On Thu, 22 Sep 2016 11:43:08 +0200
Boris Brezillon [off-list ref] wrote:

On Thu, 22 Sep 2016 09:39:04 +0200
Boris Brezillon [off-list ref] wrote:

quoted

On Thu, 22 Sep 2016 09:07:46 +0200
Uwe Kleine-K?nig [off-list ref] wrote:
  

quoted

On Wed, Sep 21, 2016 at 12:44:14PM +0200, Nicolas Ferre wrote:    

quoted

From: Alexey Starikovskiy <redacted>

The problem with previous code was it rounded values in wrong
place and produced wrong baud rate in some cases.

Signed-off-by: Alexey Starikovskiy <redacted>
[nicolas.ferre at atmel.com: port to newer kernel and add commit log]
Signed-off-by: Nicolas Ferre <redacted>
---
 drivers/tty/serial/atmel_serial.c | 10 ++++++----
 include/linux/atmel_serial.h      |  1 +
 2 files changed, 7 insertions(+), 4 deletions(-)

diff --git a/drivers/tty/serial/atmel_serial.c b/drivers/tty/serial/atmel_serial.c
index 5f550d9feed9..fd8aa1f4ba78 100644
--- a/drivers/tty/serial/atmel_serial.c
+++ b/drivers/tty/serial/atmel_serial.c

@@ -2170,13 +2170,15 @@ static void atmel_set_termios(struct uart_port *port, struct ktermios *termios,
 	 * accurately. This feature is enabled only when using normal mode.
 	 * baudrate = selected clock / (8 * (2 - OVER) * (CD + FP / 8))
 	 * Currently, OVER is always set to 0 so we get
-	 * baudrate = selected clock (16 * (CD + FP / 8))
+	 * baudrate = selected clock / (16 * (CD + FP / 8))
+	 * then
+	 * 8 CD + FP = selected clock / (2 * baudrate)
 	 */
 	if (atmel_port->has_frac_baudrate &&
 	    (mode & ATMEL_US_USMODE) == ATMEL_US_USMODE_NORMAL) {
-		div = DIV_ROUND_CLOSEST(port->uartclk, baud);
-		cd = div / 16;
-		fp = DIV_ROUND_CLOSEST(div % 16, 2);
+		div = DIV_ROUND_CLOSEST(port->uartclk, baud * 2);
+		cd = div >> 3;
+		fp = div & ATMEL_US_FP_MASK;      

given baud = 115200 and uartclk = 5414300 this results in:

	div = DIV_ROUND_CLOSEST(5414300, 115200 * 2) = 23
	cd = 2
	fp = 7    

How about:

	div = DIV_ROUND_CLOSEST(port->uartclk, baud);
	cd = div / 16;
	fp = (div % 16) / 2;

	best_baud = port->uartclk / ((16 * cd) +  (8 * fp));

	/* Check if we can get a better approximation by rounding up. */
	if (div % 2) {
		int alt_baud, alt_fp, alt_cd;

		alt_fp = fp++;
		alt_cd = cd;
		if (alt_fp > 7) {
			alt_cd++;
			alt_fp = 0;
		}

		alt_baud = port->uartclk / ((16 * alt_cd) +  (8 *alt_fp));
		if (abs(best_baud - baud) > abs(alt_baud - baud)) {  

After a lengthy discussion that happened on IRC (#armlinux), Uwe
proved me wrong. This should actually be


		/*
		 * Calculate the Error in the time domain:
		 * Error = (RealBaudPeriod - ExpectedBaudPeriod) /
		 *	   ExpectedBaudPeriod;
		 *
		 * which after conversion to the frequency domain gives:
		 * Error = 1 - (ExpectedBaudRate/RealBaudRate);
		 *
		 * and since we want to compare 2 errors and avoid
		 * approximation, we have:
		 *
		 * if (RealBaudRate2 * (RealBaudRate1 - ExpectedBaudRate) <
		 *     RealBaudRate1 * (RealBaudRate2 - ExpectedBaudRate))
		 *	...
		 * 
		 */
		if (alt_baud * abs(best_baud - baud) >
		    best_baud * abs(alt_baud - baud))

Thanks for your patience ;-).

quoted

			best_baud = alt_baud;
			fp = alt_fp;
			cd = alt_cd;
		}
	}
  

quoted

which yields a rate of 5414300 / 46 = 117702.17. With cd = 3 and fp = 0
however the resulting rate is 5414300 / 48 = 112797.92.

Which one is better?

Okay, it seems I was wrong here. It appears that 117702.17 is better
than 112797.92, and Alexey's patch is calculating the best cd and fp
values for all cases.