i965: Move VUE map computation to once at VS compile time.

With this and the previous patch, 640x480 nexuiz is running 0.169118%
+/- 0.0863696% faster (n=121).  On a VS state change microbenchmark,
performance is increased 8.28645% +/- 0.460478% (n=52).

v2: Fix CACHE_NEW_VS comment.

Reviewed-by: Kenneth Graunke <[email protected]>

12 files changed, 42 insertions, 49 deletions

diff --git a/src/mesa/drivers/dri/i965/brw_clip.c b/src/mesa/drivers/dri/i965/brw_clip.c
index 5b5f551948e..d411208f2c9 100644
--- a/src/mesa/drivers/dri/i965/brw_clip.c
+++ b/src/mesa/drivers/dri/i965/brw_clip.c
@@ -69,7 +69,7 @@ static void compile_clip_prog( struct brw_context *brw,
    c.func.single_program_flow = 1;
 
    c.key = *key;
-   brw_compute_vue_map(&c.vue_map, intel, brw->vs.prog_data);
+   c.vue_map = brw->vs.prog_data->vue_map;
 
    /* nr_regs is the number of registers filled by reading data from the VUE.
     * This program accesses the entire VUE, so nr_regs needs to be the size of
diff --git a/src/mesa/drivers/dri/i965/brw_context.h b/src/mesa/drivers/dri/i965/brw_context.h
index c6860a76d50..503585c70c8 100644
--- a/src/mesa/drivers/dri/i965/brw_context.h
+++ b/src/mesa/drivers/dri/i965/brw_context.h
@@ -381,6 +381,8 @@ struct brw_gs_prog_data {
 };
 
 struct brw_vs_prog_data {
+   struct brw_vue_map vue_map;
+
    GLuint curb_read_length;
    GLuint urb_read_length;
    GLuint total_grf;
@@ -1045,9 +1047,6 @@ void brw_upload_cs_urb_state(struct brw_context *brw);
 int brw_disasm (FILE *file, struct brw_instruction *inst, int gen);
 
 /* brw_vs.c */
-void brw_compute_vue_map(struct brw_vue_map *vue_map,
-                         const struct intel_context *intel,
-                         const struct brw_vs_prog_data *prog_data);
 gl_clip_plane *brw_select_clip_planes(struct gl_context *ctx);
 
 /* brw_wm.c */
diff --git a/src/mesa/drivers/dri/i965/brw_gs.c b/src/mesa/drivers/dri/i965/brw_gs.c
index c6132df26d1..bfca169f7cb 100644
--- a/src/mesa/drivers/dri/i965/brw_gs.c
+++ b/src/mesa/drivers/dri/i965/brw_gs.c
@@ -56,7 +56,7 @@ static void compile_gs_prog( struct brw_context *brw,
    memset(&c, 0, sizeof(c));
    
    c.key = *key;
-   brw_compute_vue_map(&c.vue_map, intel, brw->vs.prog_data);
+   c.vue_map = brw->vs.prog_data->vue_map;
    c.nr_regs = (c.vue_map.num_slots + 1)/2;
 
    mem_ctx = NULL;
diff --git a/src/mesa/drivers/dri/i965/brw_sf.c b/src/mesa/drivers/dri/i965/brw_sf.c
index 6c28d773c59..6e63583e546 100644
--- a/src/mesa/drivers/dri/i965/brw_sf.c
+++ b/src/mesa/drivers/dri/i965/brw_sf.c
@@ -63,7 +63,7 @@ static void compile_sf_prog( struct brw_context *brw,
    brw_init_compile(brw, &c.func, mem_ctx);
 
    c.key = *key;
-   brw_compute_vue_map(&c.vue_map, intel, brw->vs.prog_data);
+   c.vue_map = brw->vs.prog_data->vue_map;
    c.urb_entry_read_offset = brw_sf_compute_urb_entry_read_offset(intel);
    c.nr_attr_regs = (c.vue_map.num_slots + 1)/2 - c.urb_entry_read_offset;
    c.nr_setup_regs = c.nr_attr_regs;
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
index 917c927a125..f9eed61d92c 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_emit.cpp
@@ -96,7 +96,7 @@ vec4_visitor::setup_attributes(int payload_reg)
 
    prog_data->urb_read_length = (nr_attributes + 1) / 2;
 
