Remove unused code in arm_conv/depthwise/

* Removed header files in arm_conv/depthwise

* Resolves MLCE-990

Change-Id: Iacddd80e2d83ff0fbafb817014f90c5bc80dab3c
Signed-off-by: Pablo Marquez Tello <pablo.tello@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8946
Reviewed-by: Andrew Mundy <Andrew.mundy@arm.com>
Reviewed-by: Viet-Hoa Do <viet-hoa.do@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 0 insertions, 1388 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_multiplier.hpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_multiplier.hpp
deleted file mode 100644
index bb580e605a..0000000000
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_multiplier.hpp
+++ /dev/null
@@ -1,473 +0,0 @@
-/*
- * 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.
- */
-
-#pragma once
-
-#include "src/core/NEON/kernels/arm_gemm/utils.hpp"
-
-#ifdef CYCLE_PROFILING
-#include "profiler.hpp"
-#endif
-
-#include <limits>
-
-namespace arm_conv {
-namespace depthwise {
-
-template <class strategy>
-class DepthwiseDepthfirstGenericWithMultiplierBase :
-  public DepthwiseCommon<typename strategy::input_type,
-                         typename strategy::weight_type,
-                         typename strategy::return_type>
-{
-  protected:
-
-  using TInput = typename strategy::input_type;
-  using TWeight = typename strategy::weight_type;
-  using TOutput = typename strategy::return_type;
-  using TAccum = typename strategy::bias_type;
-
-  unsigned int kernel_points(void) const
-  {
-    return this->m_args.kernel_rows * this->m_args.kernel_cols;
-  }
-
-  unsigned int input_rows(void) const
-  {
-    return (strategy::output_rows() - 1) * this->m_args.stride_rows + this->m_args.kernel_rows;
-  }
-
-  unsigned int input_cols(void) const
-  {
-    return (strategy::output_cols() - 1) * this->m_args.stride_cols + this->m_args.kernel_cols;
-  }
-
-  size_t sizeof_inptr_array(void) const
-  {
-    return sizeof(TInput *) * kernel_points() * strategy::output_rows();
-  }
-
-  size_t sizeof_input_samples(void) const
-  {
-    // We have a sample for each kernel point, for each point of the output array.
-    return sizeof(TInput) * kernel_points() *
-                            strategy::output_rows() *
-                            strategy::output_col_regs() *
-                            (16 / sizeof(TAccum));
-  }
-
-  size_t sizeof_outptr_array(void) const
-  {
-    return sizeof(TOutput *) * strategy::output_rows() * strategy::output_cols();
-  }
-
-  size_t sizeof_output_buffer(unsigned int n_channels) const
-  {
-    const unsigned int vl = arm_gemm::utils::get_vector_length<TOutput>(strategy::vl_type);
-    const auto rounded_channels = arm_gemm::roundup(n_channels, vl);
-    return sizeof(TOutput) * rounded_channels;
-  }
-
-  void pack_weights(TWeight *buffer, const TWeight *weights, size_t ld_weight_col, size_t ld_weight_row) const
-  {
-    const unsigned int vl = arm_gemm::utils::get_vector_length<TAccum>(strategy::vl_type);
-    ld_weight_col = (ld_weight_col == 0) ? this->m_args.channel_multiplier * this->m_args.input_channels : ld_weight_col;
-    ld_weight_row = (ld_weight_row == 0) ? this->m_args.kernel_cols * ld_weight_col : ld_weight_row;
-
-    for (unsigned int in_c = 0; in_c < this->m_args.input_channels; in_c++)
-    {
-      for (unsigned int n = 0; n < this->m_args.channel_multiplier; n += vl)
-      {
-        const unsigned int out_c = in_c * this->m_args.channel_multiplier + n;
-        const unsigned int todo = std::min(vl, this->m_args.channel_multiplier - n);
-
-        // Copy each of the weights in turn
-        auto weights_row = weights + out_c;
-        for (unsigned int i = 0; i < this->m_args.kernel_rows; i++)
-        {
-          auto weights_col = weights_row;
-
-          for (unsigned int j = 0; j < this->m_args.kernel_cols; j++)
-          {
-            for (unsigned int m = 0; m < todo; m++)
-            {
-              buffer[m] = weights_col[m];
-            }
-            buffer += vl;
-
-            weights_col += ld_weight_col;
-          }
-
-          weights_row += ld_weight_row;
-        }
-      }
-    }
-  }
-
-  void execute_tiles(
-    std::function<void(const TInput **, TOutput **, const TWeight *, unsigned int, unsigned int)> tile_fn,
-    const TInput pad_value,
-    const unsigned int batches,
-    const unsigned int input_height,
-    const unsigned int input_width,
-    const unsigned int input_channels,
-    const PaddingValues &padding,
-    const void *const _input,
-    const size_t ld_input_col,
-    const size_t ld_input_row,
-    const size_t ld_input_batch,
-    const void *const parameters,
-    const unsigned int output_height,
-    const unsigned int output_width,
-    void *const _output,
-    const size_t ld_output_col,
-    const size_t ld_output_row,
-    const size_t ld_output_batch,
-    void *const _working_space,
-    const unsigned int thread_id,
-    const unsigned int n_threads
-  ) const
-  {
-#ifdef CYCLE_PROFILING
-    arm_gemm::profiler prof;
-#endif
-
-    // Determine what portion of the work to do.
-    const unsigned int n_rows_per_thread = arm_gemm::iceildiv(output_height, n_threads);
-    const int start_out_height = std::min(thread_id * n_rows_per_thread, output_height);
-    const int end_out_height = std::min(start_out_height + n_rows_per_thread, output_height);
-
-    // Need a stride over blocks of parameters
-    const unsigned int vl = arm_gemm::utils::get_vector_length<TAccum>(strategy::vl_type);
-    const unsigned int param_stride = arm_gemm::roundup(this->m_args.channel_multiplier, vl) * kernel_points();
-
-    // Cast input and output pointers into the right types
-    const TInput *const inptr = static_cast<const TInput *>(_input);
-    TOutput *const outptr = static_cast<TOutput *>(_output);
-
-    // Allocate portions of the working space
-    uint8_t *working_space = static_cast<uint8_t *>(_working_space) +
-                             get_working_size(thread_id, input_channels);
-
-    const TInput **inptrs = reinterpret_cast<const TInput **>(working_space);
-    working_space += sizeof_inptr_array();
-
-    // To simplify the kernel, we process padded or non-NCHW-ordered input into
-    // a form which can be consumed by the kernel. This data is stored here and
-    // passed into the kernel as an array of N pointers (one per row of the
-    // input).
