ESP8266 HW Timer1 硬體計時器之規劃與改寫－MultiTimers
硬體計時器 hw_timer1，我們若不用它似乎對不起單晶片，在運用上也會綁手綁腳的甚至於無法正確實現我們要的功能。因此筆者就來改寫 hw_timer.c，之後，也再創建 PWM 於此計時函式的基礎上。當然，任何不是自己寫的函式庫用上 hw timer1 則會造成衝突，也就只能嘗試改寫第三方用到 timer1 的程式內容成使用此處的函式了。
目的
計時器只有一個，如何來讓多個不同功能取向的程式共用。例如三相馬達，必須三個不同相位的 PWM 來驅動；擴展 GPIO 必須輪詢每一個用上的 IO 通道；LCD display 也有所謂 refresh rate 來保持畫面或更新畫面，其他軟體定時器函式庫根源也是硬體計時器 timer1／沒查證／八九不離十，例如 Ticker。官方 non-os-sdk 其實已配置好 timer1，但應是只做半套，即，只將 timer1 用在 pwm 上面並稱它為虛擬的硬體 pwm，即，ESP8266 並無硬體 pwm 模組，但用硬體 timer1 來專用在軟體 pwm 上。並且此軟體 pwm 本身就已實作多個 pwm 共享一個 timer 的管理了。其實若當初有完善規劃現在也不會有 timer1 的相關使用疑問或衝突了。
故此處的目的，就是最大化共用 hw timer1。
程式碼列於下。
原則上用上較少的計時器及較大的週期會較準確。並且 milisecond 等級的會較準確；而 microsecond 的仍是稍準確。事實上，單單就是看前述原則。
而若用上較多的計時器，則當中最大週期的計時器誤差會較大，原因就是累積性誤差。
其次每一顆計時器也不會有固定誤差，這也是此實作品的先天限制。
簡單講，用於實務上並且於程式內有參數可調整誤差補償，則此多重計時器仍是相當實用準確的。
程式中的測試範例用上 9 個計時器，最多可設定 15 個，若改一下參數則可用上 2 的 9 次方 -1 顆計時器，只要記憶體夠用，當然這只是理論上及程式支援上，筆者未親試。而設計上，參數改為最多 6 顆才是符合最佳設計。
以本例而言，
[1310] [1320] [1330] [1340] [1350] [2360] [2370] [3380] [5390]
[1318] [1328] [1340] [1348] [1358] [2379] [2391] [3408] [5436]
誤差：
8us, 8us, 10us, 8us, 8us, 19us, 21us, 28us, 46us
相信使用者就知道該怎麼做了：用上較少（必要數量）的計時器；並於主系統中能共用同一顆計時器就共用之，即，將計時器的誤差分攤到於主程式中，則時間準確度將最大化。
#ifndef _c_HW_TIMER_
#define _c_HW_TIMER_
typedef enum {
DIVDED_BY_1 = 0, // timer clock
DIVDED_BY_16 = 4, // divided by 16
DIVDED_BY_256 = 8, // divided by 256
} time_predived_mode;
typedef enum { // timer interrupt mode
TM_LEVEL_INT = 1, // level interrupt
TM_EDGE_INT = 0, // edge interrupt
} time_int_mode;
typedef enum {
FRC1_SOURCE = 0,
NMI_SOURCE = 1,
} frc1_timer_source_type;
// not special, seems to prevent from calculation overflowing
// the t uses 1us
#define US_TO_RTC_TIMER_TICKS(t) \
((t) ? \
(((t) > 0x35A) ? \
(((t)>>2) * ((APB_CLK_FREQ>>4)/250000) + ((t)&0x3) * ((APB_CLK_FREQ>>4)/1000000)) : \
(((t) *(APB_CLK_FREQ>>4)) / 1000000)) : \
0)
#define FRC1_ENABLE_TIMER BIT7
#define FRC1_AUTO_LOAD BIT6
// multiple timers (max 6) share a single HW Timer.
// timers can be synced at the same time(but not suggest, rather, do in client) and can be offset(1.67s max).
// B syncs to A which means B's isr is going to issued once after B's sync call, at the same time by A's isr issuing(both timing still unaffect).
// B's offset sync call which means the moment A's isr issued, and then offset-value of time past would issue B's isr.
// so, once B's sync called, B's isr would not according to B's timing, once isr issued will back to normal timing(reloaded).
// since the timeout-reload and isr servicing have the payload around 4us in average(div-16 used).
// the lib constraints user to only between 30us(33.3KHz) to 1.67s(0.6Hz) a safe range, with 1us resolution.
