[Mesa-dev] [PATCH] glsl: do not reset prog->TransformFeedback.BufferStride

[Timothy Arceri]

On 22/06/17 02:41, Juan A. Suarez Romero wrote:
> On Wed, 2017-06-21 at 20:24 +1000, Timothy Arceri wrote:
>> On 21/06/17 18:13, Juan A. Suarez Romero wrote:
>>> link_xfb_stride_layout_qualifiers() can be called multiple times, and
>>> each time we call prog->TransformFeedback.BufferStride is reset to 0.
>>>
>>> Thus it is loosing the values set in previous call.
>>>
>>> Do not perform such reset.
>>>
>>> Fixes:
>>> KHR-GL45.enhanced_layouts.xfb_stride_of_empty_list
>>> KHR-GL45.enhanced_layouts.xfb_stride_of_empty_list_and_api
>>>
>>> Signed-off-by: Juan A. Suarez Romero <jasuarez at igalia.com>
>>> ---
>>>    src/compiler/glsl/linker.cpp | 4 ----
>>>    1 file changed, 4 deletions(-)
>>>
>>> diff --git a/src/compiler/glsl/linker.cpp b/src/compiler/glsl/linker.cpp
>>> index adfa3b7b1d..1fe0ccc496 100644
>>> --- a/src/compiler/glsl/linker.cpp
>>> +++ b/src/compiler/glsl/linker.cpp
>>> @@ -1623,10 +1623,6 @@ link_xfb_stride_layout_qualifiers(struct gl_context *ctx,
>>>                                      struct gl_shader **shader_list,
>>>                                      unsigned num_shaders)
>>>    {
>>> -   for (unsigned i = 0; i < MAX_FEEDBACK_BUFFERS; i++) {
>>> -      prog->TransformFeedback.BufferStride[i] = 0;
>>> -   }
>>
>> I think this probably needs to be moved rather than just deleted. I
>> think the idea is to reset things to 0 in case we are re-linking an
>> existing program with different shaders attached.
>>
> 
> Checking how this was done prior the refactoring in 4d65f68 (mesa/glsl:
> move TransformFeedbackBufferStride to gl_shader), actually the reset
> should be done if the new shaders provide an explicit xfb_stride.
> Otherwise we should kept the old value.
> 

No, that's not how it worked. The related change in 4d65f68 was:

     for (unsigned i = 0; i < MAX_FEEDBACK_BUFFERS; i++) {
-      linked_shader->info.TransformFeedback.BufferStride[i] = 0;
+      prog->TransformFeedback.BufferStride[i] = 0;
     }

There was no check previously we just reset everything.

The problem is that previously each stage had its own copy of a 
BufferStride array. There was no need for this as xfb only applies to 
the last stage in the pipeline before the fragment shader.

4d65f68 changed it so that we share a single BufferStride array across 
the program instead, this allowed simplifications elsewhere and reduced 
the amount of data needed for shader cache. I assume the issue is that 
if we have a fragment shader attached then the array would get 
incorrectly reset.

Please see my reply to v2 for a suggested fix. Please let me know if 
that doesn't fix it.


> I'm sending a new version.
> 
> 
>>> -
>>>       for (unsigned i = 0; i < num_shaders; i++) {
>>>          struct gl_shader *shader = shader_list[i];
>>>    
>>>
>>
>>