[Mesa-dev] [PATCH v2 14/16] i965: Add fs_visitor::run_vs() to generate scalar vertex shader code

[Kristian Høgsberg]

On Fri, Nov 14, 2014 at 4:21 PM, Kenneth Graunke <kenneth at whitecape.org> wrote:
> On Thursday, November 13, 2014 04:28:20 PM Kristian Høgsberg wrote:
>> This patch uses the previous refactoring to add a new run_vs() method
>> that generates vertex shader code using the scalar visitor and
>> optimizer.
>>
>> Signed-off-by: Kristian Høgsberg <krh at bitplanet.net>
>> ---
>>  src/mesa/drivers/dri/i965/brw_fs.cpp         |  99 ++++++++-
>>  src/mesa/drivers/dri/i965/brw_fs.h           |  21 +-
>>  src/mesa/drivers/dri/i965/brw_fs_visitor.cpp | 303
> ++++++++++++++++++++++++++-
>>  3 files changed, 412 insertions(+), 11 deletions(-)
>>
>> diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp
> b/src/mesa/drivers/dri/i965/brw_fs.cpp
>> index 4dce0a2..8007977 100644
>> --- a/src/mesa/drivers/dri/i965/brw_fs.cpp
>> +++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
>> @@ -1828,6 +1828,56 @@ fs_visitor::assign_urb_setup()
>>        urb_start + prog_data->num_varying_inputs * 2;
>>  }
>>
>> +void
>> +fs_visitor::assign_vs_urb_setup()
>> +{
>> +   brw_vs_prog_data *vs_prog_data = (brw_vs_prog_data *) prog_data;
>> +   int grf, count, slot, channel, attr;
>> +
>> +   assert(stage == MESA_SHADER_VERTEX);
>> +   count = _mesa_bitcount_64(vs_prog_data->inputs_read);
>> +   if (vs_prog_data->uses_vertexid || vs_prog_data->uses_instanceid)
>> +      count++;
>> +
>> +   /* Each attribute is 4 regs. */
>> +   this->first_non_payload_grf =
>> +      payload.num_regs + prog_data->curb_read_length + count * 4;
>> +
>> +   unsigned vue_entries =
>> +      MAX2(count, vs_prog_data->base.vue_map.num_slots);
>> +
>> +   vs_prog_data->base.urb_entry_size = ALIGN(vue_entries, 4) / 4;
>> +   vs_prog_data->base.urb_read_length = (count + 1) / 2;
>> +
>> +   assert(vs_prog_data->base.urb_read_length <= 15);
>
> What enforces this assertion?  It obviously needs to be true, since
> the "Vertex URB Entry Read Length" field in 3DSTATE_VS has a maximum
> value of 15 in SIMD8 mode.
>
> But this looks like the exact same calculation done in
> vec4_vs_visitor::setup_attributes, and the limit there is [1, 63].
>
> The documentation indicates that the API may allow more vertex attributes
> than can practically be supported by pushing VUEs in as payload registers:
>
> "The EU limit of 128 GRFs imposes a maximum limit of 30 elements per vertex
>  pushed into the payload, though the practical limit may be lower. If input
>  vertices exceed the practical limit, software must decide between resorting
>  to pulling elements during thread execution or dropping back to SIMD4x2
>  dispatch."
>
> I don't see any code in this series for falling back, nor for pulling VUEs
> out of the URB (and avoiding overwriting them before they're loaded), so
> I expect this would simply assert fail on a vertex shader with too many
> inputs.

We only advertise the spec required minimum of 16 vertex attributes
right now.  The 3DSTATE_VS field counts pairs of vertex attributes
which means we can handle up to 30 in scalar mode.

Kristian