes(), framework::dataset::make("DataType",
+                                                                                                          DataType::F32)),
+                                                                                                  grouped_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -168,19 +160,17 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<float>, framework::DatasetMode:
 TEST_SUITE_END()
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::GroupedIm2ColSmallShapes(), framework::dataset::make("DataType",
-                                                                                                             DataType::F16)),
-                                                                                                     grouped_args),
-                                                                                             framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::GroupedIm2ColSmallShapes(), framework::dataset::make("DataType",
+                                                                                                     DataType::F16)),
+                                                                                             grouped_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::GroupedIm2ColLargeShapes(), framework::dataset::make("DataType",
-                                                                                                                 DataType::F16)),
-                                                                                                         grouped_args),
-                                                                                                 framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::GroupedIm2ColLargeShapes(), framework::dataset::make("DataType",
+                                                                                                         DataType::F16)),
+                                                                                                 grouped_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
@@ -188,19 +178,17 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<half>, framework::DatasetMode::
 TEST_SUITE_END()
 
 TEST_SUITE(QASYMM8)
-FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::GroupedIm2ColSmallShapes(), framework::dataset::make("DataType",
-                                                                                                                DataType::QASYMM8)),
-                                                                                                        grouped_args),
-                                                                                                framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunSmall, CLIm2ColFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(datasets::GroupedIm2ColSmallShapes(), framework::dataset::make("DataType",
+                                                                                                        DataType::QASYMM8)),
+                                                                                                grouped_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
 }
 
-FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::GroupedIm2ColLargeShapes(), framework::dataset::make("DataType",
-                                                                                                                    DataType::QASYMM8)),
-                                                                                                            grouped_args),
-                                                                                                    framework::dataset::make("ChannelsFirstOutputNHWC", true)))
+FIXTURE_DATA_TEST_CASE(RunLarge, CLIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::GroupedIm2ColLargeShapes(), framework::dataset::make("DataType",
+                                                                                                            DataType::QASYMM8)),
+                                                                                                    grouped_args))
 {
     // Validate output
     validate(CLAccessor(_target), _reference);
diff --git a/tests/validation/NEON/Im2Col.cpp b/tests/validation/NEON/Im2Col.cpp
index 0ea68bf49d..5a2b46a550 100644
--- a/tests/validation/NEON/Im2Col.cpp
+++ b/tests/validation/NEON/Im2Col.cpp
@@ -78,16 +78,14 @@ using NEIm2ColFixture = Im2ColValidationFixture<Tensor, Accessor, NEIm2Col, T, f
 
 TEST_SUITE(Float)
 TEST_SUITE(FP32)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)),
-                                                                                                      conv_args),
-                                                                                              framework::dataset::make("ChannelsFirstOutputNHWC", false)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)),
+                                                                                              conv_args))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)),
-                                                                                                          conv_args),
-                                                                                                  framework::dataset::make("ChannelsFirstOutputNHWC", false)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F32)),
+                                                                                                  conv_args))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -97,16 +95,14 @@ TEST_SUITE_END()
 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
 TEST_SUITE(FP16)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)),
-                                                                                                     conv_args),
-                                                                                             framework::dataset::make("ChannelsFirstOutputNHWC", false)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)),
+                                                                                             conv_args))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)),
-                                                                                                         conv_args),
-                                                                                                 framework::dataset::make("ChannelsFirstOutputNHWC", false)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::F16)),
+                                                                                                 conv_args))
 {
     // Validate output
     validate(Accessor(_target), _reference);
@@ -118,16 +114,14 @@ TEST_SUITE_END()
 TEST_SUITE_END()
 
 TEST_SUITE(QASYMM8)
-FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
-                                                                                                        conv_args),
-                                                                                                framework::dataset::make("ChannelsFirstOutputNHWC", false)))
+FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                conv_args))
 {
     // Validate output
     validate(Accessor(_target), _reference);
 }
-FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
-                                                                                                            conv_args),
-                                                                                                    framework::dataset::make("ChannelsFirstOutputNHWC", false)))
+FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", DataType::QASYMM8)),
+                                                                                                    conv_args))
 {
     // Validate output
     validate(Accessor(_target), _reference);
diff --git a/tests/validation/fixtures/Im2ColFixture.h b/tests/validation/fixtures/Im2ColFixture.h
index b5e83a9872..809bafd0b2 100644
--- a/tests/validation/fixtures/Im2ColFixture.h
+++ b/tests/validation/fixtures/Im2ColFixture.h
@@ -50,7 +50,7 @@ class Im2ColValidationFixture : public framework::Fixture
 public:
     template <typename...>
     void setup(TensorShape input_shape, DataType data_type, const Size2D &kernel_dims, const PadStrideInfo &conv_info, const QuantizationInfo &quant_info, const DataLayout &data_layout,
-               unsigned int num_groups, bool channels_first_output_nhwc)
+               unsigned int num_groups)
     {
         _kernel_dims = kernel_dims;
         _conv_info   = conv_info;
@@ -70,7 +70,7 @@ public:
         const TensorShape output_shape = compute_im2col_conv_shape(&input_info, _kernel_dims, _conv_info, _has_bias, Size2D(1U, 1U), batch_size_on_z && _num_groups == 1, _num_groups);
         _target                        = compute_target(input_shape, output_shape, data_type);
 
-        compute_reference(input_shape, output_shape, data_type, channels_first_output_nhwc);
+        compute_reference(input_shape, output_shape, data_type);
     }
 
 protected:
@@ -109,7 +109,7 @@ protected:
         return dst;
     }
 
-    void compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, DataType data_type, bool channels_first_output_nhwc)
+    void compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, DataType data_type)
     {
         // Create reference
         SimpleTensor<T> src{ input_shape, data_type, 1, _quant_info, _data_layout };
@@ -118,7 +118,7 @@ protected:
         // Fill reference
         fill(src);
 
-        reference::im2col<T>(src, _reference, _kernel_dims, _conv_info, _has_bias, _num_groups, channels_first_output_nhwc);
+        reference::im2col<T>(src, _reference, _kernel_dims, _conv_info, _has_bias, _num_groups);
     }
     TensorType       _target{};
     SimpleTensor<T>  _reference{};
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
diff --git a/tests/validation/reference/Im2Col.h b/tests/validation/reference/Im2Col.h
index 84ee237453..f519d0e602 100644
--- a/tests/validation/reference/Im2Col.h
+++ b/tests/validation/reference/Im2Col.h
@@ -35,8 +35,7 @@ namespace validation
 namespace reference
 {
 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 = false);
+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);
 } // namespace reference
 } // namespace validation
 } // namespace test