[Mesa-dev] [PATCH 107/133] nir: Add a lowering pass for adding source modifiers where possible
Jason Ekstrand
jason at jlekstrand.net
Mon Dec 15 22:12:50 PST 2014
---
src/glsl/Makefile.sources | 1 +
src/glsl/nir/nir.h | 1 +
src/glsl/nir/nir_lower_to_source_mods.c | 181 +++++++++++++++++++++++++++++++
src/mesa/drivers/dri/i965/brw_fs_nir.cpp | 5 +
4 files changed, 188 insertions(+)
create mode 100644 src/glsl/nir/nir_lower_to_source_mods.c
diff --git a/src/glsl/Makefile.sources b/src/glsl/Makefile.sources
index 3cee2e0..12c1d83 100644
--- a/src/glsl/Makefile.sources
+++ b/src/glsl/Makefile.sources
@@ -27,6 +27,7 @@ NIR_FILES = \
$(GLSL_SRCDIR)/nir/nir_lower_io.c \
$(GLSL_SRCDIR)/nir/nir_lower_samplers.cpp \
$(GLSL_SRCDIR)/nir/nir_lower_system_values.c \
+ $(GLSL_SRCDIR)/nir/nir_lower_to_source_mods.c \
$(GLSL_SRCDIR)/nir/nir_lower_variables.c \
$(GLSL_SRCDIR)/nir/nir_lower_vec_to_movs.c \
$(GLSL_SRCDIR)/nir/nir_metadata.c \
diff --git a/src/glsl/nir/nir.h b/src/glsl/nir/nir.h
index 4fe5c68..3ce8396 100644
--- a/src/glsl/nir/nir.h
+++ b/src/glsl/nir/nir.h
@@ -1393,6 +1393,7 @@ void nir_lower_samplers(nir_shader *shader,
void nir_lower_system_values(nir_shader *shader);
void nir_lower_atomics(nir_shader *shader);
+void nir_lower_to_source_mods(nir_shader *shader);
void nir_live_variables_impl(nir_function_impl *impl);
bool nir_ssa_defs_interfere(nir_ssa_def *a, nir_ssa_def *b);
diff --git a/src/glsl/nir/nir_lower_to_source_mods.c b/src/glsl/nir/nir_lower_to_source_mods.c
new file mode 100644
index 0000000..c3f39bc
--- /dev/null
+++ b/src/glsl/nir/nir_lower_to_source_mods.c
@@ -0,0 +1,181 @@
+/*
+ * Copyright © 2014 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.
+ *
+ * Authors:
+ * Jason Ekstrand (jason at jlekstrand.net)
+ *
+ */
+
+#include "nir.h"
+
+/*
+ * This pass lowers the neg, abs, and sat operations to source modifiers on
+ * ALU operations to make things nicer for the backend. It's just much
+ * easier to not have them when we're doing optimizations.
+ */
+
+static bool
+nir_lower_to_source_mods_block(nir_block *block, void *state)
+{
+ nir_foreach_instr(block, instr) {
+ if (instr->type != nir_instr_type_alu)
+ continue;
+
+ nir_alu_instr *alu = nir_instr_as_alu(instr);
+
+ for (unsigned i = 0; i < nir_op_infos[alu->op].num_inputs; i++) {
+ if (!alu->src[i].src.is_ssa)
+ continue;
+
+ if (alu->src[i].src.ssa->parent_instr->type != nir_instr_type_alu)
+ continue;
+
+ nir_alu_instr *parent = nir_instr_as_alu(alu->src[i].src.ssa->parent_instr);
+
+ if (parent->dest.saturate)
+ continue;
+
+ switch (nir_op_infos[alu->op].input_types[i]) {
+ case nir_type_float:
+ if (parent->op != nir_op_fmov)
+ continue;
+ break;
+ case nir_type_int:
+ if (parent->op != nir_op_imov)
+ continue;
+ break;
+ default:
+ continue;
+ }
+
+ nir_instr_rewrite_src(instr, &alu->src[i].src, parent->src[0].src);
+ if (alu->src[i].abs) {
+ /* abs trumps both neg and abs, do nothing */
+ } else {
+ alu->src[i].negate = (alu->src[i].negate != parent->src[0].negate);
