mirror of
https://github.com/firestar5683/StarPilot.git
synced 2026-08-21 00:03:45 +08:00
gc old stuff in modeld
This commit is contained in:
Adeeb Shihadeh
@@ -1,101 +0,0 @@
|
||||
// clang++ -O2 repro.cc && ./a.out
|
||||
|
||||
#include <sched.h>
|
||||
#include <sys/types.h>
|
||||
#include <unistd.h>
|
||||
|
||||
#include <cstdint>
|
||||
#include <cstdio>
|
||||
#include <cstdlib>
|
||||
#include <cstring>
|
||||
#include <ctime>
|
||||
|
||||
static inline double millis_since_boot() {
|
||||
struct timespec t;
|
||||
clock_gettime(CLOCK_BOOTTIME, &t);
|
||||
return t.tv_sec * 1000.0 + t.tv_nsec * 1e-6;
|
||||
}
|
||||
|
||||
#define MODEL_WIDTH 320
|
||||
#define MODEL_HEIGHT 640
|
||||
|
||||
// null function still breaks it
|
||||
#define input_lambda(x) x
|
||||
|
||||
// this is copied from models/dmonitoring.cc, and is the code that triggers the issue
|
||||
void inner(uint8_t *resized_buf, float *net_input_buf) {
|
||||
int resized_width = MODEL_WIDTH;
|
||||
int resized_height = MODEL_HEIGHT;
|
||||
|
||||
// one shot conversion, O(n) anyway
|
||||
// yuvframe2tensor, normalize
|
||||
for (int r = 0; r < MODEL_HEIGHT/2; r++) {
|
||||
for (int c = 0; c < MODEL_WIDTH/2; c++) {
|
||||
// Y_ul
|
||||
net_input_buf[(c*MODEL_HEIGHT/2) + r] = input_lambda(resized_buf[(2*r*resized_width) + (2*c)]);
|
||||
// Y_ur
|
||||
net_input_buf[(c*MODEL_HEIGHT/2) + r + (2*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf[(2*r*resized_width) + (2*c+1)]);
|
||||
// Y_dl
|
||||
net_input_buf[(c*MODEL_HEIGHT/2) + r + ((MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf[(2*r*resized_width+1) + (2*c)]);
|
||||
// Y_dr
|
||||
net_input_buf[(c*MODEL_HEIGHT/2) + r + (3*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf[(2*r*resized_width+1) + (2*c+1)]);
|
||||
// U
|
||||
net_input_buf[(c*MODEL_HEIGHT/2) + r + (4*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf[(resized_width*resized_height) + (r*resized_width/2) + c]);
|
||||
// V
|
||||
net_input_buf[(c*MODEL_HEIGHT/2) + r + (5*(MODEL_WIDTH/2)*(MODEL_HEIGHT/2))] = input_lambda(resized_buf[(resized_width*resized_height) + ((resized_width/2)*(resized_height/2)) + (r*resized_width/2) + c]);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
float trial() {
|
||||
int resized_width = MODEL_WIDTH;
|
||||
int resized_height = MODEL_HEIGHT;
|
||||
|
||||
int yuv_buf_len = (MODEL_WIDTH/2) * (MODEL_HEIGHT/2) * 6; // Y|u|v -> y|y|y|y|u|v
|
||||
|
||||
// allocate the buffers
|
||||
uint8_t *resized_buf = (uint8_t*)malloc(resized_width*resized_height*3/2);
|
||||
float *net_input_buf = (float*)malloc(yuv_buf_len*sizeof(float));
|
||||
printf("allocate -- %p 0x%x -- %p 0x%lx\n", resized_buf, resized_width*resized_height*3/2, net_input_buf, yuv_buf_len*sizeof(float));
|
||||
|
||||
// test for bad buffers
|
||||
static int CNT = 20;
|
||||
float avg = 0.0;
|
||||
for (int i = 0; i < CNT; i++) {
|
||||
double s4 = millis_since_boot();
|
||||
inner(resized_buf, net_input_buf);
|
||||
double s5 = millis_since_boot();
|
||||
avg += s5-s4;
|
||||
}
|
||||
avg /= CNT;
|
||||
|
||||
// once it's bad, it's reliably bad
|
||||
if (avg > 10) {
|
||||
printf("HIT %f\n", avg);
|
||||
printf("BAD\n");
|
||||
|
||||
for (int i = 0; i < 200; i++) {
|
||||
double s4 = millis_since_boot();
|
||||
inner(resized_buf, net_input_buf);
|
||||
double s5 = millis_since_boot();
|
||||
printf("%.2f ", s5-s4);
|
||||
}
|
||||
printf("\n");
|
||||
|
||||
exit(0);
|
||||
}
|
||||
|
||||
// don't free so we get a different buffer each time
|
||||
//free(resized_buf);
|
||||
//free(net_input_buf);
|
||||
|
||||
return avg;
|
||||
}
|
||||
|
||||
int main() {
|
||||
while (true) {
|
||||
float ret = trial();
|
||||
