i965/vs: Move virtual GRFs with array accesses to them to scratch space.
authorEric Anholt <eric@anholt.net>
Sun, 7 Aug 2011 19:15:26 +0000 (12:15 -0700)
committerEric Anholt <eric@anholt.net>
Tue, 16 Aug 2011 20:04:42 +0000 (13:04 -0700)
src/mesa/drivers/dri/i965/brw_defines.h
src/mesa/drivers/dri/i965/brw_vec4.h
src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp

index e3823c65d1a1949e92dccb82adba7d4a71430c79..b740d87c93372d596ca0485b3b7436192181eff7 100644 (file)
@@ -640,6 +640,8 @@ enum opcode {
    FS_OPCODE_PULL_CONSTANT_LOAD,
 
    VS_OPCODE_URB_WRITE,
+   VS_OPCODE_SCRATCH_READ,
+   VS_OPCODE_SCRATCH_WRITE,
 };
 
 #define BRW_PREDICATE_NONE             0
index bb40c71e4c9d8054364cf3b35db564c984707b4d..2f171b72049ac6b256faf79103a9f9c254bbef24 100644 (file)
@@ -376,6 +376,7 @@ public:
    void setup_payload();
    void reg_allocate_trivial();
    void reg_allocate();
+   void move_grf_array_access_to_scratch();
 
    vec4_instruction *emit(enum opcode opcode);
 
@@ -424,6 +425,17 @@ public:
    int emit_vue_header_gen4(int header_mrf);
    void emit_urb_writes(void);
 
+   src_reg get_scratch_offset(vec4_instruction *inst,
+                             src_reg *reladdr, int reg_offset);
+   void emit_scratch_read(vec4_instruction *inst,
+                         dst_reg dst,
+                         src_reg orig_src,
+                         int base_offset);
+   void emit_scratch_write(vec4_instruction *inst,
+                          src_reg temp,
+                          dst_reg orig_dst,
+                          int base_offset);
+
    GLboolean try_emit_sat(ir_expression *ir);
 
    bool process_move_condition(ir_rvalue *ir);
index bc3110b0458141195256029dcaabfb6cfa16774e..57eb467567e614a40b2a538a82f20f9e6668f2aa 100644 (file)
@@ -321,7 +321,7 @@ vec4_visitor::generate_vs_instruction(vec4_instruction *instruction,
 bool
 vec4_visitor::run()
 {
-   /* Generate FS IR for main().  (the visitor only descends into
+   /* Generate VS IR for main().  (the visitor only descends into
     * functions called "main").
     */
    foreach_iter(exec_list_iterator, iter, *shader->ir) {
@@ -332,6 +332,14 @@ vec4_visitor::run()
 
    emit_urb_writes();
 
+   /* Before any optimization, push array accesses out to scratch
+    * space where we need them to be.  This pass may allocate new
+    * virtual GRFs, so we want to do it early.  It also makes sure
+    * that we have reladdr computations available for CSE, since we'll
+    * often do repeated subexpressions for those.
+    */
+   move_grf_array_access_to_scratch();
+
    if (failed)
       return false;
 
index e01318af1ab997861efd2cde5b16556231f44a54..049af6c3992c7153140f7e0d12f91a5733c085b3 100644 (file)
@@ -1805,6 +1805,169 @@ vec4_visitor::emit_urb_writes()
       c->prog_data.urb_entry_size = ALIGN(urb_entry_size, 4) / 4;
 }
 
