nopenpilot/selfdrive/modeld/models/commonmodel.h

42 lines
1009 B
C
Raw Normal View History

#pragma once
#include <cfloat>
#include <cstdlib>
#include <memory>
#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
#ifdef __APPLE__
#include <OpenCL/cl.h>
#else
#include <CL/cl.h>
#endif
#include "selfdrive/common/mat.h"
#include "selfdrive/modeld/transforms/loadyuv.h"
#include "selfdrive/modeld/transforms/transform.h"
const bool send_raw_pred = getenv("SEND_RAW_PRED") != NULL;
Torch model (#2452) * refactor draw model * rebase master * correct valid_len * rename function * rename variables * white space * rebase to master * e16c13ac-927d-455e-ae0a-81b482a2c787 * start rewriting * save proress * compiles! * oops * many fixes * seems to work * fix desires * finally cleaned * wrong std for ll * dont pulse none * compiles! * ready to test * WIP does not compile * compiles * various fixes * does something! * full 3d * not needed * draw up to 100m * fix segfault * wrong sign * fix flicker * add road edges * finish v2 packet * Added pytorch supercombo * fix rebase * no more keras * Hacky solution to the NCHW/NHWC incompatibility between SNPE and our frame data * dont break dmonitoringd, final model 229e3ce1-7259-412b-85e6-cc646d70f1d8/430 * fix hack * Revert "fix hack" This reverts commit 5550fc01a7881d065a5eddbbb42dac55ef7ec36c. * Removed axis permutation hack * Folded padding layers into conv layers * Removed the last pad layer from the dlc * Revert "Removed the last pad layer from the dlc" This reverts commit b85f24b9e1d04abf64e85901a7ff49e00d82020a. * Revert "Folded padding layers into conv layers" This reverts commit b8d1773e4e76dea481acebbfad6a6235fbb58463. * vision model: 5034ac8b-5703-4a49-948b-11c064d10880/780 temporal model: 229e3ce1-7259-412b-85e6-cc646d70f1d8/430 with permute + pool opt * fix ui drawing with clips * ./compile_torch.py 5034ac8b-5703-4a49-948b-11c064d10880/780 dfcd2375-81d8-49df-95bf-1d2d6ad86010/450 with variable history length * std::clamp * not sure how this compiled before * 2895ace6-a296-47ac-86e6-17ea800a74e5/550 * db090195-8810-42de-ab38-bb835d775d87/601 * 5m is very little * onnx runner * add onnxruntime to pipfile * run in real time without using the whole CPU * bump cereal; * add stds * set road edge opacity based on stddev * don't access the model packet in paint * convert mat.h to a c++ header file (#2499) * update tests * safety first Co-authored-by: deanlee <deanlee3@gmail.com> Co-authored-by: mitchell <mitchell@comma.ai> Co-authored-by: Comma Device <device@comma.ai> Co-authored-by: George Hotz <george@comma.ai> Co-authored-by: Adeeb Shihadeh <adeebshihadeh@gmail.com>
2020-11-11 21:31:46 -07:00
void softmax(const float* input, float* output, size_t len);
float sigmoid(float input);
class ModelFrame {
public:
ModelFrame(cl_device_id device_id, cl_context context);
~ModelFrame();
float* prepare(cl_mem yuv_cl, int width, int height, const mat3& transform, cl_mem *output);
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;
private:
Transform transform;
LoadYUVState loadyuv;
cl_command_queue q;
cl_mem y_cl, u_cl, v_cl, net_input_cl;
std::unique_ptr<float[]> input_frames;
};