Update Neon™ depthwise kernel

- Reduce duplication and simplify overall structure.
- Improve multi-threaded performance by sharing more data
  in lower-level caches.

Partially Resolves: COMPMID-5054
Signed-off-by: Ramy Elgammal <ramy.elgammal@arm.com>
Change-Id: Iac747f39b21c540122fa75218762631c4d787911
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/7449
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Andrew Mundy
Reviewed-by: Sheri Zhang <sheri.zhang@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

5 files changed, 457 insertions, 73 deletions

diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp
new file mode 100644
index 0000000000..056f08d037
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.cpp
@@ -0,0 +1,150 @@
+/*
+ * Copyright (c) 2022 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 "generic.hpp"
+
+#include <functional>
+
+namespace arm_conv {
+namespace depthwise {
+namespace interleaves {
+
+PackingArguments::PackingArguments(
+  unsigned int kernel_rows, unsigned int kernel_cols, size_t weight_element_size,
+  bool include_bias, size_t bias_element_size,
+  arm_gemm::VLType vl_type, size_t accumulator_element_size, unsigned int accumulator_depth_vl,
+  std::function<bool(unsigned int, unsigned int &, unsigned int &)> get_weight_pos
+) : kernel_rows(kernel_rows), kernel_cols(kernel_cols), weight_element_size(weight_element_size),
+    include_bias(include_bias), bias_element_size(bias_element_size),
+    vl_type(vl_type), accumulator_element_size(accumulator_element_size), accumulator_depth_vl(accumulator_depth_vl),
+    get_weight_pos(get_weight_pos)
+{
+}
+
+size_t get_storage_size_generic(const PackingArguments &packing_args, const DepthwiseArgs &args)
+{
+  // If the channel multiplier is greater than one, then we treat this as a
+  // repeated packing of `channel_multiplier`-sized problems.
+  if (args.channel_multiplier > 1)
+  {
+    DepthwiseArgs args_per_input_channel(args);
+    args_per_input_channel.input_channels = args.channel_multiplier;
+    args_per_input_channel.channel_multiplier = 1;
+
+    return args.input_channels * get_storage_size_generic(packing_args, args_per_input_channel);
+  }
+
+  const unsigned int vl =
+    packing_args.accumulator_depth_vl *
+    arm_gemm::utils::get_vector_length<uint8_t>(packing_args.vl_type) / packing_args.accumulator_element_size;
+  const unsigned int n_packs = arm_gemm::iceildiv(args.input_channels, vl);
+  const auto pack_size = (packing_args.include_bias ? packing_args.bias_element_size : 0) +
+                         packing_args.kernel_points() * packing_args.weight_element_size;
+  return n_packs * pack_size * vl;
+}
+
+void pack_parameters_generic(
+  const PackingArguments &packing_args,
+  const DepthwiseArgs &args,
+  void *buffer_raw,
+  const void *biases_raw,
+  const void *weights_raw,
+  size_t ld_weight_col,
+  size_t ld_weight_row
+)
+{
+  // Cast the pointers to byte sizes
+  auto *buffer = static_cast<uint8_t *>(buffer_raw);
+  auto *biases = static_cast<const uint8_t *>(biases_raw);
+  auto *weights = static_cast<const uint8_t *>(weights_raw);
+
+  // If the channel multiplier is greater than one, then we treat this as a
+  // repeated packing of `channel_multiplier`-sized problems.
+  if (args.channel_multiplier > 1)
+  {
+    // Get a modified copy of the depthwise arguments
+    DepthwiseArgs args_per_input_channel(args);
+    args_per_input_channel.input_channels = args.channel_multiplier;
+    args_per_input_channel.channel_multiplier = 1;
+
+    // Resolve the strides here
+    ld_weight_col = ld_weight_col ? ld_weight_col : args.input_channels * args.channel_multiplier;
+    ld_weight_row = ld_weight_row ? ld_weight_row : ld_weight_col * packing_args.kernel_cols;
+
+    auto per_input_channel_size = get_storage_size_generic(packing_args, args_per_input_channel);
+
+    for (unsigned int c = 0; c < args.input_channels; c++)
+    {
+      pack_parameters_generic(
+        packing_args, args_per_input_channel, buffer, biases, weights, ld_weight_col, ld_weight_row);
+
+      // Update the pointers
+      buffer += per_input_channel_size;
+      biases += (biases == nullptr) ? 0 : packing_args.bias_element_size * args.channel_multiplier;
+      weights += packing_args.weight_element_size * args.channel_multiplier;
+    }
+    return;
+  }
+
+  // Finalise the weight strides
+  ld_weight_col = (ld_weight_col == 0) ? args.input_channels : ld_weight_col;
+  ld_weight_row = (ld_weight_row == 0) ? packing_args.kernel_cols * ld_weight_col : ld_weight_row;
+
+  const unsigned int vl =
+    packing_args.accumulator_depth_vl *
+    arm_gemm::utils::get_vector_length<uint8_t>(packing_args.vl_type) / packing_args.accumulator_element_size;
+
+  for (unsigned int n = 0; n < args.input_channels; n += vl)
+  {
+    const unsigned int todo = std::min(vl, args.input_channels - n);
+
+    if (packing_args.include_bias)
+    {
+      if (biases != nullptr)
+      {
+        memcpy(buffer, biases, todo * packing_args.bias_element_size);
+        biases += todo * packing_args.bias_element_size;
+      }
+      else
+      {
+        memset(buffer, 0, vl * packing_args.bias_element_size);
+      }
+
+      buffer += vl * packing_args.bias_element_size;
+    }
+
+    // Copy each of the weights in turn
+    unsigned int kx, ky;
+    for (int kindex = 0; packing_args.get_weight_pos(kindex, kx, ky); kindex++)
+    {
+      const auto src_ptr = weights + (kx*ld_weight_row + ky*ld_weight_col + n) * packing_args.weight_element_size;
+      memcpy(buffer, src_ptr, todo * packing_args.weight_element_size);
+      buffer += vl * packing_args.weight_element_size;
+    }
+  }
+}
+
+}  // namespace interleaves
+}  // namespace depthwise
+}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.hpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.hpp
new file mode 100644
index 0000000000..5b5ae17806
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic.hpp
@@ -0,0 +1,80 @@
+/*
+ * Copyright (c) 2022 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"
+#include "depthwise.hpp"
+
+#include <functional>
+
+namespace arm_conv {
+namespace depthwise {
+namespace interleaves {
+
+struct PackingArguments
+{
+  const unsigned int kernel_rows;
