[U-Boot] [PATCH 2/2] spi: mxc_spi: Update pre and post divider algorithm

[Troy Kisky <troy.kisky@boundarydevices.com>]

On 5/10/2013 12:08 PM, Dirk Behme wrote:
> Am 10.05.2013 20:44, schrieb Troy Kisky:
>> On 5/9/2013 10:34 PM, Dirk Behme wrote:
>>> Am 09.05.2013 20:00, schrieb Troy Kisky:
>>>> On 5/8/2013 10:19 PM, Dirk Behme wrote:
>>>>> The spi clock divisor is of the form x * (2**y),  or  x  << y, where
>>>>> x is
>>>>> 1 to 16, and y is 0 to 15. Note the similarity with floating point
>>>>> numbers.
>>>>> Convert the desired divisor to the smallest number which is >=
>>>>> desired divisor,
>>>>> and can be represented in this form. The previous algorithm chose a
>>>>> divisor
>>>>> which could be almost twice as large as needed.
>>>>>
>>>>> Signed-off-by: Troy Kisky <troy.kisky@boundarydevices.com>
>>>>> Signed-off-by: Dirk Behme <dirk.behme@gmail.com>
>>>>> ---
>>>>>   drivers/spi/mxc_spi.c |   27 ++++++++++++---------------
>>>>>   1 file changed, 12 insertions(+), 15 deletions(-)
>>>>>
>>>>> diff --git a/drivers/spi/mxc_spi.c b/drivers/spi/mxc_spi.c
>>>>> index 3e903b3..66c2ad8 100644
>>>>> --- a/drivers/spi/mxc_spi.c
>>>>> +++ b/drivers/spi/mxc_spi.c
>>>>> @@ -128,7 +128,7 @@ static s32 spi_cfg_mxc(struct mxc_spi_slave
>>>>> *mxcs, unsigned int cs,
>>>>>           unsigned int max_hz, unsigned int mode)
>>>>>   {
>>>>>       u32 clk_src = mxc_get_clock(MXC_CSPI_CLK);
>>>>> -    s32 pre_div = 1, post_div = 0, i, reg_ctrl, reg_config;
>>>>> +    s32 pre_div = 1, post_div = 0, reg_ctrl, reg_config;
>>>>
>>>> First, I'm totally fine with the patch as it is. I'm just going to
>>>> point out things you may want to change, or
>>>> send a follow-up patch.
>>>>
>>>> Here, no need to initialize pre_div, post_div, if you delete the  if
>>>> (clk_src > max_hz) below which is not needed.
>>>
>>> Hmm, why is it not needed?
>>>
>>> If you remove the if, you *always* do at least the computation
>>>
>>> pre_div = DIV_ROUND_UP(clk_src, max_hz);
>>> post_div = fls(pre_div - 1);
>>> if (post_div > 4) {
>>>
>>> I would think that doing the initialization and the if is much
>>> cheaper than always doing above computation, even if its not needed?
>>> I would keep the if.
>>>
>>>>>       u32 ss_pol = 0, sclkpol = 0, sclkpha = 0;
>>>>>       struct cspi_regs *regs = (struct cspi_regs *)mxcs->base;
>>>>> @@ -147,27 +147,24 @@ static s32 spi_cfg_mxc(struct mxc_spi_slave
>>>>> *mxcs, unsigned int cs,
>>>>>       reg_ctrl |=  MXC_CSPICTRL_EN;
>>>>>       reg_write(&regs->ctrl, reg_ctrl);
>>>>> -    /*
>>>>> -     * The following computation is taken directly from Freescale's
>>>>> code.
>>>>> -     */
>>>>>       if (clk_src > max_hz) {
>>>> This "if" can be removed.
>>>>
>>>>>           pre_div = DIV_ROUND_UP(clk_src, max_hz);
>>>> If you subtract -1 here instead of when you set the divisor register,
>>>> the logic becomes simpler
>>>>
>>>>
>>>> pre_div = DIV_ROUND_UP(clk_src, max_hz) - 1;
>>>>
>>>> or just
>>>>
>>>> pre_div = (clk_src - 1) / max_hz;
>>>>
>>>>> -        if (pre_div > 16) {
>>>>> -            post_div = pre_div / 16;
>>>>> -            pre_div = 16;
>>>>> -        }
>>>>> -        if (post_div != 0) {
>>>>> -            for (i = 0; i < 16; i++) {
>>>>> -                if ((1 << i) >= post_div)
>>>>> -                    break;
>>>>> -            }
>>>>> -            if (i == 16) {
>>>>> +        /* fls(1) = 1, fls(0x80000000) = 32, fls(16) = 5 */
>>>>> +        post_div = fls(pre_div - 1);
>>>>> +        if (post_div > 4) {
>>>>> +            post_div -= 4;
>>>>> +
>>>>> +            if (post_div >= 16) {
>>>>>                   printf("Error: no divider for the freq: %d\n",
>>>>>                       max_hz);
>>>>>                   return -1;
>>>>>               }
>>>>> -            post_div = i;
>>>>> +            pre_div = (pre_div + (1 << post_div) - 1) >> post_div;
>>>>
>>>> This would become
>>>> pre_div >>= post_div;
>>>>> +
>>>>> +        } else {
>>>>> +            post_div = 0;
>>>>>           }
>>>>> +
>>>>>       }
>>>>>       debug("pre_div = %d, post_div=%d\n", pre_div, post_div);
>>>>
>>>> And
>>>>
>>>> MXC_CSPICTRL_PREDIV(pre_div - 1);
>>>>
>>>> would return to
>>>>
>>>> MXC_CSPICTRL_PREDIV(pre_div);
>>>
>>> Well, where to do the -1 mainly depends on how we want to interpret
>>> the output of
>>>
>>> debug("pre_div = %d, post_div=%d\n", pre_div, post_div);
>> I'd change to
>>
>> debug("actual div = %d, pre_div = %d, post_div=%d\n", (pre_div + 1) <<
>> post_div, pre_div, post_div);
>>
>> to eliminate confusion.
>>
>>
>>>
>>> There are two options how to understand this, at least the pre_div =
>>> %d part:
>>>
>>> a) print the pre_div as the real divider used in a human readable
>>> format. I.e. if we divide by /15 then print 15
>>>
>>> or
>>>
>>> b) print the pre_div value we write to the register. I.e. if we
>>> divide by /15 then print 14
>>>
>>> Up to now we are doing (a) and calculate the register value after
>>> the debug print. Your proposal would switch this to (b). Anyway, if
>>> it makes the algorithm simpler I'd agree that it's fine to switch to
>>> (b).
>>>
>>> To summarize, this would become then:
>>>
>>> s32 pre_div = 1, post_div = 0, reg_ctrl, reg_config;
>>>
>>> ...
>>
>> If you keep the "if", pre_div = 0
>
> Ah, yes, thanks.
>
>> Also, I'd change s32 to unsigned or u32.
>>
>> But  usually the if will be true, so why keep it?
>
> For cases where clk_src <= max_hz. Then don't do the calculation 
> because it's unnecessary.

