Crop Kernel Decoupling

Change-Id: I1f39ba6a255847f4d21837a3dce4f867322203e6
Signed-off-by: alerah01 <alex.rahlis@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/6799
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>

11 files changed, 545 insertions, 146 deletions

diff --git a/Android.bp b/Android.bp
index 5654b840fe..4d5f2bdf68 100644
--- a/Android.bp
+++ b/Android.bp
@@ -453,6 +453,10 @@ cc_library_static {
         "src/cpu/kernels/add/generic/sve2/qasymm8.cpp",
         "src/cpu/kernels/add/generic/sve2/qasymm8_signed.cpp",
         "src/cpu/kernels/add/generic/sve2/qsymm16.cpp",
+        "src/cpu/kernels/crop/generic/neon/fp16.cpp",
+        "src/cpu/kernels/crop/generic/neon/fp32.cpp",
+        "src/cpu/kernels/crop/generic/neon/impl.cpp",
+        "src/cpu/kernels/crop/generic/neon/integer.cpp",
         "src/cpu/kernels/elementwise/sve/elementwise.cpp",
         "src/cpu/kernels/elementwise/sve/elementwise_unary.cpp",
         "src/cpu/kernels/floor/neon/fp16.cpp",
diff --git a/filelist.json b/filelist.json
index d7847480a4..040ff51333 100644
--- a/filelist.json
+++ b/filelist.json
@@ -1066,7 +1066,13 @@
           "common": [
             "src/core/NEON/kernels/NECropKernel.cpp",
             "src/runtime/NEON/functions/NECropResize.cpp"
-          ]
+          ],
+          "neon": {
+            "common":  [ "src/cpu/kernels/crop/generic/neon/impl.cpp" ],
+            "fp32":    [ "src/cpu/kernels/crop/generic/neon/fp32.cpp" ],
+            "fp16":    [ "src/cpu/kernels/crop/generic/neon/fp16.cpp" ],
+            "integer": [ "src/cpu/kernels/crop/generic/neon/integer.cpp"   ]
+          }  
         }
       },
       "Deconv2d": {
diff --git a/src/core/NEON/kernels/NECropKernel.cpp b/src/core/NEON/kernels/NECropKernel.cpp
index fabbd64305..729402116d 100644
--- a/src/core/NEON/kernels/NECropKernel.cpp
+++ b/src/core/NEON/kernels/NECropKernel.cpp
@@ -31,136 +31,92 @@
 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
 #include "src/core/CPP/Validate.h"
 #include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.h"
 #include "src/core/helpers/AutoConfiguration.h"
 #include "src/core/helpers/WindowHelpers.h"
 #include "src/core/utils/helpers/bit_ops.h"
+#include "src/cpu/kernels/crop/generic/neon/list.h"
 
 namespace arm_compute
 {
 namespace
 {
-template <typename T>
-inline float32x4_t load_as_f32(T *ptr)
+struct CropSelectorData
 {
-    ARM_COMPUTE_UNUSED(ptr);
-    ARM_COMPUTE_ERROR("Type not supported.");
-}
-
-template <>
-inline float32x4_t load_as_f32(float *ptr)
-{
-    return wrapper::vloadq(ptr);
-}
-
-template <>
-inline float32x4_t load_as_f32(int32_t *ptr)
-{
-    return vcvtq_f32_s32(wrapper::vloadq(ptr));
-}
+    DataType dt;
+};
 
-template <>
-inline float32x4_t load_as_f32(uint32_t *ptr)
-{
-    return vcvtq_f32_u32(wrapper::vloadq(ptr));
-}
+using CropSelectorPtr = std::add_pointer<bool(const CropSelectorData &data)>::type;
+using CropUKernelPtr  = std::add_pointer<void(const ITensor *, const ITensor *, float *, Coordinates, int32_t, int32_t, int32_t, bool, bool)>::type;
 
-template <>
-inline float32x4_t load_as_f32(int16_t *ptr)
+struct CropUKernel
 {
-    return vcvtq_f32_s32(vmovl_s16(wrapper::vload(ptr)));
-}
+    const char           *name;
+    const CropSelectorPtr is_selected;
+    CropUKernelPtr        ukernel;
+};
 