-    TInput *rearranged_input = reinterpret_cast<TInput *>(working_space);
-    working_space += sizeof_input_samples();
-
-    TOutput **outptr_array = reinterpret_cast<TOutput **>(working_space);
-    working_space += sizeof_outptr_array();
-
-    TOutput *const output_buffer = reinterpret_cast<TOutput *>(working_space);
-
-    // TODO Dynamically change the input pointer array in cases where we could
-    // read directly from the input tensor; for now though assume we will
-    // always read from the sample array.
-    {
-      auto my_inptrs = inptrs;
-      auto my_input_samples = rearranged_input;
-
-      // For each kernel point; for each row of output; for each register of
-      // values containing a QUAD of source values.
-      const unsigned int quad_length = 16 / sizeof(TAccum);
-
-      for (auto p = 0u; p < kernel_points() * strategy::output_rows(); p++)
-      {
-        *(my_inptrs)++ = my_input_samples;
-        my_input_samples += arm_gemm::roundup(strategy::output_cols(), quad_length);
-      }
-    }
-
-    // For each output tile, construct the requisite set of pointers and call
-    // into the kernel.
-    for (unsigned int batch = 0; batch < batches; batch++)
-    {
-      // Get batch pointers
-      const auto inptr_batch = inptr + batch * ld_input_batch;
-      const auto outptr_batch = outptr + batch * ld_output_batch;
-
-      for (int start_out_i = start_out_height;
-           start_out_i < end_out_height;
-           start_out_i += static_cast<int>(strategy::output_rows()))
-      {
-        const int end_out_i = std::min(start_out_i + static_cast<int>(strategy::output_rows()), end_out_height);
-        const int start_in_i = start_out_i * this->m_args.stride_rows - padding.top;
-        const int end_in_i = start_in_i + input_rows();
-
-        // Compute top/bottom padding
-        const auto pad_top = static_cast<unsigned int>(-std::min(start_in_i, 0));
-        const auto pad_bottom = static_cast<unsigned int>(-std::min(static_cast<int>(input_height) - end_in_i, 0));
-        const unsigned int valid_output_rows = std::min(
-          end_out_i - start_out_i,
-          static_cast<int>(output_height) - start_out_i
-        );
-
-        const int pad_rows = pad_top + pad_bottom;
-
-        for (int start_out_j = 0; start_out_j < static_cast<int>(output_width);)
-        {
-          const int start_in_j = start_out_j * this->m_args.stride_cols - this->m_args.padding.left;
-          const int pad_left = -std::min(0, start_in_j);
-
-          const int end_out_j = start_out_j + strategy::output_cols();
-          const int end_in_j = start_in_j + input_cols();
-
-          const auto pad_right = static_cast<unsigned int>(-std::min(static_cast<int>(input_width) - end_in_j, 0));
-          const unsigned int valid_output_cols = std::min(
-            end_out_j - start_out_j,
-            static_cast<int>(output_width) - start_out_j
-          );
-
-          const int pad_cols = pad_left + pad_right;
-
-          // Construct the output pointer array.
-          TOutput **outptr_pos = outptr_array;
-          for (auto i = 0u; i < valid_output_rows; i++)
-          {
-            unsigned int j = 0u;
-            TOutput *colptr = outptr_batch + (start_out_i + i) * ld_output_row + start_out_j * ld_output_col;
-            for (; j < valid_output_cols; j++)
-            {
-              *(outptr_pos++) = colptr;
-               colptr += ld_output_col;
-            }
-            for (; j < strategy::output_cols(); j++)
-            {
-              *(outptr_pos++) = output_buffer;
-            }
-          }
-          for (auto i = valid_output_rows; i < strategy::output_rows(); i++)
-          {
-            for (auto j = 0u; j < strategy::output_cols(); j++)
-            {
-              *(outptr_pos++) = output_buffer;
-            }
-          }
-
-          start_out_j += strategy::output_cols();
-
-          const TWeight *params = static_cast<const TWeight *>(parameters);
-
-          // Fill the input samples with padding. We can do this outside of
-          // the channel loop, as the position of padding isn't going to
-          // change as a function of channel.
-          for (auto i = 0u; i < kernel_points() * strategy::output_rows() * strategy::output_cols(); i++)
-          {
-            rearranged_input[i] = pad_value;
-          }
-
-          // Loop over the input channels
-          for (unsigned int in_c = 0; in_c < input_channels; in_c++)
-          {
-            auto inptr_row = inptr_batch + in_c +
-                             (start_in_i + pad_top) * ld_input_row +
-                             (start_in_j + pad_left) * ld_input_col;
-
-            // Construct the array of input samples; for each point of the
-            // kernel we provide an input value for each output point.
-            auto input_samples = rearranged_input;
-            for (auto ki = 0u; ki < this->m_args.kernel_rows; ki++)
-            {
-              for (auto kj = 0u; kj < this->m_args.kernel_cols; kj++)
-              {
-                // Copy the pointer for the input samples associated with this
-                // kernel point. Then update the main pointer to account for
-                // this point.
-                auto point_input_samples = input_samples;
-                input_samples += strategy::output_rows() * strategy::output_cols();
-
-                int ii = static_cast<int>(ki) - static_cast<int>(pad_top);
-                for (auto oi = 0u;
-                     oi < strategy::output_rows() &&
-                     ii < static_cast<int>(input_rows()) - pad_rows;
-                     oi++, ii += this->m_args.stride_rows)
-                {
-                  if (0 <= ii) // Fill in values only if this row is in range.
-                  {
-                    int ij = static_cast<int>(kj) - static_cast<int>(pad_left);
-                    for (auto oj = 0u;
-                         oj < strategy::output_cols() &&
-                         ij < static_cast<int>(input_cols()) - pad_cols;
-                         oj++, ij += this->m_args.stride_cols)
-                    {
-                      if (0 <= ij) // Sample if the point is in range.