+src_reg
+vec4_visitor::get_scratch_offset(vec4_instruction *inst,
+                                src_reg *reladdr, int reg_offset)
+{
+   /* Because we store the values to scratch interleaved like our
+    * vertex data, we need to scale the vec4 index by 2.
+    */
+   int message_header_scale = 2;
+
+   /* Pre-gen6, the message header uses byte offsets instead of vec4
+    * (16-byte) offset units.
+    */
+   if (intel->gen < 6)
+      message_header_scale *= 16;
+
+   if (reladdr) {
+      src_reg index = src_reg(this, glsl_type::int_type);
+
+      vec4_instruction *add = emit(BRW_OPCODE_ADD,
+                                  dst_reg(index),
+                                  *reladdr,
+                                  src_reg(reg_offset));
+      /* Move our new instruction from the tail to its correct place. */
+      add->remove();
+      inst->insert_before(add);
+
+      vec4_instruction *mul = emit(BRW_OPCODE_MUL, dst_reg(index),
+                                  index, src_reg(message_header_scale));
+      mul->remove();
+      inst->insert_before(mul);
+
+      return index;
+   } else {
+      return src_reg(reg_offset * message_header_scale);
+   }
+}
+
+/**
+ * Emits an instruction before @inst to load the value named by @orig_src
+ * from scratch space at @base_offset to @temp.
+ */
+void
+vec4_visitor::emit_scratch_read(vec4_instruction *inst,
+                               dst_reg temp, src_reg orig_src,
+                               int base_offset)
+{
+   int reg_offset = base_offset + orig_src.reg_offset;
+   src_reg index = get_scratch_offset(inst, orig_src.reladdr, reg_offset);
+
+   vec4_instruction *scratch_read_inst = emit(VS_OPCODE_SCRATCH_READ,
+                                             temp, index);
+
+   scratch_read_inst->base_mrf = 14;
+   scratch_read_inst->mlen = 1;
+   /* Move our instruction from the tail to its correct place. */
+   scratch_read_inst->remove();
+   inst->insert_before(scratch_read_inst);
+}
+
+/**
+ * Emits an instruction after @inst to store the value to be written
+ * to @orig_dst to scratch space at @base_offset, from @temp.
+ */
+void
+vec4_visitor::emit_scratch_write(vec4_instruction *inst,
+                                src_reg temp, dst_reg orig_dst,
+                                int base_offset)
+{
+   int reg_offset = base_offset + orig_dst.reg_offset;
+   src_reg index = get_scratch_offset(inst, orig_dst.reladdr, reg_offset);
+
+   dst_reg dst = dst_reg(brw_writemask(brw_vec8_grf(0, 0),
+                                      orig_dst.writemask));
+   vec4_instruction *scratch_write_inst = emit(VS_OPCODE_SCRATCH_WRITE,
+                                              dst, temp, index);
+   scratch_write_inst->base_mrf = 13;
+   scratch_write_inst->mlen = 2;
+   scratch_write_inst->predicate = inst->predicate;
+   /* Move our instruction from the tail to its correct place. */
+   scratch_write_inst->remove();
+   inst->insert_after(scratch_write_inst);
+}
+
+/**
+ * We can't generally support array access in GRF space, because a
+ * single instruction's destination can only span 2 contiguous
+ * registers.  So, we send all GRF arrays that get variable index
+ * access to scratch space.
+ */
+void
+vec4_visitor::move_grf_array_access_to_scratch()
+{
+   int scratch_loc[this->virtual_grf_count];
+
+   for (int i = 0; i < this->virtual_grf_count; i++) {
+      scratch_loc[i] = -1;
+   }
+
+   /* First, calculate the set of virtual GRFs that need to be punted
+    * to scratch due to having any array access on them, and where in
+    * scratch.
+    */
+   foreach_list(node, &this->instructions) {
+      vec4_instruction *inst = (vec4_instruction *)node;
+
+      if (inst->dst.file == GRF && inst->dst.reladdr &&
+         scratch_loc[inst->dst.reg] == -1) {
+        scratch_loc[inst->dst.reg] = c->last_scratch;
+        c->last_scratch += this->virtual_grf_sizes[inst->dst.reg] * 8 * 4;
+      }
+
+      for (int i = 0 ; i < 3; i++) {
+        src_reg *src = &inst->src[i];
+
+        if (src->file == GRF && src->reladdr &&
+            scratch_loc[src->reg] == -1) {
+           scratch_loc[src->reg] = c->last_scratch;
+           c->last_scratch += this->virtual_grf_sizes[src->reg] * 8 * 4;
+        }
+      }
+   }
+
+   /* Now, for anything that will be accessed through scratch, rewrite
+    * it to load/store.  Note that this is a _safe list walk, because
+    * we may generate a new scratch_write instruction after the one
+    * we're processing.
+    */
+   foreach_list_safe(node, &this->instructions) {
+      vec4_instruction *inst = (vec4_instruction *)node;
+
+      /* Set up the annotation tracking for new generated instructions. */
+      base_ir = inst->ir;
+      current_annotation = inst->annotation;
+
+      if (inst->dst.file == GRF && scratch_loc[inst->dst.reg] != -1) {
+        src_reg temp = src_reg(this, glsl_type::vec4_type);
+
+        emit_scratch_write(inst, temp, inst->dst, scratch_loc[inst->dst.reg]);
+
+        inst->dst.file = temp.file;
+        inst->dst.reg = temp.reg;
+        inst->dst.reg_offset = temp.reg_offset;
+        inst->dst.reladdr = NULL;
+      }
+
+      for (int i = 0 ; i < 3; i++) {
+        if (inst->src[i].file != GRF || scratch_loc[inst->src[i].reg] == -1)
+           continue;
+
+        dst_reg temp = dst_reg(this, glsl_type::vec4_type);
+
+        emit_scratch_read(inst, temp, inst->src[i],
+                          scratch_loc[inst->src[i].reg]);
+
+        inst->src[i].file = temp.file;
+        inst->src[i].reg = temp.reg;
+        inst->src[i].reg_offset = temp.reg_offset;
+        inst->src[i].reladdr = NULL;
+      }
+   }
+}
+
+
 vec4_visitor::vec4_visitor(struct brw_vs_compile *c,
                           struct gl_shader_program *prog,
                           struct brw_shader *shader)