[Mesa-dev] [PATCH 4/4] radeonsi: add support for interpolateAt functions

[Dave Airlie]

On 25 July 2015 at 01:15, Marek Olšák <maraeo at gmail.com> wrote:
> On Wed, Jul 22, 2015 at 12:51 AM, Dave Airlie <airlied at gmail.com> wrote:
>> From: Dave Airlie <airlied at redhat.com>
>>
>> This is part of ARB_gpu_shader5, and this passes
>> all the piglit tests currently available.
>>
>> Signed-off-by: Dave Airlie <airlied at redhat.com>
>> ---
>>  docs/GL3.txt                             |   2 +-
>>  src/gallium/drivers/radeonsi/si_shader.c | 232 ++++++++++++++++++++++++++++++-
>>  2 files changed, 232 insertions(+), 2 deletions(-)
>>
>> diff --git a/docs/GL3.txt b/docs/GL3.txt
>> index 4f6c415..d74ae63 100644
>> --- a/docs/GL3.txt
>> +++ b/docs/GL3.txt
>> @@ -107,7 +107,7 @@ GL 4.0, GLSL 4.00:
>>    - Geometry shader instancing                         DONE (r600, radeonsi, llvmpipe, softpipe)
>>    - Geometry shader multiple streams                   DONE ()
>>    - Enhanced per-sample shading                        DONE (r600, radeonsi)
>> -  - Interpolation functions                            DONE (r600)
>> +  - Interpolation functions                            DONE (r600, radeonsi)
>>    - New overload resolution rules                      DONE
>>    GL_ARB_gpu_shader_fp64                               DONE (nvc0, radeonsi, llvmpipe, softpipe)
>>    GL_ARB_sample_shading                                DONE (i965, nv50, nvc0, r600, radeonsi)
>> diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
>> index c5d80f0..0c01c90 100644
>> --- a/src/gallium/drivers/radeonsi/si_shader.c
>> +++ b/src/gallium/drivers/radeonsi/si_shader.c
>> @@ -2263,6 +2263,225 @@ static void si_llvm_emit_ddxy(
>>         emit_data->output[0] = lp_build_gather_values(gallivm, result, 4);
>>  }
>>
>> +/* return 4 values - v2i32 DDX, v2i32 DDY */
>> +static LLVMValueRef si_llvm_emit_ddxy_interp(
>> +       struct lp_build_tgsi_context * bld_base,
>> +       LLVMValueRef interp_ij)
>
> Is there any chance we could simplify this by using the DDX/DDY
> instructions directly here? For example:
>
> result = lp_build_emit_llvm_unary(bld_base, TGSI_OPCODE_DDX, arg0);

Not really, I thought about trying to combine the functions a few
time, but nothing fell out that I thought was simpler.

The interp version just does a ddxy on an input i/j pair, not a tgsi
src. It didn't seem useful to convert things to a tgsi src,
just to pipe it back in. I've also no idea what the overhead on LDS
store/loads is, but this patch definitely tries to minimise
how many times we hit it, whereas using the DDX/Y interface would hit
LDS stores 4 times, this only hits it twice.

Dave.