+ alu->src[i].abs |= parent->src[0].abs;
+ }
+
+ for (int j = 0; j < 4; ++j) {
+ if (!nir_alu_instr_channel_used(alu, i, j))
+ continue;
+ alu->src[i].swizzle[j] = parent->src[0].swizzle[alu->src[i].swizzle[j]];
+ }
+
+ if (parent->dest.dest.ssa.uses->entries == 0 &&
+ parent->dest.dest.ssa.if_uses->entries == 0)
+ nir_instr_remove(&parent->instr);
+ }
+
+ switch (alu->op) {
+ case nir_op_fsat:
+ alu->op = nir_op_fmov;
+ alu->dest.saturate = true;
+ break;
+ case nir_op_ineg:
+ alu->op = nir_op_imov;
+ alu->src[0].negate = !alu->src[0].negate;
+ break;
+ case nir_op_fneg:
+ alu->op = nir_op_fmov;
+ alu->src[0].negate = !alu->src[0].negate;
+ break;
+ case nir_op_iabs:
+ alu->op = nir_op_imov;
+ alu->src[0].abs = true;
+ alu->src[0].negate = false;
+ break;
+ case nir_op_fabs:
+ alu->op = nir_op_fmov;
+ alu->src[0].abs = true;
+ alu->src[0].negate = false;
+ break;
+ default:
+ break;
+ }
+
+ /* We've covered sources. Now we're going to try and saturate the
+ * destination if we can.
+ */
+
+ if (!alu->dest.dest.is_ssa)
+ continue;
+
+ if (alu->dest.dest.ssa.if_uses->entries != 0)
+ continue;
+
+ bool all_children_are_sat = true;
+ struct set_entry *entry;
+ set_foreach(alu->dest.dest.ssa.uses, entry) {
+ const nir_instr *child = entry->key;
+ if (child->type != nir_instr_type_alu) {
+ all_children_are_sat = false;
+ continue;
+ }
+
+ nir_alu_instr *child_alu = nir_instr_as_alu(child);
+ if (child_alu->src[0].negate || child_alu->src[0].abs) {
+ all_children_are_sat = false;
+ continue;
+ }
+
+ if (child_alu->op != nir_op_fsat &&
+ !(child_alu->op == nir_op_fmov && child_alu->dest.saturate)) {
+ all_children_are_sat = false;
+ continue;
+ }
+ }
+
+ if (!all_children_are_sat)
+ continue;
+
+ alu->dest.saturate = true;
+
+ set_foreach(alu->dest.dest.ssa.uses, entry) {
+ nir_alu_instr *child_alu = nir_instr_as_alu((nir_instr *)entry->key);
+ child_alu->op = nir_op_fmov;
+ child_alu->dest.saturate = false;
+ /* We could propagate the dest of our instruction to the
+ * destinations of the uses here. However, one quick round of
+ * copy propagation will clean that all up and then we don't have
+ * the complexity.
+ */
+ }
+ }
+
+ return true;
+}
+
+static void
+nir_lower_to_source_mods_impl(nir_function_impl *impl)
+{
+ nir_foreach_block(impl, nir_lower_to_source_mods_block, NULL);
+}
+
+void
+nir_lower_to_source_mods(nir_shader *shader)
+{
+ nir_foreach_overload(shader, overload) {
+ if (overload->impl)
+ nir_lower_to_source_mods_impl(overload->impl);
+ }
+}
diff --git a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
index 67714ec..ec1ef28 100644
--- a/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs_nir.cpp
@@ -66,6 +66,11 @@ fs_visitor::emit_nir_code()
nir_remove_dead_variables(nir);
nir_validate_shader(nir);
+
+ nir_lower_to_source_mods(nir);
+ nir_validate_shader(nir);
+ nir_copy_prop(nir);
+ nir_validate_shader(nir);
nir_convert_from_ssa(nir);
nir_validate_shader(nir);
nir_lower_vec_to_movs(nir);
--
2.2.0
More information about the mesa-dev
mailing list