nir: Add a bool to float32 lowering pass
authorJason Ekstrand <jason.ekstrand@intel.com>
Mon, 22 Oct 2018 21:44:59 +0000 (16:44 -0500)
committerJason Ekstrand <jason@jlekstrand.net>
Mon, 14 Jan 2019 19:27:06 +0000 (19:27 +0000)
From @jekstrand's nir-1-bit-bool branch, with improved ior/inot lowering.

ior: fmax instead of fadd allows removing the fsat.

inot: seq(x, 0) can be better than fsub(1, x). On a2xx, it works better
with the scalar instruction set.

Reviewed-by: Jonathan Marek <jonathan@marek.ca>
src/compiler/Makefile.sources
src/compiler/nir/meson.build
src/compiler/nir/nir.h
src/compiler/nir/nir_lower_bool_to_float.c [new file with mode: 0644]

index ec62c2fbb42a8201e6cc67ed28244fb9933b9f80..66c1018be7a5271716b9ce0eff8c601034003949 100644 (file)
@@ -232,6 +232,7 @@ NIR_FILES = \
        nir/nir_lower_atomics_to_ssbo.c \
        nir/nir_lower_bitmap.c \
        nir/nir_lower_bit_size.c \
+       nir/nir_lower_bool_to_float.c \
        nir/nir_lower_bool_to_int32.c \
        nir/nir_lower_clamp_color_outputs.c \
        nir/nir_lower_clip.c \
index e252f64539e67262ff2826b8810bc05dd540183a..20a26a262558f557abb17c98907a4bc4c74ca4c4 100644 (file)
@@ -114,6 +114,7 @@ files_libnir = files(
   'nir_lower_alpha_test.c',
   'nir_lower_atomics_to_ssbo.c',
   'nir_lower_bitmap.c',
+  'nir_lower_bool_to_float.c',
   'nir_lower_bool_to_int32.c',
   'nir_lower_clamp_color_outputs.c',
   'nir_lower_clip.c',
index 3cb2d166cb3d83ff3678a861ef981047502fd565..7a7e910ba49b2dd3eb104a8e045c64af3b083226 100644 (file)
@@ -2950,6 +2950,7 @@ void nir_lower_alpha_test(nir_shader *shader, enum compare_func func,
                           bool alpha_to_one);
 bool nir_lower_alu(nir_shader *shader);
 bool nir_lower_alu_to_scalar(nir_shader *shader);
+bool nir_lower_bool_to_float(nir_shader *shader);
 bool nir_lower_bool_to_int32(nir_shader *shader);
 bool nir_lower_load_const_to_scalar(nir_shader *shader);
 bool nir_lower_read_invocation_to_scalar(nir_shader *shader);
diff --git a/src/compiler/nir/nir_lower_bool_to_float.c b/src/compiler/nir/nir_lower_bool_to_float.c
new file mode 100644 (file)
index 0000000..52fc55c
--- /dev/null
@@ -0,0 +1,178 @@
+/*
+ * Copyright © 2018 Intel Corporation
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "nir.h"
+#include "nir_builder.h"
+
+static bool
+assert_ssa_def_is_not_1bit(nir_ssa_def *def, UNUSED void *unused)
+{
+   assert(def->bit_size > 1);
+   return true;
+}
+
+static bool
+rewrite_1bit_ssa_def_to_32bit(nir_ssa_def *def, void *_progress)
+{
+   bool *progress = _progress;
+   if (def->bit_size == 1) {
+      def->bit_size = 32;
+      *progress = true;
+   }
+   return true;
+}
+
+static bool
+lower_alu_instr(nir_builder *b, nir_alu_instr *alu)
+{
+   const nir_op_info *op_info = &nir_op_infos[alu->op];
+
+   b->cursor = nir_before_instr(&alu->instr);
+
+   /* Replacement SSA value */
+   nir_ssa_def *rep = NULL;
+   switch (alu->op) {
+   case nir_op_b2f32: alu->op = nir_op_fmov; break;
+   case nir_op_b2i32: alu->op = nir_op_fmov; break;
+   case nir_op_f2b1:
+   case nir_op_i2b1:
+      rep = nir_sne(b, nir_ssa_for_alu_src(b, alu, 0),
+                       nir_imm_float(b, 0));
+      break;
+
+   case nir_op_flt: alu->op = nir_op_slt; break;
+   case nir_op_fge: alu->op = nir_op_sge; break;
+   case nir_op_feq: alu->op = nir_op_seq; break;
+   case nir_op_fne: alu->op = nir_op_sne; break;
+   case nir_op_ilt: alu->op = nir_op_slt; break;
+   case nir_op_ige: alu->op = nir_op_sge; break;
+   case nir_op_ieq: alu->op = nir_op_seq; break;
+   case nir_op_ine: alu->op = nir_op_sne; break;