-template <>
-inline float32x4_t load_as_f32(uint16_t *ptr)
+static const CropUKernel available_kernels[] =
 {
-    return vcvtq_f32_u32(vmovl_u16(wrapper::vload(ptr)));
-}
-
-template <>
-inline float32x4_t load_as_f32(uint8_t *ptr)
-{
-    return vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8(wrapper::vload(ptr)))));
-}
-
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-template <>
-inline float32x4_t load_as_f32(float16_t *ptr)
-{
-    return vcvt_f32_f16(wrapper::vload(ptr));
-}
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-
-template <typename T>
-inline void in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
-                                  int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
-{
-    // Reverse elements if width flipped.
-    if(is_width_flipped)
     {
-        // Collapse first dimension if possible.
-        if(input_has_single_channel)
-        {
-            int32_t     x = output_width_start;
-            Coordinates negative_offset(input_offset);
-            negative_offset.set(1, negative_offset[1] - window_step_x + 1);
-            for(; x <= output_width_limit - window_step_x; x += window_step_x, negative_offset[1] -= window_step_x)
-            {
-                auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(negative_offset)));
-
-                in = wrapper::vrev64(in);
-                in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in));
+        "fp16_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::F16; },
+        REGISTER_FP16_NEON(arm_compute::cpu::fp16_in_bounds_crop_window)
+    },
+    {
+        "f32_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::F32; },
+        REGISTER_FP32_NEON(arm_compute::cpu::fp32_in_bounds_crop_window)
+    },
+    {
+        "u8_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::U8; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::u8_in_bounds_crop_window)
+    },
+    {
+        "u16_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::U16; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::u16_in_bounds_crop_window)
+    },
+    {
+        "u32_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::U32; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::u32_in_bounds_crop_window)
+    },
+    {
+        "s8_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::S8; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::s8_in_bounds_crop_window)
+    },
+    {
+        "s16_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::S16; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::s16_in_bounds_crop_window)
+    },
+    {
+        "s32_neon_crop",
+        [](const CropSelectorData & data) { return data.dt == DataType::S32; },
+        REGISTER_INTEGER_NEON(arm_compute::cpu::s32_in_bounds_crop_window)
+    },
+};
 
-                wrapper::vstore(output_ptr + x, in);
-            }
-            input_offset[1] = negative_offset[1] + window_step_x - 1;
-            for(; x < output_width_limit; ++x, --input_offset[1])
-            {
-                *(output_ptr + x) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
-            }
-        }
-        else
-        {
-            for(int32_t x = output_width_start; x < output_width_limit; ++x, --input_offset[1])
-            {
-                input_offset.set(0, 0);
-                int32_t c = 0;
-                for(; c <= static_cast<int32_t>(input->info()->dimension(0)) - window_step_x; c += window_step_x, input_offset[0] += window_step_x)
-                {
-                    auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
-                    wrapper::vstore(output_ptr + x * output->info()->dimension(0) + c, in);
-                }
-                for(; c < static_cast<int32_t>(input->info()->dimension(0)); ++c, ++input_offset[0])
-                {
-                    *(output_ptr + x * output->info()->dimension(0) + c) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
-                }
-            }
-        }
-    }
-    else
+/** Micro-kernel selector
+ *
+ * @param[in] data Selection data passed to help pick the appropriate micro-kernel
+ *
+ * @return A matching micro-kernel else nullptr
+ */
+const CropUKernel *get_implementation(const CropSelectorData &data)
+{
+    for(const auto &uk : available_kernels)
     {
-        // Use memcpy if the elements don't need converting to float.
-        if(std::is_same<T, float>::value)
+        if(uk.is_selected(data))
         {
-            memcpy(static_cast<void *>(output_ptr + output_width_start * output->info()->dimension(0)),
-                   reinterpret_cast<const void *>(input->ptr_to_element(input_offset)),
-                   (output_width_limit - output_width_start) * output->info()->dimension(0) * output->info()->element_size());
-        }
-        else
-        {
-            int32_t x                = 0;
-            int32_t limit            = (output_width_limit - output_width_start) * static_cast<int32_t>(output->info()->dimension(0));
-            float *output_start_ptr = output_ptr + output_width_start * output->info()->dimension(0);
-            for(; x <= limit - window_step_x; x += window_step_x, input_offset[0] += window_step_x)
-            {
-                auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
-                wrapper::vstore(output_start_ptr + x, in);
-            }
-            for(; x < limit; ++x, ++input_offset[0])
-            {
-                *(output_start_ptr + x) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
-            }
+            return &uk;
         }
     }
+
+    return nullptr;
 }
 
