[Intel-gfx] [PATCH 06/13] drm/i915: Use binary search when looking up forcewake domains

[Tvrtko Ursulin]

On 29/09/2016 17:16, Chris Wilson wrote:
> On Thu, Sep 29, 2016 at 04:35:49PM +0100, Tvrtko Ursulin wrote:
>> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>>
>> Instead of the existing linear seach, now that we have sorted
>> range tables, we can do a binary search on them for some
>> potential miniscule performance gain, but more importantly
>> for elegance and code size. Hopefully the perfomance gain is
>> sufficient to offset the function calls which were not there
>> before.
>>
>> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
>> ---
>>   drivers/gpu/drm/i915/intel_uncore.c | 28 ++++++++++++++++++++--------
>>   1 file changed, 20 insertions(+), 8 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/i915/intel_uncore.c b/drivers/gpu/drm/i915/intel_uncore.c
>> index bee1482a5ece..ae5edaea16f7 100644
>> --- a/drivers/gpu/drm/i915/intel_uncore.c
>> +++ b/drivers/gpu/drm/i915/intel_uncore.c
>> @@ -26,6 +26,7 @@
>>   #include "i915_vgpu.h"
>>   
>>   #include <linux/pm_runtime.h>
>> +#include <linux/bsearch.h>
>>   
>>   #define FORCEWAKE_ACK_TIMEOUT_MS 50
>>   
>> @@ -589,20 +590,31 @@ struct intel_forcewake_range
>>   	enum forcewake_domains domains;
>>   };
>>   
>> +static int fw_range_cmp(const void *key, const void *elt)
>> +{
>> +	struct intel_forcewake_range *entry =
>> +		(struct intel_forcewake_range *)elt;
>> +	u32 offset = (u32)((unsigned long)key);
>> +
>> +	if (offset < entry->start)
>> +		return -1;
>> +	else if (offset > entry->end)
>> +		return 1;
>> +	else
>> +		return 0;
>> +}
>> +
>>   static enum forcewake_domains
>>   find_fw_domain(u32 offset, const struct intel_forcewake_range *ranges,
>>   	       unsigned int num_ranges)
>>   {
>> -	unsigned int i;
>> -	struct intel_forcewake_range *entry =
>> -		(struct intel_forcewake_range *)ranges;
>> +	struct intel_forcewake_range *entry;
>>   
>> -	for (i = 0; i < num_ranges; i++, entry++) {
>> -		if (offset >= entry->start && offset <= entry->end)
>> -			return entry->domains;
>> -	}
>> +	entry = bsearch((void *)(unsigned long)offset, (const void *)ranges,
>> +			num_ranges, sizeof(struct intel_forcewake_range),
>> +			fw_range_cmp);
> How much for bsearch() to be turned into a generator macro?

By default it is a small code size win (128 bytes). It makes 
find_fw_domain a function with an inlined comparator (so one function 
call less per search iteration than using library bsearch) and inlines 
is_gen8_shadowed completely.

Forcing find_fw_domain to be fully inline adds approximately 1k.

I am not sure - you think it is worth doing some of the above? Function 
calls are supposed to be cheap so perhaps just with the default 
inlining, but then it is either pushing the core patch or having a local 
copy of a macro.

>> -	return -1;
>> +	return entry ? entry->domains : -1;
>>   }
> Looks ok, maybe pass in the default value to return if !entry, saves the
> double check.

It goes away later in the series so it is fine. I just wanted to be 
gradual so it is easy to review.

Regards,

Tvrtko