i965/fs: Add an instruction scheduler.

Improves performance of my GLSL demo by 5.1% (+/- 1.4%, n=7).  It also
reschedules the giant multiply tree at the end of
glsl-fs-convolution-1 so that we end up not spilling registers,
producing the expected level of performance.

4 files changed, 479 insertions, 0 deletions

diff --git a/src/mesa/drivers/dri/i965/Makefile b/src/mesa/drivers/dri/i965/Makefile
index 7c3ac0c14ef..b05ba35d65f 100644
--- a/src/mesa/drivers/dri/i965/Makefile
+++ b/src/mesa/drivers/dri/i965/Makefile
@@ -108,6 +108,7 @@ CXX_SOURCES = \
 	brw_fs.cpp \
 	brw_fs_channel_expressions.cpp \
 	brw_fs_reg_allocate.cpp \
+	brw_fs_schedule_instructions.cpp \
 	brw_fs_vector_splitting.cpp
 
 ASM_SOURCES = 
diff --git a/src/mesa/drivers/dri/i965/brw_fs.cpp b/src/mesa/drivers/dri/i965/brw_fs.cpp
index 5b595a3c0e7..c24060b8c6f 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.cpp
+++ b/src/mesa/drivers/dri/i965/brw_fs.cpp
@@ -3696,6 +3696,8 @@ brw_wm_fs_emit(struct brw_context *brw, struct brw_wm_compile *c)
 	 progress = v.dead_code_eliminate() || progress;
       } while (progress);
 
+      v.schedule_instructions();
+
       if (0) {
 	 /* Debug of register spilling: Go spill everything. */
 	 int virtual_grf_count = v.virtual_grf_next;
diff --git a/src/mesa/drivers/dri/i965/brw_fs.h b/src/mesa/drivers/dri/i965/brw_fs.h
index 65c8a1e020f..7c991f32658 100644
--- a/src/mesa/drivers/dri/i965/brw_fs.h
+++ b/src/mesa/drivers/dri/i965/brw_fs.h
@@ -436,6 +436,8 @@ public:
    bool dead_code_eliminate();
    bool remove_duplicate_mrf_writes();
    bool virtual_grf_interferes(int a, int b);
+   void schedule_instructions();
+
    void generate_code();
    void generate_fb_write(fs_inst *inst);
    void generate_linterp(fs_inst *inst, struct brw_reg dst,
diff --git a/src/mesa/drivers/dri/i965/brw_fs_schedule_instructions.cpp b/src/mesa/drivers/dri/i965/brw_fs_schedule_instructions.cpp
new file mode 100644
index 00000000000..00aa99de7f1
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_fs_schedule_instructions.cpp
@@ -0,0 +1,474 @@
+/*
+ * Copyright © 2010 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:
+ *    Eric Anholt <[email protected]>
+ *
+ */
+
+extern "C" {
+
+#include <sys/types.h>
+
+#include "main/macros.h"
+#include "main/shaderobj.h"
+#include "main/uniforms.h"
+#include "program/prog_optimize.h"
+#include "program/register_allocate.h"
+#include "program/sampler.h"
+#include "program/hash_table.h"
+#include "brw_context.h"
+#include "brw_eu.h"
+#include "brw_wm.h"
+#include "talloc.h"
+}
+#include "brw_fs.h"
+#include "../glsl/glsl_types.h"
+#include "../glsl/ir_optimization.h"
+#include "../glsl/ir_print_visitor.h"
+
+/** @file brw_fs_schedule_instructions.cpp
+ *
+ * List scheduling of FS instructions.
+ *
+ * The basic model of the list scheduler is to take a basic block,
+ * compute a DAG of the dependencies (RAW ordering with latency, WAW
+ * ordering, WAR ordering), and make a list of the DAG heads.
+ * Heuristically pick a DAG head, then put all the children that are
+ * now DAG heads into the list of things to schedule.
+ *
+ * The heuristic is the important part.  We're trying to be cheap,
+ * since actually computing the optimal scheduling is NP complete.
+ * What we do is track a "current clock".  When we schedule a node, we
+ * update the earliest-unblocked clock time of its children, and
+ * increment the clock.  Then, when trying to schedule, we just pick
+ * the earliest-unblocked instruction to schedule.
+ *
+ * Note that often there will be many things which could execute
+ * immediately, and there are a range of heuristic options to choose
+ * from in picking among those.