printf("got %f\n", ret);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1,2 +0,0 @@
|
||||
#!/usr/bin/env bash
|
||||
clang++ -I /home/batman/one/external/tensorflow/include/ -L /home/batman/one/external/tensorflow/lib -Wl,-rpath=/home/batman/one/external/tensorflow/lib main.cc -ltensorflow
|
||||
@@ -1,69 +0,0 @@
|
||||
#include <cassert>
|
||||
#include <cstdio>
|
||||
#include <cstdlib>
|
||||
#include "tensorflow/c/c_api.h"
|
||||
|
||||
void* read_file(const char* path, size_t* out_len) {
|
||||
FILE* f = fopen(path, "r");
|
||||
if (!f) {
|
||||
return NULL;
|
||||
}
|
||||
fseek(f, 0, SEEK_END);
|
||||
long f_len = ftell(f);
|
||||
rewind(f);
|
||||
|
||||
char* buf = (char*)calloc(f_len, 1);
|
||||
assert(buf);
|
||||
|
||||
size_t num_read = fread(buf, f_len, 1, f);
|
||||
fclose(f);
|
||||
|
||||
if (num_read != 1) {
|
||||
free(buf);
|
||||
return NULL;
|
||||
}
|
||||
|
||||
if (out_len) {
|
||||
*out_len = f_len;
|
||||
}
|
||||
|
||||
return buf;
|
||||
}
|
||||
|
||||
static void DeallocateBuffer(void* data, size_t) {
|
||||
free(data);
|
||||
}
|
||||
|
||||
int main(int argc, char* argv[]) {
|
||||
TF_Buffer* buf;
|
||||
TF_Graph* graph;
|
||||
TF_Status* status;
|
||||
char *path = argv[1];
|
||||
|
||||
// load model
|
||||
{
|
||||
size_t model_size;
|
||||
char tmp[1024];
|
||||
snprintf(tmp, sizeof(tmp), "%s.pb", path);
|
||||
printf("loading model %s\n", tmp);
|
||||
uint8_t *model_data = (uint8_t *)read_file(tmp, &model_size);
|
||||
buf = TF_NewBuffer();
|
||||
buf->data = model_data;
|
||||
buf->length = model_size;
|
||||
buf->data_deallocator = DeallocateBuffer;
|
||||
printf("loaded model of size %d\n", model_size);
|
||||
}
|
||||
|
||||
// import graph
|
||||
status = TF_NewStatus();
|
||||
graph = TF_NewGraph();
|
||||
TF_ImportGraphDefOptions *opts = TF_NewImportGraphDefOptions();
|
||||
TF_GraphImportGraphDef(graph, buf, opts, status);
|
||||
TF_DeleteImportGraphDefOptions(opts);
|
||||
TF_DeleteBuffer(buf);
|
||||
if (TF_GetCode(status) != TF_OK) {
|
||||
printf("FAIL: %s\n", TF_Message(status));
|
||||
} else {
|
||||
printf("SUCCESS\n");
|
||||
}
|
||||
}
|
||||
@@ -1,8 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
import sys
|
||||
import tensorflow as tf
|
||||
|
||||
with open(sys.argv[1], "rb") as f:
|
||||
graph_def = tf.compat.v1.GraphDef()
|
||||
graph_def.ParseFromString(f.read())
|
||||
#tf.io.write_graph(graph_def, '', sys.argv[1]+".try")
|
||||
@@ -1,39 +0,0 @@
|
||||
#!/usr/bin/env python3
|
||||
# type: ignore
|
||||
|
||||
import os
|
||||
import time
|
||||
import numpy as np
|
||||
|
||||
import cereal.messaging as messaging
|
||||
from openpilot.system.manager.process_config import managed_processes
|
||||
|
||||
|
||||
N = int(os.getenv("N", "5"))
|
||||
TIME = int(os.getenv("TIME", "30"))
|
||||
|
||||
if __name__ == "__main__":
|
||||
sock = messaging.sub_sock('modelV2', conflate=False, timeout=1000)
|
||||
|
||||
execution_times = []
|
||||
|
||||
for _ in range(N):
|
||||
os.environ['LOGPRINT'] = 'debug'
|
||||
managed_processes['modeld'].start()
|
||||
time.sleep(5)
|
||||
|
||||
t = []
|
||||
start = time.monotonic()
|
||||
while time.monotonic() - start < TIME:
|
||||
msgs = messaging.drain_sock(sock, wait_for_one=True)
|
||||
for m in msgs:
|
||||
t.append(m.modelV2.modelExecutionTime)
|
||||
|
||||
execution_times.append(np.array(t[10:]) * 1000)
|
||||
managed_processes['modeld'].stop()
|
||||
|
||||
print("\n\n")
|
||||
print(f"ran modeld {N} times for {TIME}s each")
|
||||
for _, t in enumerate(execution_times):
|
||||
print(f"\tavg: {sum(t)/len(t):0.2f}ms, min: {min(t):0.2f}ms, max: {max(t):0.2f}ms")
|
||||
print("\n\n")
|
||||
Reference in New Issue
Block a user