-   unsigned vue_entries = MAX2(nr_attributes, c->vue_map.num_slots);
+   unsigned vue_entries = MAX2(nr_attributes, c->prog_data.vue_map.num_slots);
 
    if (intel->gen == 6)
       c->prog_data.urb_entry_size = ALIGN(vue_entries, 8) / 8;
diff --git a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
index 5dfe1c1354c..f9a08a011f2 100644
--- a/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
+++ b/src/mesa/drivers/dri/i965/brw_vec4_visitor.cpp
@@ -2252,8 +2252,6 @@ vec4_visitor::emit_urb_writes()
 
    /* FINISHME: edgeflag */
 
-   brw_compute_vue_map(&c->vue_map, intel, &c->prog_data);
-
    /* First mrf is the g0-based message header containing URB handles and such,
     * which is implied in VS_OPCODE_URB_WRITE.
     */
@@ -2265,8 +2263,8 @@ vec4_visitor::emit_urb_writes()
 
    /* Set up the VUE data for the first URB write */
    int slot;
-   for (slot = 0; slot < c->vue_map.num_slots; ++slot) {
-      emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]);
+   for (slot = 0; slot < c->prog_data.vue_map.num_slots; ++slot) {
+      emit_urb_slot(mrf++, c->prog_data.vue_map.slot_to_vert_result[slot]);
 
       /* If this was max_usable_mrf, we can't fit anything more into this URB
        * WRITE.
@@ -2281,16 +2279,16 @@ vec4_visitor::emit_urb_writes()
    vec4_instruction *inst = emit(VS_OPCODE_URB_WRITE);
    inst->base_mrf = base_mrf;
    inst->mlen = align_interleaved_urb_mlen(brw, mrf - base_mrf);
-   inst->eot = (slot >= c->vue_map.num_slots);
+   inst->eot = (slot >= c->prog_data.vue_map.num_slots);
 
    /* Optional second URB write */
    if (!inst->eot) {
       mrf = base_mrf + 1;
 
-      for (; slot < c->vue_map.num_slots; ++slot) {
+      for (; slot < c->prog_data.vue_map.num_slots; ++slot) {
 	 assert(mrf < max_usable_mrf);
 
-         emit_urb_slot(mrf++, c->vue_map.slot_to_vert_result[slot]);
+         emit_urb_slot(mrf++, c->prog_data.vue_map.slot_to_vert_result[slot]);
       }
 
       current_annotation = "URB write";
diff --git a/src/mesa/drivers/dri/i965/brw_vs.c b/src/mesa/drivers/dri/i965/brw_vs.c
index be82177f40d..ca205cdf79a 100644
--- a/src/mesa/drivers/dri/i965/brw_vs.c
+++ b/src/mesa/drivers/dri/i965/brw_vs.c
@@ -57,13 +57,13 @@ static inline void assign_vue_slot(struct brw_vue_map *vue_map,
  * prog_data->userclip and prog_data->outputs_written in their key
  * (generated by CACHE_NEW_VS_PROG).
  */
-void
-brw_compute_vue_map(struct brw_vue_map *vue_map,
-                    const struct intel_context *intel,
-		    const struct brw_vs_prog_data *prog_data)
+static void
+brw_compute_vue_map(struct brw_vs_compile *c)
 {
-   bool userclip_active = prog_data->userclip;
-   GLbitfield64 outputs_written = prog_data->outputs_written;
+   struct brw_context *brw = c->func.brw;
+   const struct intel_context *intel = &brw->intel;
+   struct brw_vue_map *vue_map = &c->prog_data.vue_map;
+   GLbitfield64 outputs_written = c->prog_data.outputs_written;
    int i;
 