// the timers timeout under a span of time of PAYLOAD_US(4us) will be served as well at a call of main isr.
// here is the scenario, suppose there is B timer within 4us to be served after A having served at this call,
// B will be served as well at this call with a delay. if C got to be served within 4us after B served,
// delay and serve C as well and so on. So there are cases stay at a main ISR call too long, but not too bad.
// so, the minimal period as a constraint for every timer and the max 6 timers could ease such situations.
// such a strategy keeps timing accuracy and minimally affect to entire main system.
// for efficiency, the hw counter reg is 23-bit, we have 9, say, 3 bits as for an identifier tailing to the counter value.
#define C_HW_TIMER_DEBUG 0
#define SYNC_WAITING 0 // if enabled, sync could avoid mem-fault however, timings could affected at a syncing
#define TIMER_NUM 9 // up to number of timers simul used(recommended max 6, allowable 15)
#define ID_MASK_VAL 0x0F // to extract id from counter value. note that id is from 1 to TIMER_NUM.(recommended 0x07)
#define ID_MASK_BIT 4 // number of bits to extract id from counter value(recommended 3)
#define PAYLOAD_US 4 // payload in us to substract(recommended 4us)
#define LO_BOUND_US 30
#define UP_BOUND_US 1670000
#define PAYLOAD_CNT ((PAYLOAD_US)*5) // payload in counts to substract, since 1us=5x0.2us; 1us has 5 counts
#define LO_BOUND_CNT ((LO_BOUND_US)*5)
#define UP_BOUND_CNT ((UP_BOUND_US)*5)
#define OP_GET_CNT(x) ((x)>>ID_MASK_BIT)
#define OP_GET_ID(x) ((x)&ID_MASK_VAL)
#define OP_MERGE_CNTID(cnt, id) (((cnt)<<ID_MASK_BIT)|(id))
#define OP_HEAP_TOP_CNT OP_GET_CNT(timer_obj.current[1])
#define OP_HEAP_TOP_ID OP_GET_ID(timer_obj.current[1])
#define INTRUDE_TIME_CNT 100 // the counts(20us*5) safely/enough-time for access data
#define CNT_TO_US(x) (((x)+2)/5)
#define US_TO_CNT(x) ((x)*5)
#if C_HW_TIMER_DEBUG
ICACHE_RAM_ATTR int hw_timer_overflow, hw_timer_stay_time_us1, hw_timer_stay_time_us;
#endif
#if SYNC_WAITING
/*ICACHE_RAM_ATTR */bool sync_spin=false; // do not know the attrs affection
#endif
class cHwTimer{
typedef void (*void_func1)();
typedef struct{
unsigned reload[TIMER_NUM+1]; // num of counts for each timer to be reloaded(payload is evaled).
// since we support one-shot, it's better as to use this array
// and the reload[0] to be always 0 for the procedure reloading 0,
// which automatically stop.
// simply to say, set reload[i] to 0(should use loc to deallocate) or
// the current[i] being reloaded from reload[0] yields current[i] stop counting.
unsigned current[TIMER_NUM+1]; // current nit counts for each timer to be reloaded(incl. payload is evaled).
// this array is dedicated for using in heap funcs,
// so, the current[0] is the current number of running timers.
// timers occupy current[1] to current[number_of_timers],
// which are current[1] to current[ current[0] ].
// note that this array is entirely maintained by heap funcs called in main isr,
// but there are cases we have to intrude on this array(danger).
void_func1 isr[TIMER_NUM+1]; // callback functions.
// the up two arrays have distinct purpose of the [0] element,
// so is the isr[0] which used for the current isr going to run.
// Besides, isr[i] equals zero uniquely denotes the timer slot is empty.
} sHwTimerObj;
static sHwTimerObj timer_obj;
static void main_isr();
static void default_isr();
static void min_heap_insertion(unsigned key, unsigned *t);
static unsigned min_heap_removal(unsigned *t);
char loc; // identifies loc in array, please note that the loc starts from 1 to TIMER_NUM.
// loc 0 is reserved for one-shot identification.
// loc -1 identifies for uninitialized value.
void_func1 func_for_pause;
void arm(unsigned period_us, void(*isr)(), bool periodically){
for (unsigned i=1; i<=TIMER_NUM; i++){
// please especially note that we use isr[i] to identify whether a timer slot is empty.
