Introduce batching engine and shadow batching

bin/segment2.c

@@ -2761,14 +2761,24 @@ const Gfx dl_draw_quad_verts_4567[] = {
     gsSPEndDisplayList(),
 };
 
+#define G_CC_SHADOW TEXEL0, 0, SHADE, 0, TEXEL0, 0, PRIMITIVE, 0
+
 const Gfx dl_shadow_begin[] = {
     gsDPPipeSync(),
     gsSPClearGeometryMode(G_LIGHTING | G_CULL_BACK),
-    gsDPSetCombineMode(G_CC_MODULATEIFADEA, G_CC_MODULATEIFADEA),
+    gsDPSetCombineMode(G_CC_SHADOW, G_CC_SHADOW),
     gsSPTexture(0xFFFF, 0xFFFF, 0, G_TX_RENDERTILE, G_ON),
     gsSPEndDisplayList(),
 };
 
+const Gfx dl_shadow_end[] = {
+    gsDPPipeSync(),
+    gsSPTexture(0xFFFF, 0xFFFF, 0, G_TX_RENDERTILE, G_OFF),
+    gsSPSetGeometryMode(G_LIGHTING | G_CULL_BACK),
+    gsDPSetCombineMode(G_CC_SHADE, G_CC_SHADE),
+    gsSPEndDisplayList(),
+};
+
 #ifdef HD_SHADOWS
 const Gfx dl_shadow_circle[] = {
     gsSPDisplayList(dl_shadow_begin),
@@ -2810,13 +2820,9 @@ static const Vtx vertex_shadow[] = {
 };
 
 // 0x02014638 - 0x02014660
-const Gfx dl_shadow_end[] = {
+const Gfx dl_shadow_tris[] = {
     gsSPVertex(vertex_shadow, 4, 0),
     gsSP2Triangles( 0,  2,  1, 0x0,  1,  2,  3, 0x0),
-    gsDPPipeSync(),
-    gsSPTexture(0xFFFF, 0xFFFF, 0, G_TX_RENDERTILE, G_OFF),
-    gsSPSetGeometryMode(G_LIGHTING | G_CULL_BACK),
-    gsDPSetCombineMode(G_CC_SHADE, G_CC_SHADE),
     gsSPEndDisplayList(),
 };
 

include/sm64.h

@@ -70,12 +70,25 @@ enum RenderLayers {
     LAYER_OCCLUDE_SILHOUETTE_OPAQUE,
     LAYER_OCCLUDE_SILHOUETTE_ALPHA,
 #endif
+    LAYER_CLD,
     LAYER_TRANSPARENT_DECAL,
     LAYER_TRANSPARENT,
     LAYER_TRANSPARENT_INTER,
     LAYER_COUNT
 };
 
+enum BatchTransparentDecal {
+    BATCH_TRANSPARENT_DECAL_SHADOW_CIRCLE,
+    BATCH_TRANSPARENT_DECAL_SHADOW_SQUARE,
+    BATCH_TRANSPARENT_DECAL_COUNT,
+};
+
+enum BatchCld {
+    BATCH_CLD_SHADOW_CIRCLE,
+    BATCH_CLD_SHADOW_SQUARE,
+    BATCH_CLD_COUNT,
+};
+
 #define LAYER_FIRST                         LAYER_FORCE
 #define LAYER_LAST                          (LAYER_COUNT - 1)
 

