#368 GBA pre-shift z by OT_SHIFT, pre-shift mesh vertex coords, fog m…

…ath optims
XProger · Dec 3, 2022 · 90e03fa · 90e03fa
1 parent b1a5590
commit 90e03fa
Show file tree

Hide file tree

Showing 17 changed files with 102 additions and 110 deletions.
diff --git a/src/fixed/common.h b/src/fixed/common.h
@@ -205,7 +205,7 @@
 // the maximum of active enemies
     #define MAX_ENEMIES 3
 // visibility distance
-    #define VIEW_DIST (1024 * 10)
+    #define VIEW_DIST (10 << 10)
 // skip collideSpheres for enemies
     #define FAST_HITMASK
 #endif
@@ -224,7 +224,7 @@
 // set the maximum number of simultaneously played channels
     #define SND_CHANNELS 4
 // visibility distance
-    #define VIEW_DIST (1024 * 10)
+    #define VIEW_DIST (10 << 10)
 // skip collideSpheres for enemies
     #define FAST_HITMASK
 #endif
@@ -239,7 +239,7 @@
 // the maximum of active enemies
     #define MAX_ENEMIES 3
 // visibility distance
-    #define VIEW_DIST (1024 * 10)
+    #define VIEW_DIST (10 << 10)
 // skip collideSpheres for enemies
     #define FAST_HITMASK
 #endif
@@ -457,10 +457,12 @@ extern int32 fps;
 #define FOV_SHIFT       3
 #define FOG_SHIFT       1
 #define FOG_MAX         VIEW_DIST
-#define FOG_MIN         (FOG_MAX - (8192 >> FOG_SHIFT))
+#define FOG_MIN         (FOG_MAX - 4096)
 #define VIEW_MIN_F      (64 << FIXED_SHIFT)
 #define VIEW_MAX_F      (VIEW_DIST << FIXED_SHIFT)
 
+#define MESH_SHIFT      2
+
 #define TEX_ATTR_AKILL  1
 
 #define NOT_ENEMY       -0x4000     // default hp for non enemies

diff --git a/src/fixed/room.h b/src/fixed/room.h
@@ -667,23 +667,19 @@ bool traceX(const Location &from, Location &to, bool accurate)
     if (!d.x)
         return true;
 
-    int32 dx = abs(d.x) >> 3;
-    ASSERT(dx < DIV_TABLE_SIZE);
-    dx = FixedInvU(dx);
-
-    d.y = (d.y * dx) >> (16 + 3 - TRACE_SHIFT);
-    d.z = (d.z * dx) >> (16 + 3 - TRACE_SHIFT);
+    d.y = (d.y << TRACE_SHIFT) / d.x;
+    d.z = (d.z << TRACE_SHIFT) / d.x;
 
     vec3i p = from.pos;
 
     Room* room = from.room;
 
     if (d.x < 0)
     {
-        d.x = -1024;
+        d.x = 1024;
         p.x &= ~1023;
-        p.y -= d.y * (p.x - from.pos.x) >> TRACE_SHIFT;
-        p.z -= d.z * (p.x - from.pos.x) >> TRACE_SHIFT;
+        p.y += d.y * (p.x - from.pos.x) >> TRACE_SHIFT;
+        p.z += d.z * (p.x - from.pos.x) >> TRACE_SHIFT;
 
         while (p.x > to.pos.x)
         {
@@ -694,7 +690,7 @@ bool traceX(const Location &from, Location &to, bool accurate)
             TRACE_CHECK(nextRoom, p.x - 1, p.y, p.z);
 
             room = nextRoom;
-            p += d;
+            p -= d;
         }
     }
     else
@@ -729,23 +725,19 @@ bool traceZ(const Location &from, Location &to, bool accurate)
     if (!d.z)
         return true;
 
-    int32 dz = abs(d.z) >> 3;
-    ASSERT(dz < DIV_TABLE_SIZE);
-    dz = FixedInvU(dz);
-
-    d.x = (d.x * dz) >> (16 + 3 - TRACE_SHIFT);
-    d.y = (d.y * dz) >> (16 + 3 - TRACE_SHIFT);
+    d.x = (d.x << TRACE_SHIFT) / d.z;
+    d.y = (d.y << TRACE_SHIFT) / d.z;
 
     vec3i p = from.pos;
 
     Room* room = from.room;
 
     if (d.z < 0)
     {
-        d.z = -1024;
+        d.z = 1024;
         p.z &= ~1023;
-        p.x -= d.x * (p.z - from.pos.z) >> TRACE_SHIFT;
-        p.y -= d.y * (p.z - from.pos.z) >> TRACE_SHIFT;
+        p.x += d.x * (p.z - from.pos.z) >> TRACE_SHIFT;
+        p.y += d.y * (p.z - from.pos.z) >> TRACE_SHIFT;
 
         while (p.z > to.pos.z)
         {
@@ -756,7 +748,7 @@ bool traceZ(const Location &from, Location &to, bool accurate)
             TRACE_CHECK(nextRoom, p.x, p.y, p.z - 1);
 
             room = nextRoom;
-            p += d;
+            p -= d;
         }
     }
     else

diff --git a/src/platform/gba/asm/common_asm.inc b/src/platform/gba/asm/common_asm.inc
@@ -71,21 +71,16 @@
 
