Improve bitbang driver speed

martingingras9 · ‎2014-11-24

Posted on November 24, 2014 at 21:06

Hi,

I am using the STM32F2xx cube API to implement a raw bitbang usb driver for the FTDI ft245. I am currently achieving speed of 5000 bytes/s. I would like to improve the speed. The bitbang code is simple as:

<..>

__disable_irq();

// set data pins

HAL_GPIO_WritePin(FT_D0_PORT, FT_D0_PIN, (CHECK_BIT(data, 0) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

HAL_GPIO_WritePin(FT_D1_PORT, FT_D1_PIN, (CHECK_BIT(data, 1) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

HAL_GPIO_WritePin(FT_D2_PORT, FT_D2_PIN, (CHECK_BIT(data, 2) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

HAL_GPIO_WritePin(FT_D3_PORT, FT_D3_PIN, (CHECK_BIT(data, 3) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

HAL_GPIO_WritePin(FT_D4_PORT, FT_D4_PIN, (CHECK_BIT(data, 4) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

HAL_GPIO_WritePin(FT_D5_PORT, FT_D5_PIN, (CHECK_BIT(data, 5) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

HAL_GPIO_WritePin(FT_D6_PORT, FT_D6_PIN, (CHECK_BIT(data, 6) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

HAL_GPIO_WritePin(FT_D7_PORT, FT_D7_PIN, (CHECK_BIT(data, 7) == 0) ? GPIO_PIN_RESET : GPIO_PIN_SET);

__enable_irq();

<...>

The pins are configured as follow:

FT_DX_PORT, FT_DX_PIN, GPIO_MODE_INPUT, GPIO_NOPULL, GPIO_SPEED_HIGH

is there anything I can do or example I could follow or inspire myself? (beside implementing some asm?)

thanx

Mart

#bitband-bit-bang

Tesla DeLorean · ‎2014-11-24

Posted on November 25, 2014 at 04:40

Architecturally BSRR is a 32-bit register

The assignment would permitted a rotate, harder to do in C

Remember the interactions with the APB/AHB will be less efficient than operations on processor registers, and that peripheral registers are ''volatile'' and will be re-read, often unnecessarily with compound modification. Consider loading settings into a temporary register, and modify that before storing the finished value back.

Tips, Buy me a coffee, or three.. PayPal Venmo
Up vote any posts that you find helpful, it shows what's working..

mrjaner · ‎2014-11-24

Posted on November 25, 2014 at 08:38

why doesnt stm32cube have any usb samples for stm32f4-discovery? there is a FS USB port on

waclawek.jan · ‎2014-11-24

Posted on November 25, 2014 at 08:45

The barrel shift in ARM is ''free'', i.e. rotation can be added to any instruction without timing penalty. Thumb(2) places restrictions on this, but still rotation is quite effective.

This:

void foo(uint8_t a) __attribute__((noinline));

void foo(uint8_t a) {

uint32_t t;

t = a & 0x0F;

t = (t >> 2) | (t << (32 - 2)); // ROR 2

t = t | ((t >> 16) | (t << (32 - 16))); // + ROR 16

GPIOD->BSRR = t; // this needs augmented stm32f4xx.h

}

compiles in gcc into

800019c: f000 000f and.w r0, r0, #15

80001a0: f44f 6340 mov.w r3, #3072 ; 0xc00

80001a4: ea4f 00b0 mov.w r0, r0, ror #2

80001a8: f2c4 0302 movt r3, #16386 ; 0x4002

80001ac: ea40 4030 orr.w r0, r0, r0, ror #16

80001b0: 6198 str r0, [r3, #24]

80001b2: 4770 bx lr

which runs in 9 cycles including the ''call'' and ''ret''; and much of the time is consumed to load the register address which in real program would often get optimized for multiple accesses to the same peripheral.

JW

waclawek.jan · ‎2014-11-24

Posted on November 25, 2014 at 08:47

> why doesnt stm32cube have any usb samples for stm32f4-discovery? there is a FS USB port

on

1. don't hijack other threads, start your own

2. don't start it in this forum as it is a cube-related question - rather, start it in the dedicated [DEAD LINK /public/STe2ecommunities/mcu/Lists/STM32Java/AllItems.aspx]''software tools'' subforum

mrjaner · ‎2014-11-24

Posted on November 25, 2014 at 08:58

if you read my posts, i started but nobody answer to me. but i will start again

Tesla DeLorean · ‎2014-11-25

Posted on November 25, 2014 at 16:37

if you read my posts, i started but nobody answer to me. but i will start again

Forums depend on others to answer the posts, if no one present has an answer, it will be left unanswered.

Tips, Buy me a coffee, or three.. PayPal Venmo
Up vote any posts that you find helpful, it shows what's working..

martingingras9 · ‎2014-11-27

Posted on November 27, 2014 at 16:14

Hi,

I understand what Clive1 said about ''design ahead of time''. If all ports were aligned, it will be easy to write directly in the GPIOX->BSSR register. Now, according to the current state, if I have the value 0x6 to transmit:

Port D

0 1 1 0

PIN_1 PIN_0 PIN_15 PIN_14

I dont see how can I can do it with doing a stupid HAL_GPIO_WritePin for each like I am doing in order to save some cycle and improve performance...

Mart

waclawek.jan · ‎2014-11-27

Posted on November 27, 2014 at 17:00

> 0 1 1 0

> PIN_1 PIN_0 PIN_15 PIN_14

>

> I dont see how can I can do it with doing a stupid HAL_GPIO_WritePin for each like I am

> doing in order to save some cycle and improve performance...

But why would you use HAL_GPIO_WritePin(), once it's stupid?

I gave you the solution above, what is in it that is not clear?

JW

martingingras9 · ‎2014-12-01

Posted on December 01, 2014 at 20:35

Hi Mr. waclawek,

obviously, I dont understand the mechanic behind it.

However, testing the provided function, I ended up with a value of of 0x80018001 in ''t''. However, using HAL_GPIO_WritePin, I have 0x00008001 in the GPIOD->ODR register which is correct:

input: 0x46

Port E

0 1 0 0

PIN_7 PIN_8 PIN_9 PIN_10

Port D

0 1 1 0

PIN_1 PIN_0 PIN_15 PIN_14

Mart

martingingras9 · ‎2014-12-01

Posted on December 01, 2014 at 20:47

I made a mistake:

input: 0x46

Port E

0 1 0 0

PIN_10 PIN_9 PIN_8 PIN_7

Port D

0 1 1 0

PIN_1 PIN_0 PIN_15 PIN_14