add modificationsdp-narrowing

1 files changed, 34 insertions, 0 deletions

diff --git a/src/intel/compiler/brw_vec4_nir.cpp b/src/intel/compiler/brw_vec4_nir.cpp
index 32c5bd40406..8fc350fd060 100644
--- a/src/intel/compiler/brw_vec4_nir.cpp
+++ b/src/intel/compiler/brw_vec4_nir.cpp
@@ -1346,6 +1346,40 @@ vec4_visitor::try_VF_for_constants(const float const1[4], src_reg *op1,
    num_combined = merge_constants(const2, mask, combined, num_combined,
                                   op2_swiz, no_modification);
 
+   if (num_combined > 4) {
+      /* Try really, really hard to fit both vectors into a single VF
+       * constant.  Try two transformations to find more redundancy.
+       *
+       * - If all the components of one constant have the same sign, that
+       *   operand will get either an abs() or -abs() source modifier applied.
+       *   The components of that constant don't need the same sign as a
+       *   component from the other vector to match.  This helps cases like
+       *   const1 = { 6.0, 7.0, 4.0, 1.0}, const2 = { -6.0, 1.0, 7.0, -4.0 }
+       *   fit.
+       *
+       * - If both constants have mixed signs, try appling a negation to one
+       *   of the operands.  This helps cases like const1 = { 6.0, -7.0, 12.0,
+       *   21.0 }, const2 = { -6.0, 7.0, -12.0, -21.0 } fit.
+       */
+      m1 = same_sign(const1, mask);
+      if (m1 != no_modification) {
+         num_combined = merge_constants(const2, mask, combined, 0,
+                                        op2_swiz, no_modification);
+         num_combined = merge_constants(const1, mask, combined, num_combined,
+                                        op1_swiz, m1);
+      } else {
+         m2 = same_sign(const2, mask);
+
+         if (m2 == no_modification)
+            m2 = neg_modification;
+
+         num_combined = merge_constants(const1, mask, combined, 0,
+                                        op1_swiz, no_modification);
+         num_combined = merge_constants(const2, mask, combined, num_combined,
+                                        op2_swiz, m2);
+      }
+   }
+
    /* Too many values for a single VF. */
    if (num_combined > 4)
       return false;