[Mesa-dev] [PATCH 16/78] i965/nir/vec4: Implement store_output intrinsic

[Eduardo Lima Mitev]

On 07/23/2015 05:20 AM, Jason Ekstrand wrote:
> On Wed, Jul 22, 2015 at 4:37 AM, Eduardo Lima Mitev <elima at igalia.com> wrote:
>> On 07/13/2015 01:57 PM, Jason Ekstrand wrote:
>>> On Wed, Jul 8, 2015 at 11:54 PM, Eduardo Lima Mitev <elima at igalia.com> wrote:
>>>> On 06/30/2015 06:51 PM, Jason Ekstrand wrote:
>>>>> On Fri, Jun 26, 2015 at 1:06 AM, Eduardo Lima Mitev <elima at igalia.com> wrote:
>>>>>> The index into the output_reg array where to store the destination register is
>>>>>> fetched from the nir_outputs map built during nir_setup_outputs stage.
>>>>>>
>>>>>> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=89580
>>>>>> ---
>>>>>>  src/mesa/drivers/dri/i965/brw_vec4_nir.cpp | 17 +++++++++++++++--
>>>>>>  1 file changed, 15 insertions(+), 2 deletions(-)
>>>>>>
>>>>>> diff --git a/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp b/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
>>>>>> index 8a2d335..55d4490 100644
>>>>>> --- a/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
>>>>>> +++ b/src/mesa/drivers/dri/i965/brw_vec4_nir.cpp
>>>>>> @@ -520,10 +520,23 @@ vec4_visitor::nir_emit_intrinsic(nir_intrinsic_instr *instr)
>>>>>>     }
>>>>>>
>>>>>>     case nir_intrinsic_store_output_indirect:
>>>>>> +      has_indirect = true;
>>>>>>        /* fallthrough */
>>>>>> -   case nir_intrinsic_store_output:
>>>>>> -      /* @TODO: Not yet implemented */
>>>>>> +   case nir_intrinsic_store_output: {
>>>>>> +      int offset = instr->const_index[0];
>>>>>> +      int output = nir_outputs[offset];
>>>>>> +
>>>>>> +      src = get_nir_src(instr->src[0], nir_output_types[offset]);
>>>>>> +      dest = dst_reg(src);
>>>>>> +
>>>>>> +      dest.writemask = brw_writemask_for_size(instr->num_components);
>>>>>> +
>>>>>> +      if (has_indirect)
>>>>>> +         dest.reladdr = new(mem_ctx) src_reg(get_nir_src(instr->src[1]));
>>>>>> +
>>>>>> +      output_reg[output] = dest;
>>>>>
>>>>> I'm very confused about the amount of indirection going on here.  It
>>>>> seems to me that we should be setting these outputs up in
>>>>> setup_outputs() rather than storring off a map from ints to other ints
>>>>> and setting it up here.  I didn't make this comment on the patch for
>>>>> setup_outputs() because I wanted to wait to see it used before I
>>>>> commented on it.
>>>>>
>>>>> I'm guessing you did it this way because the nir_assign_var_locations
>>>>> is giving you bogus values.  If so, then it might be better to just
>>>>> assign variable locations in setup_outputs() rather than having a
>>>>> remap table.  The whole point of nir_lower_io is to make IO easy for
>>>>> the back-end.  If you need a re-map table, then it's no longer making
>>>>> it easy and we need to think more about what's going on.
>>>>> --Jason
>>>>>
>>>>
>>>> That double indirection felt bad since the beginning, but it was needed
>>>> to store the original variable's location (var->data.location). Let me
>>>> explain:
>>>>
>>>> We are (re)using the plumbering in vec4_visitor to setup URB, so the
>>>> only thing we need to do is to store the out register in "output_reg"
>>>> map at the correct location. And that location is given by the original
>>>> location in the shader (var->data.location).
>>>>
>>>> So, in this case, "nir_assign_var_locations" pass, which constructs
>>>> var->data.driver_location, is not useful to us, except to give us
>>>> consecutive indexes to construct the other map we have, the type map,
>>>> which is needed to carry the correct type from the original variable to
>>>> the output register.
>>>
>>> If nir_assign_var_locations isn't doing anything for you, don't call
>>> it.  You'll need to do something with var->data.driver_location.  If
>>> what you really want is var->data.location, then just copy that to
>>> var->data.driver_location when you do nir_setup_outputs.  Or
>>> (depending on how the URB setup works, I don't actually know), put the
>>> actual URB location in var->data.driver_location when you walk the
>>> outputs.
>>>
>>> From there, you have two options.  One would be to setup output_reg at
>>> the same time with the correct types right away and emit a MOV when
>>> you get a store_output.  (Copy propagation should clean up the MOV.)
>>> For what it's worth, I don't think the type matters; a URB write just
>>> writes data to something so as long as you don't have a type mismatch
>>> in a MOV, the hardware won't care.
>>>
>>> The other option, would be to directly emit the URB write in
>>> store_output.  At the moment, it may be better to take the first
>>> option since that better matches what the FS does right now.  But both
>>> should work fine.
>>>
>>
>> Thanks for these hints, they were very useful.
>>
>> I rewrote the implementation of store_output intrinsic to avoid the
>> setup phase completely. The type, as you suggested, was not important as
>> long as they match while MOVing the contents of output_reg. To guarantee
>> that, I had to patch the emit_urb_slot() to guarantee the types always
>> match. This code is shared with vec4_visitor, so it makes sense to move
>> the safeguards there instead of having both backends provide the correct
>> register type in output_reg entries.
>> For reference, this is the patch that implements it:
>> https://github.com/Igalia/mesa/commit/8c703937f285c0b3a1e7bf6681c7ed7fe09815aa
> 
> Seems reasonable.
> 
>> I also put var->data.location in const_index[1] of the intrinsic op, and
>> disabled nir_assign_var_locations() for output variables, since I don't
>> need var->data.driver_location. I could have used const_index[0], but I
>> prefer to leave driver_location there, and use const_index[1], to avoid
>> breaking any driver that rely on current layout of const_index (like
>> FS-nir). I think it is a safer approach.
> 
> You're not going to break anything by going through the output
> variables and setting driver_location equal to location.  The whole
> point of driver_location is to store some backend-specific index for
> the variable.  In other words, to do exactly what you're doing.  The
> assign_var_locations calls are simply convenience functions for
> setting the driver_location field.  In other words, using
> driver_location and const_index[0] is *exactly* what you should do.
> 

