i965: Move payload reg setup to compile, not lookup time.

Payload reg setup on gen6 depends more on the dispatch width as well
as the uses_depth, computes_depth, and other flags.  That's something
we want to decide at compile time, not at cache lookup.  As a bonus,
the fragment shader program cache lookup should be cheaper now that
there's less to compute for the hash key.

src/mesa/drivers/dri/i965/brw_fs.cpp

@@ -1955,7 +1955,7 @@ fs_visitor::emit_interpolation_setup_gen6()
    emit(fs_inst(BRW_OPCODE_MOV, this->pixel_y, int_pixel_y));
 
    this->current_annotation = "compute 1/pos.w";
-   this->wpos_w = fs_reg(brw_vec8_grf(c->key.source_w_reg, 0));
+   this->wpos_w = fs_reg(brw_vec8_grf(c->source_w_reg, 0));
    this->pixel_w = fs_reg(this, glsl_type::float_type);
    emit_math(FS_OPCODE_RCP, this->pixel_w, wpos_w);
 
@@ -1983,17 +1983,17 @@ fs_visitor::emit_fb_writes()
       nr += 2;
    }
 
-   if (c->key.aa_dest_stencil_reg) {
+   if (c->aa_dest_stencil_reg) {
       emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
-		   fs_reg(brw_vec8_grf(c->key.aa_dest_stencil_reg, 0))));
+		   fs_reg(brw_vec8_grf(c->aa_dest_stencil_reg, 0))));
    }
 
    /* Reserve space for color. It'll be filled in per MRT below. */
    int color_mrf = nr;
    nr += 4;
 
-   if (c->key.source_depth_to_render_target) {
-      if (c->key.computes_depth) {
+   if (c->source_depth_to_render_target) {
+      if (c->computes_depth) {
 	 /* Hand over gl_FragDepth. */
 	 assert(this->frag_depth);
 	 fs_reg depth = *(variable_storage(this->frag_depth));
@@ -2002,13 +2002,13 @@ fs_visitor::emit_fb_writes()
       } else {
 	 /* Pass through the payload depth. */
 	 emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
-		      fs_reg(brw_vec8_grf(c->key.source_depth_reg, 0))));
+		      fs_reg(brw_vec8_grf(c->source_depth_reg, 0))));
       }
    }
 
-   if (c->key.dest_depth_reg) {
+   if (c->dest_depth_reg) {
       emit(fs_inst(BRW_OPCODE_MOV, fs_reg(MRF, nr++),
-		   fs_reg(brw_vec8_grf(c->key.dest_depth_reg, 0))));
+		   fs_reg(brw_vec8_grf(c->dest_depth_reg, 0))));
    }
 
    fs_reg color = reg_undef;
@@ -2458,7 +2458,7 @@ fs_visitor::generate_pull_constant_load(fs_inst *inst, struct brw_reg dst)
 void
 fs_visitor::assign_curb_setup()
 {
-   c->prog_data.first_curbe_grf = c->key.nr_payload_regs;
+   c->prog_data.first_curbe_grf = c->nr_payload_regs;
    c->prog_data.curb_read_length = ALIGN(c->prog_data.nr_params, 8) / 8;
 
    /* Map the offsets in the UNIFORM file to fixed HW regs. */

src/mesa/drivers/dri/i965/brw_wm.c

@@ -119,6 +119,62 @@ brw_wm_non_glsl_emit(struct brw_context *brw, struct brw_wm_compile *c)
    brw_wm_emit(c);
 }
 
