#407 32X S/F/FT/FTA/GT/GTA SH-2 rasterization asm, transform optims, …

…change input mapping

src/fixed/common.cpp

@@ -42,11 +42,7 @@ EWRAM_DATA ExtraInfoLara playersExtra[MAX_PLAYERS];
     #define LEVEL_INFO(name, title, track, secrets) { #name, NULL, title, track, secrets }
 #endif
 
-#ifdef __3DO__ // TODO fix the title scren on 3DO
-EWRAM_DATA LevelID gLevelID = LVL_TR1_1;
-#else
 EWRAM_DATA LevelID gLevelID = LVL_TR1_TITLE;
-#endif
 
 const LevelInfo gLevelInfo[LVL_MAX] = {
 // TR1
@@ -97,7 +93,7 @@ int32 rand_draw()
 
 #ifdef USE_DIV_TABLE
 EWRAM_DATA ALIGN16 divTableInt divTable[DIV_TABLE_SIZE] = { // must be at EWRAM start
-    0xFFFF, 0xFFFF, 0x8000, 0x5555, 0x4000, 0x3333, 0x2AAA, 0x2492,
+    0x0000, 0x7FFF, 0x7FFF, 0x5555, 0x4000, 0x3333, 0x2AAA, 0x2492,
     0x2000, 0x1C71, 0x1999, 0x1745, 0x1555, 0x13B1, 0x1249, 0x1111,
     0x1000, 0x0F0F, 0x0E38, 0x0D79, 0x0CCC, 0x0C30, 0x0BA2, 0x0B21,
     0x0AAA, 0x0A3D, 0x09D8, 0x097B, 0x0924, 0x08D3, 0x0888, 0x0842,

src/fixed/common.h

@@ -9,7 +9,6 @@
 #endif
 
 #if defined(_WIN32)
-    #define MODE4
     #define USE_DIV_TABLE
 
     #define MODE4
@@ -218,6 +217,12 @@ typedef unsigned int       uint32;
 typedef uint16             divTableInt;
 #endif
 
+#if defined(__32X__)
+    typedef int8 ColorIndex;
+#else
+    typedef uint8 ColorIndex;
+#endif
+
 //#include <new>
 inline void* operator new(size_t, void *ptr)
 {
@@ -826,8 +831,6 @@ struct RoomVertex
 {
 #if defined(__3DO__)
     uint16 xyz565;
-#elif defined(__32X__)
-    uint8 g, z, y, x;
 #else
     uint8 x, y, z, g;
 #endif

src/fixed/game.h

@@ -201,6 +201,7 @@ void gameLoadLevel(const void* data)
         //resetLara(0, 9, _vec3i(49669, 7680, 57891), ANGLE_0); // first door
         //resetLara(0, 10, _vec3i(43063, 7168, 61198), ANGLE_0); // transp
         //resetLara(0, 14, _vec3i(20215, 6656, 52942), ANGLE_90 + ANGLE_45); // bridge
+        //resetLara(0, 25, _vec3i(8789, 5632, 80173), 0); // portal
         //resetLara(0, 17, _vec3i(16475, 6656, 59845), ANGLE_90); // bear
         //resetLara(0, 26, _vec3i(24475, 6912, 83505), ANGLE_90); // switch timer 1
         //resetLara(0, 35, _vec3i(35149, 2048, 74189), ANGLE_90); // switch timer 2

src/fixed/item.h

@@ -29,6 +29,9 @@ int32 alignOffset(int32 a, int32 b)
 
 void* soundPlay(int16 id, const vec3i* pos)
 {
+#ifdef __32X__ // TODO
+    return NULL;
+#endif
     if (!gSettings.audio_sfx)
         return NULL;
 

src/fixed/lara.h

@@ -2739,12 +2739,12 @@ struct Lara : ItemObj
         }
     #elif defined(__32X__)
         // 6 buttons
-        if (keys & IK_A) input |= IN_WEAPON;
-        if (keys & IK_B) input |= IN_ACTION;
-        if (keys & IK_C) input |= IN_JUMP;
-        if (keys & IK_X) input |= IN_LOOK;
+        if (keys & IK_A) input |= IN_ACTION;
+        if (keys & IK_B) input |= IN_JUMP;
+        if (keys & IK_C) input |= IN_WEAPON;
+        if (keys & IK_X) input |= IN_WALK;
         if (keys & IK_Y) input |= IN_UP | IN_DOWN;
-        if (keys & IK_Z) input |= IN_WALK;
+        if (keys & IK_Z) input |= IN_LOOK;
     #elif defined(__GBA__) || defined(_WIN32)
         int32 ikA, ikB;
 

src/platform/32x/Makefile

@@ -88,4 +88,4 @@ $(BUILD)/%.o: ../../fixed/%.cpp
 	$(SHXX) $(SHCCFLAGS) $(INCPATH) -o $@ $<
 
 clean:
-	$(RM) $(BUILD)/* *.32x *.elf
+	$(RM) $(BUILD)/* $(TARGET).32x $(TARGET).elf

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

@@ -39,14 +39,18 @@
 
 #define FACE_TYPE_F     1
 
-#define VERTEX_X        0
-#define VERTEX_Y        2
-#define VERTEX_Z        4
-#define VERTEX_G        6
-#define VERTEX_CLIP     7
-#define VERTEX_T        8
-#define VERTEX_PREV     12
-#define VERTEX_NEXT     13
+#define VERTEX_X                0
+#define VERTEX_Y                2
+#define VERTEX_Z                4
+#define VERTEX_G                6
+#define VERTEX_CLIP             7
+#define VERTEX_T                8
+#define VERTEX_PREV             12
+#define VERTEX_NEXT             13
+#define VERTEX_PADDING          14
+
+#define VERTEX_SIZEOF_SHIFT     4
+#define VERTEX_SIZEOF           (1 << VERTEX_SIZEOF_SHIFT)
 
