From 2c8f55a22b2dc25bc82d3dab0aebe04e61e82d9f Mon Sep 17 00:00:00 2001
From: XProger <xproger@list.ru>
Date: Fri, 21 Jan 2022 04:24:52 +0300
Subject: [PATCH] #368 GBA TEX_2PX akill texturing, fix FTA/GTA unaligned VRAM
 write

---
 src/platform/gba/rasterizeFTA.s  | 21 +++++++++++++++------
 src/platform/gba/rasterizeGTA.s  | 15 ++++++++++++---
 src/platform/gba/transformMesh.s |  8 --------
 3 files changed, 27 insertions(+), 17 deletions(-)

diff --git a/src/platform/gba/rasterizeFTA.s b/src/platform/gba/rasterizeFTA.s
index 4a8b7ec..23d3713 100644
--- a/src/platform/gba/rasterizeFTA.s
+++ b/src/platform/gba/rasterizeFTA.s
@@ -60,13 +60,14 @@ SP_RDX = 8
 SP_RDT = 12
 
 .macro PUT_PIXELS
+#ifndef TEX_2PX
     tex indexA, t
     add t, dtdx
 
     tex indexB, t
     add t, dtdx
 
-  // cheap non-accurate alpha test, skip pixels pair if one or both are transparent
+    // cheap non-accurate alpha test, skip pixels pair if one or both are transparent
     ands indexA, #255
     andnes indexB, #255
     orrne indexB, indexA, indexB, lsl #8   // indexB = indexA | (indexB << 8)
@@ -75,6 +76,14 @@ SP_RDT = 12
     orrne indexA, indexB, lsl #8
     strneh indexA, [tmp]
     add tmp, #2
+#else
+    tex indexA, t
+    add t, dtdx, lsl #1
+    cmp indexA, #0
+    ldrneb indexA, [LMAP, indexA]
+    strneb indexA, [tmp]
+    add tmp, #2
+#endif
 .endm
 
 .global rasterizeFTA_asm
@@ -215,12 +224,11 @@ rasterizeFTA_asm:
     ldrb indexA, [TILE, indexA]
 
     cmp indexA, #0
-    ldrneb indexB, [tmp, #-1]        // read pal index from VRAM (byte)
+    ldrneb indexB, [tmp, #-1]!      // read pal index from VRAM (byte)
     ldrneb indexA, [LMAP, indexA]
     orrne indexB, indexA, lsl #8
-    strneh indexB, [tmp]
-
-    add tmp, #1
+    strneh indexB, [tmp], #2
+    addeq tmp, #1
     add t, dtdx
 
     subs width, #1              // width--
@@ -240,7 +248,8 @@ rasterizeFTA_asm:
     ldrneb indexA, [LMAP, indexA]
     ldrneb indexB, [tmp, width]
     orrne indexB, indexA, indexB, lsl #8
-    strneh indexB, [tmp, width]
+    addne indexA, tmp, width
+    strneh indexB, [indexA, #-1]
 
     subs width, #1              // width--
       beq .scanline_end         // if (width == 0)
diff --git a/src/platform/gba/rasterizeGTA.s b/src/platform/gba/rasterizeGTA.s
index 487c45a..17b4012 100644
--- a/src/platform/gba/rasterizeGTA.s
+++ b/src/platform/gba/rasterizeGTA.s
@@ -88,13 +88,14 @@ SP_RDT = 20
     bic LMAP, g, #255
     add g, dgdx
 
+#ifndef TEX_2PX
     tex indexA, t
     add t, dtdx
 
     tex indexB, t
     add t, dtdx
 
-  // cheap non-accurate alpha test, skip pixels pair if one or both are transparent
+    // cheap non-accurate alpha test, skip pixels pair if one or both are transparent
     ands indexA, #255
     andnes indexB, #255
     orrne indexB, indexA, indexB, lsl #8   // indexB = indexA | (indexB << 8)
@@ -102,6 +103,14 @@ SP_RDT = 20
     ldrneb indexB, [LMAP, indexB, lsr #8]
     orrne indexA, indexB, lsl #8
     strneh indexA, [ptr]
+#else
+    tex indexA, t
+    add t, dtdx, lsl #1
+    cmp indexA, #0
+    ldrneb indexA, [LMAP, indexA]
+    strneb indexA, [ptr]
+#endif
+
     add ptr, #2
 .endm
 
@@ -270,9 +279,9 @@ rasterizeGTA_asm:
     bic LMAP, g, #255
     ldrb indexA, [LMAP, indexA]
 
-    ldrb indexB, [ptr, #-1]      // read pal index from VRAM (byte)
+    ldrb indexB, [ptr, #-1]!     // read pal index from VRAM (byte)
     orr indexB, indexA, lsl #8
-    strh indexB, [ptr]
+    strh indexB, [ptr], #1
 
 .skip_left:
     add ptr, #1
diff --git a/src/platform/gba/transformMesh.s b/src/platform/gba/transformMesh.s
index 6bf7f5c..a1240a6 100644
--- a/src/platform/gba/transformMesh.s
+++ b/src/platform/gba/transformMesh.s
@@ -121,18 +121,10 @@ transformMesh_asm:
     add y, y, #(FRAME_HEIGHT >> 1)
 
     // store the result
-#if 1
-    // 2+2+2+2 = 8 ticks
     strh x, [res], #2
     strh y, [res], #2
     strh z, [res], #2
     strh vg, [res], #2
-#else
-    // 1+1+3 = 5 ticks SLOWER! WTF?
-    orr x, x, y, lsl #16
-    orr z, z, vg, lsl #16
-    stmia res!, {x, z}
-#endif
 
     subs count, #1
     bne .loop