+static void
+brw_wm_payload_setup(struct brw_context *brw,
+		     struct brw_wm_compile *c)
+{
+   struct intel_context *intel = &brw->intel;
+   bool uses_depth = (c->fp->program.Base.InputsRead &
+		      (1 << FRAG_ATTRIB_WPOS)) != 0;
+
+   if (intel->gen >= 6) {
+      /* R0-1: masks, pixel X/Y coordinates. */
+      c->nr_payload_regs = 2;
+      /* R2: only for 32-pixel dispatch.*/
+      /* R3-4: perspective pixel location barycentric */
+      c->nr_payload_regs += 2;
+      /* R5-6: perspective pixel location bary for dispatch width != 8 */
+      if (!c->fp->isGLSL) { /* dispatch_width != 8 */
+	 c->nr_payload_regs += 2;
+      }
+      /* R7-10: perspective centroid barycentric */
+      /* R11-14: perspective sample barycentric */
+      /* R15-18: linear pixel location barycentric */
+      /* R19-22: linear centroid barycentric */
+      /* R23-26: linear sample barycentric */
+
+      /* R27: interpolated depth if uses source depth */
+      if (uses_depth) {
+	 c->source_depth_reg = c->nr_payload_regs;
+	 c->nr_payload_regs++;
+	 if (!c->fp->isGLSL) { /* dispatch_width != 8 */
+	    /* R28: interpolated depth if not 8-wide. */
+	    c->nr_payload_regs++;
+	 }
+      }
+      /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W.
+       */
+      if (uses_depth) {
+	 c->source_w_reg = c->nr_payload_regs;
+	 c->nr_payload_regs++;
+	 if (!c->fp->isGLSL) { /* dispatch_width != 8 */
+	    /* R30: interpolated W if not 8-wide. */
+	    c->nr_payload_regs++;
+	 }
+      }
+      /* R31: MSAA position offsets. */
+      /* R32-: bary for 32-pixel. */
+      /* R58-59: interp W for 32-pixel. */
+
+      if (c->fp->program.Base.OutputsWritten &
+	  BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
+	 c->source_depth_to_render_target = GL_TRUE;
+	 c->computes_depth = GL_TRUE;
+      }
+   } else {
+      brw_wm_lookup_iz(intel, c);
+   }
+}
 
 /**
  * All Mesa program -> GPU code generation goes through this function.
@@ -167,6 +223,8 @@ static void do_wm_prog( struct brw_context *brw,
 
    brw_init_compile(brw, &c->func);
 
+   brw_wm_payload_setup(brw, c);
+
    /* temporary sanity check assertion */
    ASSERT(fp->isGLSL == brw_wm_is_glsl(&c->fp->program));
 
@@ -220,12 +278,10 @@ static void do_wm_prog( struct brw_context *brw,
 static void brw_wm_populate_key( struct brw_context *brw,
 				 struct brw_wm_prog_key *key )
 {
-   struct intel_context *intel = &brw->intel;
    struct gl_context *ctx = &brw->intel.ctx;
    /* BRW_NEW_FRAGMENT_PROGRAM */
    const struct brw_fragment_program *fp = 
       (struct brw_fragment_program *)brw->fragment_program;
-   GLboolean uses_depth = (fp->program.Base.InputsRead & (1 << FRAG_ATTRIB_WPOS)) != 0;
    GLuint lookup = 0;
    GLuint line_aa;
    GLuint i;
@@ -285,57 +341,9 @@ static void brw_wm_populate_key( struct brw_context *brw,
       }
    }
 
-   if (intel->gen >= 6) {
-      /* R0-1: masks, pixel X/Y coordinates. */
-      key->nr_payload_regs = 2;
-      /* R2: only for 32-pixel dispatch.*/
-      /* R3-4: perspective pixel location barycentric */
-      key->nr_payload_regs += 2;
-      /* R5-6: perspective pixel location bary for dispatch width != 8 */
-      if (!fp->isGLSL) { /* dispatch_width != 8 */
-	 key->nr_payload_regs += 2;
-      }
-      /* R7-10: perspective centroid barycentric */
-      /* R11-14: perspective sample barycentric */
-      /* R15-18: linear pixel location barycentric */
-      /* R19-22: linear centroid barycentric */
-      /* R23-26: linear sample barycentric */
-
-      /* R27: interpolated depth if uses source depth */
-      if (uses_depth) {
-	 key->source_depth_reg = key->nr_payload_regs;
-	 key->nr_payload_regs++;
-	 if (!fp->isGLSL) { /* dispatch_width != 8 */
-	    /* R28: interpolated depth if not 8-wide. */
-	    key->nr_payload_regs++;
-	 }
-      }
-      /* R29: interpolated W set if GEN6_WM_USES_SOURCE_W.
-       */
-      if (uses_depth) {
-	 key->source_w_reg = key->nr_payload_regs;
-	 key->nr_payload_regs++;
-	 if (!fp->isGLSL) { /* dispatch_width != 8 */
-	    /* R30: interpolated W if not 8-wide. */
-	    key->nr_payload_regs++;
-	 }
-      }
-      /* R31: MSAA position offsets. */
-      /* R32-: bary for 32-pixel. */
-      /* R58-59: interp W for 32-pixel. */
-
-      if (fp->program.Base.OutputsWritten & BITFIELD64_BIT(FRAG_RESULT_DEPTH)) {
-	 key->source_depth_to_render_target = GL_TRUE;
-	 key->computes_depth = GL_TRUE;
-      }
-
-   } else {
-      brw_wm_lookup_iz(intel,
-	      	       line_aa,
-		       lookup,
-		       uses_depth,
-		       key);
-   }
+   key->iz_lookup = lookup;
+   key->line_aa = line_aa;
+   key->stats_wm = brw->intel.stats_wm;
 