 #define VIEW_DIST       (1024 * 10)   // max = DIV_TABLE_END << PROJ_SHIFT
 #define FOG_SHIFT       1
@@ -56,20 +60,70 @@
 #define VIEW_MAX        (VIEW_DIST)
 #define VIEW_OFF        4096
 
-#define CLIP_LEFT       (1 << 0)
-#define CLIP_RIGHT      (1 << 1)
-#define CLIP_TOP        (1 << 2)
-#define CLIP_BOTTOM     (1 << 3)
-#define CLIP_FAR        (1 << 4)
-#define CLIP_NEAR       (1 << 5)
+#define CLIP_FRAME      (1 << 0)
+#define CLIP_LEFT       (1 << 1)
+#define CLIP_RIGHT      (1 << 2)
+#define CLIP_TOP        (1 << 3)
+#define CLIP_BOTTOM     (1 << 4)
+#define CLIP_FAR        (1 << 5)
+#define CLIP_NEAR       (1 << 6)
 
 #define VP_MINX         0
 #define VP_MINY         4
 #define VP_MAXX         8
 #define VP_MAXY         12
 
-.macro shlr14 reg
-        shll2   \reg
-        shlr16  \reg
-        //exts.w  reg, reg      // skip this because of mov.w
-.endm
+#define FRAME_WIDTH     320
+#define FRAME_HEIGHT    224
+
+.macro align_fetch
+    .p2alignw 2, 0x0009
+.endm
+
+.macro shlr14 x
+        shll2   \x
+        shlr16  \x
+        //exts.w  x, x      // skip this because of mov.w
+.endm
+
+// int32 >> 12
+// 1. shar x 12 => 12 op
+// 2. (int32(int16(x >> 16)) << 4) | (x >> 12) => 8 op (require an extra register)
+.macro shar12 x, t
+        swap.w  \x, \t
+        exts.w  \t, \t
+        shll2   \t
+        shll2   \t
+        shlr8   \x
+        shlr2   \x
+        shlr2   \x
+        or      \t, \x
+.endm
+
+// out = uv * f
+// uv and out regs must be different
+// destructive for uv reg
+.macro scaleUV uv, out, f
+        muls.w  \uv, \f
+        shlr16  \uv
+        sts     MACL, \out      // v = int16(uv) * f (16-bit shift)
+        muls.w  \uv, \f
+        sts     MACL, \uv       // u = int16(uv >> 16) * f (16-bit shift)
+        shlr16  \uv
+        xtrct   \uv, \out       // out = uint16(v >> 16) | (u & 0xFFFF0000)
+.endm
+
+// UUuuVVvv -> 0000VVUU
+.macro getUV uv, index
+        swap.b  \uv, \index     // UUuuvvVV
+        swap.w  \index, \index  // vvVVUUuu
+        shll8   \index          // VVUUuu00
+        shlr16  \index          // 0000VVUU
+.endm
+
+// index (r0) = gLightmap[index]
+// in index 0..255
+// in lightmap one of 32 gLightmap slices
+.macro lit lightmap, index
+        mov.b   @(\index, \lightmap), \index
+.endm

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

@@ -1,4 +1,5 @@
 #include "common.i"
+SEG_RASTER
 
 #define type    r0
 #define proc    r1
@@ -9,45 +10,46 @@
 #define pixel   flags
 #define y       type
 
-.text
 .align 4
 .global _rasterize_asm
 _rasterize_asm:
         mov     flags, type
         shll2   type
-        swap.w  type, type
-        and     #15, type
+        shlr16  type
+        extu.b  type, proc
 
-        cmp/eq  #FACE_TYPE_F, type
-        bf/s    0f
+        cmp/eq  #FACE_TYPE_F, type      // cmp/eq #imm is 8-bit
+        bf/s    .getProc
         mov     L, R
         extu.b  flags, R
 
-0:      // proc = table[type]
-        mov     type, proc
+.getProc: // proc = table[type]
         mova    var_table, type
         shll2   proc
         mov.l   @(type, proc), proc
 
-        // pixel = fb + y * 320
+        // pixel = fb + y * 320 = fb + y * 256 + y * 64
         mov.w   @(VERTEX_Y, L), y
         mov.l   var_fb, pixel
         shll8   y
         add     y, pixel        // pixel += y * 256
-        shlr2   y
+        shar    y
+        shar    y
         jmp     @proc
         add     y, pixel        // pixel += y * 64
-        nop
 
+.align 2
 var_fb:
-        .long 0x24000200
+        // overwrite image frame buffer address has the same
+        // write per but allow transparent write for byte & word
+        .long 0x24020200
 var_table:
-        .long _rasterizeS_c
-        .long _rasterizeF_c
-        .long _rasterizeFT_c
-        .long _rasterizeFTA_c
-        .long _rasterizeGT_c
-        .long _rasterizeGTA_c
+        .long _rasterizeS_asm
+        .long _rasterizeF_asm
+        .long _rasterizeFT_asm
+        .long _rasterizeFT_asm
+        .long _rasterizeGT_asm
+        .long _rasterizeGT_asm
         .long _rasterizeSprite_c
         .long _rasterizeFillS_c
         .long _rasterizeLineH_c