// and we use who uses reload[0] as to identify one-shot(the id field is 0 and reload from reload[0] which is 0)
if (!timer_obj.isr[i]){ // search for an empty slot
unsigned cnt=US_TO_RTC_TIMER_TICKS(period_us-PAYLOAD_US);
timer_obj.isr[i]=isr;
func_for_pause=isr;
cnt=OP_MERGE_CNTID(cnt, i);
if (periodically) timer_obj.reload[i]=cnt;
else timer_obj.reload[i]=i; // id without count is used to deallocate
loc=i;
min_heap_insertion(cnt, timer_obj.current);
return;
}
}
};
cHwTimer(){};
public:
cHwTimer(unsigned period_us, void(*isr)(), bool periodically=true): loc(-1), func_for_pause(default_isr){
if (!isr || period_us<LO_BOUND_US || period_us>UP_BOUND_US || (timer_obj.current[0]==TIMER_NUM)) return;
arm(period_us, isr, periodically);
if (!(RTC_REG_READ(FRC1_CTRL_ADDRESS)&FRC1_ENABLE_TIMER)){ // timer is disabled, regarded as a brand new use.
timer_obj.isr[0]=default_isr;
ETS_FRC_TIMER1_INTR_ATTACH(main_isr, NULL);
/////ETS_FRC_TIMER1_NMI_INTR_ATTACH(main_isr);
RTC_REG_WRITE(FRC1_CTRL_ADDRESS, FRC1_ENABLE_TIMER /*| FRC1_AUTO_LOAD*/ | DIVDED_BY_16 | TM_EDGE_INT);
TM1_EDGE_INT_ENABLE();
ETS_FRC1_INTR_ENABLE();
}
};
cHwTimer(const cHwTimer &a){loc=a.loc; func_for_pause=a.func_for_pause;};
bool Sync(const cHwTimer &ref, unsigned offset_delay_us){
// this func might fail, if so, affect nothing
// this func might fault, if so, mem-fault or wdt-reset
if (loc>0 && ref.loc>0){
// this step can be dangerous or long-hold because main isr may alter this array.
unsigned i, j;
j=10; // for at most delay 200us
do {
// there is a threshold time to ensure we have much time to do something.
// so we check the minimum time of the upcoming timer, which longer than 20us.
// in the following if-condition, we can use flag to prevent main isr from access array and fault,
// however main isr spin-waiting could disturb all timer timings, of course could wdt-reset.
#if SYNC_WAITING
sync_spin=true;
#endif
if ((i=timer_obj.current[0]) && (OP_HEAP_TOP_CNT>INTRUDE_TIME_CNT)){
unsigned m, n, z;
while (i){
z=OP_GET_ID(timer_obj.current[i]);
if (z==ref.loc) m=i;
else if (z==loc) n=i;
--i;
}
if ((z=OP_GET_CNT(timer_obj.current[m])+US_TO_CNT(offset_delay_us))<=UP_BOUND_CNT){
timer_obj.current[n]=OP_MERGE_CNTID(z+1, loc); // need not count on payload
#if SYNC_WAITING
sync_spin=false;
#endif
return true;
}
#if SYNC_WAITING
sync_spin=false;
#endif
}
#if SYNC_WAITING
sync_spin=false;
#endif
delayMicroseconds(20);
} while (--j);
}
return false;
};
bool isActive(){return (loc>0) && (timer_obj.reload[loc]>loc);};
void Pause(){
#if C_HW_TIMER_DEBUG
printf("\r\n\r\ntimer(%d) paused\r\n\r\n", loc);
#endif
if (isActive()) timer_obj.isr[loc]=default_isr;
};
void Resume(){
#if C_HW_TIMER_DEBUG
printf("\r\n\r\ntimer(%d) resumed\r\n\r\n", loc);
#endif
if (isActive()) timer_obj.isr[loc]=func_for_pause;
};
void Stop(){
#if C_HW_TIMER_DEBUG
printf("\r\n\r\ntimer(%d) stopped\r\n\r\n", loc);
#endif
if (loc>0){
timer_obj.reload[loc]=loc;
loc=-1;
}
};
~cHwTimer(){Stop();};
// the following 3 funcs are user functions for easy access,
// in addition, if you want to not bind to objects,
// you can comment off the destructor,
// such that timers can still run even objects are destructed.