+   case nir_op_ult: alu->op = nir_op_slt; break;
+   case nir_op_uge: alu->op = nir_op_sge; break;
+
+   case nir_op_ball_fequal2:  alu->op = nir_op_fall_equal2; break;
+   case nir_op_ball_fequal3:  alu->op = nir_op_fall_equal3; break;
+   case nir_op_ball_fequal4:  alu->op = nir_op_fall_equal4; break;
+   case nir_op_bany_fnequal2: alu->op = nir_op_fany_nequal2; break;
+   case nir_op_bany_fnequal3: alu->op = nir_op_fany_nequal3; break;
+   case nir_op_bany_fnequal4: alu->op = nir_op_fany_nequal4; break;
+   case nir_op_ball_iequal2:  alu->op = nir_op_fall_equal2; break;
+   case nir_op_ball_iequal3:  alu->op = nir_op_fall_equal3; break;
+   case nir_op_ball_iequal4:  alu->op = nir_op_fall_equal4; break;
+   case nir_op_bany_inequal2: alu->op = nir_op_fany_nequal2; break;
+   case nir_op_bany_inequal3: alu->op = nir_op_fany_nequal3; break;
+   case nir_op_bany_inequal4: alu->op = nir_op_fany_nequal4; break;
+
+   case nir_op_bcsel: alu->op = nir_op_fcsel; break;
+
+   case nir_op_imov: alu->op = nir_op_fmov; break;
+   case nir_op_iand: alu->op = nir_op_fmul; break;
+   case nir_op_ixor: alu->op = nir_op_sne; break;
+   case nir_op_ior: alu->op = nir_op_fmax; break;
+
+   case nir_op_inot:
+      rep = nir_seq(b, nir_ssa_for_alu_src(b, alu, 0),
+                       nir_imm_float(b, 0));
+      break;
+
+   default:
+      assert(alu->dest.dest.ssa.bit_size > 1);
+      for (unsigned i = 0; i < op_info->num_inputs; i++)
+         assert(alu->src[i].src.ssa->bit_size > 1);
+      return false;
+   }
+
+   if (rep) {
+      /* We've emitted a replacement instruction */
+      nir_ssa_def_rewrite_uses(&alu->dest.dest.ssa, nir_src_for_ssa(rep));
+      nir_instr_remove(&alu->instr);
+   } else {
+      if (alu->dest.dest.ssa.bit_size == 1)
+         alu->dest.dest.ssa.bit_size = 32;
+   }
+
+   return true;
+}
+
+static bool
+nir_lower_bool_to_float_impl(nir_function_impl *impl)
+{
+   bool progress = false;
+
+   nir_builder b;
+   nir_builder_init(&b, impl);
+
+   nir_foreach_block(block, impl) {
+      nir_foreach_instr_safe(instr, block) {
+         switch (instr->type) {
+         case nir_instr_type_alu:
+            progress |= lower_alu_instr(&b, nir_instr_as_alu(instr));
+            break;
+
+         case nir_instr_type_load_const: {
+            nir_load_const_instr *load = nir_instr_as_load_const(instr);
+            if (load->def.bit_size == 1) {
+               nir_const_value value = load->value;
+               for (unsigned i = 0; i < load->def.num_components; i++)
+                  load->value.f32[i] = value.b[i] ? 1.0 : 0.0;
+               load->def.bit_size = 32;
+               progress = true;
+            }
+            break;
+         }
+
+         case nir_instr_type_intrinsic:
+         case nir_instr_type_ssa_undef:
+         case nir_instr_type_phi:
+         case nir_instr_type_tex:
+            nir_foreach_ssa_def(instr, rewrite_1bit_ssa_def_to_32bit,
+                                &progress);
+            break;
+
+         default:
+            nir_foreach_ssa_def(instr, assert_ssa_def_is_not_1bit, NULL);
+         }
+      }
+   }
+
+   if (progress) {
+      nir_metadata_preserve(impl, nir_metadata_block_index |
+                                  nir_metadata_dominance);
+   }
+
+   return progress;
+}
+
+bool
+nir_lower_bool_to_float(nir_shader *shader)
+{
+   bool progress = false;
+
+   nir_foreach_function(function, shader) {
+      if (function->impl && nir_lower_bool_to_float_impl(function->impl))
+         progress = true;
+   }
+
+   return progress;
+}