openpilot_comma/selfdrive/modeld/models/dmonitoring.cc

#include <string.h>
#include "dmonitoring.h"
#include "common/mat.h"
#include "common/timing.h"
#include "common/params.h"

#include <libyuv.h>

#define MODEL_WIDTH 320
#define MODEL_HEIGHT 640
#define FULL_W 852 // should get these numbers from camerad

#if defined(QCOM) || defined(QCOM2)
#define input_lambda(x) (x - 128.f) * 0.0078125f
#else
#define input_lambda(x) x // for non SNPE running platforms, assume keras model instead has lambda layer
#endif

void dmonitoring_init(DMonitoringModelState* s) {
#if defined(QCOM) || defined(QCOM2)
  const char* model_path = "../../models/dmonitoring_model_q.dlc";
#else
  const char* model_path = "../../models/dmonitoring_model.dlc";
#endif

  int runtime = USE_DSP_RUNTIME;
  s->m = new DefaultRunModel(model_path, (float*)&s->output, OUTPUT_SIZE, runtime);
  s->is_rhd = Params().read_db_bool("IsRHD");
}

template <class T>
static inline T *get_buffer(std::vector<T> &buf, const size_t size) {
  if (buf.size() < size) buf.resize(size);
  return buf.data();
}

static inline auto get_yuv_buf(std::vector<uint8_t> &buf, const int width, int height) {
  uint8_t *y = get_buffer(buf, width * height * 3 / 2);
  uint8_t *u = y + width * height;
  uint8_t *v = u + (width /2) * (height / 2);
  return std::make_tuple(y, u, v);
}

DMonitoringResult dmonitoring_eval_frame(DMonitoringModelState* s, void* stream_buf, int width, int height) {
  uint8_t *raw_buf = (uint8_t*) stream_buf;
  uint8_t *raw_y_buf = raw_buf;
  uint8_t *raw_u_buf = raw_y_buf + (width * height);
  uint8_t *raw_v_buf = raw_u_buf + ((width/2) * (height/2));

#ifndef QCOM2
  const int cropped_width = height/2;
  const int cropped_height = height;
  const int global_x_offset = 0;
  const int global_y_offset = 0;
  const int crop_x_offset = width - cropped_width;
  const int crop_y_offset = 0;
#else
  const int full_width_tici = 1928;
  const int full_height_tici = 1208;
  const int adapt_width_tici = 668;

  const int cropped_height = adapt_width_tici / 1.33;
  const int cropped_width = cropped_height / 2;
  const int global_x_offset = full_width_tici / 2 - adapt_width_tici / 2;
  const int global_y_offset = full_height_tici / 2 - cropped_height / 2;
  const int crop_x_offset = adapt_width_tici - cropped_width + 32;
  const int crop_y_offset = -196;
#endif

  int resized_width = MODEL_WIDTH;
  int resized_height = MODEL_HEIGHT;

