[Mesa-dev] [PATCH v2 1/1] intel: print performance debug messages to `logcat' on Android.

[Chad Versace]

On 10/04/2012 02:37 PM, Eric Anholt wrote:
> Chad Versace <chad.versace at linux.intel.com> writes:
> 
>> On 10/01/2012 04:25 PM, Oliver McFadden wrote:
>>> On Mon, Oct 01, 2012 at 02:55:31PM -0700, Kenneth Graunke wrote:
>>>> On 10/01/2012 02:35 PM, Eric Anholt wrote:
>>>>> Oliver McFadden <oliver.mcfadden at linux.intel.com> writes:
>>>>>> diff --git a/src/mesa/drivers/dri/intel/intel_context.c b/src/mesa/drivers/dri/intel/intel_context.c
>>>>>> index 25334da..7eb2706 100644
>>>>>> --- a/src/mesa/drivers/dri/intel/intel_context.c
>>>>>> +++ b/src/mesa/drivers/dri/intel/intel_context.c
>>>>>> @@ -744,6 +744,11 @@ intelInitContext(struct intel_context *intel,
>>>>>>     intelInitExtensions(ctx);
>>>>>>  
>>>>>>     INTEL_DEBUG = driParseDebugString(getenv("INTEL_DEBUG"), debug_control);
>>>>>> +
>>>>>> +#ifdef HAVE_ANDROID_PLATFORM
>>>>>> +   INTEL_DEBUG |= DEBUG_PERF;
>>>>>> +#endif /* HAVE_ANDROID_PLATFORM */
>>>>>> +
>>>>>>     if (INTEL_DEBUG & DEBUG_BUFMGR)
>>>>>>        dri_bufmgr_set_debug(intel->bufmgr, true);
>>>>>
>>>>> I don't see why the driver should be logging debug messages by default.
>>>>> You want to be able to actually ship this driver for non-developers.
>>>>
>>>> At least normally, this causes a _ton_ of spew which can actually
>>>> introduce significant amounts of printf overhead.
>>>>
>>>
>>> The objective is to inform the application developer when they are doing
>>> something stupid; most of them will not go through the trouble of
>>> looking up how to turn on Intel performance debugging.
>>>
>>> I would like to see this upstream, but I can understand if you would
>>> prefer this patch be carried internally.
>>>
>>> In any case, Android logcat is not exactly known for it's sparseness; in
>>> fact the opposite.  I have seen such stupidity as "Hi, I'm still
>>> connected to the 'FUBAR' network!" every few seconds.  :-(
>>
>> To second what Oliver said. Applications that spew firehose-fashion into
>> Android's logcat are not misbehaved. For example, each time the Dalvik VM
>> performs garbage collection it spews gc statistics into logcat. What's
>> considered the norm on desktop Linux is not necessarily the norm the Android.
>>
>> I approve of this patch's intent. (Though I did have an issue with it in another
>> email).
> 
> Except that we use DEBUG_PERF to signal that we should do potentially
> slow work to tell the developer about potential issues.  We don't want
> to limit driver performance to print messages when nobody's reading.

Ah, you're right. I've seen that occur when retrieving from the shader cache.
Now I'm convinced that it's a bad idea.