IQ.Pilot Release Commit @ 5bc9cd3

selfdrive/modeld/models/README.md

@@ -0,0 +1,66 @@
+## Neural networks in openpilot
+To view the architecture of the ONNX networks, you can use [netron](https://netron.app/)
+
+## Driving Model (vision model + temporal policy model)
+### Vision inputs (Full size: 799906 x float32)
+* **image stream**
+  * Two consecutive images (256 * 512 * 3 in RGB) recorded at 20 Hz : 393216 = 2 * 6 * 128 * 256
+    * Each 256 * 512 image is represented in YUV420 with 6 channels : 6 * 128 * 256
+      * Channels 0,1,2,3 represent the full-res Y channel and are represented in numpy as Y[::2, ::2], Y[::2, 1::2], Y[1::2, ::2], and Y[1::2, 1::2]
+      * Channel 4 represents the half-res U channel
+      * Channel 5 represents the half-res V channel
+* **wide image stream**
+  * Two consecutive images (256 * 512 * 3 in RGB) recorded at 20 Hz : 393216 = 2 * 6 * 128 * 256
+    * Each 256 * 512 image is represented in YUV420 with 6 channels : 6 * 128 * 256
+      * Channels 0,1,2,3 represent the full-res Y channel and are represented in numpy as Y[::2, ::2], Y[::2, 1::2], Y[1::2, ::2], and Y[1::2, 1::2]
+      * Channel 4 represents the half-res U channel
+      * Channel 5 represents the half-res V channel
+### Policy inputs
+* **desire**
+  * one-hot encoded buffer to command model to execute certain actions, bit needs to be sent for the past 5 seconds (at 20FPS) : 100 * 8
+* **traffic convention**
+  * one-hot encoded vector to tell model whether traffic is right-hand or left-hand traffic : 2
+* **lateral control params**
+  * speed and steering delay for predicting the desired curvature: 2
+* **previous desired curvatures**
+  * vector of previously predicted desired curvatures: 100 * 1
+* **feature buffer**
+  * a buffer of intermediate features including the current feature to form a 5 seconds temporal context (at 20FPS) : 100 * 512
+
+
+### Driving Model output format (Full size: XXX x float32)
+Refer to **slice_outputs** and **parse_vision_outputs/parse_policy_outputs** in modeld.
+
+
+## Driver Monitoring Model
+* .onnx model can be run with onnx runtimes
+* .dlc file is a pre-quantized model and only runs on qualcomm DSPs
+
+### input format
+* single image W = 1440 H = 960 luminance channel (Y) from the planar YUV420 format:
+  * full input size is 1440 * 960 = 1382400
+  * normalized ranging from 0.0 to 1.0 in float32 (onnx runner) or ranging from 0 to 255 in uint8 (snpe runner)
+* camera calibration angles (roll, pitch, yaw) from liveCalibration: 3 x float32 inputs
+
+### output format
+* 84 x float32 outputs = 2 + 41 * 2 ([parsing example](https://github.com/commaai/openpilot/blob/22ce4e17ba0d3bfcf37f8255a4dd1dc683fe0c38/selfdrive/modeld/models/dmonitoring.cc#L33))
+  * for each person in the front seats (2 * 41)
+    * face pose: 12 = 6 + 6
+      * face orientation [pitch, yaw, roll] in camera frame: 3
+      * face position [dx, dy] relative to image center: 2
+      * normalized face size: 1
+      * standard deviations for above outputs: 6
+    * face visible probability: 1
+    * eyes: 20 = (8 + 1) + (8 + 1) + 1 + 1
+      * eye position and size, and their standard deviations: 8
+      * eye visible probability: 1
+      * eye closed probability: 1
+    * wearing sunglasses probability: 1
+    * face occluded probability: 1
+    * touching wheel probability: 1
+    * paying attention probability: 1
+    * (deprecated) distracted probabilities: 2
+    * using phone probability: 1
+    * distracted probability: 1
+  * common outputs 1
+    * left hand drive probability: 1

