new: 16bit particle variant, adjust examples

examples/18_particles/main.c

@@ -43,10 +43,10 @@ int main()
   rdpq_init();
   //rdpq_debug_start();
 
+  uint64_t rspTimeTPX = 0;
+  uint64_t rdpTimeBusy = 0;
   #if RSPQ_PROFILE
     rspq_profile_data_t profile_data = (rspq_profile_data_t){};
-    uint64_t rdpTimeBusy = 0;
-    uint64_t rspTimeTPX = 0;
     rspq_profile_start();
   #endif
 
@@ -65,13 +65,20 @@ int main()
   // Meaning you only have to allocate an buffer of arbitrary size here and fill it with data.
   uint32_t particleCountMax = 100'000;
   uint32_t particleCount = 2000;
+
   // NOTE: just like with vertices, particles are interleaved in pairs of 2.
-  // So one TPXParticle struct always contains 2 particles.
+  // So one TPXParticleS8 struct always contains 2 particles.
   // If you need an odd number, just set the second particle size to 0.
-  uint32_t allocSize = sizeof(TPXParticle) * particleCountMax / 2;
-  TPXParticle *particles = malloc_uncached(allocSize);
+  uint32_t allocSize = sizeof(TPXParticleS8) * particleCountMax / 2;
+  TPXParticleS8 *particlesS8 = malloc_uncached(allocSize);
   debugf("Particle-Buffer %ldkb\n", allocSize / 1024);
-  generate_particles_random(particles, particleCount);
+
+  // Additionally, a 16bit version of particles is available.
+  // This one takes up more space (24 bytes vs 16 bytes per pair) and is slightly slower.
+  // In return, it can cover a larger range which can be useful for 3D sprites placed in a scene.
+  // The 8bit variant should be preferred when possible (e.g. in local particle effects)
+  allocSize = sizeof(TPXParticleS16) * particleCountMax / 2;
+  TPXParticleS16 *particlesS16 = malloc_uncached(allocSize);
 
   // Now some regular 3D stuff, not related to particles.
   T3DModel *model = t3d_model_load("rom://scene.t3dm");
@@ -111,6 +118,7 @@ int main()
   float time = 0;
   bool needRebuild = true;
   int frameIdx = 0;
+  bool measureTime = false;
 
   for(;;)
   {
@@ -132,6 +140,10 @@ int main()
     if(joypad.btn.c_up)partSizeY += deltaTime * 0.6f;
     if(joypad.btn.c_down)partSizeY -= deltaTime * 0.6f;
 
+  #if RSPQ_PROFILE
+    measureTime = joypad.btn.z;
+  #endif
+
     partSizeX = fmaxf(0.01f, fminf(1.0f, partSizeX));
     partSizeY = fmaxf(0.01f, fminf(1.0f, partSizeY));
 
@@ -173,42 +185,44 @@ int main()
 
     // A few example particles systems.
     // This will modify the particle buffer on the CPU side.
+    bool isS16 = false;
     switch(example)
     {
       case 0: // Random
         time += deltaTime * 0.2f;
         particleRot = (T3DVec3){{time,time*0.77f,time*1.42f}};
         particleMatScale = (T3DVec3){{partMatScaleVal, partMatScaleVal, partMatScaleVal}};
 
-        if(needRebuild)generate_particles_random(particles, particleCount);
+        if(needRebuild)generate_particles_random(particlesS8, particleCount);
         rdpq_set_env_color((color_t){0xFF, 0xFF, 0xFF, 0xFF});
-      break;
+        break;
       case 1: // Flame
         particleRot = (T3DVec3){{0,0,0}};
         if(!joypad.btn.z)time += deltaTime * 1.0f;
         particleCount = 128;
         float posX = fm_cosf(time) * 80.0f;
         float posZ = fm_sinf(2*time) * 40.0f;
 