+ */
+
+class schedule_node : public exec_node
+{
+public:
+   schedule_node(fs_inst *inst)
+   {
+      this->inst = inst;
+      this->child_array_size = 0;
+      this->children = NULL;
+      this->child_latency = NULL;
+      this->child_count = 0;
+      this->parent_count = 0;
+      this->unblocked_time = 0;
+
+      int chans = 8;
+      int math_latency = 22;
+
+      switch (inst->opcode) {
+      case FS_OPCODE_RCP:
+	 this->latency = 1 * chans * math_latency;
+	 break;
+      case FS_OPCODE_RSQ:
+	 this->latency = 2 * chans * math_latency;
+	 break;
+      case FS_OPCODE_SQRT:
+      case FS_OPCODE_LOG2:
+	 /* full precision log.  partial is 2. */
+	 this->latency = 3 * chans * math_latency;
+	 break;
+      case FS_OPCODE_EXP2:
+	 /* full precision.  partial is 3, same throughput. */
+	 this->latency = 4 * chans * math_latency;
+	 break;
+      case FS_OPCODE_POW:
+	 this->latency = 8 * chans * math_latency;
+	 break;
+      case FS_OPCODE_SIN:
+      case FS_OPCODE_COS:
+	 /* minimum latency, max is 12 rounds. */
+	 this->latency = 5 * chans * math_latency;
+	 break;
+      default:
+	 this->latency = 2;
+	 break;
+      }
+   }
+
+   fs_inst *inst;
+   schedule_node **children;
+   int *child_latency;
+   int child_count;
+   int parent_count;
+   int child_array_size;
+   int unblocked_time;
+   int latency;
+};
+
+class instruction_scheduler {
+public:
+   instruction_scheduler(fs_visitor *v, void *mem_ctx, int virtual_grf_count)
+   {
+      this->v = v;
+      this->mem_ctx = talloc_new(mem_ctx);
+      this->virtual_grf_count = virtual_grf_count;
+      this->instructions.make_empty();
+      this->instructions_to_schedule = 0;
+   }
+
+   ~instruction_scheduler()
+   {
+      talloc_free(this->mem_ctx);
+   }
+   void add_barrier_deps(schedule_node *n);
+   void add_dep(schedule_node *before, schedule_node *after, int latency);
+
+   void add_inst(fs_inst *inst);
+   void calculate_deps();
+   void schedule_instructions(fs_inst *next_block_header);
+
+   void *mem_ctx;
+
+   int instructions_to_schedule;
+   int virtual_grf_count;
+   exec_list instructions;
+   fs_visitor *v;
+};
+
+void
+instruction_scheduler::add_inst(fs_inst *inst)
+{
+   schedule_node *n = new(mem_ctx) schedule_node(inst);
+
+   assert(!inst->is_head_sentinel());
+   assert(!inst->is_tail_sentinel());
+
+   this->instructions_to_schedule++;
+
+   inst->remove();
+   instructions.push_tail(n);
+}
+
+/**
+ * Add a dependency between two instruction nodes.
+ *
+ * The @after node will be scheduled after @before.  We will try to
+ * schedule it @latency cycles after @before, but no guarantees there.
+ */
+void
+instruction_scheduler::add_dep(schedule_node *before, schedule_node *after,
+			       int latency)
+{
+   if (!before || !after)
+      return;
+
+   assert(before != after);
+
+   for (int i = 0; i < before->child_count; i++) {
+      if (before->children[i] == after) {
+	 before->child_latency[i] = MAX2(before->child_latency[i], latency);
+	 return;
+      }
+   }
+
+   if (before->child_array_size <= before->child_count) {
+      if (before->child_array_size < 16)
+	 before->child_array_size = 16;
+      else
+	 before->child_array_size *= 2;
+
+      before->children = talloc_realloc(mem_ctx, before->children,
+					schedule_node *,
+					before->child_array_size);
+      before->child_latency = talloc_realloc(mem_ctx, before->child_latency,
+					     int, before->child_array_size);
+   }
+
+   before->children[before->child_count] = after;
+   before->child_latency[before->child_count] = latency;
+   before->child_count++;
+   after->parent_count++;
+}
+
+/**
+ * Sometimes we really want this node to execute after everything that
+ * was before it and before everything that followed it.  This adds
+ * the deps to do so.
