Call list and state improvements

src/accum.c

@@ -72,6 +72,7 @@ static void draw_screen(GXTexObj *texture, float value)
     GX_SetNumChans(1);
     GX_SetChanCtrl(GX_COLOR0A0, GX_DISABLE, GX_SRC_VTX, GX_SRC_VTX,
                    0, GX_DF_NONE, GX_AF_NONE);
+    glparamstate.dirty.bits.dirty_tev = 1;
 
     GX_SetCullMode(GX_CULL_NONE);
     glparamstate.dirty.bits.dirty_cull = 1;

src/arrays.cpp

@@ -116,6 +116,9 @@ static TemplateSelectionInfo select_template(GLenum type,
         info.format.type = num_components == 1 ? GX_TEX_S : GX_TEX_ST;
         info.format.size = gl_type_to_gx_size(type);
         info.same_type = num_components <= 2;
+        /* The hardware does not support sending more than 2 texture
+         * coordinates */
+        if (num_components > 2) info.format.num_components = 2;
         break;
     case GX_VA_CLR0:
     case GX_VA_CLR1:

src/call_lists.c

@@ -53,12 +53,6 @@ typedef struct
 {
     CommandType type;
     union {
-        struct GXDisplayList {
-            void *list;
-            u32 size;
-            struct client_state cs;
-        } gxlist;
-
         GLuint gllist; // glCallList
 
         GLenum cap; // glEnable, glDisable
@@ -109,7 +103,7 @@ typedef struct
         struct DrawGeometry {
             GLenum mode;
             uint16_t count;
-            struct client_state cs;
+            union client_state cs;
             u32 list_size;
             void *gxlist;
         } draw_geometry;
@@ -138,6 +132,8 @@ static GXColor s_current_color;
 static float s_current_normal[3];
 static bool s_last_draw_used_indexed_data = false;
 static uint16_t s_last_draw_sync_token = 0;
+static union client_state s_last_client_state;
+static bool s_last_client_state_is_valid = false;
 
 #define BUFFER_IS_VALID(buffer) (((uint32_t)buffer) > 1)
 #define LIST_IS_USED(index) BUFFER_IS_VALID(call_lists[index].head)
@@ -200,13 +196,12 @@ static Command *new_command(CommandBuffer **head)
     }
 }
 
-static void execute_draw_geometry_list(struct DrawGeometry *dg)
+static void setup_draw_geometry(struct DrawGeometry *dg,
+                                bool uses_indexed_data)
 {
-    static uint16_t counter = 0;
     u8 vtxindex = GX_VTXFMT0;
     GXColor current_color;
 
-    bool uses_indexed_data = !dg->cs.normal_enabled || !dg->cs.color_enabled;
     if (uses_indexed_data && s_last_draw_used_indexed_data) {
         bool data_changed = false;
         /* If the indexed data has changed, we need to wait until the previous
@@ -254,6 +249,17 @@ static void execute_draw_geometry_list(struct DrawGeometry *dg)
      * so we won't add them unless they are enabled. */
 
     GX_InvVtxCache();
+}
+
+static void execute_draw_geometry_list(struct DrawGeometry *dg)
+{
+    bool uses_indexed_data = !dg->cs.normal_enabled || !dg->cs.color_enabled;
+    if (!s_last_client_state_is_valid ||
+        s_last_client_state.as_int != dg->cs.as_int) {
+        setup_draw_geometry(dg, uses_indexed_data);
+        s_last_client_state = dg->cs;
+        s_last_client_state_is_valid = true;
+    }
 
     GX_CallDispList(dg->gxlist, dg->list_size);
 
@@ -273,7 +279,7 @@ static void flat_draw_geometry(void *cb_data)
 
 static void run_draw_geometry(struct DrawGeometry *dg)
 {
-    struct client_state cs;
+    union client_state cs;
 
     /* Update the drawing mode on the list. This required peeping into
      * GX_Begin() code. */
@@ -304,9 +310,11 @@ static void run_draw_geometry(struct DrawGeometry *dg)
     glparamstate.draw_count++;
 
     if (glparamstate.stencil.enabled) {
+        s_last_client_state_is_valid = false;
         _ogx_gpu_resources_push();
         _ogx_stencil_draw(flat_draw_geometry, dg);
         _ogx_gpu_resources_pop();
+        s_last_client_state_is_valid = false;
     }
 }
 
@@ -771,6 +779,10 @@ void glCallList(GLuint id)
     }
 
 done:
+    /* Until we find a reliable mechanism to ensure that the client state has
+     * been preserved, avoid reusing it across different lists. */
+    s_last_client_state_is_valid = false;
+
     if (must_decrement) {
         glparamstate.current_call_list.execution_depth--;
     }

src/clip.c

@@ -165,13 +165,13 @@ void _ogx_clip_setup_tev()
 void _ogx_clip_enabled(int plane)
 {
     glparamstate.clip_plane_mask |= 1 << plane;
-    glparamstate.dirty.bits.dirty_clip_planes = 1;
+    glparamstate.dirty.bits.dirty_tev = 1;
 }
 
 void _ogx_clip_disabled(int plane)
 {
     glparamstate.clip_plane_mask &= ~(1 << plane);
-    glparamstate.dirty.bits.dirty_clip_planes = 1;
+    glparamstate.dirty.bits.dirty_tev = 1;
 }
 
 void glClipPlane(GLenum plane, const GLdouble *equation)
@@ -192,4 +192,6 @@ void glClipPlane(GLenum plane, const GLdouble *equation)
     guMtx44Inverse(mv, mv_inverse);
     ClipPlane p0 = { equation[0], equation[1], equation[2], equation[3] };
     mtx44_multiply(p0, mv_inverse, *p);
+
+    glparamstate.dirty.bits.dirty_tev = 1;
 }

src/efb.c

@@ -88,6 +88,7 @@ void _ogx_efb_restore_texobj(GXTexObj *texobj)
     GX_SetNumChans(0);
     GX_SetTevOp(GX_TEVSTAGE0, GX_REPLACE);
     GX_SetTevOrder(GX_TEVSTAGE0, GX_TEXCOORD0, GX_TEXMAP0, GX_COLORNULL);
+    glparamstate.dirty.bits.dirty_tev = 1;
 
     GX_SetCullMode(GX_CULL_NONE);
     glparamstate.dirty.bits.dirty_cull = 1;