    vue_map->num_slots = 0;
@@ -118,7 +118,7 @@ brw_compute_vue_map(struct brw_vue_map *vue_map,
        */
       assign_vue_slot(vue_map, VERT_RESULT_PSIZ);
       assign_vue_slot(vue_map, VERT_RESULT_HPOS);
-      if (userclip_active) {
+      if (c->key.userclip_active) {
          assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST0);
          assign_vue_slot(vue_map, VERT_RESULT_CLIP_DIST1);
       }
@@ -218,6 +218,8 @@ do_vs_prog(struct brw_context *brw,
       c.prog_data.inputs_read |= VERT_BIT_EDGEFLAG;
    }
 
+   brw_compute_vue_map(&c);
+
    /* Put dummy slots into the VUE for the SF to put the replaced
     * point sprite coords in.  We shouldn't need these dummy slots,
     * which take up precious URB space, but it would mean that the SF
diff --git a/src/mesa/drivers/dri/i965/brw_vs.h b/src/mesa/drivers/dri/i965/brw_vs.h
index 8814251a84c..490fcc0622b 100644
--- a/src/mesa/drivers/dri/i965/brw_vs.h
+++ b/src/mesa/drivers/dri/i965/brw_vs.h
@@ -92,7 +92,6 @@ struct brw_vs_compile {
 
    GLuint nr_inputs;
 
-   struct brw_vue_map vue_map;
    GLuint first_output;
    GLuint last_scratch;
 
diff --git a/src/mesa/drivers/dri/i965/brw_vs_emit.c b/src/mesa/drivers/dri/i965/brw_vs_emit.c
index 07fc0af6a9d..4bdd366ff34 100644
--- a/src/mesa/drivers/dri/i965/brw_vs_emit.c
+++ b/src/mesa/drivers/dri/i965/brw_vs_emit.c
@@ -173,7 +173,6 @@ static inline bool can_use_direct_mrf(int vert_result,
  */
 static void brw_vs_alloc_regs( struct brw_vs_compile *c )
 {
-   struct brw_context *brw = c->func.brw;
    struct intel_context *intel = &c->func.brw->intel;
    GLuint i, reg = 0, slot;
    int attributes_in_vue;
@@ -326,13 +325,12 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c )
 
    /* Allocate outputs.  The non-position outputs go straight into message regs.
     */
-   brw_compute_vue_map(&c->vue_map, intel, &c->prog_data);
    c->first_output = reg;
 
    first_reladdr_output = get_first_reladdr_output(&c->vp->program);
 
-   for (slot = 0; slot < c->vue_map.num_slots; slot++) {
-      int vert_result = c->vue_map.slot_to_vert_result[slot];
+   for (slot = 0; slot < c->prog_data.vue_map.num_slots; slot++) {
+      int vert_result = c->prog_data.vue_map.slot_to_vert_result[slot];
       assert(vert_result < Elements(c->regs[PROGRAM_OUTPUT]));
       if (can_use_direct_mrf(vert_result, first_reladdr_output, slot)) {
          c->regs[PROGRAM_OUTPUT][vert_result] = brw_message_reg(slot + 1);
@@ -405,7 +403,7 @@ static void brw_vs_alloc_regs( struct brw_vs_compile *c )
    /* The VS VUEs are shared by VF (outputting our inputs) and VS, so size
     * them to fit the biggest thing they need to.
     */
-   attributes_in_vue = MAX2(c->vue_map.num_slots, c->nr_inputs);
+   attributes_in_vue = MAX2(c->prog_data.vue_map.num_slots, c->nr_inputs);
 
    if (intel->gen == 6) {
       /* Each attribute is 32 bytes (2 vec4s), so dividing by 8 gives us the
@@ -1678,12 +1676,12 @@ static void emit_vertex_write( struct brw_vs_compile *c)
    }
 