+ */
+void
+instruction_scheduler::add_barrier_deps(schedule_node *n)
+{
+   schedule_node *prev = (schedule_node *)n->prev;
+   schedule_node *next = (schedule_node *)n->next;
+
+   if (prev) {
+      while (!prev->is_head_sentinel()) {
+	 add_dep(prev, n, 0);
+	 prev = (schedule_node *)prev->prev;
+      }
+   }
+
+   if (next) {
+      while (!next->is_tail_sentinel()) {
+	 add_dep(n, next, 0);
+	 next = (schedule_node *)next->next;
+      }
+   }
+}
+
+void
+instruction_scheduler::calculate_deps()
+{
+   schedule_node *last_grf_write[virtual_grf_count];
+   schedule_node *last_mrf_write[BRW_MAX_MRF];
+   schedule_node *last_conditional_mod = NULL;
+
+   /* The last instruction always needs to still be the last
+    * instruction.  Either it's flow control (IF, ELSE, ENDIF, DO,
+    * WHILE) and scheduling other things after it would disturb the
+    * basic block, or it's FB_WRITE and we should do a better job at
+    * dead code elimination anyway.
+    */
+   schedule_node *last = (schedule_node *)instructions.get_tail();
+   add_barrier_deps(last);
+
+   memset(last_grf_write, 0, sizeof(last_grf_write));
+   memset(last_mrf_write, 0, sizeof(last_mrf_write));
+
+   /* top-to-bottom dependencies: RAW and WAW. */
+   foreach_iter(exec_list_iterator, iter, instructions) {
+      schedule_node *n = (schedule_node *)iter.get();
+      fs_inst *inst = n->inst;
+
+      /* read-after-write deps. */
+      for (int i = 0; i < 3; i++) {
+	 if (inst->src[i].file == GRF) {
+	    if (last_grf_write[inst->src[i].reg]) {
+	       add_dep(last_grf_write[inst->src[i].reg], n,
+		       last_grf_write[inst->src[i].reg]->latency);
+	    }
+	 } else if (inst->src[i].file != BAD_FILE &&
+		    inst->src[i].file != IMM &&
+		    inst->src[i].file != UNIFORM) {
+	    assert(inst->src[i].file != MRF);
+	    add_barrier_deps(n);
+	 }
+      }
+
+      for (int i = 0; i < inst->mlen; i++) {
+	 /* It looks like the MRF regs are released in the send
+	  * instruction once it's sent, not when the result comes
+	  * back.
+	  */
+	 if (last_mrf_write[inst->base_mrf + i]) {
+	    add_dep(last_mrf_write[inst->base_mrf + i], n,
+		    last_mrf_write[inst->base_mrf + i]->latency);
+	 }
+      }
+
+      if (inst->predicated) {
+	 assert(last_conditional_mod);
+	 add_dep(last_conditional_mod, n, last_conditional_mod->latency);
+      }
+
+      /* write-after-write deps. */
+      if (inst->dst.file == GRF) {
+	 if (last_grf_write[inst->dst.reg]) {
+	    add_dep(last_grf_write[inst->dst.reg], n,
+		    last_grf_write[inst->dst.reg]->latency);
+	 }
+	 last_grf_write[inst->dst.reg] = n;
+      } else if (inst->dst.file == MRF) {
+	 if (last_mrf_write[inst->dst.hw_reg]) {
+	    add_dep(last_mrf_write[inst->dst.hw_reg], n,
+		    last_mrf_write[inst->dst.hw_reg]->latency);
+	 }
+	 last_mrf_write[inst->dst.hw_reg] = n;
+      } else if (inst->dst.file != BAD_FILE) {
+	 add_barrier_deps(n);
+      }
+
+      if (inst->mlen > 0) {
+	 for (int i = 0; i < v->implied_mrf_writes(inst); i++) {
+	    if (last_mrf_write[inst->base_mrf + i]) {
+	       add_dep(last_mrf_write[inst->base_mrf + i], n,
+		       last_mrf_write[inst->base_mrf + i]->latency);
+	    }
+	    last_mrf_write[inst->base_mrf + i] = n;
+	 }
+      }
+
+      if (inst->conditional_mod) {
+	 add_dep(last_conditional_mod, n, 0);
+	 last_conditional_mod = n;
+      }
+   }
+
+   /* bottom-to-top dependencies: WAR */
+   memset(last_grf_write, 0, sizeof(last_grf_write));
+   memset(last_mrf_write, 0, sizeof(last_mrf_write));
+   last_conditional_mod = NULL;
+
+   exec_node *node;
+   exec_node *prev;
+   for (node = instructions.get_tail(), prev = node->prev;
+	!node->is_head_sentinel();
+	node = prev, prev = node->prev) {
+      schedule_node *n = (schedule_node *)node;
+      fs_inst *inst = n->inst;
+
+      /* write-after-read deps. */
+      for (int i = 0; i < 3; i++) {
+	 if (inst->src[i].file == GRF) {
+	    if (last_grf_write[inst->src[i].reg]) {
+	       add_dep(n, last_grf_write[inst->src[i].reg], n->latency);
+	    }
+	 } else if (inst->src[i].file != BAD_FILE &&
+		    inst->src[i].file != IMM &&
+		    inst->src[i].file != UNIFORM) {
+	    assert(inst->src[i].file != MRF);
+	    add_barrier_deps(n);
+	 }
+      }
+
+      for (int i = 0; i < inst->mlen; i++) {
+	 /* It looks like the MRF regs are released in the send
+	  * instruction once it's sent, not when the result comes
+	  * back.