 .equ FIXED_SHIFT,   14
 .equ PROJ_SHIFT,    4
-.equ OT_SHIFT,      4
+.equ MESH_SHIFT,    2
 
-.equ VIEW_DIST,     (1024 * 10)
-.equ FOG_SHIFT,     1
-.equ FOG_MAX,       VIEW_DIST
-.equ FOG_MIN,       (FOG_MAX - (8192 >> FOG_SHIFT))
-.equ VIEW_MIN,      (64)
-.equ VIEW_MAX,      (VIEW_DIST)
+.equ VIEW_MIN,      64
+.equ VIEW_MAX,      (10 << 10)
 .equ VIEW_OFF,      4096
+.equ FOG_SHIFT,     4
+.equ FOG_MIN,       (VIEW_MAX - 4096)
 
-.equ OT_SIZE,       641
-
-.equ VIEW_MIN_F,    (VIEW_MIN << FIXED_SHIFT)
-.equ VIEW_MAX_F,    (VIEW_MAX << FIXED_SHIFT)
-.equ VIEW_OFF_F,    (VIEW_OFF << FIXED_SHIFT)
+.equ OT_SHIFT,      4
+.equ OT_SIZE,       ((VIEW_MAX >> OT_SHIFT) + 1)
 
 .equ MAX_CAUSTICS,   32
 .equ MAX_RAND_TABLE, 32

diff --git a/src/platform/gba/asm/faceAddMeshQuads.s b/src/platform/gba/asm/faceAddMeshQuads.s
@@ -89,7 +89,7 @@ faceAddMeshQuads_asm:
     add depth, vg0, vg1, lsl #16
     add depth, vg2, lsl #16
     add depth, vg3, lsl #16
-    lsr depth, #(16 + 2 + OT_SHIFT)
+    lsr depth, #(16 + 2)
 
     // faceAdd
     rsb vp0, vertices, vp0, lsr #3

diff --git a/src/platform/gba/asm/faceAddMeshTriangles.s b/src/platform/gba/asm/faceAddMeshTriangles.s
@@ -80,7 +80,7 @@ faceAddMeshTriangles_asm:
     lsl vg0, #16
     add depth, vg0, vg1, lsl #16
     add depth, vg2, lsl #17
-    lsr depth, #(16 + 2 + OT_SHIFT)
+    lsr depth, #(16 + 2)
 
     // faceAdd
     rsb vp0, vertices, vp0, lsr #3

diff --git a/src/platform/gba/asm/faceAddRoomQuads.s b/src/platform/gba/asm/faceAddRoomQuads.s
@@ -92,7 +92,7 @@ faceAddRoomQuads_asm:
 
     CCW .skip
 
-    // depth = MAX_Z4
+    // depth (vz0) = MAX_Z4
     ldrh vz0, [vp0, #VERTEX_Z]
     ldrh vz1, [vp1, #VERTEX_Z]
     ldrh vz2, [vp2, #VERTEX_Z]
@@ -103,7 +103,6 @@ faceAddRoomQuads_asm:
     movlt vz0, vz2
     cmp vz0, vz3
     movlt vz0, vz3
-    mov depth, vz0, lsr #OT_SHIFT
 
     // faceAdd
     rsb vp0, vertices, vp0, lsr #3

diff --git a/src/platform/gba/asm/faceAddRoomTriangles.s b/src/platform/gba/asm/faceAddRoomTriangles.s
@@ -83,15 +83,14 @@ faceAddRoomTriangles_asm:
 
     CCW .skip
 
-    // depth = MAX_Z3
+    // depth (vz0) = MAX_Z3
     ldrh vz0, [vp0, #VERTEX_Z]
     ldrh vz1, [vp1, #VERTEX_Z]
     ldrh vz2, [vp2, #VERTEX_Z]
     cmp vz0, vz1
     movlt vz0, vz1
     cmp vz0, vz2
     movlt vz0, vz2
-    mov depth, vz0, lsr #OT_SHIFT
 
     // faceAdd
     rsb vp0, vertices, vp0, lsr #3

diff --git a/src/platform/gba/asm/sphereIsVisible.s b/src/platform/gba/asm/sphereIsVisible.s
@@ -14,13 +14,14 @@ vz      .req r13
 m       .req r14
 tmp     .req m
 vp      .req m
-vMinXY  .req z
-vMaxXY  .req r
+dz      .req vz
+minXY   .req z
+maxXY   .req r
 
-rMinX   .req vx
-rMaxX   .req x
-rMinY   .req vy
-rMaxY   .req y
+minX    .req vx
+maxX    .req x
+minY    .req vy
+maxY    .req y
 