-                      {
-                        point_input_samples[oj] = *(inptr_row + ii*ld_input_row + ij*ld_input_col);
-                      }
-                    }
-                  }
-
-                  point_input_samples += strategy::output_cols();
-                }
-              }
-            }
-
-            tile_fn(inptrs, outptr_array, params, in_c, in_c*this->m_args.channel_multiplier);
-
-            // Progress the output pointers
-            TOutput **outptr_pos = outptr_array;
-            for (auto i = 0u; i < strategy::output_rows() * strategy::output_cols(); i++)
-            {
-              outptr_pos[i] += this->m_args.channel_multiplier;
-            }
-
-            // Progress the pointer into the parameters
-            params += param_stride;
-          }
-        }
-      }
-    }
-  }
-
-  public:
-  DepthwiseDepthfirstGenericWithMultiplierBase(const DepthwiseArgs &args) : DepthwiseCommon<TInput, TWeight, TOutput>(args)
-  {
-  }
-
-  DepthwiseDepthfirstGenericWithMultiplierBase(DepthwiseDepthfirstGenericWithMultiplierBase &) = delete;
-  DepthwiseDepthfirstGenericWithMultiplierBase &operator=(DepthwiseDepthfirstGenericWithMultiplierBase &) = delete;
-
-  size_t get_storage_size(void) const override
-  {
-    const unsigned int vl = arm_gemm::utils::get_vector_length<TAccum>(strategy::vl_type);
-    const auto rounded_channels = this->m_args.input_channels * arm_gemm::roundup(this->m_args.channel_multiplier, vl);
-    return kernel_points() * rounded_channels * sizeof(TWeight);
-  }
-
-  size_t get_working_size(const unsigned int n_threads, const unsigned int n_channels) const override
-  {
-    const unsigned int n_output_channels = n_channels * this->m_args.channel_multiplier;
-    return n_threads * (sizeof_inptr_array() +
-                        sizeof_input_samples() +
-                        sizeof_outptr_array() +
-                        sizeof_output_buffer(n_output_channels));
-  }
-};
-
-template <class strategy>
-class DepthwiseDepthfirstGenericWithMultiplier : public DepthwiseDepthfirstGenericWithMultiplierBase<strategy>
-{
-  using TInput = typename strategy::input_type;
-  using TWeight = typename strategy::weight_type;
-  using TOutput = typename strategy::return_type;
-  using TAccum = typename strategy::bias_type;
-
-  using Parent = DepthwiseDepthfirstGenericWithMultiplierBase<strategy>;
-
-  const TAccum *m_biases;  // Pointer to bias vector
-
-  public:
-  DepthwiseDepthfirstGenericWithMultiplier(const DepthwiseArgs &args)
-    : Parent(args), m_biases(nullptr)
-  {
-  }
-
-  DepthwiseDepthfirstGenericWithMultiplier(DepthwiseDepthfirstGenericWithMultiplier &) = delete;
-  DepthwiseDepthfirstGenericWithMultiplier &operator=(DepthwiseDepthfirstGenericWithMultiplier &) = delete;
-
-  void pack_parameters(void *buffer, const void *biases, const void *weights, size_t ld_weight_col, size_t ld_weight_row) override
-  {
-    m_biases = static_cast<const TAccum *>(biases);
-    Parent::pack_weights(static_cast<TAccum *>(buffer), static_cast<const TWeight *>(weights), ld_weight_col, ld_weight_row);
-  }
-
-  using DepthwiseDepthfirstGenericWithMultiplierBase<strategy>::execute;
-  void execute(
-    const unsigned int batches,
-    const unsigned int input_height,
-    const unsigned int input_width,
-    const unsigned int input_channels,
-    const PaddingValues &padding,
-    const void *const _input,
-    const size_t ld_input_col,
-    const size_t ld_input_row,
-    const size_t ld_input_batch,
-    const void *const parameters,
-    const unsigned int output_height,
-    const unsigned int output_width,
-    void *const _output,
-    const size_t ld_output_col,
-    const size_t ld_output_row,
-    const size_t ld_output_batch,
-    void *const _working_space,
-    const unsigned int thread_id,
-    const unsigned int n_threads
-  ) const override
-  {
-    strategy strat(this->m_args.cpu_info);
-#ifdef CYCLE_PROFILING
-    arm_gemm::profiler prof;
-#endif
-
-    // Compute activation values
-    TAccum activation_min, activation_max;
-    std::tie(activation_min, activation_max) = get_default_activation_values<TAccum>();
-
-    switch (this->m_args.activation.type)
-    {
-      case arm_gemm::Activation::Type::BoundedReLU:
-        activation_max = static_cast<TAccum>(this->m_args.activation.param1);
-        // Fall through
-      case arm_gemm::Activation::Type::ReLU:
-        activation_min = static_cast<TAccum>(0);
-        break;
-      default:
-        break;
-    }
-
-    // Get a function to call for each point of the output
-    auto tile_fn = [&] (const TInput **inptrs,
-                        TOutput **outptrs,
-                        const TWeight *weights,
-                        const unsigned int,
-                        const unsigned int start_output_channel) {
-#ifdef CYCLE_PROFILING
-      auto p = prof.ScopedProfiler(PROFILE_KERNEL, (unsigned long)(strategy::output_rows() * strategy::output_cols() * this->m_args.channel_multiplier * this->m_args.kernel_rows * this->m_args.kernel_cols));
-#endif
-      strat.kernel(
-        inptrs, outptrs, weights,
-        m_biases ? m_biases + start_output_channel : nullptr,
-        this->kernel_points(), this->m_args.channel_multiplier,
-        activation_min, activation_max
-      );
-    };
-
-    Parent::execute_tiles(
-      tile_fn, 0.0f,
-      batches, input_height, input_width, input_channels, padding,
-      _input, ld_input_col, ld_input_row, ld_input_batch,
-      parameters,
-      output_height, output_width,
-      _output, ld_output_col, ld_output_row, ld_output_batch,
-      _working_space, thread_id, n_threads
-    );
-  }
-};
-
-}  // namespace depthwise
-}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_multiplier_quantized.hpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_multiplier_quantized.hpp
deleted file mode 100644
index d42382e208..0000000000
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_multiplier_quantized.hpp
+++ /dev/null
@@ -1,127 +0,0 @@
-/*
- * 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.