// such case is suitable for one-shot; non-oneshot hence has to be stopped explicitly.
static void StopAll(){for (int i=1; i<=TIMER_NUM; i++) timer_obj.reload[i]=i;};
static void PauseAll(){for (int i=0; i<=TIMER_NUM; i++) timer_obj.isr[i]=default_isr;};
static void ResumeAll(){/*can not resume all*/};
};
ICACHE_RAM_ATTR cHwTimer::sHwTimerObj cHwTimer::timer_obj={
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
};
IRAM_ATTR void cHwTimer::default_isr(){
#if C_HW_TIMER_DEBUG
//printf("\r\n-\r\n");
#endif
}
IRAM_ATTR void cHwTimer::main_isr(){
#if C_HW_TIMER_DEBUG
hw_timer_stay_time_us1=micros();
static bool gate=false;
if (gate){
hw_timer_overflow++;
gate=false;
}
else gate=true;
#endif
#if SYNC_WAITING
while (sync_spin) yield();
#endif
unsigned discard_min, discard_min_cnt, i, j, diff;
#if 0 ///// read the old version to know about what new version does ///// old version
while (timer_obj.current[0] && (OP_HEAP_TOP_CNT<=PAYLOAD_CNT)){//////////////////////////////////
timer_obj.isr[0]();
////delayMicroseconds(CNT_TO_US(OP_HEAP_TOP_CNT));
diff=micros()+CNT_TO_US(OP_HEAP_TOP_CNT);/////
discard_min=min_heap_removal(timer_obj.current); // discard the used min one
timer_obj.isr[0]=timer_obj.isr[OP_GET_ID(discard_min)]; // update the isr, which will run next time
discard_min_cnt=OP_GET_CNT(discard_min); // it will join to calculation
for (i=timer_obj.current[0]; i; i--){ // update all counters
j=OP_GET_ID(timer_obj.current[i]); // get an id which needs to update
timer_obj.current[i]=OP_MERGE_CNTID(OP_GET_CNT(timer_obj.current[i])-discard_min_cnt, j);
}
// add the new one that just discard
i=timer_obj.reload[OP_GET_ID(discard_min)];
if (i>TIMER_NUM) min_heap_insertion(i, timer_obj.current); // i possibly be 0 or id which are for one-aspect tackling
else timer_obj.isr[OP_GET_ID(discard_min)]=0; // it must be the one-shot or stopped timer which must deallocated
if (micros()<(diff-2)) delayMicroseconds(diff-micros());/////
#if C_HW_TIMER_DEBUG
else hw_timer_overflow++;
#endif
}//////////////////////////////////
#else //////////////////////////////////////////////////////////////////// new version
if (timer_obj.current[0] && (OP_HEAP_TOP_CNT<=PAYLOAD_CNT)){///////////////////////////
discard_min_cnt=0;
while (timer_obj.current[0] && (OP_HEAP_TOP_CNT<=(PAYLOAD_CNT+discard_min_cnt))){
timer_obj.isr[0]();
diff=micros()+CNT_TO_US(OP_HEAP_TOP_CNT-discard_min_cnt); // substract past-time equals to amount time need to spend
discard_min=min_heap_removal(timer_obj.current); // discard the used min one
timer_obj.isr[0]=timer_obj.isr[OP_GET_ID(discard_min)]; // update the isr, which will run next time
discard_min_cnt=OP_GET_CNT(discard_min); // update the past-time that is going to be(to substract outside this loop)
j=OP_GET_ID(discard_min);
// add the new one that just discard
i=timer_obj.reload[j];
if (i>TIMER_NUM) min_heap_insertion(OP_MERGE_CNTID(OP_GET_CNT(i)+discard_min_cnt, OP_GET_ID(i)), timer_obj.current);
else timer_obj.isr[j]=0; // it must be the one-shot or stopped timer which must deallocated
if (micros()<(diff-2)) delayMicroseconds(diff-micros());
#if C_HW_TIMER_DEBUG
else hw_timer_overflow++;
#endif
}
for (i=timer_obj.current[0]; i; i--){ // update all counters
j=OP_GET_ID(timer_obj.current[i]); // get an id which needs to update
