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, 251 insertions, 0 deletions

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
new file mode 100644
index 0000000000..07ce0d3b55
--- /dev/null
+++ b/src/core/NEON/kernels/arm_conv/depthwise/depthwise_depthfirst_multiplier_quantized.hpp
@@ -0,0 +1,251 @@
+/*
+ * 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