- */
-
-#pragma once
-
-#include "src/core/NEON/kernels/arm_gemm/utils.hpp"
-
-#ifdef CYCLE_PROFILING
-#include "profiler.hpp"
-#endif
-
-#include "depthwise_depthfirst_generic_multiplier.hpp"
-
-namespace arm_conv {
-namespace depthwise {
-
-template <class strategy>
-class DepthwiseDepthfirstGenericWithMultiplierQuantized : public DepthwiseDepthfirstGenericWithMultiplierBase<strategy>
-{
-  using TInput = typename strategy::input_type;
-  using TWeight = typename strategy::weight_type;
-  using TOutput = typename strategy::return_type;
-  using TAccum = typename strategy::bias_type;
-
-  using Parent = DepthwiseDepthfirstGenericWithMultiplierBase<strategy>;
-
-  arm_gemm::Requantize32 m_qp;
-
-  public:
-  DepthwiseDepthfirstGenericWithMultiplierQuantized(const DepthwiseArgs &args, const arm_gemm::Requantize32 &qp)
-    : Parent(args), m_qp(qp)
-  {
-  }
-
-  DepthwiseDepthfirstGenericWithMultiplierQuantized(DepthwiseDepthfirstGenericWithMultiplierQuantized &) = delete;
-  DepthwiseDepthfirstGenericWithMultiplierQuantized &operator=(DepthwiseDepthfirstGenericWithMultiplierQuantized &) = delete;
-
-  void pack_parameters(void *buffer, const void *biases, const void *weights, size_t ld_weight_col, size_t ld_weight_row) override
-  {
-    m_qp.bias = static_cast<const TAccum *>(biases);
-    Parent::pack_weights(static_cast<TWeight *>(buffer), static_cast<const TWeight *>(weights), ld_weight_col, ld_weight_row);
-  }
-
-  using Parent::execute;
-  void execute(
-    const unsigned int batches,
-    const unsigned int input_height,
-    const unsigned int input_width,
-    const unsigned int input_channels,
-    const PaddingValues &padding,
-    const void *const _input,
-    const size_t ld_input_col,
-    const size_t ld_input_row,
-    const size_t ld_input_batch,
-    const void *const parameters,
-    const unsigned int output_height,
-    const unsigned int output_width,
-    void *const _output,
-    const size_t ld_output_col,
-    const size_t ld_output_row,
-    const size_t ld_output_batch,
-    void *const _working_space,
-    const unsigned int thread_id,
-    const unsigned int n_threads
-  ) const override
-  {
-    strategy strat(this->m_args.cpu_info);
-#ifdef CYCLE_PROFILING
-    arm_gemm::profiler prof;
-#endif
-
-    // Get a function to call for each point of the output
-    auto tile_fn = [&] (const TInput **inptrs,
-                        TOutput **outptrs,
-                        const TWeight *weights,
-                        const unsigned int,
-                        const unsigned int start_output_channel) {
-#ifdef CYCLE_PROFILING
-      auto p = prof.ScopedProfiler(PROFILE_KERNEL, (unsigned long)(strategy::output_rows() * strategy::output_cols() * this->m_args.channel_multiplier * this->m_args.kernel_rows * this->m_args.kernel_cols));
-#endif
-      strat.kernel(
-        inptrs, outptrs, weights,
-        m_qp.bias == nullptr ? nullptr : m_qp.bias + start_output_channel,
-        this->kernel_points(),
-        this->m_args.channel_multiplier,
-        m_qp.per_channel_left_shifts == nullptr ? nullptr : m_qp.per_channel_left_shifts + start_output_channel,
-        m_qp.per_channel_muls == nullptr ? nullptr : m_qp.per_channel_muls + start_output_channel,
-        m_qp.per_channel_right_shifts == nullptr ? nullptr : m_qp.per_channel_right_shifts + start_output_channel,
-        m_qp
-      );
-    };
-
-    Parent::execute_tiles(
-      tile_fn, m_qp.a_offset,
-      batches, input_height, input_width, input_channels, padding,
-      _input, ld_input_col, ld_input_row, ld_input_batch,
-      parameters,
-      output_height, output_width,
-      _output, ld_output_col, ld_output_row, ld_output_batch,
-      _working_space, thread_id, n_threads
-    );
-  }
-};
-
-}  // namespace depthwise
-}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_quantized.hpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_quantized.hpp
deleted file mode 100644
index cfb0d4bc05..0000000000
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_quantized.hpp
+++ /dev/null
@@ -1,125 +0,0 @@
-/*
- * 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.
- */
-
-#pragma once
-
-#include "depthwise_depthfirst_generic.hpp"
-
-#include "arm_gemm.hpp"
-#include "src/core/NEON/kernels/arm_gemm/utils.hpp"
-
-#ifdef CYCLE_PROFILING
-#include "profiler.hpp"
-#endif
-
-using arm_gemm::Requantize32;
-
-namespace arm_conv {
-namespace depthwise {
-
-template <class Strategy, unsigned OutputRows, unsigned int OutputCols>
-class DepthwiseDepthfirstGenericQuantized : public DepthwiseDepthfirstGenericBase<Strategy, OutputRows, OutputCols>
-{
-  using Parent = DepthwiseDepthfirstGenericBase<Strategy, OutputRows, OutputCols>;
-  using TInput = typename Parent::TInput;
-  using TAccum = typename Parent::TAccum;
-  using TOutput = typename Parent::TOutput;
-
-  Requantize32 m_qp;
-
-  public:
-  DepthwiseDepthfirstGenericQuantized(const DepthwiseArgs &args, const Requantize32 &qp)
-    : Parent(args), m_qp(qp)
-  {
-  }
-
-  DepthwiseDepthfirstGenericQuantized(DepthwiseDepthfirstGenericQuantized &) = delete;
-  DepthwiseDepthfirstGenericQuantized &operator=(DepthwiseDepthfirstGenericQuantized &) = delete;
-
-  void pack_parameters(void *buffer, const void *biases, const void *weights, size_t ld_weight_col, size_t ld_weight_row) override
-  {
-    m_qp.bias = static_cast<const TAccum *>(biases);
-    Parent::pack_parameters(buffer, biases, weights, ld_weight_col, ld_weight_row);
-  }
-
-  using DepthwiseDepthfirstGenericBase<Strategy, OutputRows, OutputCols>::execute;
-  void execute(
-    const unsigned int batches,
-    const unsigned int input_height,
-    const unsigned int input_width,
-    const unsigned int input_channels,
-    const PaddingValues &padding,
-    const void *const _input,
-    const size_t ld_input_col,
-    const size_t ld_input_row,
-    const size_t ld_input_batch,
-    const void *const parameters,
-    const unsigned int output_height,
-    const unsigned int output_width,
-    void *const _output,
-    const size_t ld_output_col,
-    const size_t ld_output_row,
-    const size_t ld_output_batch,
-    void *const _working_space,
-    const unsigned int thread_id,
-    const unsigned int n_threads
-  ) const override
-  {
-    Strategy strat(this->m_args.cpu_info);
-#ifdef CYCLE_PROFILING
-    arm_gemm::profiler prof;
-#endif
-
-    // Create a function to initialise the input buffer
-    const auto initialise_input_buffer = [this] (TInput *const buffer, const unsigned int n) {
-      std::memset(buffer, static_cast<TInput>(m_qp.a_offset), n * sizeof(TInput));
-    };
-
-    // Create a function to execute a tile of work
-    const auto tile_fn = [&] (const TInput *const *const inptrs, TOutput *const * const outptrs) {
-#ifdef CYCLE_PROFILING
-      auto p = prof.ScopedProfiler(
-        PROFILE_KERNEL,
-        (unsigned long) (OutputRows * OutputCols * this->m_args.kernel_rows* this->m_args.kernel_cols)
-      );
-#endif
-      strat.kernel(inptrs, outptrs, parameters, m_qp,
-                   this->m_args.kernel_rows * this->m_args.kernel_cols,
-                   this->m_args.input_channels);
-    };
-
-    // Call into a parent utility function to do the actual work.
