quantized_conv2d_kn2row.cpp

/* Copyright 2018 The Blueoil Authors. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include <cassert>
#include <cstdio>

#include "de10_nano.h"
#include "func/impl/quantized_conv2d_kn2row.h"
#include "global.h"
#include "network.h"
#include "pack_input_to_qwords.h"
#include "time_measurement.h"

namespace
{
const unsigned int in_nbits = 2;
const unsigned int byte_nbits = 8;
} // namespace

namespace dlk
{

namespace impl
{

void QuantizedConv2DKn2Row(const kn2row_input_t& input,
                                  const kernel_t& kernel,
                                  const binary_convolution_parameters &p) {
  using namespace dlk;

  convolution_parameters cp = p.normal_conv_params;
  const T_UINT out_c = cp.output_channels;

  const T_UINT num_qinput_per_qword = (NBIT_QDYPE / MAX_NBIT_QINPUT);
  const T_UINT num_qkernel_per_qword = (NBIT_QDYPE / MAX_NBIT_KERNEL);

  const T_UINT k_h = cp.kernel_height;
  const T_UINT k_w = cp.kernel_width;
  const T_UINT k_c = cp.kernel_depth;
  const T_UINT k_c_by_word =
      (k_c + (num_qkernel_per_qword - 1)) / num_qkernel_per_qword;
  const T_UINT k_n = out_c;

  const T_UINT in_h = cp.input_height;
  const T_UINT in_w = cp.input_width;
  const T_UINT in_c = k_c;
  const T_UINT in_c_by_word = k_c_by_word;
  const T_UINT in_size = in_h * in_w * in_c_by_word * MAX_NBIT_QINPUT;

  const T_UINT out_h = cp.output_height;
  const T_UINT out_w = cp.output_width;
  const T_UINT out_size = out_h * out_w * out_c;

  const bool out_c_less_than_num_pe = (out_c < NUM_PE);

#ifdef DLK_DEBUG
  const bool in_c_less_than_word_size = (in_c < WORD_SIZE);
  const bool in_c_excess_the_max = (in_c > MAX_IN_C);
  const bool ts_activated = (p.thresholds != nullptr);

  if (in_c_less_than_word_size)
  {
    std::printf("[WORNING] in_c(%u) less than word_size(%u)\n", in_c,
                WORD_SIZE);
  }
  else if (out_c_less_than_num_pe)
  {
    std::printf("[WORNING] out_c(%u) less than num_pe(%u)\n", out_c, NUM_PE);
  }
  else if (in_c_excess_the_max)
  {
    std::printf("[Fatal Error] in_c(%u) excess the max(%u)\n", in_c, MAX_IN_C);
  }
  else if (ts_activated)
  {
    std::printf("[Fatal Error] Now threshold skipping is not supported\n");
  }

  assert(!in_c_less_than_word_size);
  assert(!in_c_excess_the_max);
  assert(!ts_activated);
#endif

  std::copy(input.data(), input.data() + input.size(), p.device_input_buf);

  if (out_c_less_than_num_pe)
  {

    const T_UINT out_c_aligend_with_num_pe =
        ((k_n + (NUM_PE - 1)) / NUM_PE) * NUM_PE;
    T_UINT input_byte_size =
        (cp.input_height * cp.input_width * cp.kernel_depth * in_nbits) /
        byte_nbits;

    T_UINT output_byte_size_full =
        out_h * out_w * out_c * sizeof(BIN_CONV_OUTPUT);
    T_UINT output_byte_size_partial =
        out_h * out_w * out_c_aligend_with_num_pe * sizeof(BIN_CONV_OUTPUT);

    p.dma_input_buffer->sync_size(input_byte_size);
    p.dma_output_buffer->sync_size(output_byte_size_full);

    p.dma_input_buffer->sync_for_device();
    p.dma_output_buffer->sync_for_device();