src/engine/batch_list.c

@@ -0,0 +1,30 @@
+#include "batch_list.h"
+
+#include "game/segment2.h"
+
+static const struct BatchDisplayLists BatchesTransparentDecal[] = {
+    [ BATCH_TRANSPARENT_DECAL_SHADOW_CIRCLE ] = { dl_shadow_circle, dl_shadow_end },
+    [ BATCH_TRANSPARENT_DECAL_SHADOW_SQUARE ] = { dl_shadow_square, dl_shadow_end },
+};
+STATIC_ASSERT(BATCH_TRANSPARENT_DECAL_COUNT == sizeof(BatchesTransparentDecal) / sizeof(*BatchesTransparentDecal), "Mismatch in transparent decal batch count");
+
+static const struct BatchDisplayLists BatchesCLD[] = {
+    [ BATCH_CLD_SHADOW_CIRCLE ] = { dl_shadow_circle, dl_shadow_end },
+    [ BATCH_CLD_SHADOW_SQUARE ] = { dl_shadow_square, dl_shadow_end },
+};
+STATIC_ASSERT(BATCH_CLD_COUNT == sizeof(BatchesCLD) / sizeof(*BatchesCLD), "Mismatch in CLD batch count");
+
+static inline struct BatchArray* batch_array_alloc(struct AllocOnlyPool *pool, int count, const struct BatchDisplayLists* dls) {
+    struct BatchArray* batches = alloc_only_pool_alloc(pool, sizeof(struct BatchArray) + count * sizeof(struct DisplayListLinks));
+    batches->count = count;
+    batches->batchDLs = dls;
+    return batches;
+}
+
+struct BatchArray* batch_list_objects_alloc_xlu_decal(struct AllocOnlyPool *pool) {
+    return batch_array_alloc(pool, BATCH_TRANSPARENT_DECAL_COUNT, BatchesTransparentDecal);
+}
+
+struct BatchArray* batch_list_objects_alloc_cld(struct AllocOnlyPool *pool) {
+    return batch_array_alloc(pool, BATCH_CLD_COUNT, BatchesCLD);
+}

src/engine/batch_list.h

@@ -0,0 +1,18 @@
+#pragma once
+
+#include "sm64.h"
+#include "engine/graph_node.h"
+
+struct BatchArray* batch_list_objects_alloc_xlu_decal(struct AllocOnlyPool *pool);
+struct BatchArray* batch_list_objects_alloc_cld(struct AllocOnlyPool *pool);
+
+static inline struct BatchArray* batch_list_objects_alloc(struct AllocOnlyPool *pool, enum RenderLayers layer) {
+    switch (layer) {
+        case LAYER_TRANSPARENT_DECAL:
+            return batch_list_objects_alloc_xlu_decal(pool);
+        case LAYER_CLD:
+            return batch_list_objects_alloc_cld(pool);
+        default:
+            return 0;
+    }
+}

src/engine/graph_node.c

@@ -8,6 +8,7 @@
 #include "game/rendering_graph_node.h"
 #include "game/area.h"
 #include "geo_layout.h"
+#include "batch_list.h"
 
 /**
  * Initialize a geo node with a given type. Sets all links such that there
@@ -122,6 +123,13 @@ struct GraphNodeMasterList *init_graph_node_master_list(struct AllocOnlyPool *po
 
         if (on) {
             graphNode->node.flags |= GRAPH_RENDER_Z_BUFFER;
+            for (int layer = 0; layer < LAYER_COUNT; layer++) {
+                graphNode->layers[layer].objects = batch_list_objects_alloc(pool, layer);
+            }
+        } else {
+            for (int layer = 0; layer < LAYER_COUNT; layer++) {
+                graphNode->layers[layer].objects = NULL;
+            }
         }
     }
 

src/engine/graph_node.h

@@ -140,15 +140,35 @@ struct DisplayListNode {
     struct DisplayListNode *next;
 };
 
+struct DisplayListLinks {
+    struct DisplayListNode* head;
+    struct DisplayListNode* tail;
+};
+
+struct BatchDisplayLists {
+    const void* startDl;
+    const void* endDl;
+};
+
+struct BatchArray {
+    int count;
+    const struct BatchDisplayLists* batchDLs;
+    struct DisplayListLinks batches[0];
+};
+
+struct MasterLayer {
+    struct DisplayListLinks list;
+    struct BatchArray* objects;
+};
+
 /** GraphNode that manages the 8 top-level display lists that will be drawn
  *  Each list has its own render mode, so for example water is drawn in a
  *  different master list than opaque objects.
  *  It also sets the z-buffer on before rendering and off after.
  */
 struct GraphNodeMasterList {
-    /*0x00*/ struct GraphNode node;
-    /*0x14*/ struct DisplayListNode *listHeads[LAYER_COUNT];
-    /*0x34*/ struct DisplayListNode *listTails[LAYER_COUNT];
+    struct GraphNode node;
+    struct MasterLayer layers[LAYER_COUNT];
 };
 
 /** Simply used as a parent to group multiple children.