-    Parent::execute_tiles(
-      tile_fn, initialise_input_buffer,
-      batches, input_height, input_width, input_channels, padding,
-      _input, ld_input_col, ld_input_row, ld_input_batch,
-      output_height, output_width,
-      _output, ld_output_col, ld_output_row, ld_output_batch,
-      _working_space, thread_id, n_threads
-    );
-  }
-};
-
-}  // namespace depthwise
-}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_multiplier_quantized.hpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_multiplier_quantized.hpp
deleted file mode 100644
index 07ce0d3b55..0000000000
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_multiplier_quantized.hpp
+++ /dev/null
@@ -1,251 +0,0 @@
-/*
- * 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.
- */
-
-#pragma once
-
-#include "depthwise_depthfirst_multiplier.hpp"
-
-namespace arm_conv {
-namespace depthwise {
-
-template <class strategy>
-class DepthwiseDepthfirstWithMultiplierQuantized :
-  public DepthwiseCommon<typename strategy::input_type,
-                         typename strategy::weight_type,
-                         typename strategy::return_type>
-{
-  using Parent = DepthwiseCommon<typename strategy::input_type,
-                                 typename strategy::weight_type,
-                                 typename strategy::return_type>;
-  using TInput = typename strategy::input_type;
-  using TWeight = typename strategy::weight_type;
-  using TOutput = typename strategy::return_type;
-
-  const arm_gemm::Requantize32 m_qp;
-
-  size_t sizeof_output_buffer(unsigned int n_channels) const
-  {
-    const unsigned int vl = arm_gemm::utils::get_vector_length<typename strategy::return_type>(strategy::vl_type);
-    const auto rounded_channels = arm_gemm::roundup(n_channels, vl);
-    return sizeof(typename strategy::return_type) * rounded_channels;
-  }
-
-  public:
-  DepthwiseDepthfirstWithMultiplierQuantized(const DepthwiseArgs &args, const arm_gemm::Requantize32 &qp)
-    : Parent(args), m_qp(qp)
-  {
-  }
-
-  DepthwiseDepthfirstWithMultiplierQuantized(DepthwiseDepthfirstWithMultiplierQuantized &) = delete;
-  DepthwiseDepthfirstWithMultiplierQuantized &operator=(DepthwiseDepthfirstWithMultiplierQuantized &) = delete;
-
-  size_t get_storage_size(void) const override
-  {
-    // We produce VL<int32_t> channels at a time, for each of these blocks of
-    // channels we store a vector of biases, weights (complicated) and
-    // requantize parameters.
-    const unsigned int iter_length =
-      arm_gemm::utils::get_vector_length<int32_t>(strategy::vl_type);
-    const unsigned int n_iters =
-      this->m_args.input_channels * arm_gemm::iceildiv(this->m_args.channel_multiplier, iter_length);
-
-    // Compute the cost of storing the weights
-    const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(strategy::kernel_cols, 4u);
-
-    return n_iters * iter_length * (
-        sizeof(int32_t) +  // Bias
-        4 * n_dots_per_kernel_row * strategy::kernel_rows * sizeof(TWeight) +  // Weights
-        2 * sizeof(int32_t)  // Requantisation parameters
-    );
-  }
-
-  // We'll want an optimised version of this, but for now a C++ implementation
-  // is probably sufficient.
-  void pack_parameters(void *_buffer, const void *_biases, const void *_weights, size_t ld_weight_col, size_t ld_weight_row) override
-  {
-    auto buffer = static_cast<uint8_t *>(_buffer);
-    auto biases = static_cast<const int32_t *>(_biases);
-    auto weights = static_cast<const TWeight *>(_weights);
-    auto requant_muls = m_qp.per_channel_muls;
-    auto requant_shifts = m_qp.per_channel_right_shifts;
-
-    const unsigned int iter_length =
-      arm_gemm::utils::get_vector_length<int32_t>(strategy::vl_type);
-    const unsigned int n_iters_per_input_channel =
-      arm_gemm::iceildiv(this->m_args.channel_multiplier, iter_length);
-
-    const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(strategy::kernel_cols, 4u);
-
-    const size_t iter_stride = iter_length * (
-        sizeof(int32_t) +  // Bias
-        4 * n_dots_per_kernel_row * strategy::kernel_rows * sizeof(int8_t) +  // Weights
-        2 * sizeof(int32_t)  // Requantisation parameters
-    );
-
-    ld_weight_col = (ld_weight_col == 0) ? this->m_args.input_channels * this->m_args.channel_multiplier : ld_weight_col;
-    ld_weight_row = (ld_weight_row == 0) ? this->m_args.kernel_cols * ld_weight_col : ld_weight_row;
-
-    for (unsigned int input_channel = 0; input_channel < this->m_args.input_channels; input_channel++)
-    {
-      auto buffer_input_channel = buffer + input_channel * n_iters_per_input_channel * iter_stride;
-      auto weights_input_channel = weights + input_channel * this->m_args.channel_multiplier;
-
-      for (unsigned int iter = 0; iter < n_iters_per_input_channel; iter++)
-      {
-        // Get a pointer to the start of this portion of the buffer; consequently
-        // derive pointers to the bias, weight and requantisation portions of
-        // this frame.
-        auto buffer_base = buffer_input_channel + iter_stride * iter;
-        auto buffer_biases = reinterpret_cast<int32_t *>(buffer_base);
-        auto buffer_weights = buffer_base + sizeof(int32_t) * iter_length;
-        auto buffer_requant_mul = reinterpret_cast<int32_t *>(
-          buffer_weights + strategy::kernel_rows * n_dots_per_kernel_row * 4 * iter_length);
-        auto buffer_requant_shift = buffer_requant_mul + iter_length;
-        auto weights_base = weights_input_channel + iter * iter_length;
-
-        // Hence work through the data for this iteration, on a
-        // channel-by-channel basis.
-        const auto this_iter_length = std::min<unsigned int>(
-          iter_length, this->m_args.channel_multiplier - iter * iter_length
-        );
-        for (unsigned int i = 0; i < this_iter_length; i++)
-        {
-          auto weights_channel = weights_base + i;
-
-          // Read the bias value, we modify this as we read the weights.
-          auto bias_value = biases == nullptr ? 0 : *(biases++);
-          int32_t elements_sum = 0;
-
-          // Read through the kernel; for each row, marshal together as many dot
-          // product terms as are required.