-        simulate_particles_fire(particles, particleCount, posX, posZ);
+        simulate_particles_fire(particlesS8, particleCount, posX, posZ);
         particleMatScale = (T3DVec3){{0.9f, partMatScaleVal, 0.9f}};
         particlePos.y = partMatScaleVal * 130.0f;
         rdpq_set_env_color((color_t){0xFF, 0xFF, 0xFF, 0xFF});
-      break;
+        break;
       case 2: // Grass
         time += deltaTime * 1.0f;
         particleRot = (T3DVec3){{0,0,0}};
         particlePos.y = 0;
         if(needRebuild) {
-          particleCount = simulate_particles_grass(particles, particleCount);
+          particleCount = simulate_particles_grass(particlesS16, particleCount);
         }
         particleMatScale = (T3DVec3){{partMatScaleVal, partSizeY * 2.9f, partMatScaleVal}};
         rdpq_set_env_color(blend_colors(
           (color_t){0xAA, 0xFF, 0x55, 0xFF},
           (color_t){0xFF, 0xAA, 0x55, 0xFF},
           fm_sinf(time)*0.5f+0.5f
         ));
-      break;
+        isS16 = true;
+        break;
     }
     needRebuild = false;
 
@@ -267,10 +281,29 @@ int main()
     // This can only scale particles down, so the range is 0.0 - 1.0.
     tpx_state_set_scale(partSizeX, partSizeY);
 
+    if(measureTime) {
+      rspq_wait();
+      rspq_highpri_begin();
+      wait_ms(2);
+      rspTimeTPX = get_ticks();
+    }
+
     // Now draw particles. internally this will load, transform and draw them in one go on the RSP.
     // While the ucode can only handle a 344 at a time, this function will automatically batch them
     // so you can specify an arbitrary amount of particles (as long as it's an even count)
-    tpx_particle_draw(particles, particleCount);
+    if(isS16) {
+      tpx_particle_draw_s16(particlesS16, particleCount);
+    } else {
+      tpx_particle_draw_s8(particlesS8, particleCount);
+    }
+
+    if(measureTime)
+    {
+      rspq_highpri_end();
+      rspq_highpri_sync();
+      rspTimeTPX = get_ticks() - rspTimeTPX;
+      rspTimeTPX = TICKS_TO_US(rspTimeTPX);
+    }
 
     // Make sure end up at the same stack level as before.
     tpx_matrix_pop(1);
@@ -283,18 +316,18 @@ int main()
     t3d_debug_printf(20,  30, "[C] %.2f %.2f", partSizeX, partSizeY);
     t3d_debug_printf(220, 18, "FPS: %.2f", display_get_fps());
 
-    #if RSPQ_PROFILE
+    if(measureTime)
+    {
       double timePerPart = 0;
       if(particleCount > 0) {
         timePerPart = (double)rspTimeTPX / (double)particleCount * 1000;
       }
       t3d_debug_printf(20, 240-34, "RSP/tpx: %6lldus %.1f", rspTimeTPX, timePerPart);
       //t3d_debug_printf(20, 240-34, "RSP/tpx: %6lldus", rspTimeTPX);
       t3d_debug_printf(20, 240-24, "RDP    : %6lldus", rdpTimeBusy);
-    #else
+    } else {
       t3d_debug_printf(20, 240-24, "[L/R]: %s", EXAMPLE_NAMES[example]);
-    #endif
-
+    }
     rdpq_detach_show();
 
     #if RSPQ_PROFILE

examples/18_particles/partSim.h

@@ -5,7 +5,7 @@ static int currentPart  = 0;
 /**
  * Basic static particles with random positions and colors.
  */
-static void generate_particles_random(TPXParticle *particles, uint32_t count) {
+static void generate_particles_random(TPXParticleS8 *particles, uint32_t count) {
   for (int i = 0; i < count; i++) {
     int p = i / 2;
     int8_t *ptPos = i % 2 == 0 ? particles[p].posA : particles[p].posB;
@@ -47,7 +47,7 @@ static int noise_2d(int x, int y) {
  * Static particles simulating grass.
  * This will create a random grid of 3 particles stacked on top of each other representing grass-blades.
  */
-static int simulate_particles_grass(TPXParticle *particles, uint32_t partCount) {
+static int simulate_particles_grass(TPXParticleS16 *particles, uint32_t partCount) {
 