    /* Move variable-addressed, non-overflow outputs to their MRFs. */
-   for (slot = len_vertex_header; slot < c->vue_map.num_slots; ++slot) {
+   for (slot = len_vertex_header; slot < c->prog_data.vue_map.num_slots; ++slot) {
       if (slot >= MAX_SLOTS_IN_FIRST_URB_WRITE)
          break;
 
       int mrf = slot + 1;
-      int vert_result = c->vue_map.slot_to_vert_result[slot];
+      int vert_result = c->prog_data.vue_map.slot_to_vert_result[slot];
       if (c->regs[PROGRAM_OUTPUT][vert_result].file ==
           BRW_GENERAL_REGISTER_FILE) {
          brw_MOV(p, brw_message_reg(mrf),
@@ -1691,7 +1689,7 @@ static void emit_vertex_write( struct brw_vs_compile *c)
       }
    }
 
-   eot = (slot >= c->vue_map.num_slots);
+   eot = (slot >= c->prog_data.vue_map.num_slots);
 
    /* Message header, plus the (first part of the) VUE. */
    msg_len = 1 + slot;
@@ -1712,14 +1710,14 @@ static void emit_vertex_write( struct brw_vs_compile *c)
 		 0, 		/* urb destination offset */
 		 BRW_URB_SWIZZLE_INTERLEAVE);
 
-   if (slot < c->vue_map.num_slots) {
+   if (slot < c->prog_data.vue_map.num_slots) {
       /* Not all of the vertex outputs/results fit into the MRF.
        * Move the overflowed attributes from the GRF to the MRF and
        * issue another brw_urb_WRITE().
        */
       GLuint mrf = 1;
-      for (; slot < c->vue_map.num_slots; ++slot) {
-         int vert_result = c->vue_map.slot_to_vert_result[slot];
+      for (; slot < c->prog_data.vue_map.num_slots; ++slot) {
+         int vert_result = c->prog_data.vue_map.slot_to_vert_result[slot];
          /* move from GRF to MRF */
          brw_MOV(p, brw_message_reg(mrf),
                  c->regs[PROGRAM_OUTPUT][vert_result]);
diff --git a/src/mesa/drivers/dri/i965/gen6_sf_state.c b/src/mesa/drivers/dri/i965/gen6_sf_state.c
index c4e7c4c6488..95ed1f74ce4 100644
--- a/src/mesa/drivers/dri/i965/gen6_sf_state.c
+++ b/src/mesa/drivers/dri/i965/gen6_sf_state.c
@@ -112,7 +112,6 @@ upload_sf_state(struct brw_context *brw)
 {
    struct intel_context *intel = &brw->intel;
    struct gl_context *ctx = &intel->ctx;
-   struct brw_vue_map vue_map;
    uint32_t urb_entry_read_length;
    /* BRW_NEW_FRAGMENT_PROGRAM */
    uint32_t num_outputs = _mesa_bitcount_64(brw->fragment_program->Base.InputsRead);
@@ -129,8 +128,8 @@ upload_sf_state(struct brw_context *brw)
    uint32_t point_sprite_origin;
 
    /* CACHE_NEW_VS_PROG */
-   brw_compute_vue_map(&vue_map, intel, brw->vs.prog_data);
-   urb_entry_read_length = (vue_map.num_slots + 1)/2 - urb_entry_read_offset;
+   urb_entry_read_length = ((brw->vs.prog_data->vue_map.num_slots + 1) / 2 -
+			    urb_entry_read_offset);
    if (urb_entry_read_length == 0) {
       /* Setting the URB entry read length to 0 causes undefined behavior, so
        * if we have no URB data to read, set it to 1.
@@ -301,9 +300,10 @@ upload_sf_state(struct brw_context *brw)
        */
       assert(input_index < 16 || attr == input_index);
 