selfdrive/modeld/models/commonmodel.cc

@@ -0,0 +1,64 @@
+#include "selfdrive/modeld/models/commonmodel.h"
+
+#include <cmath>
+#include <cstring>
+
+#include "common/clutil.h"
+
+DrivingModelFrame::DrivingModelFrame(cl_device_id device_id, cl_context context, int _temporal_skip) : ModelFrame(device_id, context) {
+  input_frames = std::make_unique<uint8_t[]>(buf_size);
+  temporal_skip = _temporal_skip;
+  input_frames_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, buf_size, NULL, &err));
+  img_buffer_20hz_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (temporal_skip+1)*frame_size_bytes, NULL, &err));
+  region.origin = temporal_skip * frame_size_bytes;
+  region.size = frame_size_bytes;
+  last_img_cl = CL_CHECK_ERR(clCreateSubBuffer(img_buffer_20hz_cl, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err));
+
+  loadyuv_init(&loadyuv, context, device_id, MODEL_WIDTH, MODEL_HEIGHT);
+  init_transform(device_id, context, MODEL_WIDTH, MODEL_HEIGHT);
+}
+
+cl_mem* DrivingModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) {
+  run_transform(yuv_cl, MODEL_WIDTH, MODEL_HEIGHT, frame_width, frame_height, frame_stride, frame_uv_offset, projection);
+
+  for (int i = 0; i < temporal_skip; i++) {
+    CL_CHECK(clEnqueueCopyBuffer(q, img_buffer_20hz_cl, img_buffer_20hz_cl, (i+1)*frame_size_bytes, i*frame_size_bytes, frame_size_bytes, 0, nullptr, nullptr));
+  }
+  loadyuv_queue(&loadyuv, q, y_cl, u_cl, v_cl, last_img_cl);
+
+  copy_queue(&loadyuv, q, img_buffer_20hz_cl, input_frames_cl, 0, 0, frame_size_bytes);
+  copy_queue(&loadyuv, q, last_img_cl, input_frames_cl, 0, frame_size_bytes, frame_size_bytes);
+
+  // NOTE: Since thneed is using a different command queue, this clFinish is needed to ensure the image is ready.
+  clFinish(q);
+  return &input_frames_cl;
+}
+
+DrivingModelFrame::~DrivingModelFrame() {
+  deinit_transform();
+  loadyuv_destroy(&loadyuv);
+  CL_CHECK(clReleaseMemObject(input_frames_cl));
+  CL_CHECK(clReleaseMemObject(img_buffer_20hz_cl));
+  CL_CHECK(clReleaseMemObject(last_img_cl));
+  CL_CHECK(clReleaseCommandQueue(q));
+}
+
+
+MonitoringModelFrame::MonitoringModelFrame(cl_device_id device_id, cl_context context) : ModelFrame(device_id, context) {
+  input_frames = std::make_unique<uint8_t[]>(buf_size);
+  input_frame_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, buf_size, NULL, &err));
+
+  init_transform(device_id, context, MODEL_WIDTH, MODEL_HEIGHT);
+}
+
+cl_mem* MonitoringModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) {
+  run_transform(yuv_cl, MODEL_WIDTH, MODEL_HEIGHT, frame_width, frame_height, frame_stride, frame_uv_offset, projection);
+  clFinish(q);
+  return &y_cl;
+}
+
+MonitoringModelFrame::~MonitoringModelFrame() {
+  deinit_transform();
+  CL_CHECK(clReleaseMemObject(input_frame_cl));
+  CL_CHECK(clReleaseCommandQueue(q));
+}