 inline void out_of_bounds_crop_window(const ITensor *output, float *output_ptr, float extrapolation_value,
@@ -234,8 +190,7 @@ inline void execute_window(const ITensor *input, const ITensor *output, Coordina
 } // namespace
 
 NECropKernel::NECropKernel()
-    : _input(nullptr), _crop_boxes(nullptr), _box_ind(nullptr), _output(nullptr), _start(), _end(), _crop_box_ind(0), _extrapolation_value(0), _rows_out_of_bounds(), _cols_out_of_bounds(),
-      _in_bounds_crop_function(nullptr)
+    : _input(nullptr), _crop_boxes(nullptr), _box_ind(nullptr), _output(nullptr), _start(), _end(), _crop_box_ind(0), _extrapolation_value(0), _rows_out_of_bounds(), _cols_out_of_bounds()
 {
 }
 
@@ -250,40 +205,14 @@ void NECropKernel::configure(const ITensor *input, const ITensor *crop_boxes, co
     _output              = output;
     _crop_box_ind        = crop_box_ind;
     _extrapolation_value = extrapolation_value;
-
-    switch(input->info()->data_type())
-    {
-        case DataType::F32:
-            _in_bounds_crop_function = &in_bounds_crop_window<float>;
-            break;
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-        case DataType::F16:
-            _in_bounds_crop_function = &in_bounds_crop_window<float16_t>;
-            break;
-#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */
-        case DataType::U32:
-            _in_bounds_crop_function = &in_bounds_crop_window<uint32_t>;
-            break;
-        case DataType::S32:
-            _in_bounds_crop_function = &in_bounds_crop_window<int32_t>;
-            break;
-        case DataType::U16:
-            _in_bounds_crop_function = &in_bounds_crop_window<uint16_t>;
-            break;
-        case DataType::S16:
-            _in_bounds_crop_function = &in_bounds_crop_window<int16_t>;
-            break;
-        case DataType::U8:
-            _in_bounds_crop_function = &in_bounds_crop_window<uint8_t>;
-            break;
-        default:
-            ARM_COMPUTE_ERROR("Datatype not supported");
-    }
 }
 
 Status NECropKernel::validate(const ITensorInfo *input, const ITensorInfo *crop_boxes, const ITensorInfo *box_ind, const ITensorInfo *output, uint32_t crop_box_ind, float extrapolation_value)
 {
     ARM_COMPUTE_UNUSED(extrapolation_value);
+    const auto *uk = get_implementation(CropSelectorData{ input->data_type() });
+    ARM_COMPUTE_RETURN_ERROR_ON(uk == nullptr || uk->ukernel == nullptr);
+
     ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::U16, DataType::S16, DataType::F16, DataType::U32, DataType::S32, DataType::F32);
     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(input, DataLayout::NHWC);
@@ -369,10 +298,12 @@ void NECropKernel::run(const Window &window, const ThreadInfo &info)
     ARM_COMPUTE_ERROR_ON(_input->info()->has_padding());
     ARM_COMPUTE_ERROR_ON(_output->info()->has_padding());
 