    Measurement::Start("QConv2D kn2row tiling");
    de10_nano::qconv_kn2row_tiling(
        p.device_input_phys_addr, p.device_output_phys_addr, kernel.data(),
        p.thresholds, in_w, in_h, in_c_by_word, MAX_NBIT_QINPUT, out_w, out_h,
        out_c_aligend_with_num_pe, k_w, k_h, cp.padding,
        cp.stride_along_height);
    Measurement::Stop();

    p.dma_output_buffer->sync_size(output_byte_size_partial);
    p.dma_output_buffer->sync_for_cpu();

    unsigned idx_out_src = 0;
    unsigned idx_out_dst = 0;

    for (unsigned h = 0; h < out_h; h++)
      for (unsigned w = 0; w < out_w; w++)
        for (unsigned c = 0; c < out_c_aligend_with_num_pe; c++)
        {
          if (c < out_c)
          {
            p.device_output_buf[idx_out_dst++] =
                p.device_output_buf[idx_out_src];
          }
          idx_out_src++;
        }
  }
  else
  {
    T_UINT input_byte_size =
        (cp.input_height * cp.input_width * cp.kernel_depth * in_nbits) /
        byte_nbits;

    T_UINT output_byte_size = out_h * out_w * out_c * sizeof(BIN_CONV_OUTPUT);

    p.dma_input_buffer->sync_size(input_byte_size);
    p.dma_output_buffer->sync_size(output_byte_size);

    p.dma_input_buffer->sync_for_device();
    p.dma_output_buffer->sync_for_device();

    Measurement::Start("QConv2D kn2row tiling");
    de10_nano::qconv_kn2row_tiling(
        p.device_input_phys_addr, p.device_output_phys_addr, kernel.data(),
        p.thresholds, in_w, in_h, in_c_by_word, MAX_NBIT_QINPUT, out_w, out_h,
        out_c, k_w, k_h, cp.padding, cp.stride_along_height);
    Measurement::Stop();

    p.dma_output_buffer->sync_size(output_byte_size);
    p.dma_input_buffer->sync_size(input_byte_size);

    p.dma_input_buffer->sync_for_cpu();
    p.dma_output_buffer->sync_for_cpu();
  }
}


void TCAConv2d(const kn2row_input_t& input,
    const kernel_t& kernel,
    const binary_convolution_parameters &p) {

  using namespace dlk;

  convolution_parameters cp = p.normal_conv_params;
  const T_UINT b = 32;
  const T_UINT out_c = ((cp.output_channels + b - 1) / b) * b;

  const T_UINT num_qkernel_per_qword = (NBIT_QDYPE / MAX_NBIT_KERNEL);

  const T_UINT k_h = cp.kernel_height;
  const T_UINT k_w = cp.kernel_width;
  const T_UINT k_c = cp.kernel_depth;

  const T_UINT in_h = cp.input_height;
  const T_UINT in_w = cp.input_width;

  const T_UINT out_h = cp.output_height;
  const T_UINT out_w = cp.output_width;

  const auto effective_kernel_depth = ((cp.kernel_depth + b - 1) / b) * b;

    T_UINT input_byte_size =
        (cp.input_height * cp.input_width * effective_kernel_depth * in_nbits) /
        byte_nbits;

    T_UINT output_byte_size = out_h * out_w * out_c * sizeof(BIN_CONV_OUTPUT);
    if (p.thresholds != NULL) {
      output_byte_size /= 8;
    }

    Measurement::Start("Sync UDMABuf Input");
    p.dma_input_buffer->sync_size(input_byte_size);
    p.dma_input_buffer->sync_for_device();
    Measurement::Stop();

    Measurement::Start("Conv2D TCA");
    de10_nano::RunTCA(p.device_input_phys_addr, p.device_output_phys_addr, p.device_kernel_phys_addr, p.thresholds, in_w, in_h,
      k_c, MAX_NBIT_QINPUT, out_w, out_h, out_c, k_w, k_h, cp.padding, cp.stride_along_height);
    Measurement::Stop();

    Measurement::Start("Sync UDMABuf Output");
    p.dma_output_buffer->sync_size(output_byte_size);
    p.dma_output_buffer->sync_for_cpu();
    Measurement::Stop();
}


} // namespace impl

} // namespace dlk