GCC rychle kopirovani pameti
Ondrej
ondrej-hwl na hawsys.cz
Sobota Červen 9 21:40:51 CEST 2018
A kdyby se dalo
struct { } a, b;
a = b?
Ondřej
Dne 9.6.2018 v 11:20 Jaroslav Buchta napsal(a):
> To je mi jasne, ze ten kod neni univerzalni ale je to urceno specialne
> pro kopirovani jedne struktury dlouhe cca 100B v casove kriticke ISR ,
> ktera je zarovnana na 4B a tim padem OK. Predpokladam tedy, ze
> instrukcim ldrd a strd neni potreba zarovnat adresu modulo 8 ? Ten
> druhy while ma byt samozrejme take if, ale prelozi se to stejne.
> memcpy trvalo tusim pres 10us a tato funkce to zvladne pod 3us (STM32F3)
>
> Dne 09.06.2018 v 8:07 Ondrej napsal(a):
>> Ještě jeden dodatek: co když bude adresa SRC 1 a adresa DST 3 a LEN
>> 15. Bude váš kód fungovat nebo vyhodí výjimku Unaligned memory access
>> nebo zkopíruje něco někam kam nemá?
>>
>>
>> Ondřej
>>
>>
>> Dne 9.6.2018 v 7:33 Ondrej napsal(a):
>>> Nevím jak GCC ale Visual Studio a Borland měli na memcopy také dost
>>> odlišný názor. Řešení VS bylo asi 3x rychlejší, ale ten ASM kód byl
>>> nepochopitelný - jako kdyby inicializoval nějaké DMA. A to stále
>>> bavíme o kódu pro Windows.
>>>
>>> Ondřej
>>>
>>> Dne 8.6.2018 v 12:09 Jaroslav Buchta napsal(a):
>>>> Resenim je dat fci fastcopy do jineho modulu ale stejne by me
>>>> zajimalo, jestli to je normalni.
>>>>
>>>> Dne 08.06.2018 v 12:01 Jaroslav Buchta napsal(a):
>>>>> Zdravim, dnes se zase nestiham divit, nevim, na cem je to jak
>>>>> zavisle ale bezna funkce memcpy vede jak v optimalizovanem tak v
>>>>> neoptimalizovanem prekladu k hloupemu kopirovani po 1B
>>>>>
>>>>> Zkusil jsem si napsat takovouto fci
>>>>>
>>>>> void fastcpy(void *dst, void *src, int len)
>>>>> {
>>>>> while (len >= 8)
>>>>> {
>>>>> *(uint64_t*)dst = *(uint64_t*)src;
>>>>> dst += 8;
>>>>> src += 8;
>>>>> len -= 8;
>>>>> }
>>>>> while (len >= 4)
>>>>> {
>>>>> *(uint32_t*)dst = *(uint32_t*)src;
>>>>> dst += 4;
>>>>> src += 4;
>>>>> len -= 4;
>>>>> }
>>>>> if (len >= 2)
>>>>> {
>>>>> *(uint16_t*)dst = *(uint16_t*)src;
>>>>> dst += 2;
>>>>> src += 2;
>>>>> len -= 2;
>>>>> }
>>>>> if (len != 0)
>>>>> {
>>>>> *(uint8_t*)dst = *(uint8_t*)src;
>>>>> }
>>>>> }
>>>>>
>>>>> ktera vede ke kyzenemu vysledku, je to vyrazne rychlejsi, ovsem
>>>>> POUZE, KDYZ NENI NASTAVENA OPTIMALIZACE O3, tam nejak prekladac
>>>>> usoudi, ze je to kopirovani bloku a vola memcpy... Je to vubec
>>>>> mozne, nebo delam nejakou chybu???
>>>>>
>>>>> Jak mam vynutit volani te funkce? (krome obchuzky, ze pro ni
>>>>> zakazu optimalizaci...)