   int dist = 3;
   int heightParts = 3;
@@ -56,14 +56,14 @@ static int simulate_particles_grass(TPXParticle *particles, uint32_t partCount)
   int p = 0;
   for(int y=heightParts-1; y>=0; --y)
   {
-    int8_t ptPosX = -(dist * sideLen) / 2;
+    int16_t ptPosX = -(dist * sideLen) / 2;
     for(int x=0; x<sideLen; ++x)
     {
-      int8_t ptPosZ = -(dist * sideLen) / 2;
+      int16_t ptPosZ = -(dist * sideLen) / 2;
       for(int z=0; z<sideLen; ++z)
       {
-        int8_t *ptPos = tpx_buffer_get_pos(particles, p);
-        uint8_t *ptColor = tpx_buffer_get_rgba(particles, p);
+        int16_t *ptPos = tpx_buffer_s16_get_pos(particles, p);
+        uint8_t *ptColor = tpx_buffer_s16_get_rgba(particles, p);
 
         int rnd = noise_2d(x, z);
         float height = fm_sinf((x + z) * 0.1f) * 0.5f + 0.5f;
@@ -75,7 +75,7 @@ static int simulate_particles_grass(TPXParticle *particles, uint32_t partCount)
         ptPos[0] = ptPosX + ((rnd % 3) - 1);
         ptPos[1] = y + height;
         ptPos[2] = ptPosZ + ((rnd % 3) - 1);
-        *tpx_buffer_get_size(particles, p) = size;
+        *tpx_buffer_s16_get_size(particles, p) = size;
 
         ptPosZ += dist;
 
@@ -119,7 +119,7 @@ static void gradient_fire(uint8_t *color, float t) {
  * This will simulate particles over time by moving them up and changing their color.
  * The current position is used to spawn new particles, so it can move over time leaving a trail behind.
  */
-static void simulate_particles_fire(TPXParticle *particles, uint32_t partCount, float posX, float posZ) {
+static void simulate_particles_fire(TPXParticleS8 *particles, uint32_t partCount, float posX, float posZ) {
   uint32_t p = currentPart / 2;
   if(currentPart % (1+(rand() % 3)) == 0) {
     int8_t *ptPos = currentPart % 2 == 0 ? particles[p].posA : particles[p].posB;

examples/19_particles_tex/main.c

@@ -65,10 +65,10 @@ int main()
   rdpq_init();
   //rdpq_debug_start();
 
+  uint64_t rdpTimeBusy = 0;
+  uint64_t rspTimeTPX = 0;
   #if RSPQ_PROFILE
     rspq_profile_data_t profile_data = (rspq_profile_data_t){};
-    uint64_t rdpTimeBusy = 0;
-    uint64_t rspTimeTPX = 0;
     rspq_profile_start();
   #endif
 
@@ -87,10 +87,17 @@ int main()
   // There is no special struct for textured particles compared to colored ones.
   // The only difference is that the alpha channel of the color is used for the texture offset.
   // You can still define a global alpha value via the CC ofc.
-  uint32_t allocSize = sizeof(TPXParticle) * particleCountMax / 2;
-  TPXParticle *particles = malloc_uncached(allocSize);
+  uint32_t allocSize = sizeof(TPXParticleS8) * particleCountMax / 2;
+  TPXParticleS8 *particlesS8 = malloc_uncached(allocSize);
   debugf("Particle-Buffer %ldkb\n", allocSize / 1024);
 
+  // Additionally, a 16bit version of particles is available.
+  // This one takes up more space (24 bytes vs 16 bytes per pair) and is slightly slower.
+  // In return, it can cover a larger range which can be useful for 3D sprites placed in a scene.
+  // The 8bit variant should be preferred when possible (e.g. in local particle effects)
+  allocSize = sizeof(TPXParticleS16) * particleCountMax / 2;
+  TPXParticleS16 *particlesS16 = malloc_uncached(allocSize);
+
   sprite_t *texTest[] = {
       sprite_load("rom://tex8.i8.sprite"),
       sprite_load("rom://tex16.i8.sprite"),
@@ -137,6 +144,7 @@ int main()
   float time = 0;
   float timeTile = 0;
   bool needRebuild = true;
+  bool measureTime = false;
   int frameIdx = 0;
 
   for(;;)
@@ -160,6 +168,10 @@ int main()
     if(joypad.btn.c_up)partSizeY += deltaTime * 0.6f;
     if(joypad.btn.c_down)partSizeY -= deltaTime * 0.6f;
 