+  const unsigned int kernel_cols;
+  const size_t weight_element_size;
+  const bool include_bias;
+  const size_t bias_element_size;
+  arm_gemm::VLType vl_type;
+  const size_t accumulator_element_size;
+  const unsigned int accumulator_depth_vl;
+  std::function<bool(unsigned int, unsigned int &, unsigned int &)> get_weight_pos;
+
+  unsigned int kernel_points(void) const { return kernel_cols * kernel_rows; }
+
+  PackingArguments(
+    unsigned int kernel_rows,
+    unsigned int kernel_cols,
+    size_t weight_element_size,
+    bool include_bias,
+    size_t bias_element_size,
+    arm_gemm::VLType vl_type,
+    size_t accumulator_element_size,
+    unsigned int accumulator_depth_vl,
+    std::function<bool(unsigned int, unsigned int &, unsigned int &)> get_weight_pos
+  );
+};
+
+size_t get_storage_size_generic(
+  const PackingArguments &packing_args,
+  const DepthwiseArgs &args
+);
+
+void pack_parameters_generic(
+  const PackingArguments &packing_args,
+  const DepthwiseArgs &args,
+  void *buffer_raw,
+  const void *biases_raw,
+  const void *weights_raw,
+  size_t ld_weight_col,
+  size_t ld_weight_row
+);
+
+}  // namespace interleaves
+}  // namespace depthwise
+}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp
new file mode 100644
index 0000000000..a6389054d1
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.cpp
@@ -0,0 +1,161 @@
+/*
+ * Copyright (c) 2022 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 "generic_quantized_dot_product.hpp"
+#include <cstdint>
+
+namespace arm_conv {
+namespace depthwise {
+namespace interleaves {
+namespace quantized {
+
+size_t get_storage_size(
+  const DepthwiseArgs &args,
+  const arm_gemm::VLType vl_type,
+  const unsigned int accumulator_depth_vl
+)
+{
+  // 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 = accumulator_depth_vl * arm_gemm::utils::get_vector_length<int32_t>(vl_type);
+  const unsigned int n_iters = args.input_channels * arm_gemm::iceildiv(args.channel_multiplier, iter_length);
+
+  // Compute the cost of storing the weights
+  const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(args.kernel_cols, 4u);
+
+  return n_iters * iter_length * (
+    sizeof(int32_t) +  // Bias
+    4 * n_dots_per_kernel_row * args.kernel_rows * sizeof(int8_t) +  // Weights
+    2 * sizeof(int32_t)  // Requantisation parameters
+  );
+}
+
+template <typename T>
+void pack_parameters(
+  void *_buffer, const int32_t *biases,
+  const T *weights, size_t ld_weight_col, size_t ld_weight_row,
+  const DepthwiseArgs &args,
+  const arm_gemm::Requantize32 &qp,
+  const arm_gemm::VLType vl_type,
+  const unsigned int accumulator_depth_vl
+)
+{
+  auto buffer = static_cast<uint8_t *>(_buffer);
+  auto requant_muls = qp.per_channel_muls;
+  auto requant_shifts = qp.per_channel_right_shifts;
+
+  const unsigned int iter_length = accumulator_depth_vl * arm_gemm::utils::get_vector_length<int32_t>(vl_type);
+  const unsigned int n_iters_per_input_channel = arm_gemm::iceildiv(args.channel_multiplier, iter_length);
+  const unsigned int n_dots_per_kernel_row = arm_gemm::iceildiv(args.kernel_cols, 4u);
+
+  const size_t iter_stride = iter_length * (
+      sizeof(int32_t) +  // Bias
+      4 * n_dots_per_kernel_row * args.kernel_rows * sizeof(T) +  // Weights
+      2 * sizeof(int32_t)  // Requantisation parameters
+  );
+
+  ld_weight_col = (ld_weight_col == 0) ? args.input_channels * args.channel_multiplier : ld_weight_col;
+  ld_weight_row = (ld_weight_row == 0) ? args.kernel_cols * ld_weight_col : ld_weight_row;
+
+  for (unsigned int input_channel = 0; input_channel < 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 * 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 + args.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, 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 < args.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 < args.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 - qp.a_offset * elements_sum +
+          args.kernel_rows * args.kernel_cols * qp.a_offset * qp.b_offset;
+
+        // Write out the requantisation parameters
+        *(buffer_requant_mul++) = qp.per_channel_requant ? *(requant_muls++) : qp.per_layer_mul;
+        *(buffer_requant_shift++) = qp.per_channel_requant ? *(requant_shifts++) : qp.per_layer_right_shift;
+      }
+    }
+  }
+}
+
+template void pack_parameters(void *, const int32_t *, const int8_t *, size_t, size_t, const DepthwiseArgs &, const arm_gemm::Requantize32 &, arm_gemm::VLType, unsigned int);
+template void pack_parameters(void *, const int32_t *, const uint8_t *, size_t, size_t, const DepthwiseArgs &, const arm_gemm::Requantize32 &, arm_gemm::VLType, unsigned int);
+
+}  // namespace quantized
+}  // namespace interleaves
+}  // namespace depthwise
+}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.hpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.hpp
new file mode 100644
index 0000000000..779d67d3f4
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/generic_quantized_dot_product.hpp
@@ -0,0 +1,53 @@
+/*
+ * Copyright (c) 2022 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 "generic.hpp"
+
+namespace arm_conv {
+namespace depthwise {
+namespace interleaves {
+namespace quantized {
+
+size_t get_storage_size(
+  const DepthwiseArgs &args,
+  arm_gemm::VLType vl_type,
+  unsigned int accumulator_depth_vl=1
+);
+
+template <typename T>
+void pack_parameters(
+  void *buffer, const int32_t *biases,
+  const T *weights, size_t ld_weight_col, size_t ld_weight_row,
+  const DepthwiseArgs &args,
+  const arm_gemm::Requantize32 &qp,
+  arm_gemm::VLType vl_type,
+  unsigned int accumulator_depth_vl
+);
+
+}  // namespace quantized
+}  // namespace interleaves
+}  // namespace depthwise
+}  // namespace arm_conv
diff --git a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/list.hpp b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/list.hpp
index cb49a243af..76f38eb335 100644
--- a/src/core/NEON/kernels/arm_conv/depthwise/interleaves/list.hpp
+++ b/src/core/NEON/kernels/arm_conv/depthwise/interleaves/list.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021 Arm Limited.
+ * Copyright (c) 2021-2022 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -29,90 +29,30 @@ namespace depthwise {
 