    /* BRW_NEW_WM_INPUT_DIMENSIONS */
    key->proj_attrib_mask = brw->wm.input_size_masks[4-1];

src/mesa/drivers/dri/i965/brw_wm.h

@@ -59,16 +59,9 @@
 #define AA_ALWAYS    2
 
 struct brw_wm_prog_key {
-   GLuint source_depth_reg:3;
-   GLuint source_w_reg:3;
-   GLuint aa_dest_stencil_reg:3;
-   GLuint dest_depth_reg:3;
-   GLuint nr_payload_regs:4;
-   GLuint computes_depth:1;	/* could be derived from program string */
-   GLuint source_depth_to_render_target:1;
+   GLuint stats_wm:1;
    GLuint flat_shade:1;
    GLuint linear_color:1;  /**< linear interpolation vs perspective interp */
-   GLuint runtime_check_aads_emit:1;
    GLuint nr_color_regions:5;
    GLuint render_to_fbo:1;
 
@@ -81,6 +74,8 @@ struct brw_wm_prog_key {
 
    GLushort drawable_height;
    GLbitfield64 vp_outputs_written;
+   GLuint iz_lookup;
+   GLuint line_aa;
    GLuint program_string_id:32;
 };
 
@@ -204,6 +199,15 @@ struct brw_wm_compile {
       PASS2_DONE
    } state;
 
+   GLuint source_depth_reg:3;
+   GLuint source_w_reg:3;
+   GLuint aa_dest_stencil_reg:3;
+   GLuint dest_depth_reg:3;
+   GLuint nr_payload_regs:4;
+   GLuint computes_depth:1;	/* could be derived from program string */
+   GLuint source_depth_to_render_target:1;
+   GLuint runtime_check_aads_emit:1;
+
    /* Initial pass - translate fp instructions to fp instructions,
     * simplifying and adding instructions for interpolation and
     * framebuffer writes.
@@ -306,11 +310,8 @@ void brw_wm_print_insn( struct brw_wm_compile *c,
 void brw_wm_print_program( struct brw_wm_compile *c,
 			   const char *stage );
 
-void brw_wm_lookup_iz( struct intel_context *intel,
-		       GLuint line_aa,
-		       GLuint lookup,
-		       GLboolean ps_uses_depth,
-		       struct brw_wm_prog_key *key );
+void brw_wm_lookup_iz(struct intel_context *intel,
+		      struct brw_wm_compile *c);
 
 GLboolean brw_wm_is_glsl(const struct gl_fragment_program *fp);
 void brw_wm_glsl_emit(struct brw_context *brw, struct brw_wm_compile *c);

src/mesa/drivers/dri/i965/brw_wm_emit.c

@@ -1387,8 +1387,8 @@ static void emit_aa( struct brw_wm_compile *c,
 		     GLuint reg )
 {
    struct brw_compile *p = &c->func;
-   GLuint comp = c->key.aa_dest_stencil_reg / 2;
-   GLuint off = c->key.aa_dest_stencil_reg % 2;
+   GLuint comp = c->aa_dest_stencil_reg / 2;
+   GLuint off = c->aa_dest_stencil_reg % 2;
    struct brw_reg aa = offset(arg1[comp], off);
 
    brw_push_insn_state(p);
@@ -1420,7 +1420,7 @@ void emit_fb_write(struct brw_wm_compile *c,
 
    /* Reserve a space for AA - may not be needed:
     */
-   if (c->key.aa_dest_stencil_reg)
+   if (c->aa_dest_stencil_reg)
       nr += 1;
 