+#if RSPQ_PROFILE
+    measureTime = joypad.btn.z;
+#endif
+
     partSizeX = fmaxf(0.01f, fminf(1.0f, partSizeX));
     partSizeY = fmaxf(0.01f, fminf(1.0f, partSizeY));
 
@@ -200,6 +212,7 @@ int main()
       camTarget.v[2] = camPos.v[2] + camDir.v[2];
     }
 
+    bool is16Bit = false;
     bool isSpriteRot = false;
     switch(example)
     {
@@ -211,7 +224,7 @@ int main()
         float posX = fm_cosf(time) * 80.0f;
         float posZ = fm_sinf(2*time) * 40.0f;
 
-        simulate_particles_fire(particles, particleCount, posX, posZ);
+        simulate_particles_fire(particlesS8, particleCount, posX, posZ);
         particleMatScale = (T3DVec3){{0.9f, partMatScaleVal, 0.9f}};
         particlePos.y = partMatScaleVal * 130.0f;
         rdpq_set_env_color((color_t){0xFF, 0xFF, 0xFF, 0xFF});
@@ -223,18 +236,19 @@ int main()
         particleRot = (T3DVec3){{0,0,0}};
         particlePos.y = 0;
         if(needRebuild) {
-          particleCount = simulate_particles_coins(particles, particleCount);
+          particleCount = simulate_particles_coins(particlesS16, particleCount);
         }
         particleMatScale = (T3DVec3){{partMatScaleVal, partSizeY * 2.9f, partMatScaleVal}};
         rdpq_set_env_color((color_t){0xFF, 0xFF, 0xFF, 0xFF});
+        is16Bit = true;
       break;
       default: // Random
         time += deltaTime * 0.2f;
         timeTile += deltaTime * 0.1f;
         particleRot = (T3DVec3){{time,time*0.77f,time*1.42f}};
         particleMatScale = (T3DVec3){{partMatScaleVal, partMatScaleVal, partMatScaleVal}};
 
-        if(needRebuild)generate_particles_random(particles, particleCount);
+        if(needRebuild)generate_particles_random(particlesS8, particleCount);
         rdpq_set_env_color((color_t){0xFF, 0xFF, 0xFF, 0xFF});
         isSpriteRot = true;
       break;
@@ -326,7 +340,26 @@ int main()
       case 5: tpx_state_set_tex_params((int16_t)tileIdx, 0); break;
     }
 
-    tpx_particle_draw_tex(particles, particleCount);
+    if(measureTime) {
+      rspq_wait();
+      rspq_highpri_begin();
+      wait_ms(2);
+      rspTimeTPX = get_ticks();
+    }
+
+    if(is16Bit) {
+      tpx_particle_draw_tex_s16(particlesS16, particleCount);
+    } else {
+      tpx_particle_draw_tex_s8(particlesS8, particleCount);
+    }
+
+    if(measureTime)
+    {
+      rspq_highpri_end();
+      rspq_highpri_sync();
+      rspTimeTPX = get_ticks() - rspTimeTPX;
+      rspTimeTPX = TICKS_TO_US(rspTimeTPX);
+    }
 
     tpx_matrix_pop(1);
 
@@ -335,17 +368,18 @@ int main()
     t3d_debug_printf(20,  30, "[C] %.2f %.2f", partSizeX, partSizeY);
     t3d_debug_printf(220, 18, "FPS: %.2f", display_get_fps());
 
-    #if RSPQ_PROFILE
+    if(measureTime)
+    {
       double timePerPart = 0;
       if(particleCount > 0) {
         timePerPart = (double)rspTimeTPX / (double)particleCount * 1000;
       }
       t3d_debug_printf(20, 240-34, "RSP/tpx: %6lldus %.1f", rspTimeTPX, timePerPart);
       //t3d_debug_printf(20, 240-34, "RSP/tpx: %6lldus", rspTimeTPX);
       t3d_debug_printf(20, 240-24, "RDP    : %6lldus", rdpTimeBusy);
-    #else
+    } else {
       t3d_debug_printf(20, 240-24, "[L/R]: %s", EXAMPLE_NAMES[example]);
-    #endif
+    }
 
     rdpq_detach_show();