selfdrive/modeld/models/commonmodel.h

@@ -0,0 +1,97 @@
+#pragma once
+
+#include <cfloat>
+#include <cstdlib>
+#include <cassert>
+
+#include <memory>
+
+#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+#ifdef __APPLE__
+#include <OpenCL/cl.h>
+#else
+#include <CL/cl.h>
+#endif
+
+#include "common/mat.h"
+#include "selfdrive/modeld/transforms/loadyuv.h"
+#include "selfdrive/modeld/transforms/transform.h"
+
+class ModelFrame {
+public:
+  ModelFrame(cl_device_id device_id, cl_context context) {
+    q = CL_CHECK_ERR(clCreateCommandQueue(context, device_id, 0, &err));
+  }
+  virtual ~ModelFrame() {}
+  virtual cl_mem* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) { return NULL; }
+  uint8_t* buffer_from_cl(cl_mem *in_frames, int buffer_size) {
+    CL_CHECK(clEnqueueReadBuffer(q, *in_frames, CL_TRUE, 0, buffer_size, input_frames.get(), 0, nullptr, nullptr));
+    clFinish(q);
+    return &input_frames[0];
+  }
+
+  int MODEL_WIDTH;
+  int MODEL_HEIGHT;
+  int MODEL_FRAME_SIZE;
+  int buf_size;
+
+protected:
+  cl_mem y_cl, u_cl, v_cl;
+  Transform transform;
+  cl_command_queue q;
+  std::unique_ptr<uint8_t[]> input_frames;
+
+  void init_transform(cl_device_id device_id, cl_context context, int model_width, int model_height) {
+    y_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, model_width * model_height, NULL, &err));
+    u_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (model_width / 2) * (model_height / 2), NULL, &err));
+    v_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, (model_width / 2) * (model_height / 2), NULL, &err));
+    transform_init(&transform, context, device_id);
+  }
+
+  void deinit_transform() {
+    transform_destroy(&transform);
+    CL_CHECK(clReleaseMemObject(v_cl));
+    CL_CHECK(clReleaseMemObject(u_cl));
+    CL_CHECK(clReleaseMemObject(y_cl));
+  }
+
+  void run_transform(cl_mem yuv_cl, int model_width, int model_height, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection) {
+    transform_queue(&transform, q,
+        yuv_cl, frame_width, frame_height, frame_stride, frame_uv_offset,
+        y_cl, u_cl, v_cl, model_width, model_height, projection);
+  }
+};
+
+class DrivingModelFrame : public ModelFrame {
+public:
+  DrivingModelFrame(cl_device_id device_id, cl_context context, int _temporal_skip);
+  ~DrivingModelFrame();
+  cl_mem* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection);
+
+  const int MODEL_WIDTH = 512;
+  const int MODEL_HEIGHT = 256;
+  const int MODEL_FRAME_SIZE = MODEL_WIDTH * MODEL_HEIGHT * 3 / 2;
+  const int buf_size = MODEL_FRAME_SIZE * 2; // 2 frames are temporal_skip frames apart
+  const size_t frame_size_bytes = MODEL_FRAME_SIZE * sizeof(uint8_t);
+
+private:
+  LoadYUVState loadyuv;
+  cl_mem img_buffer_20hz_cl, last_img_cl, input_frames_cl;
+  cl_buffer_region region;
+  int temporal_skip;
+};
+
+class MonitoringModelFrame : public ModelFrame {
+public:
+  MonitoringModelFrame(cl_device_id device_id, cl_context context);
+  ~MonitoringModelFrame();
+  cl_mem* prepare(cl_mem yuv_cl, int frame_width, int frame_height, int frame_stride, int frame_uv_offset, const mat3& projection);
+
+  const int MODEL_WIDTH = 1440;
+  const int MODEL_HEIGHT = 960;
+  const int MODEL_FRAME_SIZE = MODEL_WIDTH * MODEL_HEIGHT;
+  const int buf_size = MODEL_FRAME_SIZE;
+
+private:
+  cl_mem input_frame_cl;
+};

selfdrive/modeld/models/commonmodel.pxd

@@ -0,0 +1,27 @@
+# distutils: language = c++
+
+from msgq.visionipc.visionipc cimport cl_device_id, cl_context, cl_mem
+
+cdef extern from "common/mat.h":
+  cdef struct mat3:
+    float v[9]
+
+cdef extern from "common/clutil.h":
+  cdef unsigned long CL_DEVICE_TYPE_DEFAULT
+  cl_device_id cl_get_device_id(unsigned long)
+  cl_context cl_create_context(cl_device_id)
+  void cl_release_context(cl_context)
+
+cdef extern from "selfdrive/modeld/models/commonmodel.h":
+  cppclass ModelFrame:
+    int buf_size
+    unsigned char * buffer_from_cl(cl_mem*, int);
+    cl_mem * prepare(cl_mem, int, int, int, int, mat3)
+
+  cppclass DrivingModelFrame:
+    int buf_size
+    DrivingModelFrame(cl_device_id, cl_context, int)
+
+  cppclass MonitoringModelFrame:
+    int buf_size
+    MonitoringModelFrame(cl_device_id, cl_context)

