nopenpilot/selfdrive/modeld/models/commonmodel.cc

73 lines
2.6 KiB
C++
Raw Normal View History

2021-06-08 13:09:30 -06:00
#include "selfdrive/modeld/models/commonmodel.h"
#include <algorithm>
#include <cassert>
#include <cmath>
#include <cstring>
#include "selfdrive/common/clutil.h"
#include "selfdrive/common/mat.h"
#include "selfdrive/common/timing.h"
ModelFrame::ModelFrame(cl_device_id device_id, cl_context context) {
input_frames = std::make_unique<float[]>(buf_size);
q = CL_CHECK_ERR(clCreateCommandQueue(context, device_id, 0, &err));
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));
net_input_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE, MODEL_FRAME_SIZE * sizeof(float), NULL, &err));
transform_init(&transform, context, device_id);
loadyuv_init(&loadyuv, context, device_id, MODEL_WIDTH, MODEL_HEIGHT);
}
float* ModelFrame::prepare(cl_mem yuv_cl, int frame_width, int frame_height, const mat3 &projection, cl_mem *output) {
transform_queue(&this->transform, q,
yuv_cl, frame_width, frame_height,
y_cl, u_cl, v_cl, MODEL_WIDTH, MODEL_HEIGHT, projection);
if (output == NULL) {
loadyuv_queue(&loadyuv, q, y_cl, u_cl, v_cl, net_input_cl);
std::memmove(&input_frames[0], &input_frames[MODEL_FRAME_SIZE], sizeof(float) * MODEL_FRAME_SIZE);
CL_CHECK(clEnqueueReadBuffer(q, net_input_cl, CL_TRUE, 0, MODEL_FRAME_SIZE * sizeof(float), &input_frames[MODEL_FRAME_SIZE], 0, nullptr, nullptr));
clFinish(q);
return &input_frames[0];
} else {
loadyuv_queue(&loadyuv, q, y_cl, u_cl, v_cl, *output, true);
// NOTE: Since thneed is using a different command queue, this clFinish is needed to ensure the image is ready.
clFinish(q);
return NULL;
}
}
ModelFrame::~ModelFrame() {
transform_destroy(&transform);
loadyuv_destroy(&loadyuv);
CL_CHECK(clReleaseMemObject(net_input_cl));
CL_CHECK(clReleaseMemObject(v_cl));
CL_CHECK(clReleaseMemObject(u_cl));
CL_CHECK(clReleaseMemObject(y_cl));
CL_CHECK(clReleaseCommandQueue(q));
}
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) {
const float max_val = *std::max_element(input, input + len);
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
float denominator = 0;
for(int i = 0; i < len; i++) {
float const v_exp = expf(input[i] - max_val);
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
denominator += v_exp;
output[i] = v_exp;
}
const float inv_denominator = 1. / denominator;
for(int i = 0; i < len; i++) {
output[i] *= inv_denominator;
}
}
float sigmoid(float input) {
return 1 / (1 + expf(-input));
}