  auto [cropped_y_buf, cropped_u_buf, cropped_v_buf] = get_yuv_buf(s->cropped_buf, cropped_width, cropped_height);
  if (!s->is_rhd) {
    for (int r = 0; r < cropped_height/2; r++) {
      memcpy(cropped_y_buf + 2*r*cropped_width, raw_y_buf + (2*r + global_y_offset + crop_y_offset)*width + global_x_offset + crop_x_offset, cropped_width);
      memcpy(cropped_y_buf + (2*r+1)*cropped_width, raw_y_buf + (2*r + global_y_offset + crop_y_offset + 1)*width + global_x_offset + crop_x_offset, cropped_width);
      memcpy(cropped_u_buf + r*(cropped_width/2), raw_u_buf + (r + (global_y_offset + crop_y_offset)/2)*width/2 + (global_x_offset + crop_x_offset)/2, cropped_width/2);
      memcpy(cropped_v_buf + r*(cropped_width/2), raw_v_buf + (r + (global_y_offset + crop_y_offset)/2)*width/2 + (global_x_offset + crop_x_offset)/2, cropped_width/2);
    }
  } else {
    auto [premirror_cropped_y_buf, premirror_cropped_u_buf, premirror_cropped_v_buf] = get_yuv_buf(s->premirror_cropped_buf, cropped_width, cropped_height);
    for (int r = 0; r < cropped_height/2; r++) {
      memcpy(premirror_cropped_y_buf + (2*r)*cropped_width, raw_y_buf + (2*r + global_y_offset + crop_y_offset)*width + global_x_offset, cropped_width);
      memcpy(premirror_cropped_y_buf + (2*r+1)*cropped_width, raw_y_buf + (2*r + global_y_offset + crop_y_offset + 1)*width + global_x_offset, cropped_width);
      memcpy(premirror_cropped_u_buf + r*(cropped_width/2), raw_u_buf + (r + (global_y_offset + crop_y_offset)/2)*width/2 + global_x_offset/2, cropped_width/2);
      memcpy(premirror_cropped_v_buf + r*(cropped_width/2), raw_v_buf + (r + (global_y_offset + crop_y_offset)/2)*width/2 + global_x_offset/2, cropped_width/2);
    }
    libyuv::I420Mirror(premirror_cropped_y_buf, cropped_width,
                       premirror_cropped_u_buf, cropped_width/2,
                       premirror_cropped_v_buf, cropped_width/2,
                       cropped_y_buf, cropped_width,
                       cropped_u_buf, cropped_width/2,
                       cropped_v_buf, cropped_width/2,
                       cropped_width, cropped_height);
  }

  auto [resized_y_buf, resized_u_buf, resized_v_buf] = get_yuv_buf(s->resized_buf, resized_width, resized_height);
  libyuv::FilterMode mode = libyuv::FilterModeEnum::kFilterBilinear;
  libyuv::I420Scale(cropped_y_buf, cropped_width,
                    cropped_u_buf, cropped_width/2,
                    cropped_v_buf, cropped_width/2,
                    cropped_width, cropped_height,
                    resized_y_buf, resized_width,
                    resized_u_buf, resized_width/2,
                    resized_v_buf, resized_width/2,
                    resized_width, resized_height,
                    mode);

  // prerotate to be cache aware
  auto [resized_buf_rot, resized_u_buf_rot, resized_v_buf_rot] = get_yuv_buf(s->resized_buf_rot, resized_width, resized_height);
  uint8_t *resized_y_buf_rot = resized_buf_rot;
  libyuv::I420Rotate(resized_y_buf, resized_width,
                     resized_u_buf, resized_width/2,
                     resized_v_buf, resized_width/2,
                     resized_y_buf_rot, resized_height,
                     resized_u_buf_rot, resized_height/2,
                     resized_v_buf_rot, resized_height/2,
                     // negative height causes a vertical flip to match previous
                     resized_width, -resized_height, libyuv::kRotate90);

  int yuv_buf_len = (MODEL_WIDTH/2) * (MODEL_HEIGHT/2) * 6; // Y|u|v -> y|y|y|y|u|v
  float *net_input_buf = get_buffer(s->net_input_buf, yuv_buf_len);
  // one shot conversion, O(n) anyway
  // yuvframe2tensor, normalize
  for (int c = 0; c < MODEL_WIDTH/2; c++) {
    for (int r = 0; r < MODEL_HEIGHT/2; r++) {
      // Y_ul
      net_input_buf[(c*MODEL_HEIGHT/2) + r + (0*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf_rot[(2*r) + (2*c)*resized_height]);
      // Y_dl
      net_input_buf[(c*MODEL_HEIGHT/2) + r + (1*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf_rot[(2*r+1) + (2*c)*resized_height]);
      // Y_ur
      net_input_buf[(c*MODEL_HEIGHT/2) + r + (2*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf_rot[(2*r) + (2*c+1)*resized_height]);
      // Y_dr
      net_input_buf[(c*MODEL_HEIGHT/2) + r + (3*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf_rot[(2*r+1) + (2*c+1)*resized_height]);
      // U
      net_input_buf[(c*MODEL_HEIGHT/2) + r + (4*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf_rot[(resized_width*resized_height) + r + (c*resized_height/2)]);
      // V
      net_input_buf[(c*MODEL_HEIGHT/2) + r + (5*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf_rot[(resized_width*resized_height) + ((resized_width/2)*(resized_height/2)) + r + (c*resized_height/2)]);
    }
  }

