Re: [PATCH 09/37] clk: renesas: rzg2l: fix computation formula

[claudiu beznea <claudiu.beznea@tuxon.dev>]

Hi, Geert,

On 27.09.2023 11:00, Geert Uytterhoeven wrote:
> Hi Claudiu,
> 
> On Tue, Sep 26, 2023 at 4:44 PM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
>> On Tue, Sep 26, 2023 at 1:47 PM claudiu beznea <claudiu.beznea@tuxon.dev> wrote:
>>> On 14.09.2023 15:55, Geert Uytterhoeven wrote:
>>>> On Tue, Sep 12, 2023 at 6:52 AM Claudiu <claudiu.beznea@tuxon.dev> wrote:
>>>>> From: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com>
>>>>>
>>>>> According to hardware manual of RZ/G2L (r01uh0914ej0130-rzg2l-rzg2lc.pdf)
>>>>> the computation formula for PLL rate is as follows:
>>>>>
>>>>> Fout = ((m + k/65536) * Fin) / (p * 2^s)
>>>>>
>>>>> and k has values in range [-32768, 32767]. Dividing k by 65536 with
>>>>> integer variables leads all the time to zero. Thus we may have slight
>>>>> differences b/w what has been set vs. what is displayed. Thus,
>>>>> get rid of this and decompose the formula before dividing k by 65536.
>>>>>
>>>>> Fixes: ef3c613ccd68a ("clk: renesas: Add CPG core wrapper for RZ/G2L SoC")
>>>>> Signed-off-by: Claudiu Beznea <claudiu.beznea.uj@bp.renesas.com>
>>>>
>>>> Thanks for your patch!
>>>>
>>>>> --- a/drivers/clk/renesas/rzg2l-cpg.c
>>>>> +++ b/drivers/clk/renesas/rzg2l-cpg.c
>>>>> @@ -696,18 +696,22 @@ static unsigned long rzg2l_cpg_pll_clk_recalc_rate(struct clk_hw *hw,
>>>>>         struct pll_clk *pll_clk = to_pll(hw);
>>>>>         struct rzg2l_cpg_priv *priv = pll_clk->priv;
>>>>>         unsigned int val1, val2;
>>>>> -       unsigned int mult = 1;
>>>>> -       unsigned int div = 1;
>>>>> +       unsigned int div;
>>>>> +       u64 rate;
>>>>> +       s16 kdiv;
>>>>>
>>>>>         if (pll_clk->type != CLK_TYPE_SAM_PLL)
>>>>>                 return parent_rate;
>>>>>
>>>>>         val1 = readl(priv->base + GET_REG_SAMPLL_CLK1(pll_clk->conf));
>>>>>         val2 = readl(priv->base + GET_REG_SAMPLL_CLK2(pll_clk->conf));
>>>>> -       mult = MDIV(val1) + KDIV(val1) / 65536;
>>>>> +       kdiv = KDIV(val1);
>>>>>         div = PDIV(val1) << SDIV(val2);
>>>>>
>>>>> -       return DIV_ROUND_CLOSEST_ULL((u64)parent_rate * mult, div);
>>>>> +       rate = (u64)MDIV(val1) * parent_rate;
>>>>> +       rate += ((long long)parent_rate * kdiv) / 65536;
>>>>
>>>> As the division is a binary shift, you can use the mul_u64_u32_shr() helper,
>>>> and incorporate the sdiv shift at the same time:
>>>>
>>>>     rate += mul_u64_u32_shr(parent_rate, KDIV(val1), 16 + SDIV(val2));
>>
>>  [1]^
>>
>>>>
>>>> You can save a multiplication by premultiplying mdiv by 65536:
>>>>
>>>>     rate = mul_u64_u32_shr(parent_rate, (MDIV(val1) << 16)) + KDIV(val1),
>>>>                            16 + SDIV(val2));
>>
>> [2]^
>>
>>>
>>> Looking again at this: KDIV (aka DIV_K) could have negative values thus
>>> mul_u64_u32_shr() cannot be used here.
>>
>> That means you can indeed not use [1].

You're right. Thanks for the input!

>>
>> But you can still use [2], as MDIV() must be in the range 64..533[3],
>> so "(MDIV(val1) << 16)) + (s16)KDIV(val1)" is always positive.
>> Note that you do need the cast to s16 (which I had missed before), or
>> the intermediate variable kdiv of type s16 (like in your patch).
> 
> Or include the cast to a signed type in the definition of KDIV().
> 
> Gr{oetje,eeting}s,
> 
>                         Geert
>