psx_gpu: switch enhancement to 2048 width

[pcsx_rearmed.git] / plugins / gpu_neon / psx_gpu / psx_gpu.c

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.c b/plugins/gpu_neon/psx_gpu/psx_gpu.c
index 98aacc3..1385e2e 100644 (file)
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c
@@ -47,7 +47,8 @@ u32 zero_block_spans = 0;
 u32 texture_cache_loads = 0;
 u32 false_modulated_blocks = 0;
 
-u32 reciprocal_table[512];
+/* double size for enhancement */
+u32 reciprocal_table[512 * 2];
 
 
 typedef s32 fixed_type;
@@ -453,7 +454,7 @@ void setup_blocks_shaded_untextured_undithered_unswizzled_indirect(
 
 void flush_render_block_buffer(psx_gpu_struct *psx_gpu)
 {
-  if((psx_gpu->interlace_mode & RENDER_INTERLACE_ENABLED) &&
+  if((psx_gpu->render_mode & RENDER_INTERLACE_ENABLED) &&
    (psx_gpu->primitive_type == PRIMITIVE_TYPE_SPRITE))
   {
     u32 num_blocks_dest = 0;
@@ -463,7 +464,7 @@ void flush_render_block_buffer(psx_gpu_struct *psx_gpu)
     u16 *vram_ptr = psx_gpu->vram_ptr;
     u32 i;
 
-    if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+    if(psx_gpu->render_mode & RENDER_INTERLACE_ODD)
     {
       for(i = 0; i < psx_gpu->num_blocks; i++)
       {
@@ -1872,7 +1873,7 @@ void setup_blocks_##shading##_##texturing##_##dithering##_##sw##_##target(     \
     if(span_num_blocks)                                                        \
     {                                                                          \
       y = span_edge_data->y;                                                   \
-      fb_ptr = psx_gpu->vram_ptr + span_edge_data->left_x + (y * 1024);        \
+      fb_ptr = psx_gpu->vram_out_ptr + span_edge_data->left_x + (y * 1024);    \
                                                                                \
       setup_blocks_span_initialize_##shading##_##texturing();                  \
       setup_blocks_span_initialize_##dithering(texturing);                     \
@@ -2905,8 +2906,8 @@ char *render_block_flag_strings[] =
    (triangle_y_direction_##direction_c << 4) |                                 \
    (triangle_winding_##winding << 6))                                          \
 
-void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
- u32 flags)
+static int prepare_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
+ vertex_struct *vertexes_out[3])
 {
   s32 y_top, y_bottom;
   s32 triangle_area;
@@ -2927,7 +2928,7 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
 #ifdef PROFILE
     trivial_rejects++;
 #endif
-    return;
+    return 0;
   }
 
   if(b->y < a->y)
@@ -2949,7 +2950,7 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
 #ifdef PROFILE
     trivial_rejects++;
 #endif
-    return;
+    return 0;
   }
 
   if(triangle_area < 0)
@@ -2975,7 +2976,7 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
 #ifdef PROFILE
     trivial_rejects++;
 #endif
-    return;
+    return 0;
   }
 
   if(invalidate_texture_cache_region_viewport(psx_gpu, a->x, y_top, c->x,
@@ -2984,13 +2985,28 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
 #ifdef PROFILE
     trivial_rejects++;
 #endif
-    return;
+    return 0;
   }
 
-  psx_gpu->num_spans = 0;
   psx_gpu->triangle_area = triangle_area;
   psx_gpu->triangle_winding = triangle_winding;
 
+  vertexes_out[0] = a;
+  vertexes_out[1] = b;
+  vertexes_out[2] = c;
+
+  return 1;
+}
+
+static void render_triangle_p(psx_gpu_struct *psx_gpu,
+ vertex_struct *vertex_ptrs[3], u32 flags)
+{
+  psx_gpu->num_spans = 0;
+
+  vertex_struct *a = vertex_ptrs[0];
+  vertex_struct *b = vertex_ptrs[1];
+  vertex_struct *c = vertex_ptrs[2];
+
   s32 y_delta_a = b->y - a->y;
   s32 y_delta_b = c->y - b->y;
   s32 y_delta_c = c->y - a->y;
@@ -3002,7 +3018,7 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
   compute_all_gradients(psx_gpu, a, b, c);
 
   switch(y_direction_a | (y_direction_b << 2) | (y_direction_c << 4) |
-   (triangle_winding << 6))
+   (psx_gpu->triangle_winding << 6))
   {
     triangle_case(up, up, up, negative):
     triangle_case(up, up, flat, negative):
@@ -3081,11 +3097,11 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
   spans += psx_gpu->num_spans;
 #endif
 
-  if(psx_gpu->interlace_mode & RENDER_INTERLACE_ENABLED)
+  if(unlikely(psx_gpu->render_mode & RENDER_INTERLACE_ENABLED))
   {
     u32 i;
 