+	  */
+	 add_dep(n, last_mrf_write[inst->base_mrf + i], 2);
+      }
+
+      if (inst->predicated) {
+	 if (last_conditional_mod) {
+	    add_dep(n, last_conditional_mod, n->latency);
+	 }
+      }
+
+      /* Update the things this instruction wrote, so earlier reads
+       * can mark this as WAR dependency.
+       */
+      if (inst->dst.file == GRF) {
+	 last_grf_write[inst->dst.reg] = n;
+      } else if (inst->dst.file == MRF) {
+	 last_mrf_write[inst->dst.hw_reg] = n;
+      } else if (inst->dst.file != BAD_FILE) {
+	 add_barrier_deps(n);
+      }
+
+      if (inst->mlen > 0) {
+	 for (int i = 0; i < v->implied_mrf_writes(inst); i++) {
+	    last_mrf_write[inst->base_mrf + i] = n;
+	 }
+      }
+
+      if (inst->conditional_mod)
+	 last_conditional_mod = n;
+   }
+}
+
+void
+instruction_scheduler::schedule_instructions(fs_inst *next_block_header)
+{
+   int time = 0;
+
+   /* Remove non-DAG heads from the list. */
+   foreach_iter(exec_list_iterator, iter, instructions) {
+      schedule_node *n = (schedule_node *)iter.get();
+      if (n->parent_count != 0)
+	 n->remove();
+   }
+
+   while (!instructions.is_empty()) {
+      schedule_node *chosen = NULL;
+      int chosen_time = 0;
+
+      foreach_iter(exec_list_iterator, iter, instructions) {
+	 schedule_node *n = (schedule_node *)iter.get();
+
+	 if (!chosen || n->unblocked_time < chosen_time) {
+	    chosen = n;
+	    chosen_time = n->unblocked_time;
+	 }
+      }
+
+      /* Schedule this instruction. */
+      assert(chosen);
+      chosen->remove();
+      next_block_header->insert_before(chosen->inst);
+      instructions_to_schedule--;
+
+      /* Bump the clock.  If we expected a delay for scheduling, then
+       * bump the clock to reflect that.
+       */
+      time = MAX2(time + 1, chosen_time);
+
+      /* Now that we've scheduled a new instruction, some of its
+       * children can be promoted to the list of instructions ready to
+       * be scheduled.  Update the children's unblocked time for this
+       * DAG edge as we do so.
+       */
+      for (int i = 0; i < chosen->child_count; i++) {
+	 schedule_node *child = chosen->children[i];
+
+	 child->unblocked_time = MAX2(child->unblocked_time,
+				      time + chosen->child_latency[i]);
+
+	 child->parent_count--;
+	 if (child->parent_count == 0) {
+	    instructions.push_tail(child);
+	 }
+      }
+   }
+
+   assert(instructions_to_schedule == 0);
+}
+
+void
+fs_visitor::schedule_instructions()
+{
+   fs_inst *next_block_header = (fs_inst *)instructions.head;
+   instruction_scheduler sched(this, mem_ctx, this->virtual_grf_next);
+
+   while (!next_block_header->is_tail_sentinel()) {
+      /* Add things to be scheduled until we get to a new BB. */
+      while (!next_block_header->is_tail_sentinel()) {
+	 fs_inst *inst = next_block_header;
+	 next_block_header = (fs_inst *)next_block_header->next;
+
+	 sched.add_inst(inst);
+	 if (inst->opcode == BRW_OPCODE_IF ||
+	     inst->opcode == BRW_OPCODE_ELSE ||
+	     inst->opcode == BRW_OPCODE_ENDIF ||
+	     inst->opcode == BRW_OPCODE_DO ||
+	     inst->opcode == BRW_OPCODE_WHILE ||
+	     inst->opcode == BRW_OPCODE_BREAK ||
+	     inst->opcode == BRW_OPCODE_CONTINUE) {
+	    break;
+	 }
+      }
+      sched.calculate_deps();
+      sched.schedule_instructions(next_block_header);
+   }
+
+   this->live_intervals_valid = false;
+}