[Mesa-dev] [PATCH 4/4] radeonsi: add instance divisor support

[Christian König]

Am 26.03.2013 15:56, schrieb Michel Dänzer:
> On Die, 2013-03-26 at 14:51 +0100, Christian König wrote:
>> From: Christian König <christian.koenig at amd.com>
>>
>> Signed-off-by: Christian König <christian.koenig at amd.com>
>> [...]
>> diff --git a/src/gallium/drivers/radeonsi/radeonsi_shader.h b/src/gallium/drivers/radeonsi/radeonsi_shader.h
>> index 9dae742..e09f297 100644
>> --- a/src/gallium/drivers/radeonsi/radeonsi_shader.h
>> +++ b/src/gallium/drivers/radeonsi/radeonsi_shader.h
>> @@ -111,13 +111,18 @@ struct si_shader {
>>   	unsigned		nr_cbufs;
>>   };
>>   
>> -struct si_shader_key {
>> -	unsigned		export_16bpc:8;
>> -	unsigned		nr_cbufs:4;
>> -	unsigned		color_two_side:1;
>> -	unsigned		alpha_func:3;
>> -	unsigned		flatshade:1;
>> -	float			alpha_ref;
>> +union si_shader_key {
>> +	struct {
>> +		unsigned	export_16bpc:8;
>> +		unsigned	nr_cbufs:4;
>> +		unsigned	color_two_side:1;
>> +		unsigned	alpha_func:3;
>> +		unsigned	flatshade:1;
>> +		float		alpha_ref;
>> +	} ps;
>> +	struct {
>> +		unsigned	instance_divisors[PIPE_MAX_ATTRIBS];
>> +	} vs;
>>   };
> This grows the shader key from 8 to 128 bytes. I don't suppose the
> instance divisors could be encoded in a more compact way? E.g. loading
> the divisor values from constants and only tracking which elements use a
> divisor in a bitmask in the key.

Considered that also, and I have two problems with that approach:
1. While immediates are converted to shifts & muls, dividing even by a 
constant in the shader isn't cheap.
2. I couldn't get the reciprocal instruction working on the first try 
(ok you fixed that).

How about storing only a byte for the instance_divisor? That limit's the 
divisor to a modulo of 256, but I don't think that would be so extremly 
bad. That would reduce the key to 32 bytes instead.

Christian.