-    if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+    if(psx_gpu->render_mode & RENDER_INTERLACE_ODD)
     {
       for(i = 0; i < psx_gpu->num_spans; i++)
       {
@@ -3102,6 +3118,14 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
       }
     }
   }
+  if(psx_gpu->render_mode & RENDER_DOUBLE_MODE)
+  {
+    u32 i;
+    for(i = 0; i < psx_gpu->num_spans; i++)
+    {
+      psx_gpu->span_edge_data[i].y *= 2;
+    }
+  }
 
   u32 render_state = flags &
    (RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND | 
@@ -3126,6 +3150,14 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
    (psx_gpu);
 }
 
+void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
+ u32 flags)
+{
+  vertex_struct *vertex_ptrs[3];
+  if (prepare_triangle(psx_gpu, vertexes, vertex_ptrs))
+    render_triangle_p(psx_gpu, vertex_ptrs, flags);
+}
+
 
 void texture_sprite_blocks_8bpp(psx_gpu_struct *psx_gpu);
 
@@ -4245,7 +4277,7 @@ void render_line(psx_gpu_struct *psx_gpu, vertex_struct *vertexes, u32 flags,
 
   flags &= ~RENDER_FLAGS_TEXTURE_MAP;
 
-  vram_ptr = psx_gpu->vram_ptr + (y_a * 1024) + x_a;
+  vram_ptr = psx_gpu->vram_out_ptr + (y_a * 1024) + x_a;
 
   control_mask = 0x0;
 
@@ -4435,7 +4467,6 @@ void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
   if((width == 0) || (height == 0))
     return;
 
-  flush_render_block_buffer(psx_gpu);
   invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
 
   u32 r = color & 0xFF;
@@ -4445,17 +4476,17 @@ void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
    psx_gpu->mask_msb;
   u32 color_32bpp = color_16bpp | (color_16bpp << 16);
 
-  u32 *vram_ptr = (u32 *)(psx_gpu->vram_ptr + x + (y * 1024));
+  u32 *vram_ptr = (u32 *)(psx_gpu->vram_out_ptr + x + (y * 1024));
 
   u32 pitch = 512 - (width / 2);
   u32 num_width;
 
-  if(psx_gpu->interlace_mode & RENDER_INTERLACE_ENABLED)
+  if(psx_gpu->render_mode & RENDER_INTERLACE_ENABLED)
   {
     pitch += 512;
     height /= 2;
 
-    if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+    if(psx_gpu->render_mode & RENDER_INTERLACE_ODD)
       vram_ptr += 512; 
   }
 
@@ -4482,6 +4513,47 @@ void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
   }
 }
 
+void render_block_fill_enh(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
+ u32 width, u32 height)
+{
+  if((width == 0) || (height == 0))
+    return;
+
+  u32 r = color & 0xFF;
+  u32 g = (color >> 8) & 0xFF;
+  u32 b = (color >> 16) & 0xFF;
+  u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
+   psx_gpu->mask_msb;
+  u32 color_32bpp = color_16bpp | (color_16bpp << 16);
+
+  u32 *vram_ptr = (u32 *)(psx_gpu->vram_out_ptr + x + (y * 2048));
+
+  u32 pitch = 2048 / 2 - (width / 2);
+  u32 num_width;
+
+  while(height)
+  {
+    num_width = width;
+    while(num_width)
+    {
+      vram_ptr[0] = color_32bpp;
+      vram_ptr[1] = color_32bpp;
+      vram_ptr[2] = color_32bpp;
+      vram_ptr[3] = color_32bpp;
+      vram_ptr[4] = color_32bpp;
+      vram_ptr[5] = color_32bpp;
+      vram_ptr[6] = color_32bpp;
+      vram_ptr[7] = color_32bpp;
+
+      vram_ptr += 8;
+      num_width -= 16;
+    }
+
+    vram_ptr += pitch;
+    height--;
+  }
+}
+
 void render_block_copy(psx_gpu_struct *psx_gpu, u16 *source, u32 x, u32 y,
  u32 width, u32 height, u32 pitch)
 {
@@ -4522,7 +4594,8 @@ void initialize_reciprocal_table(void)
   u32 height_reciprocal;
   s32 shift;
 
-  for(height = 1; height < 512; height++)
+  for(height = 1; height < sizeof(reciprocal_table)
+       / sizeof(reciprocal_table[0]); height++)
   {
     shift = __builtin_clz(height);
     height_normalized = height << shift;
@@ -4561,6 +4634,7 @@ void initialize_psx_gpu(psx_gpu_struct *psx_gpu, u16 *vram)
   psx_gpu->num_blocks = 0;
 
   psx_gpu->vram_ptr = vram;
+  psx_gpu->vram_out_ptr = vram;
 
   psx_gpu->texture_page_base = psx_gpu->vram_ptr;
   psx_gpu->texture_page_ptr = psx_gpu->vram_ptr;
@@ -4573,7 +4647,7 @@ void initialize_psx_gpu(psx_gpu_struct *psx_gpu, u16 *vram)
   psx_gpu->texture_mask_width = 0xFF;
   psx_gpu->texture_mask_height = 0xFF;
 
-  psx_gpu->interlace_mode = 0;
+  psx_gpu->render_mode = 0;
 
   memset(psx_gpu->vram_ptr, 0, sizeof(u16) * 1024 * 512);