__udivsi3 already has this optimization.
fls can be 2 instructions
         clz    r0,r0
         rsb    r0,r0,#32

so your savings are a subroutine call and a return
in the unlikely case that (clk_src <= max_hz)

But I don't have a real objection to leaving it, so feel free to decide.

>
>> It is just code bloat and prone to errors like yours above.
>>
>> The only logic difference between keeping/killing the "if" is when
>> clk_src == 0.  Keeping will give a divide by 1. Killing will give an
>> error.
>> I'd argue that an error is more appropriate.
>>
>>>
>>> if (clk_src > max_hz) {
>>>     pre_div = (clk_src - 1) / max_hz;
>>>     /* fls(1) = 1, fls(0x80000000) = 32, fls(16) = 5 */
>>>     post_div = fls(pre_div);
>>>     if (post_div > 4) {
>>>         post_div -= 4;
>>>         if (post_div >= 16) {
>>>             printf("Error: no divider for the freq: %d\n", max_hz);
>>>             return -1;
>>>         }
>>>         pre_div >>= post_div;
>>>     } else {
>>>         post_div = 0;
>>>     }
>>> }
>>>
>>> ...
>>>
>>> MXC_CSPICTRL_PREDIV(pre_div);
>>>
>>> Is this correct?
>>>
>>> Best regards
>>>
>>> Dirk
>>
>>
>
>