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

[Jonathan Marek]

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));
 }
-- 
2.17.1