selfdrive/modeld/models/commonmodel_pyx.pxd

@@ -0,0 +1,13 @@
+# distutils: language = c++
+
+from msgq.visionipc.visionipc cimport cl_mem
+from msgq.visionipc.visionipc_pyx cimport CLContext as BaseCLContext
+
+cdef class CLContext(BaseCLContext):
+  pass
+
+cdef class CLMem:
+  cdef cl_mem * mem
+
+  @staticmethod
+  cdef create(void*)

selfdrive/modeld/models/commonmodel_pyx.pyx

@@ -0,0 +1,74 @@
+# distutils: language = c++
+# cython: c_string_encoding=ascii, language_level=3
+
+import numpy as np
+cimport numpy as cnp
+from libc.string cimport memcpy
+from libc.stdint cimport uintptr_t
+
+from msgq.visionipc.visionipc cimport cl_mem
+from msgq.visionipc.visionipc_pyx cimport VisionBuf, CLContext as BaseCLContext
+from .commonmodel cimport CL_DEVICE_TYPE_DEFAULT, cl_get_device_id, cl_create_context, cl_release_context
+from .commonmodel cimport mat3, ModelFrame as cppModelFrame, DrivingModelFrame as cppDrivingModelFrame, MonitoringModelFrame as cppMonitoringModelFrame
+
+
+cdef class CLContext(BaseCLContext):
+  def __cinit__(self):
+    self.device_id = cl_get_device_id(CL_DEVICE_TYPE_DEFAULT)
+    self.context = cl_create_context(self.device_id)
+
+  def __dealloc__(self):
+    if self.context:
+      cl_release_context(self.context)
+
+cdef class CLMem:
+  @staticmethod
+  cdef create(void * cmem):
+    mem = CLMem()
+    mem.mem = <cl_mem*> cmem
+    return mem
+
+  @property
+  def mem_address(self):
+    return <uintptr_t>(self.mem)
+
+def cl_from_visionbuf(VisionBuf buf):
+  return CLMem.create(<void*>&buf.buf.buf_cl)
+
+
+cdef class ModelFrame:
+  cdef cppModelFrame * frame
+  cdef int buf_size
+
+  def __dealloc__(self):
+    del self.frame
+
+  def prepare(self, VisionBuf buf, float[:] projection):
+    cdef mat3 cprojection
+    memcpy(cprojection.v, &projection[0], 9*sizeof(float))
+    cdef cl_mem * data
+    data = self.frame.prepare(buf.buf.buf_cl, buf.width, buf.height, buf.stride, buf.uv_offset, cprojection)
+    return CLMem.create(data)
+
+  def buffer_from_cl(self, CLMem in_frames):
+    cdef unsigned char * data2
+    data2 = self.frame.buffer_from_cl(in_frames.mem, self.buf_size)
+    return np.asarray(<cnp.uint8_t[:self.buf_size]> data2)
+
+
+cdef class DrivingModelFrame(ModelFrame):
+  cdef cppDrivingModelFrame * _frame
+
+  def __cinit__(self, CLContext context, int temporal_skip):
+    self._frame = new cppDrivingModelFrame(context.device_id, context.context, temporal_skip)
+    self.frame = <cppModelFrame*>(self._frame)
+    self.buf_size = self._frame.buf_size
+
+cdef class MonitoringModelFrame(ModelFrame):
+  cdef cppMonitoringModelFrame * _frame
+
+  def __cinit__(self, CLContext context):
+    self._frame = new cppMonitoringModelFrame(context.device_id, context.context)
+    self.frame = <cppModelFrame*>(self._frame)
+    self.buf_size = self._frame.buf_size
+

selfdrive/modeld/models/prebuilt_check.json

@@ -0,0 +1,9 @@
+{
+  "dmonitoring_model": {
+    "outputs": {
+      "dmonitoring_model_metadata.pkl": "31a86ab7a92dc0af088b15787a440dd3b210aa662e445a15145900e559a1b5c3",
+      "dmonitoring_model_tinygrad.pkl": "806c0ea75df6bf6dfeb81b832314c68e31df5865a52d0359e6eeb76d93ad2b52"
+    },
+    "signature": "e1eeb5ce45774a816c8da2394e6ee35ebf700b71dff345e0141dbab8ff592349"
+  }
+}