[Mesa-dev] [PATCH 3/7] i965: enable varying component packing for BDW+

[Timothy Arceri]

shader-db results BDW:

total instructions in shared programs: 13192895 -> 13182437 (-0.08%)
instructions in affected programs: 827145 -> 816687 (-1.26%)
helped: 5199
HURT: 116

total cycles in shared programs: 539249342 -> 539156566 (-0.02%)
cycles in affected programs: 21894552 -> 21801776 (-0.42%)
helped: 10667
HURT: 7196

LOST:   0
GAINED: 17
---
 src/mesa/drivers/dri/i965/brw_link.cpp | 14 ++++++++++++++
 1 file changed, 14 insertions(+)

diff --git a/src/mesa/drivers/dri/i965/brw_link.cpp b/src/mesa/drivers/dri/i965/brw_link.cpp
index e4b642f49d..73dec88478 100644
--- a/src/mesa/drivers/dri/i965/brw_link.cpp
+++ b/src/mesa/drivers/dri/i965/brw_link.cpp
@@ -320,31 +320,45 @@ brw_link_shader(struct gl_context *ctx, struct gl_shader_program *shProg)
                consumer = brw_nir_optimize(consumer, compiler, c_is_scalar);
             }
 
             shProg->_LinkedShaders[i]->Program->nir = producer;
             shProg->_LinkedShaders[next]->Program->nir = consumer;
 
             next = i;
        }
     }
 
+   int prev = -1;
    for (stage = 0; stage < ARRAY_SIZE(shProg->_LinkedShaders); stage++) {
       struct gl_linked_shader *shader = shProg->_LinkedShaders[stage];
       if (!shader)
          continue;
 
       struct gl_program *prog = shader->Program;
       brw_shader_gather_info(prog->nir, prog);
 
       NIR_PASS_V(prog->nir, nir_lower_samplers, shProg);
       NIR_PASS_V(prog->nir, nir_lower_atomics, shProg);
 
+      if (brw->screen->devinfo.gen >= 8 && prev != -1) {
+         nir_compact_varyings(shProg->_LinkedShaders[prev]->Program->nir,
+                              prog->nir, ctx->API != API_OPENGL_COMPAT);
+
+         /* TODO: In future nir_compact_varyings() should update the varying
+          * input/output mask to reflect the packing it has just done. However
+          * until we have an array/matrix splitting pass we would just be
+          * duplicating the complex partial marking code in nir_gather_info().
+          */
+         brw_shader_gather_info(prog->nir, prog);
+      }
+      prev = stage;
+
       infos[stage] = &prog->nir->info;
 
       update_xfb_info(prog->sh.LinkedTransformFeedback, infos[stage]);
 
       /* Make a pass over the IR to add state references for any built-in
        * uniforms that are used.  This has to be done now (during linking).
        * Code generation doesn't happen until the first time this shader is
        * used for rendering.  Waiting until then to generate the parameters is
        * too late.  At that point, the values for the built-in uniforms won't
        * get sent to the shader.
-- 
2.13.6