Well, FS-nir relies on const_index[0] being data.driver_location. So at
the very least I have to put a condition like:

if (scalar)
   const_index[0] = var->data.driver_location
else
   const_index[0] = var->data.location

Otherwise we directly break our own FS-nir pass.

My first implementation did that, but since this is common NIR code
(theoretically) shared with other backends, putting var->data.location
in const_index[0] for all non-scalar backends seemed like a bad idea.
Specially considering that this is very dependent on the implementation
of URB file in vec4_visitor, with the output_reg intermediate map and
all. That's why I decided to play safe on pure-NIR side, having both
driver_location and location available to backends.

But if you think I can ignore this then I'm all for it too.

> --Jason
> 
>> All in all, the store_output implementation got much simpler.
>>
>>>> So, before knowing that I could modify nir_lower_io, my best shot at
>>>> transferring the original variable location was to create this
>>>> nir_outputs map. Now, what I have done is to put that value in
>>>> const_index[1] of the intrinsic instruction, which was previously
>>>> unused. What do you think?
>>>>
>>>> That removes the offset to offset map, but we still need the type map.
>>>>
>>>> About your comment on initializing the register during setup stage, I'm
>>>> a bit confused: the register that we need to store is not available
>>>> during setup stage, because we still don't have local registers allocated.
>>>
>>> What do you mean?  Because you don't have the destination of the
>>> output_write intrinsic allocated?  Even if the register has a file of
>>> BAD_FILE, you could still store the type there.  Also, as I said
>>> above, the hardware shouldn't care about the types of data.  As long
>>> as the URB write code doesn't accidentally do a float -> int
>>> conversion or something, we should be fine.
>>> --Jason
>>>
>>>>>>        break;
>>>>>> +   }
>>>>>>
>>>>>>     case nir_intrinsic_load_vertex_id:
>>>>>>        unreachable("should be lowered by lower_vertex_id()");
>>>>>> --
>>>>
>>>
>>
>