    /* I don't really understand how this achieves the color interleave
@@ -1493,9 +1493,9 @@ void emit_fb_write(struct brw_wm_compile *c,
 
    brw_pop_insn_state(p);
 
-   if (c->key.source_depth_to_render_target)
+   if (c->source_depth_to_render_target)
    {
-      if (c->key.computes_depth) 
+      if (c->computes_depth)
 	 brw_MOV(p, brw_message_reg(nr), arg2[2]);
       else 
 	 brw_MOV(p, brw_message_reg(nr), arg1[1]); /* ? */
@@ -1503,10 +1503,10 @@ void emit_fb_write(struct brw_wm_compile *c,
       nr += 2;
    }
 
-   if (c->key.dest_depth_reg)
+   if (c->dest_depth_reg)
    {
-      GLuint comp = c->key.dest_depth_reg / 2;
-      GLuint off = c->key.dest_depth_reg % 2;
+      GLuint comp = c->dest_depth_reg / 2;
+      GLuint off = c->dest_depth_reg % 2;
 
       if (off != 0) {
          brw_push_insn_state(p);
@@ -1528,8 +1528,8 @@ void emit_fb_write(struct brw_wm_compile *c,
       nr -= 2;
    }
 
-   if (!c->key.runtime_check_aads_emit) {
-      if (c->key.aa_dest_stencil_reg)
+   if (!c->runtime_check_aads_emit) {
+      if (c->aa_dest_stencil_reg)
 	 emit_aa(c, arg1, 2);
 
       fire_fb_write(c, base_reg, nr, target, eot);

src/mesa/drivers/dri/i965/brw_wm_glsl.c

@@ -290,15 +290,15 @@ static void prealloc_reg(struct brw_wm_compile *c)
     c->first_free_grf = 0;
 
     for (i = 0; i < 4; i++) {
-	if (i < (c->key.nr_payload_regs + 1) / 2)
+	if (i < (c->nr_payload_regs + 1) / 2)
             reg = brw_vec8_grf(i * 2, 0);
         else
             reg = brw_vec8_grf(0, 0);
 	set_reg(c, PROGRAM_PAYLOAD, PAYLOAD_DEPTH, i, reg);
     }
     set_reg(c, PROGRAM_PAYLOAD, PAYLOAD_W, 0,
-	    brw_vec8_grf(c->key.source_w_reg, 0));
-    reg_index += c->key.nr_payload_regs;
+	    brw_vec8_grf(c->source_w_reg, 0));
+    reg_index += c->nr_payload_regs;
 
     /* constants */
     {
@@ -384,7 +384,7 @@ static void prealloc_reg(struct brw_wm_compile *c)
        }
     }
 
-    c->prog_data.first_curbe_grf = c->key.nr_payload_regs;
+    c->prog_data.first_curbe_grf = c->nr_payload_regs;
     c->prog_data.urb_read_length = urb_read_length;
     c->prog_data.curb_read_length = c->nr_creg;
     c->emit_mask_reg = brw_uw1_reg(BRW_GENERAL_REGISTER_FILE, reg_index, 0);

src/mesa/drivers/dri/i965/brw_wm_iz.c

@@ -120,14 +120,14 @@ const struct {
  * \param line_aa  AA_NEVER, AA_ALWAYS or AA_SOMETIMES
  * \param lookup  bitmask of IZ_* flags
  */
-void brw_wm_lookup_iz( struct intel_context *intel,
-		       GLuint line_aa,
-		       GLuint lookup,
-		       GLboolean ps_uses_depth,
-		       struct brw_wm_prog_key *key )
+void brw_wm_lookup_iz(struct intel_context *intel,
+		      struct brw_wm_compile *c)
 {
    GLuint reg = 2;
    GLboolean kill_stats_promoted_workaround = GL_FALSE;
+   int lookup = c->key.iz_lookup;
+   bool uses_depth = (c->fp->program.Base.InputsRead &
+		      (1 << FRAG_ATTRIB_WPOS)) != 0;
 
    assert (lookup < IZ_BIT_MAX);
 