-          for (unsigned int ki = 0; ki < strategy::kernel_rows; ki++)
-          {
-            auto buffer_row = buffer_weights + i*4 + ki * 4 * n_dots_per_kernel_row * iter_length;
-            auto weights_row = weights_channel + ki * ld_weight_row;
-
-            unsigned int kj = 0;
-            for (; kj < strategy::kernel_cols; kj++)
-            {
-              // Determine which element to which we're writing
-              const auto dot = kj / 4;
-              const auto elem = kj % 4;
-
-              // Copy the value; include in the sum
-              const auto val = weights_row[kj * ld_weight_col];
-              buffer_row[dot * 4 * iter_length + elem] = val;
-              elements_sum += val;
-            }
-            for (; kj < 4 * n_dots_per_kernel_row; kj++)
-            {
-              const auto dot = kj / 4;
-              const auto elem = kj % 4;
-              buffer_row[dot * 4 * iter_length + elem] = 0;
-            }
-
-            buffer_row += 4 * n_dots_per_kernel_row * iter_length;
-          }
-
-          // Write back the bias and offset values
-          *(buffer_biases++) =
-            bias_value - m_qp.a_offset * elements_sum +
-            strategy::kernel_rows * strategy::kernel_cols * m_qp.a_offset * m_qp.b_offset;
-
-          // Write out the requantisation parameters
-          *(buffer_requant_mul++) = m_qp.per_channel_requant ? *(requant_muls++) : m_qp.per_layer_mul;
-          *(buffer_requant_shift++) = m_qp.per_channel_requant ? *(requant_shifts++) : m_qp.per_layer_right_shift;
-        }
-      }
-    }
-  }
-
-  size_t get_working_size(const unsigned int n_threads, const unsigned int n_channels) const override
-  {
-    const unsigned int n_output_channels = n_channels * this->m_args.channel_multiplier;
-    return n_threads * sizeof_output_buffer(n_output_channels);
-  }
-
-  using Parent::execute;
-  void execute(
-    const unsigned int batches,
-    const unsigned int input_height,
-    const unsigned int input_width,
-    const unsigned int input_channels,
-    const PaddingValues &padding,
-    const void *const _input,
-    const size_t ld_input_col,
-    const size_t ld_input_row,
-    const size_t ld_input_batch,
-    const void *const parameters,
-    const unsigned int output_height,
-    const unsigned int output_width,
-    void *const _output,
-    const size_t ld_output_col,
-    const size_t ld_output_row,
-    const size_t ld_output_batch,
-    void *const _working_space,
-    const unsigned int thread_id,
-    const unsigned int n_threads
-  ) const override
-  {
-    strategy strat(this->m_args.cpu_info);
-#ifdef CYCLE_PROFILING
-    arm_gemm::profiler prof;
-#endif
-
-    auto executefn = [strat, this] (
-      const TInput *const *const inptrs,
-      TOutput *const *const outptr_array,
-      const void *const params
-    ) {
-      strat.kernel(inptrs, outptr_array, params, this->m_args.channel_multiplier, m_qp);
-    };
-
-    // Get working space for this thread
-    uint8_t *const working_space = static_cast<uint8_t *>(_working_space) + get_working_size(1, input_channels) * thread_id;
-
-    // Determine the stride across blocks of parameters
-    const unsigned int iter_length =
-      arm_gemm::utils::get_vector_length<int32_t>(strategy::vl_type);
-    const unsigned int n_iters_per_input_channel = arm_gemm::iceildiv(this->m_args.channel_multiplier, iter_length);
-    const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(strategy::kernel_cols, 4u);
-    const size_t param_stride = n_iters_per_input_channel * iter_length * (
-        sizeof(int32_t) +  // Bias
-        4 * n_dots_per_kernel_row * strategy::kernel_rows * sizeof(int8_t) +  // Weights
-        2 * sizeof(int32_t)  // Requantisation parameters
-    );
-
-    common::depthwise_multiplier_execute<strategy>(
-      executefn, m_qp.a_offset, this->m_args,
-      batches, input_height, input_width, input_channels, padding,
-      _input, ld_input_col, ld_input_row, ld_input_batch,
-      parameters, param_stride,
-      output_height, output_width,
-      _output, ld_output_col, ld_output_row, ld_output_batch,
-      working_space, thread_id, n_threads
-    );
-  }
-};
-
-}  // namespace depthwise
-}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_quantized.hpp b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_quantized.hpp
deleted file mode 100644
index f97569e958..0000000000
--- a/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_quantized.hpp
+++ /dev/null
@@ -1,412 +0,0 @@
-/*
- * 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.
- */
-
-#pragma once
-
-#include "src/core/NEON/kernels/arm_gemm/utils.hpp"
-
-#ifdef CYCLE_PROFILING
-#include "profiler.hpp"
-#endif
-
-namespace arm_conv {
-namespace depthwise {
-
-namespace
-{
-
-// We have two sets of quantized kernels; those which use the dot-product
-// instructions and which require the biases and quantisation parameters to be
-// ravelled into weights/parameter array, and those which use the MLAL
-// instructions and which consume separate bias and quantisation parameter
-// arrays. The following code adapts these two sets of kernels to use the same
-// API - allowing the same driver loop to call them both.