  //printf("preprocess completed. %d \n", yuv_buf_len);
  //FILE *dump_yuv_file = fopen("/tmp/rawdump.yuv", "wb");
  //fwrite(raw_buf, height*width*3/2, sizeof(uint8_t), dump_yuv_file);
  //fclose(dump_yuv_file);

  // *** testing ***
  // idat = np.frombuffer(open("/tmp/inputdump.yuv", "rb").read(), np.float32).reshape(6, 160, 320)
  // imshow(cv2.cvtColor(tensor_to_frames(idat[None]/0.0078125+128)[0], cv2.COLOR_YUV2RGB_I420))

  //FILE *dump_yuv_file2 = fopen("/tmp/inputdump.yuv", "wb");
  //fwrite(net_input_buf, MODEL_HEIGHT*MODEL_WIDTH*3/2, sizeof(float), dump_yuv_file2);
  //fclose(dump_yuv_file2);

  double t1 = millis_since_boot();
  s->m->execute(net_input_buf, yuv_buf_len);
  double t2 = millis_since_boot();

  DMonitoringResult ret = {0};
  memcpy(&ret.face_orientation, &s->output[0], sizeof ret.face_orientation);
  memcpy(&ret.face_orientation_meta, &s->output[6], sizeof ret.face_orientation_meta);
  memcpy(&ret.face_position, &s->output[3], sizeof ret.face_position);
  memcpy(&ret.face_position_meta, &s->output[9], sizeof ret.face_position_meta);
  memcpy(&ret.face_prob, &s->output[12], sizeof ret.face_prob);
  memcpy(&ret.left_eye_prob, &s->output[21], sizeof ret.left_eye_prob);
  memcpy(&ret.right_eye_prob, &s->output[30], sizeof ret.right_eye_prob);
  memcpy(&ret.left_blink_prob, &s->output[31], sizeof ret.right_eye_prob);
  memcpy(&ret.right_blink_prob, &s->output[32], sizeof ret.right_eye_prob);
  memcpy(&ret.sg_prob, &s->output[33], sizeof ret.sg_prob);
  ret.face_orientation_meta[0] = softplus(ret.face_orientation_meta[0]);
  ret.face_orientation_meta[1] = softplus(ret.face_orientation_meta[1]);
  ret.face_orientation_meta[2] = softplus(ret.face_orientation_meta[2]);
  ret.face_position_meta[0] = softplus(ret.face_position_meta[0]);
  ret.face_position_meta[1] = softplus(ret.face_position_meta[1]);
  ret.dsp_execution_time = (t2 - t1) / 1000.;
  return ret;
}

void dmonitoring_publish(PubMaster &pm, uint32_t frame_id, const DMonitoringResult &res, const float* raw_pred, float execution_time){
  // make msg
  MessageBuilder msg;
  auto framed = msg.initEvent().initDriverState();
  framed.setFrameId(frame_id);
  framed.setModelExecutionTime(execution_time);
  framed.setDspExecutionTime(res.dsp_execution_time);

  framed.setFaceOrientation(res.face_orientation);
  framed.setFaceOrientationStd(res.face_orientation_meta);
  framed.setFacePosition(res.face_position);
  framed.setFacePositionStd(res.face_position_meta);
  framed.setFaceProb(res.face_prob);
  framed.setLeftEyeProb(res.left_eye_prob);
  framed.setRightEyeProb(res.right_eye_prob);
  framed.setLeftBlinkProb(res.left_blink_prob);
  framed.setRightBlinkProb(res.right_blink_prob);
  framed.setSgProb(res.sg_prob);
  if (send_raw_pred) {
    framed.setRawPred(kj::arrayPtr((const uint8_t*)raw_pred, OUTPUT_SIZE*sizeof(float)));
  }

  pm.send("driverState", msg);
}

void dmonitoring_free(DMonitoringModelState* s) {
  delete s->m;
}