intel/fs: Emit LINE+MAC for LINTERP with unaligned coordinates

On g4x through Sandy Bridge, src1 (the coordinates) of the PLN
instruction is required to be an even register number.  When it's odd
(which can happen with SIMD32), we have to emit a LINE+MAC combination
instead.  Unfortunately, we can't just fall through to the gen4 case
because the input registers are still set up for PLN which lays out the
four src1 registers differently in SIMD16 than LINE.

v2 (Jason Ekstrand):
 - Take advantage of both accumulators and emit LINE LINE MAC MAC
   (Based on a patch from Francisco Jerez)
 - Unify the gen4 and gen4x-6 cases using a loop

v3 (Jason Ekstrand):
 - Don't unify gen4 with gen4x-6 as this turns out to be more fragile
   than first thought without reworking the gen4 barycentric coordinate
   layout.

Reviewed-by: Matt Turner <mattst88@gmail.com>

diff --git a/src/intel/compiler/brw_fs_generator.cpp b/src/intel/compiler/brw_fs_generator.cpp
index c352f87c16ea68b5782a011cb6d8cc9ab957351d..2265a60d1a654f67357b19ea74e370bddd872a85 100644 (file)
--- a/src/intel/compiler/brw_fs_generator.cpp
+++ b/src/intel/compiler/brw_fs_generator.cpp
@@ -761,16 +761,61 @@ fs_generator::generate_linterp(fs_inst *inst,
 
       return true;
    } else if (devinfo->has_pln) {
-      /* From the Sandy Bridge PRM Vol. 4, Pt. 2, Section 8.3.53, "Plane":
-       *
-       *    "[DevSNB]:<src1> must be even register aligned.
-       *
-       * This restriction is lifted on Ivy Bridge.
-       */
-      assert(devinfo->gen >= 7 || (delta_x.nr & 1) == 0);
-      brw_PLN(p, dst, interp, delta_x);
+      if (devinfo->gen <= 6 && (delta_x.nr & 1) != 0) {
+         /* From the Sandy Bridge PRM Vol. 4, Pt. 2, Section 8.3.53, "Plane":
+          *
+          *    "[DevSNB]:<src1> must be even register aligned.
+          *
+          * This restriction is lifted on Ivy Bridge.
+          *
+          * This means that we need to split PLN into LINE+MAC on-the-fly.
+          * Unfortunately, the inputs are laid out for PLN and not LINE+MAC so
+          * we have to split into SIMD8 pieces.  For gen4 (!has_pln), the
+          * coordinate registers are laid out differently so we leave it as a
+          * SIMD16 instruction.
+          */
+         assert(inst->exec_size == 8 || inst->exec_size == 16);
+         assert(inst->group % 16 == 0);
 
-      return false;
+         brw_push_insn_state(p);
+         brw_set_default_exec_size(p, BRW_EXECUTE_8);
+
+         /* Thanks to two accumulators, we can emit all the LINEs and then all
+          * the MACs.  This improves parallelism a bit.
+          */
+         for (unsigned g = 0; g < inst->exec_size / 8; g++) {
+            brw_inst *line = brw_LINE(p, brw_null_reg(), interp,
+                                      offset(delta_x, g * 2));
+            brw_inst_set_group(devinfo, line, inst->group + g * 8);
+
+            /* LINE writes the accumulator automatically on gen4-5.  On Sandy
+             * Bridge and later, we have to explicitly enable it.
+             */
+            if (devinfo->gen >= 6)
+               brw_inst_set_acc_wr_control(p->devinfo, line, true);
+
+            /* brw_set_default_saturate() is called before emitting
+             * instructions, so the saturate bit is set in each instruction,
+             * so we need to unset it on the LINE instructions.
+             */
+            brw_inst_set_saturate(p->devinfo, line, false);
+         }
+
+         for (unsigned g = 0; g < inst->exec_size / 8; g++) {
+            brw_inst *mac = brw_MAC(p, offset(dst, g), suboffset(interp, 1),
+                                    offset(delta_x, g * 2 + 1));
+            brw_inst_set_group(devinfo, mac, inst->group + g * 8);
+            brw_inst_set_cond_modifier(p->devinfo, mac, inst->conditional_mod);
+         }
+
+         brw_pop_insn_state(p);
+
+         return true;
+      } else {
+         brw_PLN(p, dst, interp, delta_x);
+
+         return false;
+      }
    } else {
       i[0] = brw_LINE(p, brw_null_reg(), interp, delta_x);
       i[1] = brw_MAC(p, dst, suboffset(interp, 1), delta_y);

diff --git a/src/intel/compiler/brw_shader.cpp b/src/intel/compiler/brw_shader.cpp
index 5a3070e5d2e4f73ce2a00e4e3d7cc91a3899e79c..7911065c0991a7ea01a71fd672691a7bd67b2a96 100644 (file)
--- a/src/intel/compiler/brw_shader.cpp
+++ b/src/intel/compiler/brw_shader.cpp
@@ -988,7 +988,8 @@ backend_instruction::writes_accumulator_implicitly(const struct gen_device_info
           (devinfo->gen < 6 &&
            ((opcode >= BRW_OPCODE_ADD && opcode < BRW_OPCODE_NOP) ||
             (opcode >= FS_OPCODE_DDX_COARSE && opcode <= FS_OPCODE_LINTERP))) ||
-          (opcode == FS_OPCODE_LINTERP && !devinfo->has_pln);
+          (opcode == FS_OPCODE_LINTERP &&
+           (!devinfo->has_pln || devinfo->gen <= 6));
 }
 
 bool