+    const auto *uk = get_implementation(CropSelectorData{ _input->info()->data_type() });
+
     uint32_t    batch_index = *(reinterpret_cast<int32_t *>(_box_ind->ptr_to_element(Coordinates(_crop_box_ind))));
     Coordinates input_offset(0, _end[0] < _start[0] ? _start[0] - _cols_out_of_bounds[0] : _start[0] + _cols_out_of_bounds[0],
                              _end[1] < _start[1] ? _start[1] - _rows_out_of_bounds[0] : _start[1] + _rows_out_of_bounds[0], batch_index);
-    execute_window(_input, _output, input_offset, _extrapolation_value, _rows_out_of_bounds, _cols_out_of_bounds, _in_bounds_crop_function, _end[1] < _start[1],
+    execute_window(_input, _output, input_offset, _extrapolation_value, _rows_out_of_bounds, _cols_out_of_bounds, uk->ukernel, _end[1] < _start[1],
                    _cols_out_of_bounds[0] + _cols_out_of_bounds[1] < _output->info()->dimension(1), _cols_out_of_bounds[0] > 0, _cols_out_of_bounds[1] > 0,
                    _start[0] <= _end[0], _end[0] < _start[0]);
 }
diff --git a/src/core/NEON/kernels/NECropKernel.h b/src/core/NEON/kernels/NECropKernel.h
index 742215e22b..6c989c1d2c 100644
--- a/src/core/NEON/kernels/NECropKernel.h
+++ b/src/core/NEON/kernels/NECropKernel.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -107,8 +107,6 @@ private:
     std::array<uint32_t, 2> _rows_out_of_bounds;
     /** The number of columns out of bounds at the start and end of output. */
     std::array<uint32_t, 2> _cols_out_of_bounds;
-
-    NECropKernel::InBoundsCropFunction *_in_bounds_crop_function;
 };
 } // namespace arm_compute
 #endif /*ARM_COMPUTE_NEON_CROP_KERNEL_H */
diff --git a/src/cpu/kernels/crop/generic/neon/crop_helper.h b/src/cpu/kernels/crop/generic/neon/crop_helper.h
new file mode 100644
index 0000000000..1fe8e11e98
--- /dev/null
+++ b/src/cpu/kernels/crop/generic/neon/crop_helper.h
@@ -0,0 +1,86 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_CROP_CROP_HELPER_H
+#define SRC_CORE_NEON_KERNELS_CROP_CROP_HELPER_H
+
+#include "src/core/NEON/wrapper/wrapper.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename T>
+inline float32x4_t load_as_f32(T *ptr)
+{
+    ARM_COMPUTE_UNUSED(ptr);
+    ARM_COMPUTE_ERROR("Type not supported.");
+}
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+template <>
+inline float32x4_t load_as_f32(float16_t *ptr)
+{
+    return vcvt_f32_f16(wrapper::vload(ptr));
+}
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
+
+template <>
+inline float32x4_t load_as_f32(float *ptr)
+{
+    return wrapper::vloadq(ptr);
+}
+
+template <>
+inline float32x4_t load_as_f32(int32_t *ptr)
+{
+    return vcvtq_f32_s32(wrapper::vloadq(ptr));
+}
+
+template <>
+inline float32x4_t load_as_f32(uint32_t *ptr)
+{
+    return vcvtq_f32_u32(wrapper::vloadq(ptr));
+}
+
+template <>
+inline float32x4_t load_as_f32(int16_t *ptr)
+{
+    return vcvtq_f32_s32(vmovl_s16(wrapper::vload(ptr)));
+}
+
+template <>
+inline float32x4_t load_as_f32(uint16_t *ptr)
+{
+    return vcvtq_f32_u32(vmovl_u16(wrapper::vload(ptr)));
+}
+
+template <>
+inline float32x4_t load_as_f32(uint8_t *ptr)
+{
+    return vcvtq_f32_u32(vmovl_u16(vget_low_u16(vmovl_u8(wrapper::vload(ptr)))));
+}
+}
+} // namespace arm_compute
+
+#endif //SRC_CORE_NEON_KERNELS_CROP_CROP_HELPER_H
\ No newline at end of file
diff --git a/src/cpu/kernels/crop/generic/neon/fp16.cpp b/src/cpu/kernels/crop/generic/neon/fp16.cpp
new file mode 100644
index 0000000000..218ebba191
--- /dev/null
+++ b/src/cpu/kernels/crop/generic/neon/fp16.cpp
@@ -0,0 +1,40 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+
+#include "src/cpu/kernels/crop/generic/neon/impl.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void fp16_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<float16_t>(input, output, output_ptr, input_offset,
+                                            window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+}
+} // namespace arm_compute
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
diff --git a/src/cpu/kernels/crop/generic/neon/fp32.cpp b/src/cpu/kernels/crop/generic/neon/fp32.cpp
new file mode 100644
index 0000000000..16d0218fce
--- /dev/null
+++ b/src/cpu/kernels/crop/generic/neon/fp32.cpp
@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "src/cpu/kernels/crop/generic/neon/impl.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void fp32_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<float32_t>(input, output, output_ptr, input_offset,
+                                            window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+}
+} // namespace arm_compute
diff --git a/src/cpu/kernels/crop/generic/neon/impl.cpp b/src/cpu/kernels/crop/generic/neon/impl.cpp
new file mode 100644
index 0000000000..95ab804940
--- /dev/null
+++ b/src/cpu/kernels/crop/generic/neon/impl.cpp
@@ -0,0 +1,127 @@
+/*
+ * Copyright (c) 2018-2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "src/cpu/kernels/crop/generic/neon/impl.h"
