draw_stripe : Opti max (Pour du C)

CMakeLists.txt

@@ -18,7 +18,7 @@ set(SOURCES
   src/main.c
   src/moteur.c
   src/map.c
-  src/opti.S
+#  src/opti.S
 )
 
 set(ASSETS

src/main.c

@@ -27,8 +27,6 @@
 #error Ce code est pour FXCG50/G90+E uniquement, enlevez ce message a vos riques et périls
 #endif
 
-#define debug 0 //pour afficher les infos de debug
-
 image_t *tex_index[TINDEX_S];
 
 extern char map_test[map_w][map_h];
@@ -91,8 +89,6 @@ int main(){
 
   load_map();
 
-  image_t *frame_buffer = image_create_vram();
-
   extern image_t *tex_index[TINDEX_S];
 
   tex_index[1] = &buisson0;
@@ -108,6 +104,10 @@ int main(){
 
   prof_init();
 
+  #if debug
+  EngineTimers timers;
+  #endif
+
   while (!exit_game) {
     prof_t frame = prof_make();
     prof_enter(frame);
@@ -117,14 +117,18 @@ int main(){
     dma_memset((void*)((uint32_t)gint_vram + viewport_w*viewport_h),
                0xc4c9c4c9, viewport_w*viewport_h);
 
-    draw_walls();
-
     if(first_frame){
       main_menu();
     }
 
     keys_get();
 
+    draw_walls(
+    #if debug
+      &timers
+    #endif
+    );
+
     if (disp_frame_time == 1) dprint( 1, 10, C_BLACK, "Frame time : %d ms", frame_time);
 
     #if debug
@@ -134,6 +138,10 @@ int main(){
     dprint( 1, 50, C_BLACK, "dirY : %d", dirY);
     dprint( 1, 60, C_BLACK, "posX : %d", posX);
     dprint( 1, 70, C_BLACK, "posY : %d", posY);
+    dprint( 1, 80, C_BLACK, "Raycast time : %d", prof_time(timers.raycast_time));
+    dprint( 1, 90, C_BLACK, "Draw time : %d", prof_time(timers.draw_time));
+    timers.raycast_time = prof_make();
+    timers.draw_time = prof_make();
     #endif
 
     dupdate();

src/moteur.c

@@ -168,24 +168,40 @@ void load_map(){
 }
 
 #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){
+inline void __attribute__((always_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++){
+    register int vramSize asm("r2") = 396;
-      if(screenPos < -texSize) goto noDraw;
+    register int oneConst asm("r12");
-      for(fixed_t oldPos = screenPos; oldPos < screenPos+texSize; oldPos+=0xFFFF){
+    asm("mov #-1, r12\n"
-        gint_vram[ffloor(oldPos)*396+x] = *(uint16_t*)texDat;
+        "extu.w r12, r12"
+        : "=r" (oneConst)
+        : "r" (oneConst)
+    );
+    for(int texPos = texSampleY; texPos < 64; ++texPos){
+      if(screenPos >= -texSize){
+        int vrampos = ffloor(screenPos)*vramSize+x;
+        fixed_t oldPos = screenPos;
+        int16_t pix = *(int16_t*)texDat;
+        do{
+          gint_vram[vrampos] = pix;
+          vrampos += vramSize;
+          oldPos += oneConst;
+        }while(oldPos < screenPos+texSize);
       }
-      noDraw:
       screenPos += texSize;
       texDat += tex->stride;
     }
 }
 #endif
 
-void draw_walls(){
+void draw_walls(
+#if debug
+  EngineTimers *timers
+#endif
+){
     extern fixed_t posX;
     extern fixed_t posY;
     extern fixed_t dirX;
@@ -222,6 +238,9 @@ void draw_walls(){
     struct image_linear_map temp;
 
     for(x = 0; x < viewport_w; x++) {
+      #if debug
+      prof_enter(timers->raycast_time);
+      #endif
 
       //calculate ray position and direction
       cameraX = fdiv(fix(x*2), fix(viewport_w)) - 0xFFFF + h_offset; //x-coordinate in camera space
@@ -245,6 +264,7 @@ void draw_walls(){
       // stepping further below works. So the values can be computed as below.
       // Division through zero is prevented, even though technically that's not
       // needed in C++ with IEEE 754 floating point values. 
+      //Fcalva : It is with fp32s !
 
       rayDirX = rayDirX == 0 ? 1 : rayDirX;
       rayDirY = rayDirY == 0 ? 1 : rayDirY;
@@ -281,7 +301,7 @@ void draw_walls(){
           break;
         }
         //Otherwise check if ray has hit a wall
-        else if (map_test[mapX][mapY] > 0) {
+        if (map_test[mapX][mapY] > 0) {
           break;
         }
         //jump to next map square, either in x-direction, or in y-direction
@@ -306,6 +326,11 @@ void draw_walls(){
       if (side == 0) perpWallDist = (sideDistX - deltaDistX);
       else perpWallDist = (sideDistY - deltaDistY);
 
+      #if debug
+      prof_leave(timers->raycast_time);
+      prof_enter(timers->draw_time);
+      #endif
+
       //texturing calculations
 
       //calculate value of wallX
@@ -341,5 +366,9 @@ void draw_walls(){
       image_t *tex = tex_index[map_test[mapX][mapY]];
 
       draw_stripe(tex, texSampleY, linePos, texSize, texX, x);
+
+      #if debug
+      prof_leave(timers->draw_time);
+      #endif
   }
 }

src/moteur.h

@@ -1,24 +1,40 @@
-// Voir README.md pour license précise, par Fcalva 2023-2024 et est sous GPLv3
+// Voir README.md pour license précise, par Fcalva 2023-2024 et est sous GPLv3
-
+
-#ifndef moteur_h
+#ifndef moteur_h
-#define moteur_h
+#define moteur_h
-
+
-#include <gint/image.h>
+#include "libprof.h"
-
+#include <gint/image.h>
-//param. graphiques
+
-#define screen_w 396
+//param. graphiques
-#define screen_h 224
+#define screen_w 396
-#define viewport_w 396
+#define screen_h 224
-#define viewport_h 224
+#define viewport_w 396
-#define max_dist fix(32) //en tuiles << 16, actuellement 32
+#define viewport_h 224
-
+#define max_dist fix(32) //en tuiles << 16, actuellement 32
-#define TINDEX_S 256
+
-
+#define TINDEX_S 256
-#define asm_opti 1
+
-
+#define debug 0 //pour afficher les infos de debug
-void load_map();
+
-void end_screen();
+#define asm_opti 0
-void draw_walls();
+
-void move();
+typedef struct {
-
+
-#endif /* moteur */
+  prof_t raycast_time;
+  prof_t draw_time;
+
+} EngineTimers;
+
+void draw_stripe(image_t *tex, int texSampleY, int linePos, fixed_t texSize, int texX, int x);
+
+void load_map();
+void end_screen();
+void draw_walls(
+#if debug
+  EngineTimers *timers
+#endif
+);
+void move();
+
+#endif /* moteur */

src/opti.S

@@ -1,99 +0,0 @@
-##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