timer_obj.current[i]=OP_MERGE_CNTID(OP_GET_CNT(timer_obj.current[i])-discard_min_cnt, j);
}
}//////////////////////////////
#endif ////////////////////////////////////////////////////////////////////
if (timer_obj.current[0]){
RTC_REG_WRITE(FRC1_LOAD_ADDRESS, OP_HEAP_TOP_CNT); // load the current min asap
timer_obj.isr[0](); // service the last time of isr asap, which means needs to update var this time
discard_min=min_heap_removal(timer_obj.current); // discard the used min one
timer_obj.isr[0]=timer_obj.isr[OP_GET_ID(discard_min)]; // update the isr, which will run next time
discard_min_cnt=OP_GET_CNT(discard_min); // it will join to calculation
for (i=timer_obj.current[0]; i; i--){ // update all counters
j=OP_GET_ID(timer_obj.current[i]); // get an id which needs to update
timer_obj.current[i]=OP_MERGE_CNTID(OP_GET_CNT(timer_obj.current[i])-discard_min_cnt, j);
}
// add the new one that just discard
i=timer_obj.reload[OP_GET_ID(discard_min)];
if (i>TIMER_NUM) min_heap_insertion(i, timer_obj.current); // i possibly be 0 or id which are for one-aspect tackling
else timer_obj.isr[OP_GET_ID(discard_min)]=0; // it must be the one-shot or stopped timer which must deallocated
}
else { // if here to do, no timer remains, the last one timer which belongs to this isr has been deallocated properly
timer_obj.isr[0]();
timer_obj.isr[0]=default_isr; // since hw timer still counting and intring
RTC_REG_WRITE(FRC1_CTRL_ADDRESS, RTC_REG_READ(FRC1_CTRL_ADDRESS)&~(FRC1_ENABLE_TIMER));
TM1_EDGE_INT_DISABLE();
}
#if C_HW_TIMER_DEBUG
gate=false;
hw_timer_stay_time_us=micros()-hw_timer_stay_time_us1;
#endif
}
IRAM_ATTR void cHwTimer::min_heap_insertion(unsigned key, unsigned *t){ // t[0] must be the current size
unsigned i=++t[0], j=(i>>1);
for (; j && t[j]>key; t[i]=t[j], i=j, j>>=1);
t[i]=key;
// should use OP_GET_CNT(t[]) however, thinking about same count with diff id, A(111-001), B(111-000),
// A is larger than B(or say B is smaller than A), no matter which one larger,
// we ignore this case which still holds heap property since A==B in this fact. (in another words,)
// (either A or B is larger, will be maintained, however A==B need not maintained by heap even it's maintained.)
// and the other case A(111-uvw), B(110-xyz), no matter uvwxyz are, all lead to A>B which holds heap property.
// therefore we needn't use OP_GET_CNT(t[]) within this func.
};
IRAM_ATTR unsigned cHwTimer::min_heap_removal(unsigned *t){ // t[0] must be the current size
///if (!t[0]) return -1;
unsigned key=t[1];
unsigned j=2, i=1, k=t[0]--;
for (; j<k; i=j, j<<=1){
if (t[j]>t[j+1]) ++j;
t[i]=t[j];
}
for (j=(i>>1); j && t[j]>t[k]; t[i]=t[j], i=j, j>>=1);
t[i]=t[k];
return key;
};
#endif // _c_HW_TIMER_
/////////////////#include<ESP8266WiFi.h>system_update_cpu_freq(160);
unsigned x=0;
unsigned ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8, ay9;
void IRAM_ATTR user1(void)
{
static unsigned z=micros();
ay1=micros()-z;
z=micros();
x^=0x01;
}
void IRAM_ATTR user2(void)
{
static unsigned z=micros();
ay2=micros()-z;
z=micros();
x^=0x02;
}
void IRAM_ATTR user3(void)
{
static unsigned z=micros();
ay3=micros()-z;
z=micros();
x^=0x04;
}
void IRAM_ATTR user4(void)
{
static unsigned z=micros();
ay4=micros()-z;
