[Mesa-dev] [RFC 04/10] nir/lower_double_ops: lower eq()

[Elie Tournier]

Signed-off-by: Elie Tournier <elie.tournier at collabora.com>
---
 src/compiler/nir/nir.h                  |  3 ++-
 src/compiler/nir/nir_lower_double_ops.c | 43 +++++++++++++++++++++++++++++++++
 src/intel/compiler/brw_nir.c            |  3 ++-
 3 files changed, 47 insertions(+), 2 deletions(-)

diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
index 10dd6b1056..7b1a4655ca 100644
--- a/src/compiler/nir/nir.h
+++ b/src/compiler/nir/nir.h
@@ -2571,7 +2571,8 @@ typedef enum {
    nir_lower_dmod = (1 << 8),
    nir_lower_dabs = (1 << 9),
    nir_lower_dneg = (1 << 10),
-   nir_lower_dsign = (1 << 11)
+   nir_lower_dsign = (1 << 11),
+   nir_lower_deq = (1 << 12)
 } nir_lower_doubles_options;
 
 bool nir_lower_doubles(nir_shader *shader, nir_lower_doubles_options options);
diff --git a/src/compiler/nir/nir_lower_double_ops.c b/src/compiler/nir/nir_lower_double_ops.c
index 5c0d62b554..d3e05bf519 100644
--- a/src/compiler/nir/nir_lower_double_ops.c
+++ b/src/compiler/nir/nir_lower_double_ops.c
@@ -526,6 +526,37 @@ lower_fsign64(nir_builder *b, nir_ssa_def *src)
                                            nir_ior(b, sign, one)));
 }
 
+static nir_ssa_def *
+lower_feq64(nir_builder *b, nir_ssa_def *x, nir_ssa_def *y)
+{
+   nir_ssa_def *x_lo = nir_unpack_64_2x32_split_x(b, x);
+   nir_ssa_def *x_hi = nir_unpack_64_2x32_split_y(b, x);
+   nir_ssa_def *y_lo = nir_unpack_64_2x32_split_x(b, y);
+   nir_ssa_def *y_hi = nir_unpack_64_2x32_split_y(b, y);
+
+   nir_ssa_def *eq_lo = nir_ieq(b, x_lo, y_lo);
+   nir_ssa_def *eq_hi = nir_ieq(b, x_hi, y_hi);
+   nir_ssa_def *eq_x_lo = nir_ieq(b, x_lo, nir_imm_int(b, 0));
+   nir_ssa_def *eq_xy_hi = nir_ieq(b,
+                                   nir_ishl(b,
+                                            nir_ior(b, x_hi, y_hi),
+                                            nir_imm_int(b, 1)),
+                                   nir_imm_int(b, 0));
+   /* if x or y is a nan
+    *    return false;
+    * else
+    *    return (x_lo == y_lo) &&
+    *       ((x_hi == y_hi) ||
+    *       ((x_lo == 0) && (((x_hi | y_hi)<<1) == 0)));
+    */
+   return nir_bcsel(b, nir_ior(b, is_nan(b, x), is_nan(b, y)),
+                       nir_imm_int(b, NIR_FALSE),
+                       nir_iand(b, eq_lo,
+                                   nir_ior(b,
+                                           eq_hi,
+                                           nir_iand(b, eq_x_lo, eq_xy_hi))));
+}
+
 static bool
 lower_doubles_instr(nir_alu_instr *instr, nir_lower_doubles_options options)
 {
@@ -594,6 +625,11 @@ lower_doubles_instr(nir_alu_instr *instr, nir_lower_doubles_options options)
          return false;
       break;
 
+   case nir_op_feq:
+      if (!(options & nir_lower_deq))
+         return false;
+      break;
+
    default:
       return false;
    }
@@ -652,6 +688,13 @@ lower_doubles_instr(nir_alu_instr *instr, nir_lower_doubles_options options)
       result = lower_fsign64(&bld, src);
       break;
 
+   case nir_op_feq: {
+      nir_ssa_def *src1 = nir_fmov_alu(&bld, instr->src[1],
+                                      instr->dest.dest.ssa.num_components);
+      result = lower_feq64(&bld, src, src1);
+   }
+      break;
+
    default:
       unreachable("unhandled opcode");
    }
diff --git a/src/intel/compiler/brw_nir.c b/src/intel/compiler/brw_nir.c
index 405d756e6c..7b8b34b4ba 100644
--- a/src/intel/compiler/brw_nir.c
+++ b/src/intel/compiler/brw_nir.c
@@ -512,7 +512,8 @@ nir_optimize(nir_shader *nir, const struct brw_compiler *compiler,
                              nir_lower_dmod |
                              nir_lower_dabs |
                              nir_lower_dneg |
-                             nir_lower_dsign);
+                             nir_lower_dsign |
+                             nir_lower_deq);
       OPT(nir_lower_64bit_pack);
    } while (progress);
 
-- 
2.11.0