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.h"
+#include "src/cpu/kernels/crop/generic/neon/crop_helper.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename T>
+void in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                           int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    // Reverse elements if width flipped.
+    if(is_width_flipped)
+    {
+        // Collapse first dimension if possible.
+        if(input_has_single_channel)
+        {
+            int32_t     x = output_width_start;
+            Coordinates negative_offset(input_offset);
+            negative_offset.set(1, negative_offset[1] - window_step_x + 1);
+            for(; x <= output_width_limit - window_step_x; x += window_step_x, negative_offset[1] -= window_step_x)
+            {
+                auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(negative_offset)));
+
+                in = wrapper::vrev64(in);
+                in = wrapper::vcombine(wrapper::vgethigh(in), wrapper::vgetlow(in));
+
+                wrapper::vstore(output_ptr + x, in);
+            }
+            input_offset[1] = negative_offset[1] + window_step_x - 1;
+            for(; x < output_width_limit; ++x, --input_offset[1])
+            {
+                *(output_ptr + x) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
+            }
+        }
+        else
+        {
+            for(int32_t x = output_width_start; x < output_width_limit; ++x, --input_offset[1])
+            {
+                input_offset.set(0, 0);
+                int32_t c = 0;
+                for(; c <= static_cast<int32_t>(input->info()->dimension(0)) - window_step_x; c += window_step_x, input_offset[0] += window_step_x)
+                {
+                    auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
+                    wrapper::vstore(output_ptr + x * output->info()->dimension(0) + c, in);
+                }
+                for(; c < static_cast<int32_t>(input->info()->dimension(0)); ++c, ++input_offset[0])
+                {
+                    *(output_ptr + x * output->info()->dimension(0) + c) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
+                }
+            }
+        }
+    }
+    else
+    {
+        // Use memcpy if the elements don't need converting to float.
+        if(std::is_same<T, float>::value)
+        {
+            memcpy(static_cast<void *>(output_ptr + output_width_start * output->info()->dimension(0)),
+                   reinterpret_cast<const void *>(input->ptr_to_element(input_offset)),
+                   (output_width_limit - output_width_start) * output->info()->dimension(0) * output->info()->element_size());
+        }
+        else
+        {
+            int32_t x                = 0;
+            int32_t limit            = (output_width_limit - output_width_start) * static_cast<int32_t>(output->info()->dimension(0));
+            float *output_start_ptr = output_ptr + output_width_start * output->info()->dimension(0);
+            for(; x <= limit - window_step_x; x += window_step_x, input_offset[0] += window_step_x)
+            {
+                auto in = load_as_f32(reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
+                wrapper::vstore(output_start_ptr + x, in);
+            }
+            for(; x < limit; ++x, ++input_offset[0])
+            {
+                *(output_start_ptr + x) = static_cast<float>(*reinterpret_cast<T *>(input->ptr_to_element(input_offset)));
+            }
+        }
+    }
+}
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS)
+template void in_bounds_crop_window<float16_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                               int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+#endif /* defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && defined(ENABLE_FP16_KERNELS) */
+template void in_bounds_crop_window<float32_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                               int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+template void in_bounds_crop_window<uint8_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                             int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+template void in_bounds_crop_window<uint16_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                              int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+template void in_bounds_crop_window<uint32_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                              int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+template void in_bounds_crop_window<int8_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                            int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+template void in_bounds_crop_window<int16_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                             int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+template void in_bounds_crop_window<int32_t>(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                                             int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+}