-
-template <typename TIn, typename TWeight, typename TOut>
-using UnravelledKernFn = std::function<void(unsigned int, const TIn *const *, const TWeight *, const int32_t *, const arm_gemm::Requantize32 &, const int32_t *, const int32_t *, TOut *const *)>;
-
-template <typename TIn, typename TOut>
-using RavelledKernFn = std::function<void(const TIn *const *, TOut *const *, const void *, uint64_t, const arm_gemm::Requantize32 &)>;
-
-template <typename TIn, typename TWeight, typename TOut>
-const UnravelledKernFn<TIn, TWeight, TOut> get_unified_kernel(const UnravelledKernFn<TIn, TWeight, TOut> &f) { return f; }
-
-template <typename TIn, typename TWeight, typename TOut>
-const UnravelledKernFn<TIn, TWeight, TOut> get_unified_kernel(const RavelledKernFn<TIn, TOut> &f)
-{
-  return [f] (const unsigned int n_channels,
-              const TIn *const *const inptrs,
-              const TWeight *const weights,
-              const int32_t *,  // Bias (ravelled)
-              const arm_gemm::Requantize32 &qp,
-              const int32_t *,  // Requantisation muls (ravelled)
-              const int32_t *,  // Requantisation shifts (ravelled)
-              TOut *const *const outptrs) {
-    return f(inptrs, outptrs, weights, n_channels, qp);
-  };
-}
-
-template <typename T>
-using UnravelledPackingFn = std::function<void(unsigned int, void *, const T *, size_t, size_t)>;
-
-template <typename T>
-using RavelledPackingFn = std::function<void(unsigned int, void *, const int32_t *, const T *, const arm_gemm::Requantize32 &, size_t, size_t)>;
-
-template <typename T>
-const RavelledPackingFn<T> get_unified_packer(const UnravelledPackingFn<T> &f)
-{
-  return [f] (const unsigned int n_channels,
-              void *buffer,
-              const int32_t *,  // Bias
-              const T *weights,
-              const arm_gemm::Requantize32 &,
-              size_t ld_weight_col,
-              size_t ld_weight_row)
-  {
-    return f(n_channels, buffer, weights, ld_weight_col, ld_weight_row);
-  };
-}
-
-template <typename T>
-const RavelledPackingFn<T> get_unified_packer(const RavelledPackingFn<T> &f) { return f; }
-
-template <typename T>
-constexpr bool requires_unravelled_bias_and_quant_params(const UnravelledPackingFn<T> &) { return true; }
-
-template <typename T>
-constexpr bool requires_unravelled_bias_and_quant_params(const RavelledPackingFn<T> &) { return false; }
-
-template <class strategy>
-constexpr bool strategy_requires_unravelled_bias_and_quant_params(void)
-{
-  return requires_unravelled_bias_and_quant_params<typename strategy::weight_type>(strategy::pack_parameters);
-}
-
-}
-
-template <class strategy>
-class DepthwiseDepthfirstQuantized :
-  public DepthwiseCommon<typename strategy::input_type,
-                         typename strategy::weight_type,
-                         typename strategy::return_type>
-{
-  using TInput = typename strategy::input_type;
-  using TWeight = typename strategy::weight_type;
-  using TOutput = typename strategy::return_type;
-  using TAccum = typename strategy::bias_type;
-
-  arm_gemm::Requantize32 m_qp;
-
-  size_t sizeof_input_buffer(unsigned int n_channels) const
-  {
-    const unsigned int vl = arm_gemm::utils::get_vector_length<TInput>(strategy::vl_type);
-    const auto rounded_channels = arm_gemm::roundup(n_channels, vl);
-    return sizeof(TInput) * rounded_channels;
-  }
-
-  size_t sizeof_output_buffer(unsigned int n_channels) const
-  {
-    const unsigned int vl = arm_gemm::utils::get_vector_length<TOutput>(strategy::vl_type);
-    const auto rounded_channels = arm_gemm::roundup(n_channels, vl);
-    return sizeof(TOutput) * rounded_channels;
-  }
-
-  size_t sizeof_bias_buffer(unsigned int n_channels) const
-  {
-    if (strategy_requires_unravelled_bias_and_quant_params<strategy>())
-    {
-      return (m_qp.bias == nullptr) ? sizeof(TAccum) * n_channels : 0;
-    }
-
-    return 0;
-  }
-
-  size_t sizeof_requant_mul_buffer(unsigned int n_channels) const
-  {
-    if (strategy_requires_unravelled_bias_and_quant_params<strategy>())
-    {
-      return m_qp.per_channel_requant ? 0 : sizeof(int32_t) * n_channels;
-    }
-
-    return 0;
-  }
-
-  size_t sizeof_requant_shift_buffer(unsigned int n_channels) const
-  {
-    if (strategy_requires_unravelled_bias_and_quant_params<strategy>())
-    {
-      return m_qp.per_channel_requant ? 0 : sizeof(int32_t) * n_channels;
-    }
-
-    return 0;
-  }
-
-  public:
-  DepthwiseDepthfirstQuantized(const DepthwiseArgs &args, const arm_gemm::Requantize32 &qp)
-    : DepthwiseCommon<TInput, TWeight, TOutput>(args), m_qp(qp)
-  {
-  }
-
-  DepthwiseDepthfirstQuantized(DepthwiseDepthfirstQuantized &) = delete;
-  DepthwiseDepthfirstQuantized &operator=(DepthwiseDepthfirstQuantized &) = delete;
-
-  size_t get_storage_size(void) const override
-  {
-    return strategy::get_packed_size(this->m_args);
-  }
-
-  void pack_parameters(void *buffer, const void *const bias, const void *weights, size_t ld_weight_col, size_t ld_weight_row) override
-  {
-    if (strategy_requires_unravelled_bias_and_quant_params<strategy>())
-    {
-      m_qp.bias = static_cast<const int32_t *>(bias);
-    }
-
-    get_unified_packer<TWeight>(strategy::pack_parameters)(
-      this->m_args.input_channels,
-      buffer,
-      static_cast<const int32_t *>(bias),
-      reinterpret_cast<const TWeight *>(weights),
-      m_qp,
-      ld_weight_col,
-      ld_weight_row
-    );
-  }
-
-  size_t get_working_size(const unsigned int n_threads, const unsigned int n_channels) const override
-  {
-    const unsigned int n_output_channels = n_channels * this->m_args.channel_multiplier;
-    return n_threads * (
-      sizeof_output_buffer(n_output_channels) +
-      sizeof_input_buffer(n_channels) +
-      sizeof_bias_buffer(n_channels) +
-      sizeof_requant_mul_buffer(n_channels) +
-      sizeof_requant_shift_buffer(n_channels)
-    );
-  }
-
-  using DepthwiseCommon<typename strategy::input_type, typename strategy::weight_type, typename strategy::return_type>::execute;
-  void execute(
-    const unsigned int batches,
-    const unsigned int input_height,
-    const unsigned int input_width,
-    const unsigned int input_channels,
-    const PaddingValues &padding,
-    const void *const _input,
-    const size_t ld_input_col,
-    const size_t ld_input_row,
-    const size_t ld_input_batch,
-    const void *const parameters,
-    const unsigned int output_height,
-    const unsigned int output_width,
-    void *const _output,
-    const size_t ld_output_col,
-    const size_t ld_output_row,
-    const size_t ld_output_batch,
-    void *_working_space,
-    const unsigned int thread_id,
-    const unsigned int n_threads
-  ) const override
-  {
-    strategy strat(this->m_args.cpu_info);
-#ifdef CYCLE_PROFILING
-    arm_gemm::profiler prof;
-#endif
-    // Get a unified API for the kernel function
-    auto kernel = get_unified_kernel<TInput, TWeight, TOutput>(strat.kernel);
-
-    // Determine what portion of the work to do.