@@ -136,36 +136,36 @@ void brw_wm_lookup_iz( struct intel_context *intel,
     * statistics are enabled..." paragraph of 11.5.3.2: Early Depth
     * Test Cases [Pre-DevGT] of the 3D Pipeline - Windower B-Spec.
     */
-   if (intel->stats_wm &&
+   if (c->key.stats_wm &&
        (lookup & IZ_PS_KILL_ALPHATEST_BIT) &&
        wm_iz_table[lookup].mode == P) {
       kill_stats_promoted_workaround = GL_TRUE;
    }
 
    if (lookup & IZ_PS_COMPUTES_DEPTH_BIT)
-      key->computes_depth = 1;
+      c->computes_depth = 1;
 
-   if (wm_iz_table[lookup].sd_present || ps_uses_depth ||
+   if (wm_iz_table[lookup].sd_present || uses_depth ||
        kill_stats_promoted_workaround) {
-      key->source_depth_reg = reg;
+      c->source_depth_reg = reg;
       reg += 2;
    }
 
    if (wm_iz_table[lookup].sd_to_rt || kill_stats_promoted_workaround)
-      key->source_depth_to_render_target = 1;
+      c->source_depth_to_render_target = 1;
 
-   if (wm_iz_table[lookup].ds_present || line_aa != AA_NEVER) {
-      key->aa_dest_stencil_reg = reg;
-      key->runtime_check_aads_emit = (!wm_iz_table[lookup].ds_present &&
-				      line_aa == AA_SOMETIMES);
+   if (wm_iz_table[lookup].ds_present || c->key.line_aa != AA_NEVER) {
+      c->aa_dest_stencil_reg = reg;
+      c->runtime_check_aads_emit = (!wm_iz_table[lookup].ds_present &&
+				    c->key.line_aa == AA_SOMETIMES);
       reg++;
    }
 
    if (wm_iz_table[lookup].dd_present) {
-      key->dest_depth_reg = reg;
+      c->dest_depth_reg = reg;
       reg+=2;
    }
 
-   key->nr_payload_regs = reg;
+   c->nr_payload_regs = reg;
 }
 

src/mesa/drivers/dri/i965/brw_wm_pass0.c

@@ -380,7 +380,7 @@ static void pass0_init_payload( struct brw_wm_compile *c )
    GLuint i;
 
    for (i = 0; i < 4; i++) {
-      GLuint j = i >= (c->key.nr_payload_regs + 1) / 2 ? 0 : i;
+      GLuint j = i >= (c->nr_payload_regs + 1) / 2 ? 0 : i;
       pass0_set_fpreg_value( c, PROGRAM_PAYLOAD, PAYLOAD_DEPTH, i, 
 			     &c->payload.depth[j] );
    }

src/mesa/drivers/dri/i965/brw_wm_pass1.c

@@ -128,8 +128,7 @@ void brw_wm_pass1( struct brw_wm_compile *c )
       if (inst->opcode == WM_FB_WRITE) {
 	 track_arg(c, inst, 0, WRITEMASK_XYZW); 
 	 track_arg(c, inst, 1, WRITEMASK_XYZW); 
-	 if (c->key.source_depth_to_render_target &&
-	     c->key.computes_depth)
+	 if (c->source_depth_to_render_target && c->computes_depth)
 	    track_arg(c, inst, 2, WRITEMASK_Z); 
 	 else
 	    track_arg(c, inst, 2, 0); 

src/mesa/drivers/dri/i965/brw_wm_pass2.c

@@ -76,7 +76,7 @@ static void init_registers( struct brw_wm_compile *c )
    for (j = 0; j < c->grf_limit; j++) 
       c->pass2_grf[j].nextuse = BRW_WM_MAX_INSN;
 
-   for (j = 0; j < (c->key.nr_payload_regs + 1) / 2; j++)
+   for (j = 0; j < (c->nr_payload_regs + 1) / 2; j++)
       prealloc_reg(c, &c->payload.depth[j], i++);
 
    for (j = 0; j < c->nr_creg; j++) 
@@ -101,7 +101,7 @@ static void init_registers( struct brw_wm_compile *c )
 
    assert(nr_interp_regs >= 1);
 
-   c->prog_data.first_curbe_grf = ALIGN(c->key.nr_payload_regs, 2);
+   c->prog_data.first_curbe_grf = ALIGN(c->nr_payload_regs, 2);
    c->prog_data.urb_read_length = nr_interp_regs * 2;
    c->prog_data.curb_read_length = c->nr_creg * 2;