#407 32X add SH-2 ASM faceAdd* routines, on-chip blocks for prepare a…

…nd render functions (still slower than SDRAM), multi-CPU OT flush, fb clear on the slave CPU

src/fixed/common.h

@@ -149,6 +149,12 @@
     #define USE_FMT     (LVL_FMT_PKD)
 
     #include "32x.h"
+
+    enum MarsCmd {
+        MARS_CMD_NONE = 0,
+        MARS_CMD_CLEAR,
+        MARS_CMD_FLUSH
+    };
 #else
     #error unsupported platform
 #endif
@@ -2913,7 +2919,7 @@ void drawLevelInit();
 void drawLevelFree();
 void drawText(int32 x, int32 y, const char* text, TextAlign align);
 void drawModel(const ItemObj* item);
-void drawItem(const ItemObj* item);
+void drawSprite(const ItemObj* item);
 void drawRooms(Camera* camera);
 void drawCinematicRooms();
 void drawHUD(Lara* lara);

src/fixed/draw.h

@@ -701,15 +701,6 @@ void drawModel(const ItemObj* item)
     }
 }
 
-void drawItem(const ItemObj* item)
-{
-    if (level.models[item->type].count > 0) {
-        drawModel(item);
-    } else {
-        drawSprite(item);
-    }
-}
-
 void drawRoom(const Room* room)
 {
     setViewport(room->clip);
@@ -811,7 +802,7 @@ void drawRooms(Camera* camera)
 
     Room** visRoom = camera->view.room->getVisibleRooms();
 
-    // draw Lara first
+#ifdef DRAW_LARA_FIRST
     for (int32 i = 0; i < MAX_PLAYERS; i++)
     {
         Lara* lara = players[i];
@@ -823,7 +814,7 @@ void drawRooms(Camera* camera)
             lara->flags |= ITEM_FLAG_STATUS_INVISIBLE; // skip drawing in the general pass
         }
     }
-
+#endif
     // draw rooms and objects
     while (*visRoom)
     {
@@ -832,6 +823,7 @@ void drawRooms(Camera* camera)
         room->reset();
     }
 
+#ifdef DRAW_LARA_FIRST
     // reset visibility flags for Lara
     for (int32 i = 0; i < MAX_PLAYERS; i++)
     {
@@ -841,6 +833,7 @@ void drawRooms(Camera* camera)
             lara->flags &= ~ITEM_FLAG_STATUS;
         }
     }
+#endif
 
     setPaletteIndex(0);
     setViewport(vp);

src/fixed/enemy.h

@@ -787,7 +787,7 @@ struct Wolf : Enemy
             case STATE_STOP:
             {
                 if (nextState)
-                    nextState;
+                    return nextState;
                 return STATE_WALK;
             }
 

src/fixed/item.h

@@ -1013,7 +1013,11 @@ void ItemObj::update()
 
 void ItemObj::draw()
 {
-    drawItem(this);
+    if (level.models[type].count > 0) {
+        drawModel(this);
+    } else {
+        drawSprite(this);
+    }
 }
 
 struct ItemSave {

src/platform/32x/32x.h

@@ -155,4 +155,6 @@ extern "C"
     CacheControl(0);\
     CacheControl(SH2_CCTL_CP | SH2_CCTL_CE);
 
+#define MARS_WAIT() {while (MARS_SYS_COMM4);}
+
 #endif

src/platform/32x/asm/block_prepare.s

@@ -0,0 +1,61 @@
+#include "common.i"
+.data
+
+.global _block_prepare_start
+.global _block_prepare_end
+
+.align 4
+_block_prepare_start:
+
+#include "transformMesh.i"
+#include "transformRoom.i"
+
+.align 2
+var_gVerticesBase:
+        .long   _gVerticesBase
+var_gMatrixPtr:
+        .long   _gMatrixPtr
+var_gLightAmbient:
+        .long   _gLightAmbient
+var_divTable:
+        .long   _divTable
+var_viewportRel:
+        .long   _viewportRel
+
+#include "faceAddMeshQuads.i"
+#include "faceAddMeshTriangles.i"
+
+.align 2
+var_gVertices_fam:
+        .long   _gVertices
+var_gFacesBase_fam:
+        .long   _gFacesBase
+var_gVerticesBase_fam:
+        .long   _gVerticesBase
+const_FACE_CLIPPED_fam:
+        .long   FACE_CLIPPED
+const_FACE_TRIANGLE_fam:
+        .long   FACE_TRIANGLE
+var_gOT_fam:
+        .long   _gOT
+
+#include "faceAddRoomQuads.i"
+#include "faceAddRoomTriangles.i"
+
+.align 2
+var_gVertices_far:
+        .long   _gVertices
+var_gFacesBase_far:
+        .long   _gFacesBase
+var_gVerticesBase_far:
+        .long   _gVerticesBase
+const_FACE_CLIPPED_far:
+        .long   FACE_CLIPPED
+const_FACE_GOURAUD_far:
+        .long   FACE_GOURAUD
+const_FACE_TRIANGLE_far:
+        .long   FACE_TRIANGLE
+var_gOT_far:
+        .long   _gOT
+
+_block_prepare_end:

src/platform/32x/asm/block_render.s

@@ -0,0 +1,36 @@
+#include "common.i"
+.data
+
+.global _block_render_start
+.global _block_render_end
+
+.align 4
+_block_render_start:
+
+#include "rasterize.i"
+//#include "rasterize_dummy.i"
+#include "rasterizeS.i"
+#include "rasterizeF.i"
+
+.align 2
+var_LMAP_ADDR_fs:
+        .long   _gLightmap_base
+var_divTable_fs:
+        .long   _divTable
+var_frameWidth_fs:
+        .word   FRAME_WIDTH
+
+#include "rasterizeFT.i"
+#include "rasterizeGT.i"
+
+.align 2
+var_LMAP_ADDR:
+        .long   _gLightmap_base
+var_divTable:
+        .long   _divTable
+var_mask:
+        .word   0xFF00
+var_frameWidth:
+        .word   FRAME_WIDTH
+
+_block_render_end:

src/platform/32x/asm/common.i

@@ -1,9 +1,11 @@
-#define SEG_MATH    .text
-#define SEG_TRANS   .data
-#define SEG_FACE    .data
-#define SEG_RASTER  .data
+#ifndef H_COMMON_ASM
+#define H_COMMON_ASM
+
+#define SEG_MATH    .data
 #define SEG_PHYSICS .data
 
+//#define ON_CHIP_RENDER
+
 // Matrix:
 // int16 e00, e01, e02  // rotation
 // int16 e10, e11, e12  // rotation
@@ -32,6 +34,10 @@
 #define FIXED_SHIFT     14
 
 #define FACE_TYPE_F     1
+#define FACE_TYPE_SHIFT 14
+#define FACE_CLIPPED    (1 << 30)
+#define FACE_TRIANGLE   (1 << 31)
+#define FACE_GOURAUD    (2 << FACE_TYPE_SHIFT)
 
 #define VERTEX_X                0
 #define VERTEX_Y                2
@@ -46,6 +52,8 @@
 #define VERTEX_SIZEOF_SHIFT     4
 #define VERTEX_SIZEOF           (1 << VERTEX_SIZEOF_SHIFT)
 
+#define FACE_SIZEOF             16
+
 #define VIEW_DIST       (1024 * 10)   // max = DIV_TABLE_END << PROJ_SHIFT
 #define FOG_SHIFT       1
 #define FOG_MAX         VIEW_DIST
@@ -61,6 +69,7 @@
 #define CLIP_BOTTOM     (1 << 4)
 #define CLIP_FAR        (1 << 5)
 #define CLIP_NEAR       (1 << 6)
+#define CLIP_DISCARD    (CLIP_LEFT + CLIP_RIGHT + CLIP_TOP + CLIP_BOTTOM + CLIP_FAR + CLIP_NEAR)
 
 #define VP_MINX         0
 #define VP_MINY         4
@@ -121,3 +130,26 @@
 .macro lit lightmap, index
         mov.b   @(\index, \lightmap), \index
 .endm
+
+// (vy1 - vy0) * (vx0 - vx2) <= (vx1 - vx0) * (vy0 - vy2)
+.macro ccw vp0, vp1, vp2, vx0, vy0, vx1, vy1, vx2, vy2
+        mov.w   @\vp0+, \vx0
+        mov.w   @\vp0+, \vy0
+        mov.w   @\vp1+, \vx1
+        mov.w   @\vp1+, \vy1
+        sub     \vx0, \vx1      // vx1 -= vx0
+        sub     \vy0, \vy1      // vy1 -= vy0
+        mov.w   @\vp2+, \vx2
+        sub     \vx2, \vx0      // vx0 -= vx2
+        mov.w   @\vp2+, \vy2
+        sub     \vy2, \vy0      // vy0 -= vy2
+
+        muls.w  \vy1, \vx0
+        sts     MACL, \vx0      // vx0 *= vy1
+        muls.w  \vx1, \vy0
+        sts     MACL, \vy0      // vy0 *= vx1
+
+        cmp/ge  \vx0, \vy0      // T = (vy0 >= vx0)
+.endm
+
+#endif // H_COMMON_ASM