-    const unsigned int n_rows_per_thread = arm_gemm::iceildiv(output_height, n_threads);
-    const int start_out_height = std::min(thread_id * n_rows_per_thread, output_height);
-    const int end_out_height = std::min(start_out_height + n_rows_per_thread, output_height);
-
-    // Cast input and output pointers into the right types
-    const TInput *const inptr = static_cast<const TInput *>(_input);
-    TOutput *const outptr = static_cast<TOutput *>(_output);
-
-    // Create an array for the input pointers
-    const TInput * _inptr_array[strategy::input_rows * strategy::input_cols];
-    const TInput **const inptr_array = _inptr_array;
-
-    // Create an array for the output pointers
-    TOutput * _outptr_array[strategy::output_rows * strategy::output_cols];
-    TOutput **const outptr_array = _outptr_array;
-
-    // Allocate portions of the working space
-    uint8_t *working_space = static_cast<uint8_t *>(_working_space) + get_working_size(thread_id, input_channels);
-
-    TOutput *const output_buffer = reinterpret_cast<TOutput *>(working_space);
-    working_space += sizeof_output_buffer(input_channels * this->m_args.channel_multiplier);
-
-    TInput *const input_buffer = reinterpret_cast<TInput *>(working_space);
-    working_space += sizeof_input_buffer(input_channels);
-
-    const int32_t *const bias_ptr = (m_qp.bias == nullptr) ? reinterpret_cast<int32_t *>(working_space)
-                                                           : m_qp.bias;
-    working_space += sizeof_bias_buffer(input_channels * this->m_args.channel_multiplier);
-
-    const int32_t *const requant_mul_vec = !m_qp.per_channel_requant ? reinterpret_cast<int32_t *>(working_space)
-                                                                     : m_qp.per_channel_muls;
-    working_space += sizeof_requant_mul_buffer(input_channels * this->m_args.channel_multiplier);
-
-    const int32_t *const requant_shift_vec = !m_qp.per_channel_requant ? reinterpret_cast<int32_t *>(working_space)
-                                                                       : m_qp.per_channel_right_shifts;
-
-    if (strategy_requires_unravelled_bias_and_quant_params<strategy>())
-    {
-      // Initialise the bias buffer
-      if (m_qp.bias == nullptr)
-      {
-        for (unsigned int c = 0; c < input_channels * this->m_args.channel_multiplier; c++)
-        {
-          const_cast<int32_t *>(bias_ptr)[c] = 0;
-        }
-      }
-
-      // Initialise the requantisation parameters
-      if (!m_qp.per_channel_requant)
-      {
-        for (unsigned int c = 0; c < input_channels * this->m_args.channel_multiplier; c++)
-        {
-          const_cast<int32_t *>(requant_mul_vec)[c] = m_qp.per_layer_mul;
-          const_cast<int32_t *>(requant_shift_vec)[c] = m_qp.per_layer_right_shift;
-        }
-      }
-    }
-
-    // Initialise the input buffer
-    for (unsigned int c = 0; c < input_channels; c++)
-    {
-      input_buffer[c] = static_cast<TInput>(m_qp.a_offset);
-    }
-
-    // For each output tile, construct the requisite set of pointers and call
-    // into the kernel.
-    for (unsigned int batch = 0; batch < batches; batch++)
-    {
-      // Get batch pointers
-      const auto inptr_batch = inptr + batch * ld_input_batch;
-      const auto outptr_batch = outptr + batch * ld_output_batch;
-
-      for (int start_out_i = start_out_height;
-           start_out_i < end_out_height;
-           start_out_i += static_cast<int>(strategy::output_rows))
-      {
-        const int end_out_i = start_out_i + strategy::output_rows;
-        const int start_in_i = start_out_i * strategy::stride_rows - padding.top;
-        const int end_in_i = start_in_i + strategy::input_rows;
-
-        // Compute top/bottom padding
-        const auto pad_top = static_cast<unsigned int>(-std::min(start_in_i, 0));
-        const auto pad_bottom = static_cast<unsigned int>(-std::min(static_cast<int>(input_height) - end_in_i, 0));
-        const unsigned int valid_output_rows = std::min(
-          end_out_i - start_out_i,
-          static_cast<int>(output_height) - start_out_i
-        );
-
-        // Fill the input pointer array with padding values
-        for (auto index = 0u; index < strategy::input_rows * strategy::input_cols; index++)
-        {
-          inptr_array[index] = input_buffer;
-        }
-
-        for (int start_out_j = 0; start_out_j < static_cast<int>(output_width);)
-        {
-          const int start_in_j = start_out_j * strategy::stride_cols - this->m_args.padding.left;
-          const int pad_left = -std::min(0, start_in_j);
-
-          const int end_out_j = start_out_j + strategy::output_cols;
-          const int end_in_j = start_in_j + strategy::input_cols;
-
-          const auto pad_right = static_cast<unsigned int>(-std::min(static_cast<int>(input_width) - end_in_j, 0));
-          const unsigned int valid_output_cols = std::min(
-            end_out_j - start_out_j,
-            static_cast<int>(output_width) - start_out_j
-          );
-
-          // Construct the input pointer array - fill the array with pointers to
-          // the input buffer and then fill in the required values.
-          for (auto i = pad_top; i < strategy::input_rows - pad_bottom; i++)
-          {
-            // Can skip over the left padding because we will have either the
-            // same or less than the previous tile.
-            unsigned int j = pad_left;
-            const TInput *colptr = inptr_batch + (start_in_i + i) * ld_input_row + (start_in_j + j) * ld_input_col;
-            const TInput **ptrs = inptr_array + i * strategy::input_cols + j;
-            for (; j < strategy::input_cols - pad_right; j++)
-            {
-              *(ptrs++) = colptr;
-              colptr += ld_input_col;
-            }
-            for (; j < strategy::input_cols; j++)
-            {
-              *(ptrs++) = input_buffer;
-            }
-          }
-
-          // Construct the output pointer array.
-          TOutput **outptr_pos = outptr_array;
-          for (auto i = 0u; i < valid_output_rows; i++)
-          {
-            unsigned int j = 0u;
-            TOutput *colptr = outptr_batch + (start_out_i + i) * ld_output_row + start_out_j * ld_output_col;
-            for (; j < valid_output_cols; j++)
-            {
-              *(outptr_pos++) = colptr;
-               colptr += ld_output_col;
-            }
-            for (; j < strategy::output_cols; j++)
-            {
-              *(outptr_pos++) = output_buffer;
-            }
-          }
-          for (auto i = valid_output_rows; i < strategy::output_rows; i++)
-          {
-            for (auto j = 0u; j < strategy::output_cols; j++)
-            {
-              *(outptr_pos++) = output_buffer;
-            }
-          }
-
-          start_out_j += strategy::output_cols;
-
-#ifdef CYCLE_PROFILING
-          // TODO Work number
-          auto p = prof.ScopedProfiler(PROFILE_KERNEL, (unsigned long)(strategy::output_rows * strategy::output_cols * this->m_args.kernel_rows * this->m_args.kernel_cols));
-#endif
-          kernel(
-            this->m_args.input_channels,
-            inptr_array,
-            reinterpret_cast<const TWeight *>(parameters),
-            bias_ptr, m_qp, requant_mul_vec, requant_shift_vec,
-            outptr_array
-          );
-        }
-      }
-    }
-  }
-};
-
-}  // namespace depthwise
-}  // namespace arm_conv