-      /* _NEW_LIGHT | _NEW_PROGRAM */
+      /* CACHE_NEW_VS_PROG | _NEW_LIGHT | _NEW_PROGRAM */
       attr_overrides[input_index++] =
-         get_attr_override(&vue_map, urb_entry_read_offset, attr,
+         get_attr_override(&brw->vs.prog_data->vue_map,
+			   urb_entry_read_offset, attr,
                            ctx->VertexProgram._TwoSideEnabled);
    }
 
diff --git a/src/mesa/drivers/dri/i965/gen7_sf_state.c b/src/mesa/drivers/dri/i965/gen7_sf_state.c
index 49460b2802d..c8f4393a529 100644
--- a/src/mesa/drivers/dri/i965/gen7_sf_state.c
+++ b/src/mesa/drivers/dri/i965/gen7_sf_state.c
@@ -33,7 +33,6 @@ upload_sbe_state(struct brw_context *brw)
 {
    struct intel_context *intel = &brw->intel;
    struct gl_context *ctx = &intel->ctx;
-   struct brw_vue_map vue_map;
    uint32_t urb_entry_read_length;
    /* BRW_NEW_FRAGMENT_PROGRAM */
    uint32_t num_outputs = _mesa_bitcount_64(brw->fragment_program->Base.InputsRead);
@@ -49,8 +48,8 @@ upload_sbe_state(struct brw_context *brw)
    uint32_t point_sprite_origin;
 
    /* CACHE_NEW_VS_PROG */
-   brw_compute_vue_map(&vue_map, intel, brw->vs.prog_data);
-   urb_entry_read_length = (vue_map.num_slots + 1)/2 - urb_entry_read_offset;
+   urb_entry_read_length = ((brw->vs.prog_data->vue_map.num_slots + 1) / 2 -
+			    urb_entry_read_offset);
    if (urb_entry_read_length == 0) {
       /* Setting the URB entry read length to 0 causes undefined behavior, so
        * if we have no URB data to read, set it to 1.
@@ -114,9 +113,10 @@ upload_sbe_state(struct brw_context *brw)
        */
       assert(input_index < 16 || attr == input_index);
 
-      /* _NEW_LIGHT | _NEW_PROGRAM */
+      /* CACHE_NEW_VS_PROG | _NEW_LIGHT | _NEW_PROGRAM */
       attr_overrides[input_index++] =
-         get_attr_override(&vue_map, urb_entry_read_offset, attr,
+         get_attr_override(&brw->vs.prog_data->vue_map,
+			   urb_entry_read_offset, attr,
                            ctx->VertexProgram._TwoSideEnabled);
    }
 
diff --git a/src/mesa/drivers/dri/i965/gen7_sol_state.c b/src/mesa/drivers/dri/i965/gen7_sol_state.c
index 134153e6703..1a89503d2d5 100644
--- a/src/mesa/drivers/dri/i965/gen7_sol_state.c
+++ b/src/mesa/drivers/dri/i965/gen7_sol_state.c
@@ -239,14 +239,11 @@ upload_sol_state(struct brw_context *brw)
    struct gl_transform_feedback_object *xfb_obj =
       ctx->TransformFeedback.CurrentObject;
    bool active = xfb_obj->Active && !xfb_obj->Paused;
-   struct brw_vue_map vue_map;
-
-   /* CACHE_NEW_VS_PROG */
-   brw_compute_vue_map(&vue_map, intel, brw->vs.prog_data);
 
    if (active) {
       upload_3dstate_so_buffers(brw);
-      upload_3dstate_so_decl_list(brw, &vue_map);
+      /* CACHE_NEW_VS_PROG */
+      upload_3dstate_so_decl_list(brw, &brw->vs.prog_data->vue_map);
 
       intel->batch.needs_sol_reset = true;
    }
@@ -256,7 +253,7 @@ upload_sol_state(struct brw_context *brw)
     * MMIO register updates (current performed by the kernel at each batch
     * emit).
     */
-   upload_3dstate_streamout(brw, active, &vue_map);
+   upload_3dstate_streamout(brw, active, &brw->vs.prog_data->vue_map);
 }
 
 const struct brw_tracked_state gen7_sol_state = {