Merge branch 'dev'

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

@@ -27,10 +27,6 @@
 #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
-
-//extern uint16_t *gint_vram;
-
 image_t *tex_index[TINDEX_S];
 
 extern char map_test[map_w][map_h];
@@ -73,7 +69,7 @@ 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
 }
 
@@ -93,8 +89,6 @@ int main(){
 
   load_map();
 
-  image_t *frame_buffer = image_create_vram();
-
   extern image_t *tex_index[TINDEX_S];
 
   tex_index[1] = &buisson0;
@@ -110,6 +104,10 @@ int main(){
 
   prof_init();
 
+  #if debug
+  EngineTimers timers;
+  #endif
+
   while (!exit_game) {
     prof_t frame = prof_make();
     prof_enter(frame);
@@ -119,23 +117,31 @@ 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);
 
-    #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);
     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

@@ -167,21 +167,41 @@ 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
+#else
+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;
@@ -218,6 +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
@@ -241,6 +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;
@@ -277,7 +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
@@ -302,6 +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
@@ -336,6 +365,10 @@ 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);
+
+      #if debug
+      prof_leave(timers->draw_time);
+      #endif
   }
 }

src/moteur.h

@@ -3,6 +3,7 @@
 #ifndef moteur_h
 #define moteur_h
 
+#include "libprof.h"
 #include <gint/image.h>
 
 //param. graphiques
@@ -14,9 +15,26 @@
 
 #define TINDEX_S 256
 
+#define debug 0 //pour afficher les infos de debug
+
+#define asm_opti 0
+
+typedef struct {
+
+  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();
+void draw_walls(
+#if debug
+  EngineTimers *timers
+#endif
+);
 void move();
 
 #endif /* moteur */