z=micros();
x^=0x08;
}
void IRAM_ATTR user5(void)
{
static unsigned z=micros();
ay5=micros()-z;
z=micros();
x^=0x10;
}
void IRAM_ATTR user6(void)
{
static unsigned z=micros();
ay6=micros()-z;
z=micros();
x^=0x20;
}
void IRAM_ATTR user7(void)
{
static unsigned z=micros();
ay7=micros()-z;
z=micros();
x^=0x40;
}
void IRAM_ATTR user8(void)
{
static unsigned z=micros();
ay8=micros()-z;
z=micros();
x^=0x80;
}
void IRAM_ATTR user9(void)
{
static unsigned z=micros();
ay9=micros()-z;
z=micros();
x=0xFF;
}
void IRAM_ATTR user10(void)
{
printf("\r\n\r\none-shot\r\n\r\n");
}
cHwTimer test[9]={
cHwTimer(1310, user1),
cHwTimer(1320, user2),
cHwTimer(1330, user3),
cHwTimer(1340, user4),
cHwTimer(1350, user5),
cHwTimer(2360, user6),
cHwTimer(2370, user7),
cHwTimer(3380, user8),
cHwTimer(5390, user9)
};
void setup() {
Serial.begin(115200);
}
void loop() {
if (random(7)==1) test[random(9)].Stop(); // random(9) is 0 to 8
else if (random(13)==3) test[random(9)].Pause();
else if (random(5)==3) test[random(9)].Resume();
else if (random(2)==1) cHwTimer(1660000, user10, false);
//printf("\r\n\r\n<<%d>>\r\n\r\n", test[random(9)].Sync(test[0], 0));
delay(1000);
#if C_HW_TIMER_DEBUG
printf("[% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] overflow=%d, stay=%d, %X\r\n",\
ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8, ay9, hw_timer_overflow, hw_timer_stay_time_us, x);
#else
printf("[% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] [% 4u] overflow=%d, stay=%d, %X\r\n",\
ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8, ay9, 0, 0, x);
#endif
}
/* output
13:00:00.310 -> [ 0] [ 0] [ 0] [ 0] [ 0] [ 0] [ 0] [ 0] [ 0] overflow=0, stay=0, 0
13:00:01.303 -> [1315] [1327] [1337] [1347] [1357] [2381] [2390] [3409] [5439] overflow=0, stay=0, 40
13:00:02.329 -> [1319] [1329] [1337] [1347] [1357] [2379] [2387] [3409] [5446] overflow=0, stay=0, EB
13:00:03.322 -> [1320] [1334] [1335] [1350] [1362] [2382] [2394] [3415] [5442] overflow=0, stay=0, FD
13:00:04.316 -> [1319] [1330] [1339] [1350] [1361] [2384] [2392] [3421] [5446] overflow=0, stay=0, EE
13:00:05.342 -> [1321] [1329] [1341] [1352] [1359] [2373] [2396] [3411] [5439] overflow=0, stay=0, 20
13:00:06.336 -> [1320] [1328] [1338] [1348] [1358] [2385] [2393] [3416] [5448] overflow=0, stay=0, 9F
13:00:07.362 -> [1315] [1327] [1340] [1349] [1360] [2379] [2390] [3410] [5439] overflow=0, stay=0, 27
13:00:08.355 -> [1320] [1332] [1340] [1352] [1360] [2380] [2388] [3408] [5442] overflow=0, stay=0, 74
13:00:09.382 -> [1319] [1325] [1340] [1349] [1359] [2379] [2391] [3411] [5443] overflow=0, stay=0, 66
13:00:10.375 -> [1319] [1329] [1340] [1349] [1357] [2388] [2387] [3409] [5454] overflow=0, stay=0, DF
13:00:11.401 -> [1320] [1331] [1340] [1350] [1360] [2383] [2393] [3413] [5443] overflow=0, stay=0, CF
13:00:12.395 -> [1317] [1329] [1340] [1347] [1366] [2379] [2394] [3414] [5442] overflow=0, stay=0, 2F
13:00:13.388 -> [1318] [1328] [1340] [1348] [1358] [2379] [2391] [3408] [5436] overflow=0, stay=0, 64
13:00:14.414 -> [1315] [1330] [1340] [1343] [1353] [2379] [2385] [3405] [5429] overflow=0, stay=0, 3E
*/
另外下面再附上 15 個計時器的主程式，方便使用者評估與測試。
其中由亂數取時可看出，原則上使用 milisecond 等級的計時器，且只要不同餘（使得計時器們擠在短時間內都需要中斷處理），都可降低 main-isr 內 overflow。當然，避免不了的誤差就是 main-isr 內的運算耗時了。因此在此例中，15ms 的計時器，最大誤差來到約 0.35ms，誤差率 2.3%，若以平均來看甚至於更好，如附圖，表示 milisecond-timer 還是有相當的精確度的，並且可同時用上十五個計時器。此任務算順利完成了。
