Essai de boucle de dessin en asm (echec)

CMakeLists.txt

@@ -2,7 +2,7 @@
 # toolchain file and module path of the fxSDK
 
 cmake_minimum_required(VERSION 3.15)
-project(Copy3DEngine)
+project(Copy3DEngine LANGUAGES C ASM)
 
 include(GenerateG3A)
 include(Fxconv)
@@ -18,6 +18,7 @@ set(SOURCES
   src/main.c
   src/moteur.c
   src/map.c
+  src/opti.S
 )
 
 set(ASSETS

src/main.c

@@ -27,9 +27,7 @@
 #error Ce code est pour FXCG50/G90+E uniquement, enlevez ce message a vos riques et périls
 #endif
 
-//#define debug //pour afficher les infos de debug
+#define debug 0 //pour afficher les infos de debug
-
-//extern uint16_t *gint_vram;
 
 image_t *tex_index[TINDEX_S];
 
@@ -73,7 +71,7 @@ void keys_get(){
   if (keydown(KEY_EXIT)) exit_game = 1;
 
   #ifdef debug
-  if (keydown(KEY_TAN)) end_screen();
+  //if (keydown(KEY_TAN)) end_screen();
   #endif
 }
 
@@ -129,7 +127,7 @@ int main(){
 
     if (disp_frame_time == 1) dprint( 1, 10, C_BLACK, "Frame time : %d ms", frame_time);
 
-    #ifdef debug
+    #if debug
     dprint( 1, 20, C_BLACK, "planeX : %d", planeX);
     dprint( 1, 30, C_BLACK, "planeY : %d", planeY);
     dprint( 1, 40, C_BLACK, "dirX : %d", dirX);

src/moteur.c

@@ -167,7 +167,10 @@ void load_map(){
   spawn_gen();
 }
 
-void draw_stripe(image_t *tex, int texSampleY, int texSample, int linePos, fixed_t texSize, int texX, int x){
+#if asm_opti
+void draw_stripe(image_t *tex, int texSampleY, int linePos, fixed_t texSize, int texX, int x);
+#else
+void inline draw_stripe(image_t *tex, int texSampleY, int linePos, fixed_t texSize, int texX, int x){
     fixed_t screenPos = fix(linePos);
     uint32_t texDat = (uint32_t)tex->data + 2*texX;
     for(int texPos = texSampleY; texPos < 64; texPos++){
@@ -180,6 +183,7 @@ void draw_stripe(image_t *tex, int texSampleY, int texSample, int linePos, fixed
       texDat += tex->stride;
     }
 }
+#endif
 
 void draw_walls(){
     extern fixed_t posX;
@@ -336,6 +340,6 @@ void draw_walls(){
 
       image_t *tex = tex_index[map_test[mapX][mapY]];
 
-      draw_stripe(tex, texSampleY, texSample, linePos, texSize, texX, x);
+      draw_stripe(tex, texSampleY, linePos, texSize, texX, x);
   }
 }

src/moteur.h

@@ -14,6 +14,8 @@
 
 #define TINDEX_S 256
 
+#define asm_opti 1
+
 void load_map();
 void end_screen();
 void draw_walls();

src/opti.S

@@ -0,0 +1,99 @@
+##void draw_stripe(image_t *tex, int texSampleY, int linePos, fixed_t texSize, int texX, int x){
+##    fixed_t screenPos = fix(linePos);
+##    uint32_t texDat = (uint32_t)tex->data + 2*texX;
+##    for(int texPos = texSampleY; texPos < 64; texPos++){
+##      if(screenPos < -texSize) goto noDraw;
+##      for(fixed_t oldPos = screenPos; oldPos < screenPos+texSize; oldPos+=0xFFFF){
+##        gint_vram[ffloor(oldPos)*396+x] = *(uint16_t*)texDat;
+##      }
+##      noDraw:
+##      screenPos += texSize;
+##      texDat += tex->stride;
+##    }
+##}
+
+# r0 : vram
+# r1 : tex->stride
+# r2 : texDat
+# r3 : 0xFFFF / swapped with 64
+# r4 : tex -> x -> texX -> oldPos
+# r5 : texSampleY -> texPos
+# r6 : linePos -> screenPos
+# r7 : texSize
+# r8 : 396*2
+# r9 : -texSize
+# r10 : px save
+# @-4 : texX -> r8 (save)
+# @-8 : x -> r9 (save)
+# @-12 : r10 (save)
+
+.global _draw_stripe
+.align 4
+_draw_stripe:
+  mov.l .gint_vram, r0
+
+  add #4, r4
+  mov.l @r4, r1
+
+  add #4, r4
+  mov.l @r4, r2
+  mov.l @r15+, r4
+  shll r4
+  add r4, r2
+
+  mov.l @r15+, r4
+  shll r4
+  add r4, r0
+
+  shll16 r6
+
+  mov.l r8, @-r15
+  mov #99, r8
+  shll2 r8
+  shll r8
+
+  mov.l r9, @-r15
+  mov r7, r9
+  neg r9, r9
+
+  mov.l r10, @-r15
+
+  ds_lpstart:
+    cmp/gt r6, r9
+    bt noDraw
+    nop
+    nop
+    ds_tex_lpstart:
+      mov #-1, r3
+      mov r6, r4
+      #tex read
+      mov.w @r2, r10
+      extu.w r3, r3
+      add r7, r6
+      mov.w r10, @r0
+      add r3, r4
+      cmp/gt r4, r6
+      add r8, r0
+      bf ds_tex_lpstart
+      bt ds_tex_exit
+      nop
+    noDraw:
+    add r7, r6
+    add r8, r0
+    ds_tex_exit:
+    add r1, r2
+    mov #64, r3
+    add #1, r5
+    cmp/ge r5, r3
+    bf ds_lpstart
+    nop
+
+  mov.l @r15+, r10
+  mov.l @r15+, r9
+  mov.l @r15+, r8
+  rts
+  add #-8, r15
+
+.align 4
+.gint_vram:
+  .long _gint_vram