Integrate improved CPU depthwise convolution kernels

* Replace assembly kernels for depthwise convolution with more optimized
  ones.
* Add int8 assembly kernels.
* Fix implicit padding on optimized kernels

Resolves: COMPMID-3867, COMPMID-4361

Change-Id: I0b0867e05f61be4f368f62190d55e14d0ab3ebf2
Signed-off-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5622
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

1 files changed, 480 insertions, 0 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
new file mode 100644
index 0000000000..656e4413b2
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_generic_multiplier.hpp
@@ -0,0 +1,480 @@
+/*
+ * 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 {
+
+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;
+    if (std::numeric_limits<TAccum>::is_integer)
+    {
+      activation_min = std::numeric_limits<TAccum>::min();
+      activation_max = std::numeric_limits<TAccum>::max();
+    }
+    else
+    {
+      activation_min = static_cast<TAccum>(-std::numeric_limits<float>::infinity());
+      activation_max = static_cast<TAccum>(std::numeric_limits<float>::infinity());
+    }
+
+    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