comment off the destructor is required///////////////////////////////////////
unsigned x=0;
unsigned ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8, ay9, ay10, ay11, ay12, ay13, ay14, ay15;
void IRAM_ATTR user1(void)
{
static unsigned z=micros();
ay1=micros()-z;
z=micros();
x^=0x01;
}
void IRAM_ATTR user2(void)
{
static unsigned z=micros();
ay2=micros()-z;
z=micros();
x^=0x02;
}
void IRAM_ATTR user3(void)
{
static unsigned z=micros();
ay3=micros()-z;
z=micros();
x^=0x04;
}
void IRAM_ATTR user4(void)
{
static unsigned z=micros();
ay4=micros()-z;
z=micros();
x^=0x08;
}
void IRAM_ATTR user5(void)
{
static unsigned z=micros();
ay5=micros()-z;
z=micros();
x^=0x10;
}
void IRAM_ATTR user6(void)
{
static unsigned z=micros();
ay6=micros()-z;
z=micros();
x^=0x20;
}
void IRAM_ATTR user7(void)
{
static unsigned z=micros();
ay7=micros()-z;
z=micros();
x^=0x40;
}
void IRAM_ATTR user8(void)
{
static unsigned z=micros();
ay8=micros()-z;
z=micros();
x^=0x80;
}
void IRAM_ATTR user9(void)
{
static unsigned z=micros();
ay9=micros()-z;
z=micros();
x^=0x100;
}
void IRAM_ATTR user10(void)
{
static unsigned z=micros();
ay10=micros()-z;
z=micros();
x^=0x200;
}
void IRAM_ATTR user11(void)
{
static unsigned z=micros();
ay11=micros()-z;
z=micros();
x^=0x400;
}
void IRAM_ATTR user12(void)
{
static unsigned z=micros();
ay12=micros()-z;
z=micros();
x^=0x800;
}
void IRAM_ATTR user13(void)
{
static unsigned z=micros();
ay13=micros()-z;
z=micros();
x^=0x1000;
}
void IRAM_ATTR user14(void)
{
static unsigned z=micros();
ay14=micros()-z;
z=micros();
x^=0x2000;
}
void IRAM_ATTR user15(void)
{
static unsigned z=micros();
ay15=micros()-z;
z=micros();
x^=0x4000;
//printf("\r\n\r\none-shot\r\n\r\n");
}
void (*test[15])()={
user1, user2, user3, user4, user5, user6, user7, user8, user9, user10, user11, user12, user13, user14, user15
};
void setup() {
Serial.begin(115200);
}
void loop() {
static unsigned times[15];
if (random(13)==11){
// change all timers' timing to all non-oneshot
printf("\r\n-=change=-\r\n");
cHwTimer::StopAll();
delay(1700);
for (int i=0; i<15; i++){
times[i]=random(15000-30)+30; // 15ms at most
cHwTimer(times[i], test[i]);
}
}
#if C_HW_TIMER_DEBUG
printf("\r\n[overflow=%d, stay=%d, isr_run=%X]\r\n", hw_timer_overflow, hw_timer_stay_time_us, x);
#endif
printf("[% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u][% 5u]\r\n",\
times[0], times[1], times[2], times[3], times[4], times[5], times[6], times[7],\
times[8], times[9], times[10], times[11], times[12], times[13], times[14]);
printf("[% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d][% 5d]\r\n",\
ay1-times[0], ay2-times[1], ay3-times[2], ay4-times[3], ay5-times[4], ay6-times[5], ay7-times[6], ay8-times[7],\
ay9-times[8], ay10-times[9], ay11-times[10], ay12-times[11], ay13-times[12], ay14-times[13], ay15-times[14]);
// the % cal is diff from the upper two line is because ayi are altered by isri instantly.
printf("[% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%]"\
"[% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%][% 3.1f%%]\r\n",\
(ay1-times[0])/(float)(times[0])*100.0f, (ay2-times[1])/float(times[1])*100.0f, (ay3-times[2])/float(times[2])*100.0f,\