>>>>>
>>>>>
>>>>> ////////////////////////////////
>>>>>
>>>>> Optimalizovany kod:
>>>>>
>>>>> 100 mccState.iB = (float)(mccState.bufAdc1[2]);
>>>>> 080004c2: vstr s14, [r4, #68] ; 0x44
>>>>> 080004c6: vstr s15, [r4, #72] ; 0x48
>>>>> 102 if (!mccStateCopy.f_LockCopy)
>>>>> 080004ca: bmi.n 0x8000400 <HAL_ADC_ConvCpltCallback+24>
>>>>> 080004cc: movs r2, #72 ; 0x48
>>>>> 080004ce: mov r1, r4
>>>>> 080004d0: add.w r0, r4, #76 ; 0x4c
>>>>> 080004d4: bl 0x8005e00 <memcpy>
>>>>> 72 *(uint32_t*)dst = *(uint32_t*)src;
>>>>>
>>>>>
>>>>> Neoptimalizovany kod:
>>>>>
>>>>> 101 mccState.iC = (float)(mccState.bufAdc1[3]);
>>>>> 08000528: vmov s15, r1
>>>>> 0800052c: vcvt.f32.u32 s15, s15
>>>>> 08000530: vstr s15, [r3, #72] ; 0x48
>>>>> 102 if (!mccStateCopy.f_LockCopy)
>>>>> 08000534: ldrb.w r3, [r3, #76] ; 0x4c
>>>>> 08000538: tst.w r3, #2
>>>>> 0800053c: bne.n 0x8000470 <HAL_ADC_ConvCpltCallback+24>
>>>>> 104 fastcpy((void *)&mccStateCopy, (void
>>>>> *)&mccState, sizeof(MCCOMSTATE));
>>>>> 0800053e: ldr r0, [pc, #16] ; (0x8000550
>>>>> <HAL_ADC_ConvCpltCallback+248>)
>>>>> 08000540: movs r2, #76 ; 0x4c
>>>>> 08000542: mov r1, r0
>>>>> 08000544: add r0, r2
>>>>> 08000546: bl 0x80003f4 <fastcpy>
>>>>>
>>>>>
>>>>> fastcpy:
>>>>> 080003f4: cmp r2, #7
>>>>> 080003f6: ble.n 0x800043a <fastcpy+70>
>>>>> 62 {
>>>>> 080003f8: push {r4, r5}
>>>>> 65 *(uint64_t*)dst = *(uint64_t*)src;
>>>>> 080003fa: ldrd r4, r5, [r1], #8
>>>>> 080003fe: strd r4, r5, [r0], #8
>>>>> 68 len -= 8;
>>>>> 08000402: subs r2, #8
>>>>> 63 while (len >= 8)
>>>>> 08000404: cmp r2, #7
>>>>> 08000406: bgt.n 0x80003fa <fastcpy+6>
>>>>> 08000408: b.n 0x8000414 <fastcpy+32>
>>>>> 72 *(uint32_t*)dst = *(uint32_t*)src;
>>>>> 0800040a: ldr.w r3, [r1], #4
>>>>> 0800040e: str.w r3, [r0], #4
>>>>> 75 len -= 4;
>>>>> 08000412: subs r2, #4
>>>>> 70 while (len >= 4)
>>>>> 08000414: cmp r2, #3
>>>>> 08000416: bgt.n 0x800040a <fastcpy+22>
>>>>> 77 if (len >= 2)
>>>>> 08000418: cmp r2, #1
>>>>> 0800041a: ble.n 0x8000426 <fastcpy+50>
>>>>> 79 *(uint16_t*)dst = *(uint16_t*)src;
>>>>> 0800041c: ldrh.w r3, [r1], #2
>>>>> 08000420: strh.w r3, [r0], #2
>>>>> 82 len -= 2;
>>>>> 08000424: subs r2, #2
>>>>> 84 if (len != 0)
>>>>> 08000426: cbz r2, 0x800042c <fastcpy+56>
>>>>> 86 *(uint8_t*)dst = *(uint8_t*)src;
>>>>> 08000428: ldrb r3, [r1, #0]
>>>>> 0800042a: strb r3, [r0, #0]
>>>>> 88 }
>>>>> 0800042c: pop {r4, r5}
>>>>> 0800042e: bx lr
>>>>> 08000430: ldr.w r3, [r1], #4
>>>>> 08000434: str.w r3, [r0], #4
>>>>> 75 len -= 4;
>>>>> 08000438: subs r2, #4
>>>>> 0800043a: cmp r2, #3
>>>>> 0800043c: bgt.n 0x8000430 <fastcpy+60>
>>>>> 0800043e: cmp r2, #1
>>>>> 08000440: ble.n 0x800044c <fastcpy+88>
>>>>> 08000442: ldrh.w r3, [r1], #2
>>>>> 08000446: strh.w r3, [r0], #2
>>>>> 82 len -= 2;
>>>>> 0800044a: subs r2, #2
>>>>> 84 if (len != 0)
>>>>> 0800044c: cmp r2, #0
>>>>> 0800044e: beq.n 0x800042e <fastcpy+58>
>>>>> 86 *(uint8_t*)dst = *(uint8_t*)src;
>>>>> 08000450: ldrb r3, [r1, #0]
>>>>> 08000452: strb r3, [r0, #0]
>>>>> 08000454: bx lr
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> HW-list mailing list - sponsored by www.HW.cz
>>>>> Hw-list na list.hw.cz
>>>>> http://list.hw.cz/mailman/listinfo/hw-list
>>>>
>>>>
>>>> _______________________________________________
>>>> HW-list mailing list - sponsored by www.HW.cz
>>>> Hw-list na list.hw.cz
>>>> http://list.hw.cz/mailman/listinfo/hw-list
>>>
>>> _______________________________________________
>>> HW-list mailing list - sponsored by www.HW.cz
>>> Hw-list na list.hw.cz
>>> http://list.hw.cz/mailman/listinfo/hw-list
>>
>> _______________________________________________
>> HW-list mailing list - sponsored by www.HW.cz
>> Hw-list na list.hw.cz
>> http://list.hw.cz/mailman/listinfo/hw-list
>
>
> _______________________________________________
> HW-list mailing list - sponsored by www.HW.cz
> Hw-list na list.hw.cz
> http://list.hw.cz/mailman/listinfo/hw-list
Další informace o konferenci Hw-list