 #if defined(ARM_COMPUTE_ENABLE_SVE)
 
-class interleave_sve_u8q_3x3_dot
+struct interleave_sve_u8q_3x3_dot
 {
-  public:
-    static void pack_parameters(unsigned int, void *, const int32_t *, const uint8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
+  static void pack_parameters(unsigned int, void *, const int32_t *, const uint8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
+  static size_t get_packed_size(const DepthwiseArgs &);
 };
 
-class interleave_sve_s8q_3x3_dot
+struct interleave_sve_s8q_3x3_dot
 {
-  public:
-    static void pack_parameters(unsigned int, void *, const int32_t *, const int8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_sve_u8q_3x3_mla
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const uint8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_sve_s8q_3x3_mla
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const int8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_sve_u8q_5x5_mla
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const uint8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_sve_s8q_5x5_mla
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const int8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
+  static void pack_parameters(unsigned int, void *, const int32_t *, const int8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
+  static size_t get_packed_size(const DepthwiseArgs &);
 };
 
 #endif  // defined(ARM_COMPUTE_ENABLE_SVE)
 
-class interleave_a64_u8q_3x3_dot
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const int32_t *, const uint8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_a64_s8q_3x3_dot
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const int32_t *, const int8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_a64_u8q_3x3_mla
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const uint8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_a64_s8q_3x3_mla
-{
-  public:
-    static void pack_parameters(unsigned int, void *, const int8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
-};
-
-class interleave_a64_u8q_5x5_mla
+struct interleave_a64_u8q_3x3_dot
 {
-  public:
-    static void pack_parameters(unsigned int, void *, const uint8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
+  static void pack_parameters(unsigned int, void *, const int32_t *, const uint8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
+  static size_t get_packed_size(const DepthwiseArgs &);
 };
 
-class interleave_a64_s8q_5x5_mla
+struct interleave_a64_s8q_3x3_dot
 {
-  public:
-    static void pack_parameters(unsigned int, void *, const int8_t *, size_t, size_t);
-    static size_t get_packed_size(const DepthwiseArgs &);
+  static void pack_parameters(unsigned int, void *, const int32_t *, const int8_t *, const arm_gemm::Requantize32 &, size_t, size_t);
+  static size_t get_packed_size(const DepthwiseArgs &);
 };
 
 }  // namespace depthwise