(ay4-times[3])/float(times[3])*100.0f, (ay5-times[4])/float(times[4])*100.0f, (ay6-times[5])/float(times[5])*100.0f,\
(ay7-times[6])/float(times[6])*100.0f, (ay8-times[7])/float(times[7])*100.0f, (ay9-times[8])/float(times[8])*100.0f,\
(ay10-times[9])/float(times[9])*100.0f, (ay11-times[10])/float(times[10]*100.0f), (ay12-times[11])/float(times[11])*100.0f,\
(ay13-times[12])/float(times[12])*100.0f, (ay14-times[13])/float(times[13])*100.0f, (ay15-times[14])/float(times[14])*100.0f);
delay(1500);
}
結論
-> [overflow=0, stay=22, isr_run=5FC4]
-> [ 50][ 92][ 32][ 75][ 66][ 79][ 73][ 33][ 80][ 87][ 41][ 80][ 56][ 76][ 78]
-> [ 93][ 221][ 95][ 163][ 119][ 176][ 192][ 63][ 180][ 175][ 88][ 180][ 112][ 141][ 161]
->
-> [overflow=0, stay=22, isr_run=7D7A]
-> [ 50][ 92][ 32][ 75][ 66][ 79][ 73][ 33][ 80][ 87][ 41][ 80][ 56][ 76][ 78]
-> [ 139][ 177][ 86][ 166][ 161][ 174][ 166][ 30][ 160][ 199][ 99][ 182][ 77][ 153][ 162]
讓我們接著來看上面這個實驗結果。
用上了 15 顆計時器，並且，週期全都在 30us 至 100us 之間，這表示平均每 4.7us 便有一顆計時器需要服務，誤差竟來到了三倍XD。而 overflow 計數有 main-isr 的函式內及函式外，但已將函式內的計數部份關掉。
因此可看出，（此函式庫是使用 FRC1，並且有無 auto-reload 都一樣），一旦 counter reg 時間到中斷觸發，中斷旗標會舉起來，但 isr 並不會被岔斷，會等到離開 isr 後，檢查出又有發生了中斷，便再次進入 isr。這表明 ESP8266 將不斷地跑 isr，所有 CPU 時間全被 isr 佔用。但事實是 loop() 仍能打印，故可以肯定，在每一次作 isr 的服務之前，必定會跑一次 loop()（並且應會有 timeout 機制或掌控權轉移機制）。因此如此的行為全都是 ESP8266 boot-loader 在掌控著，亦即又表明兩件事，bootloader 是必要的，等同一個小型，standalone 的 os，及選用好的 bootloader 對於單晶片的效能有著關鍵性的影響。
其次，筆者曾試著超頻到 160MHz 圖降低 main-isr 的運算的影響，但似乎都一樣。因留待有主程式系統後再調試為佳，故此舉並不能作為改善計時準確度的主要考量。
故結論還是回到那句老話，軟體計時器的使用原則。
最後，例如需要 500us 的計時器 5 顆，作法之一我們可以想定，若時間精準度要 99%，則誤差是 500us x 0.01 得 5us，顯然不可能。在先天限制下，500us / 30us 得 16.7。那就用上此函式庫所提供的最小的 30us 的時基，在 sub-isr 內查看歴時及計數累加，到 500us – 30us 前後時，判斷是要停等觸發或下次觸發，如此以求更高的精確度。各位認為這樣可行嗎？不用跑實驗結論還是不可行，在要求精確度的前提下，一來，要停等觸發，必拖累所有 timers，別忘了最多會等上 30us 以上。要下次觸發，必又亂飄了。因此，筆者又有想法了，便是追加一顆必然的最小計時器。只要有使用者的計時器開始跑了，這顆最小計時器便會開始跑。並用此計時器來補償時間的誤差。該怎麼做來達到所有使用者計時器都有相當的準確度，就等待筆者下篇文章（如果有的話XD）。當前筆者也沒想法XD。
不過好消息是，如下用上五顆 475us~500us 之間的計時器，誤差是 3%，所以本文這個函式庫還是令筆者放心的。
11:39:05.136 -> [overflow=1400650, stay=15, isr_run=1A]
11:39:05.136 -> [ 481][ 495][ 481][ 492][ 491]
11:39:05.169 -> [ 14][ 15][ 14][ 14][ 14]
11:39:06.659 ->
11:39:06.659 -> [overflow=1404060, stay=15, isr_run=1D]
11:39:06.659 -> [ 481][ 495][ 481][ 492][ 491]
11:39:06.692 -> [ 14][ 14][ 13][ 13][ 14]
11:39:08.182 ->
11:39:08.182 -> [overflow=1407465, stay=7, isr_run=7]
11:39:08.182 -> [ 481][ 495][ 481][ 492][ 491]
11:39:08.215 -> [ 14][ 13][ 14][ 14][ 14]
11:39:09.704 ->
11:39:09.704 -> [overflow=1410869, stay=7, isr_run=1F]
11:39:09.704 -> [ 481][ 495][ 481][ 492][ 491]
11:39:09.737 -> [ 17][ 14][ 17][ 14][ 14]
補充
- 後來看到有玩家也寫了這樣的函式庫，故附了上來以參考改進此處不足之處。
- https://forum.arduino.cc/t/esp8266-timerinterrupt-library/637769
- https://github.com/khoih-prog/ESP8266TimerInterrupt
20220316 補充
- hardware timer 的改版如後面文章。
- 切記，若使用上 print，經筆者辛苦地除錯及實測，將會使 hardware timer 暫停 200ms。
- 也有會其他的東東影響 hardware timer 的，筆者未知，猜測有可能是連續的不同的 gpio 中斷且皆尚未完成所致，特例的實測下影響約在 2ms 左右。當然此情況也有可能是如同 print 一樣是某函式庫於某段 critical code 中關開 timer int 所致。
- 因此我們使用 timer 處理程式時須謹防此類例外；除了後面文章中提到的 timer int + gpio int 的做法可利用（但非解法）外似乎只能 case by case 地處置了，而且很有可能解決不了；即，一來要用計時中斷去處理某些運算的需求，一來又要避免（不了）所製造出此需求者同時去干擾中斷。
- 別處的 delay() 不會影響 timer int。
delayMicroseconds() 則會。
發佈留言