 .global sphereIsVisible_asm
 sphereIsVisible_asm:
@@ -42,38 +43,42 @@ sphereIsVisible_asm:
     mla vz, my, y, vz
     mla vz, mz, z, vz
 
-    cmp vz, #VIEW_MAX_F
+    cmp vz, #(VIEW_MAX << FIXED_SHIFT)
     bhi .fail
 
     mov x, vx, asr #FIXED_SHIFT
     mov y, vy, asr #FIXED_SHIFT
+    mov z, vz, asr #(FIXED_SHIFT + OT_SHIFT)
 
-    mov z, vz, lsr #(FIXED_SHIFT + 4)
-    add z, vz, lsr #(FIXED_SHIFT + 6)
-    divLUT tmp, z
+    add dz, z, z, lsr #2
+    divLUT tmp, dz
     mul x, tmp, x
     mul y, tmp, y
     mul r, tmp, r
 
     mov x, x, asr #(16 - PROJ_SHIFT)
-    mov y, y, lsl #(PROJ_SHIFT)
+    mov y, y, asr #(16 - PROJ_SHIFT)
 
-    sub rMinX, x, r, lsr #(16 - PROJ_SHIFT)
-    add rMaxX, x, r, lsr #(16 - PROJ_SHIFT)
-    sub rMinY, y, r, lsl #PROJ_SHIFT
-    add rMaxY, y, r, lsl #PROJ_SHIFT
+    sub minX, x, r, lsr #(16 - PROJ_SHIFT)
+    add maxX, x, r, lsr #(16 - PROJ_SHIFT)
+    sub minY, y, r, lsr #(16 - PROJ_SHIFT)
+    add maxY, y, r, lsr #(16 - PROJ_SHIFT)
 
     ldr vp, =viewportRel
-    ldmia vp, {vMinXY, vMaxXY}
-
-    cmp rMaxX, vMinXY, asr #16
-    blt .fail
-    cmp rMaxY, vMinXY, lsl #16
-    blt .fail
-    cmp rMinX, vMaxXY, asr #16
-    bgt .fail
-    cmp rMinY, vMaxXY, lsl #16
-    bgt .fail
+    ldmia vp, {minXY, maxXY}
+
+    cmp maxX, minXY, asr #16
+    ble .fail
+    cmp minX, maxXY, asr #16
+    bge .fail
+
+    lsl minXY, #16
+    lsl maxXY, #16
+
+    cmp maxY, minXY, asr #16
+    ble .fail
+    cmp minY, maxXY, asr #16
+    bge .fail
 
     mov r0, #1
     fiq_off

diff --git a/src/platform/gba/asm/transformMesh.s b/src/platform/gba/asm/transformMesh.s
@@ -58,7 +58,7 @@ transformMesh_asm:
     asr mw1, #FIXED_SHIFT
     fiq_off
     ldmia m, {mx2, my2, mz2, mw2}
-    asr mw2, #FIXED_SHIFT
+    asr mw2, #(FIXED_SHIFT + OT_SHIFT)
     fiq_on
 
 .loop:
@@ -71,40 +71,39 @@ transformMesh_asm:
     mul x, mx0, vx
     mla x, my0, vy, x
     mla x, mz0, vz, x
-    add x, mw0, x, asr #FIXED_SHIFT
+    add x, mw0, x, asr #(FIXED_SHIFT - MESH_SHIFT)
 
     // transform y
     mul y, mx1, vx
     mla y, my1, vy, y
     mla y, mz1, vz, y
-    add y, mw1, y, asr #FIXED_SHIFT
+    add y, mw1, y, asr #(FIXED_SHIFT - MESH_SHIFT)
     fiq_off
 
     // transform z
     mul z, mx2, vx
     mla z, my2, vy, z
     mla z, mz2, vz, z
-    add z, mw2, z, asr #FIXED_SHIFT
+    add z, mw2, z, asr #(FIXED_SHIFT - MESH_SHIFT + OT_SHIFT)
 
     bic vg, #CLIP_MASK  // clear clipping flags
 
     // z clipping
-    cmp z, #VIEW_MIN
-    movle z, #VIEW_MIN
+    cmp z, #(VIEW_MIN >> OT_SHIFT)
+    movle z, #(VIEW_MIN >> OT_SHIFT)
     orrle vg, #CLIP_NEAR
-    cmp z, #VIEW_MAX
-    movge z, #VIEW_MAX
+    cmp z, #(VIEW_MAX >> OT_SHIFT)
+    movge z, #(VIEW_MAX >> OT_SHIFT)
     orrge vg, #CLIP_FAR
 
     // project
-    mov dz, z, lsr #4
-    add dz, z, lsr #6
+    add dz, z, z, lsr #2
     divLUT tmp, dz
     mul dx, x, tmp
     mul dy, y, tmp
+
     asr x, dx, #(16 - PROJ_SHIFT)
     asr y, dy, #(16 - PROJ_SHIFT)
-
     add x, #(FRAME_WIDTH >> 1)
     add y, #(FRAME_HEIGHT >> 1)