[Mesa-dev] [PATCH 06/16] nir: improve convert_yuv_to_rgb when fuse_ffma=true

[Lionel Landwerlin]

I did not but then saw someone pointed out an issue with this particular 
patch.
I can do tomorrow.
Do you have link to the updated series?

Thanks,

-
Lionel

On 07/01/2019 16:54, Jonathan Marek wrote:
> Hi,
>
> Did you get a chance try this? If not, I might be able to try it 
> myself as I have Intel HW.
>
> On 12/19/18 12:34 PM, Lionel Landwerlin wrote:
>> Hey Jonathan,
>>
>> I'm kind of curious as to whether we can have a single expression 
>> that pretty much generates the same final code (through some of the 
>> algebraic lowering/optimizations).
>> I'll give it a try on Intel HW, see what it does.
>>
>> -
>> Lionel
>>
>> On 19/12/2018 16:39, Jonathan Marek wrote:
>>> When ffma is available, we can use a different arrangement of 
>>> constants to
>>> get a better result. On freedreno/ir3, this reduces the YUV->RGB to 7
>>> scalar ffma. On freedreno/a2xx, it will allow YUV->RGB to be 3 vec4 
>>> ffma.
>>>
>>> Signed-off-by: Jonathan Marek <jonathan at marek.ca>
>>> ---
>>>   src/compiler/nir/nir_lower_tex.c | 62 
>>> ++++++++++++++++++++++----------
>>>   1 file changed, 43 insertions(+), 19 deletions(-)
>>>
>>> diff --git a/src/compiler/nir/nir_lower_tex.c 
>>> b/src/compiler/nir/nir_lower_tex.c
>>> index 6a6b6c41a7..f7c821bb34 100644
>>> --- a/src/compiler/nir/nir_lower_tex.c
>>> +++ b/src/compiler/nir/nir_lower_tex.c
>>> @@ -342,25 +342,49 @@ convert_yuv_to_rgb(nir_builder *b, 
>>> nir_tex_instr *tex,
>>>                      nir_ssa_def *y, nir_ssa_def *u, nir_ssa_def *v,
>>>                      nir_ssa_def *a)
>>>   {
>>> -   nir_const_value m[3] = {
>>> -      { .f32 = { 1.0f,  0.0f,         1.59602678f, 0.0f } },
>>> -      { .f32 = { 1.0f, -0.39176229f, -0.81296764f, 0.0f } },
>>> -      { .f32 = { 1.0f,  2.01723214f,  0.0f,        0.0f } }
>>> -   };
>>> -
>>> -   nir_ssa_def *yuv =
>>> -      nir_vec4(b,
>>> -               nir_fmul(b, nir_imm_float(b, 1.16438356f),
>>> -                        nir_fadd(b, y, nir_imm_float(b, -16.0f / 
>>> 255.0f))),
>>> -               nir_channel(b, nir_fadd(b, u, nir_imm_float(b, 
>>> -128.0f / 255.0f)), 0),
>>> -               nir_channel(b, nir_fadd(b, v, nir_imm_float(b, 
>>> -128.0f / 255.0f)), 0),
>>> -               nir_imm_float(b, 0.0));
>>> -
>>> -   nir_ssa_def *red = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, 
>>> m[0]));
>>> -   nir_ssa_def *green = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, 
>>> m[1]));
>>> -   nir_ssa_def *blue = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, 
>>> m[2]));
>>> -
>>> -   nir_ssa_def *result = nir_vec4(b, red, green, blue, a);
>>> +   nir_ssa_def *result;
>>> +
>>> +
>>> +   if (b->shader->options->fuse_ffma) {
>>> +      nir_const_value m[4] = {
>>> +         { .f32 = { 1.16438356f, 1.16438356f, 1.16438356f, 0.0f } },
>>> +         { .f32 = { 0.0f,       -0.39176229f, 2.01723214f, 0.0f } },
>>> +         { .f32 = { 1.59602678f,-0.81296764f, 0.0f, 0.0f } },
>>> +      };
>>> +      static const float y_off = -16.0f * 1.16438356f / 255.0f;
>>> +      static const float sc = 128.0f / 255.0f;
>>> +
>>> +      nir_ssa_def *offset =
>>> +         nir_vec4(b,
>>> +                  nir_imm_float(b, y_off - sc * 1.59602678f),
>>> +                  nir_imm_float(b, y_off + sc * (0.81296764f + 
>>> 0.39176229f)),
>>> +                  nir_imm_float(b, y_off - sc * 2.01723214f),
>>> +                  a);
>>> +
>>> +      result = nir_ffma(b, y, nir_build_imm(b, 4, 32, m[0]),
>>> +                       nir_ffma(b, u, nir_build_imm(b, 4, 32, m[1]),
>>> +                                nir_ffma(b, v, nir_build_imm(b, 4, 
>>> 32, m[2]), offset)));
>>> +   } else {
>>> +      nir_const_value m[3] = {
>>> +         { .f32 = { 1.0f,  0.0f,         1.59602678f, 0.0f } },
>>> +         { .f32 = { 1.0f, -0.39176229f, -0.81296764f, 0.0f } },
>>> +         { .f32 = { 1.0f,  2.01723214f,  0.0f,        0.0f } }
>>> +      };
>>> +
>>> +      nir_ssa_def *yuv =
>>> +         nir_vec4(b,
>>> +                  nir_fmul(b, nir_imm_float(b, 1.16438356f),
>>> +                           nir_fadd(b, y, nir_imm_float(b, -16.0f / 
>>> 255.0f))),
>>> +                  nir_channel(b, nir_fadd(b, u, nir_imm_float(b, 
>>> -128.0f / 255.0f)), 0),
>>> +                  nir_channel(b, nir_fadd(b, v, nir_imm_float(b, 
>>> -128.0f / 255.0f)), 0),
>>> +                  nir_imm_float(b, 0.0));
>>> +
>>> +      nir_ssa_def *red = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, 
>>> m[0]));
>>> +      nir_ssa_def *green = nir_fdot4(b, yuv, nir_build_imm(b, 4, 
>>> 32, m[1]));
>>> +      nir_ssa_def *blue = nir_fdot4(b, yuv, nir_build_imm(b, 4, 32, 
>>> m[2]));
>>> +
>>> +      result = nir_vec4(b, red, green, blue, a);
>>> +   }
>>>      nir_ssa_def_rewrite_uses(&tex->dest.ssa, nir_src_for_ssa(result));
>>>   }
>>
>>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev