[Mesa-stable] [Mesa-dev] [PATCH] st/mesa: add a fallback for clear_with_quad when no vs_layer

[Jose Fonseca]

On 17/11/14 13:32, Ilia Mirkin wrote:
> Well, nvidia hw (pre-GM204; it seems like it's supported there now)
> doesn't support setting gl_Layer from VS, so it's just not an option
> there. However I do believe that the HW has enough support to handle
> color masking as well as scissors, so I might try to add a PIPE_CAP
> which indicates that pipe->clear can do it all. [First I'd have to
> double-check that the nvidia hw really can support all this stuff
> properly.]

I forgot to say it explicitely, but I meant supporting this on llvmpipe 
(or more accurately, in draw module.)

> On the subject of strings -- yeah, parameterizing strings is annoying.
> When parameters are involved, using ureg_* seems like the better way
> to go. However for static shaders like this, it'll be hard to beat the
> readability of a plain string. Either way, I was just copying the way
> that the existing (gl_Layer-using) vertex shader was implemented.
>
> Perhaps the decision of whether to do it one way or another can be
> arrived at separately from this change?
Sure. I don't have any objection this going in as is, as it is 
consistent with what was done in there already.  It  was more a remark 
for future code.

Jose

>
>    -ilia
>
> On Mon, Nov 17, 2014 at 8:20 AM, Jose Fonseca <jfonseca at vmware.com> wrote:
>> Hi Marek,
>>
>>> First of all, sorry for breaking Draw yesterday.
>> No problem. draw_llvm.c was really obfuscating the fact that it relied on the layout of the pipe_viewport_state.
>>
>>> There's also the OpenGL extension AMD_vertex_shader_layer which is
>>> supported by all GS-capable Mesa drivers except for nouveau and
>>> llvmpipe I think.
>> I see. It sounds we should try to support these things from vs then.
>>
>>> tgsi_parse is something else. Did you mean tgsi_text_translate?
>> Yes, I meant tgsi_text_translate.
>>
>>> I
>>> don't like using ureg for helper shaders unless I have to. Shaders
>>> written with ureg are harder to read and the code is longer. I've been
>>> thinking of rewriting all simple shaders to text.
>> I don't think that's a good idea.
>>
>> Besides all the reasons I said before, string constants are not convenient to parameterize.  One needs to have multiple versions of the text file, or use snprintf. either way quite ugly.
>>
>> Regarding readability, if one has a comment before the code is just as readable. This is not what's done in u_simple_shaders.c, but we use this approach elsewhere and it works quite nicely:
>>
>>     [...]
>>
>>     /* DP3 TEMP.w, SRC, SRC */
>>     ureg_DP3(tbi->prog, temp_dst_w, src, src);
>>
>>     /* RSQ TEMP.w, TEMP.w */
>>     ureg_RSQ(tbi->prog, temp_dst_w, temp_src_w);
>>
>>     [...]
>>
>>
>> If people feel strongly about writing helper TGSI shaders in plain text, then we should create a simple off-line compilation tool that compiles the TGSI and creates *.h files (similar like DirectX HLSL tools).
>>
>>
>> Jose
>>
>> PS: Apologies for quoting only parts of the email. I'm sort of limited to using Outlook Web Access for email when working from home, and need to manually quote...
>>
>> ________________________________________
>> From: Marek Olšák <maraeo at gmail.com>
>> Sent: 16 November 2014 15:51
>> To: Jose Fonseca
>> Cc: Ilia Mirkin; mesa-dev at lists.freedesktop.org; 10.3 10.4
>> Subject: Re: [Mesa-dev] [PATCH] st/mesa: add a fallback for clear_with_quad when no vs_layer
>>
>> Hi Jose,
>>
>> First of all, sorry for breaking Draw yesterday.
>>
>> On Sun, Nov 16, 2014 at 2:57 PM, Jose Fonseca <jfonseca at vmware.com> wrote:
>>>> Fun fact -- llvmpipe also needs this.
>>> I think this is because this functionality was developed with D3D10 in mind, and
>>>
>>> https://urldefense.proofpoint.com/v2/url?u=http-3A__msdn.microsoft.com_en-2Dgb_library_windows_desktop_bb509647.aspx&d=AAIBaQ&c=Sqcl0Ez6M0X8aeM67LKIiDJAXVeAw-YihVMNtXt-uEs&r=zfmBZnnVGHeYde45pMKNnVyzeaZbdIqVLprmZCM2zzE&m=W_WiVknXaiEWMUhwH9NoKnxZ7SZQGp1B1VyLSBhv0rU&s=hwsRL-R8OHz5lpkwIvnmSRlHe-ijQAReKLFMkeVDWHM&e=
>>>
>>> states that SV_RenderTargetArrayIndex "an be written from the geometry shader and read by the pixel shader."
>> There's also the OpenGL extension AMD_vertex_shader_layer which is
>> supported by all GS-capable Mesa drivers except for nouveau and
>> llvmpipe I think.
>>
>>>
>>> BTW, coding these helper shaders as text is not a great idea IMO:
>>>
>>> - The tgsi_parse thing was written for debugging purposes mainly.
>>>
>>> - Whenever there are TGSI interface changes, any breakage with ureg will cause compile errors, where as it will only be detected at runtime with TGSI
>>>
>>> - ureg module makes it really easy to write shaders.  It's not really much more work.
>> tgsi_parse is something else. Did you mean tgsi_text_translate? I
>> don't like using ureg for helper shaders unless I have to. Shaders
>> written with ureg are harder to read and the code is longer. I've been
>> thinking of rewriting all simple shaders to text.
>>
>> Marek