[Mesa-dev] [PATCH 2/9] radv: start conditionalising vertex inputs.

[Nicolai Hähnle]

On 18.04.2017 07:19, Dave Airlie wrote:
> On 18 April 2017 at 15:07, Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl> wrote:
>> On Tue, Apr 18, 2017 at 5:57 AM, Dave Airlie <airlied at gmail.com> wrote:
>>> From: Dave Airlie <airlied at redhat.com>
>>>
>>> In practice this will probably just drop draw id in a few places.
>>>
>>> Signed-off-by: Dave Airlie <airlied at redhat.com>
>>> ---
>>>  src/amd/common/ac_nir_to_llvm.c  | 42 +++++++++++++++++++++++++++-----------
>>>  src/amd/common/ac_shader_info.c  | 26 ++++++++++++++++++++++++
>>>  src/amd/common/ac_shader_info.h  |  5 +++++
>>>  src/amd/vulkan/radv_cmd_buffer.c | 44 +++++++++++++++++++++++++++++++---------
>>>  4 files changed, 95 insertions(+), 22 deletions(-)
>>>
>>> diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
>>> index dbb3b67..7161caf 100644
>>> --- a/src/amd/common/ac_nir_to_llvm.c
>>> +++ b/src/amd/common/ac_nir_to_llvm.c
>>> @@ -614,10 +614,15 @@ static void create_function(struct nir_to_llvm_context *ctx)
>>>                 break;
>>>         case MESA_SHADER_VERTEX:
>>>                 if (!ctx->is_gs_copy_shader) {
>>> -                       arg_types[arg_idx++] = const_array(ctx->v16i8, 16); /* vertex buffers */
>>> -                       arg_types[arg_idx++] = ctx->i32; // base vertex
>>> -                       arg_types[arg_idx++] = ctx->i32; // start instance
>>> -                       arg_types[arg_idx++] = ctx->i32; // draw index
>>> +                       if (ctx->shader_info->info.vs.has_vertex_buffers)
>>> +                               arg_types[arg_idx++] = const_array(ctx->v16i8, 16); /* vertex buffers */
>>> +                       if (ctx->shader_info->info.vs.needs_base_vertex_start_instance ||
>>> +                           ctx->shader_info->info.vs.needs_draw_id) {
>>> +                               arg_types[arg_idx++] = ctx->i32; // base vertex
>>> +                               arg_types[arg_idx++] = ctx->i32; // start instance
>>
>> I'm not sure we can avoid having the CP write these two to some user
>> SGPRs for indirect draws? If so, we cannot skip them.
>
> Yeah I did wonder this, it's a pity but you are probably right, I'll
> rework this patch
> to always allocate those two.

You *may* be able to instruct the CP to write those values somewhere 
into the void, i.e. to a non-existing register. It's not something 
that's documented anywhere as a thing that one could or should do, but I 
once did some experiments where this at least didn't hang the GPU, so...

Cheers,
Nicolai
-- 
Lerne, wie die Welt wirklich ist,
Aber vergiss niemals, wie sie sein sollte.