+} // namespace arm_compute
diff --git a/src/cpu/kernels/crop/generic/neon/impl.h b/src/cpu/kernels/crop/generic/neon/impl.h
new file mode 100644
index 0000000000..50f889705a
--- /dev/null
+++ b/src/cpu/kernels/crop/generic/neon/impl.h
@@ -0,0 +1,41 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_CROP_IMPL_H
+#define SRC_CORE_NEON_KERNELS_CROP_IMPL_H
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+template <typename T>
+void in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                           int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped);
+} // namespace cpu
+} // namespace arm_compute
+#endif //SRC_CORE_NEON_KERNELS_CROP_IMPL_H
diff --git a/src/cpu/kernels/crop/generic/neon/integer.cpp b/src/cpu/kernels/crop/generic/neon/integer.cpp
new file mode 100644
index 0000000000..7dbf0a7f5e
--- /dev/null
+++ b/src/cpu/kernels/crop/generic/neon/integer.cpp
@@ -0,0 +1,74 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "src/cpu/kernels/crop/generic/neon/impl.h"
+#include "src/cpu/kernels/crop/generic/neon/list.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+void u8_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                              int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<uint8_t>(input, output, output_ptr, input_offset,
+                                          window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+
+void u16_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                               int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<uint16_t>(input, output, output_ptr, input_offset,
+                                           window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+
+void u32_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                               int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<uint32_t>(input, output, output_ptr, input_offset,
+                                           window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+
+void s8_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                              int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<int8_t>(input, output, output_ptr, input_offset,
+                                         window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+
+void s16_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                               int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<int16_t>(input, output, output_ptr, input_offset,
+                                          window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+
+void s32_in_bounds_crop_window(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset,
+                               int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+{
+    return in_bounds_crop_window<int32_t>(input, output, output_ptr, input_offset,
+                                          window_step_x, output_width_start, output_width_limit, input_has_single_channel, is_width_flipped);
+}
+}
+} // namespace arm_compute
diff --git a/src/cpu/kernels/crop/generic/neon/list.h b/src/cpu/kernels/crop/generic/neon/list.h
new file mode 100644
index 0000000000..f33049304d
--- /dev/null
+++ b/src/cpu/kernels/crop/generic/neon/list.h
@@ -0,0 +1,54 @@
+/*
+ * Copyright (c) 2021 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef SRC_CORE_NEON_KERNELS_CROP_LIST_H
+#define SRC_CORE_NEON_KERNELS_CROP_LIST_H
+
+#include "arm_compute/core/Helpers.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "src/core/NEON/wrapper/wrapper.h"
+#include "src/core/common/Registrars.h"
+#include "src/cpu/kernels/crop/generic/neon/impl.h"
+
+namespace arm_compute
+{
+namespace cpu
+{
+#define DECLARE_CROP_KERNEL(func_name)                                                                       \
+    void func_name(const ITensor *input, const ITensor *output, float *output_ptr, Coordinates input_offset, \
+                   int32_t window_step_x, int32_t output_width_start, int32_t output_width_limit, bool input_has_single_channel, bool is_width_flipped)
+
+DECLARE_CROP_KERNEL(fp16_in_bounds_crop_window);
+DECLARE_CROP_KERNEL(fp32_in_bounds_crop_window);
+DECLARE_CROP_KERNEL(s8_in_bounds_crop_window);
+DECLARE_CROP_KERNEL(s16_in_bounds_crop_window);
+DECLARE_CROP_KERNEL(s32_in_bounds_crop_window);
+DECLARE_CROP_KERNEL(u8_in_bounds_crop_window);
+DECLARE_CROP_KERNEL(u16_in_bounds_crop_window);
+DECLARE_CROP_KERNEL(u32_in_bounds_crop_window);
+
+#undef DECLARE_CROP_KERNEL
+
+} // namespace cpu
+} // namespace arm_compute
+#endif //SRC_CORE_NEON_KERNELS_CROP_LIST_H