Re: [PATCH 3/3] drm/i915: go back to switch for VLV mem freq detection

[Kenneth Graunke <kenneth@whitecape.org>]

On 05/01/2013 06:45 PM, Ben Widawsky wrote:
> On Wed, May 01, 2013 at 06:38:17PM -0700, Kenneth Graunke wrote:
>> On 05/01/2013 06:26 PM, Ben Widawsky wrote:
>>> On Wed, May 01, 2013 at 04:09:22PM -0700, Jesse Barnes wrote:
>>>> We need to catch the invalid case and override it.
>>>>
>>>> Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
>>
>> Jesse, it's actually worse than that:
>>
>> Your parenthesis were off in the previous version, causing it to do
>> the multiplication by 266 before the bitwise-and:
>>
>>     800 + ((266 * (val >> 6)) & 3)
>>
>> This meant that you always got 800 or 802 Mhz.
>>
>> Also, even when corrected, your previous formula would return 1332
>> Mhz instead of 1333 Mhz, which is slightly off and doesn't work
>> since you use it in explicit switch statements.
>>
>> It might be worth mentioning in the commit message.
>>
>>>> ---
>>>>   drivers/gpu/drm/i915/intel_pm.c |   16 +++++++++++++++-
>>>>   1 file changed, 15 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
>>>> index 0f4b46e..0a3b0b3 100644
>>>> --- a/drivers/gpu/drm/i915/intel_pm.c
>>>> +++ b/drivers/gpu/drm/i915/intel_pm.c
>>>> @@ -2902,7 +2902,21 @@ static void valleyview_enable_rps(struct drm_device *dev)
>>>>   		   GEN7_RC_CTL_TO_MODE);
>>>>
>>>>   	valleyview_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS, &val);
>>>> -	dev_priv->mem_freq = 800 + (266 * (val >> 6) & 3);
>>>> +	switch ((val >> 6) & 3) {
>>>> +	case 0:
>>>> +		dev_priv->mem_freq = 800;
>>>> +		break;
>>>> +	case 1:
>>>> +		dev_priv->mem_freq = 1066;
>>>> +		break;
>>>> +	case 2:
>>>> +		dev_priv->mem_freq = 1333;
>>>> +		break;
>>>> +	case 3:
>>>> +		DRM_ERROR("invalid mem freq, assuming 800MHz\n");
>>>> +		dev_priv->mem_freq = 800;
>>>> +		break;
>>>> +	}
>>>>   	DRM_DEBUG_DRIVER("DDR speed: %d MHz", dev_priv->mem_freq);
>>>>
>>>>   	DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & 0x10 ? "yes" : "no");
>>>
>>>
>>> I guess it doesn't handle the last case, but:
>>> dev_priv->mem_freq = 800 + (266 * (val))) + val/2
>>
>> Don't use this method.  Presumably Ben meant (val >> 6) here, but
>> even with that fix this formula returns 1067 rather than 1066, which
>> is wrong.
>
> You're right about the missing shift, but otherwise the formula is
> correct. Val is 1 in the case of 1066MHz...
>
>>
>>> or
>>>
>>> u32 freqs[] = {800,1066,1333,800};
>>> dev_priv->mem_freq = freqs[(val >> 6) & 3];
>>
>> I like this table based approach best.  It's much more succinct than
>> the switch statement, and exact.  Might be worth preserving the
>> DRM_ERROR.
>>
>>> Are two ways I would have used before making a switch block :P
>>>
>>> Just a thought, but perhaps we don't want to enable RPS if we can't
>>> reliably figure out the memory freq. (case 3)
>>>
>>> Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
>>
>> For either the switch or Ben's table:
>> Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
>>
>> Incidentally, my Punit returns 3 here, but I believe the memory is
>> actually 1066.  Not sure what the deal is.

You're right - my mistake, sorry.  I had been doing this in a python 
interpreter and accidentally put 0b01 for the first val and 0b10 for the 
second val.

--Ken