mirror of
https://github.com/dragonpilot/dragonpilot.git
synced 2026-02-23 10:13:53 +08:00
630c7309a5
* template funcion fill_meta * add function get_plan_max_idx * add function get_mdn_max_idx * remove temp builder variables * get_mdn_max_idx->get_lead_data * get_pan_max_idx->get_plan_data * fill_lane_line * combine fill_lane_line&fill_path into one function * fill_lead_vw * using unique_ptr * prefer using constexpr * more constexpr,remove duplicate defined TRAJECTORY_SIZE * remove suffix _arr from variable name * misc * remove extern c * refactor model_publish * remove unused paramaters * traffic_convention to c style array * c style array:prev_desire&pulse_desire * fix error&make easy for review * const mat3 &transform * move cl_command_queue into ModelState * use maco LEAD_MHP_SELECTION * move constexpr from .h to .cc * remove #define MODEL_NAME * modeldata.h: contexpr * remove param temporal from model_init * helper function get_best_data * fix probs * int Co-authored-by: deanlee <deanlee3@gmail.com>
84 lines
3.1 KiB
C++
84 lines
3.1 KiB
C++
#include <assert.h>
|
|
#include <math.h>
|
|
#include "commonmodel.h"
|
|
#include "common/clutil.h"
|
|
#include "common/mat.h"
|
|
#include "common/timing.h"
|
|
|
|
void frame_init(ModelFrame* frame, int width, int height,
|
|
cl_device_id device_id, cl_context context) {
|
|
transform_init(&frame->transform, context, device_id);
|
|
frame->transformed_width = width;
|
|
frame->transformed_height = height;
|
|
|
|
frame->transformed_y_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE,
|
|
(size_t)frame->transformed_width*frame->transformed_height, NULL, &err));
|
|
frame->transformed_u_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE,
|
|
(size_t)(frame->transformed_width/2)*(frame->transformed_height/2), NULL, &err));
|
|
frame->transformed_v_cl = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE,
|
|
(size_t)(frame->transformed_width/2)*(frame->transformed_height/2), NULL, &err));
|
|
frame->net_input_size = ((width*height*3)/2)*sizeof(float);
|
|
frame->net_input = CL_CHECK_ERR(clCreateBuffer(context, CL_MEM_READ_WRITE,
|
|
frame->net_input_size, (void*)NULL, &err));
|
|
loadyuv_init(&frame->loadyuv, context, device_id, frame->transformed_width, frame->transformed_height);
|
|
}
|
|
|
|
float *frame_prepare(ModelFrame* frame, cl_command_queue q,
|
|
cl_mem yuv_cl, int width, int height,
|
|
const mat3 &transform) {
|
|
transform_queue(&frame->transform, q,
|
|
yuv_cl, width, height,
|
|
frame->transformed_y_cl, frame->transformed_u_cl, frame->transformed_v_cl,
|
|
frame->transformed_width, frame->transformed_height,
|
|
transform);
|
|
loadyuv_queue(&frame->loadyuv, q,
|
|
frame->transformed_y_cl, frame->transformed_u_cl, frame->transformed_v_cl,
|
|
frame->net_input);
|
|
float *net_input_buf = (float *)CL_CHECK_ERR(clEnqueueMapBuffer(q, frame->net_input, CL_TRUE,
|
|
CL_MAP_READ, 0, frame->net_input_size,
|
|
0, NULL, NULL, &err));
|
|
clFinish(q);
|
|
return net_input_buf;
|
|
}
|
|
|
|
void frame_free(ModelFrame* frame) {
|
|
transform_destroy(&frame->transform);
|
|
loadyuv_destroy(&frame->loadyuv);
|
|
CL_CHECK(clReleaseMemObject(frame->net_input));
|
|
CL_CHECK(clReleaseMemObject(frame->transformed_v_cl));
|
|
CL_CHECK(clReleaseMemObject(frame->transformed_u_cl));
|
|
CL_CHECK(clReleaseMemObject(frame->transformed_y_cl));
|
|
}
|
|
|
|
void softmax(const float* input, float* output, size_t len) {
|
|
float max_val = -FLT_MAX;
|
|
for(int i = 0; i < len; i++) {
|
|
const float v = input[i];
|
|
if( v > max_val ) {
|
|
max_val = v;
|
|
}
|
|
}
|
|
|
|
float denominator = 0;
|
|
for(int i = 0; i < len; i++) {
|
|
float const v = input[i];
|
|
float const v_exp = expf(v - max_val);
|
|
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));
|
|
}
|
|
|
|
float softplus(float input) {
|
|
return log1p(expf(input));
|
|
}
|