gpu_neon: psx_gpu: try to update texture cache instead of invalidating

[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 76bfb15..cf6e2be 100644 (file)
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c
@@ -244,6 +244,58 @@ u32 invalidate_texture_cache_region_viewport(psx_gpu_struct *psx_gpu, u32 x1,
   return mask;
 }
 
+void update_texture_cache_region(psx_gpu_struct *psx_gpu, u32 x1, u32 y1,
+ u32 x2, u32 y2)
+{
+  u32 mask = texture_region_mask(x1, y1, x2, y2);
+  u32 texture_page;
+  u8 *texture_page_ptr;
+  u16 *vram_ptr;
+  u32 texel_block;
+  u32 sub_x, sub_y;
+
+  psx_gpu->dirty_textures_8bpp_mask |= mask;
+  psx_gpu->dirty_textures_8bpp_alternate_mask |= mask;
+
+  if ((psx_gpu->dirty_textures_4bpp_mask & mask) == 0 &&
+      (x1 & 3) == 0 && (y1 & 15) == 0 && x2 - x1 < 4 && y2 - y1 < 16)
+  {
+    texture_page = ((x1 / 64) & 15) + (y1 / 256) * 16;
+    texture_page_ptr = psx_gpu->texture_4bpp_cache[texture_page];
+    texture_page_ptr += (x1 / 4 & 15) * 16*16 + (y1 / 16 & 15) * 16*16*16;
+    vram_ptr = psx_gpu->vram_ptr + x1 + y1 * 1024;
+    sub_x = 4;
+    sub_y = 16;
+
+    while(sub_y)
+    {
+      while(sub_x)
+      {
+        texel_block = *vram_ptr;
+
+        texture_page_ptr[0] = texel_block & 0xF;
+        texture_page_ptr[1] = (texel_block >> 4) & 0xF;
+        texture_page_ptr[2] = (texel_block >> 8) & 0xF;
+        texture_page_ptr[3] = texel_block >> 12;
+        
+        vram_ptr++;
+        texture_page_ptr += 4;
+
+        sub_x--;          
+      }
+
+      vram_ptr -= 4;
+      sub_x = 4;
+
+      sub_y--;
+      vram_ptr += 1024;
+    }
+  }
+  else
+  {
+    psx_gpu->dirty_textures_4bpp_mask |= mask;
+  }
+}
 
 void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
  u32 texture_page);
@@ -401,6 +453,48 @@ 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) &&
+   (psx_gpu->primitive_type == PRIMITIVE_TYPE_SPRITE))
+  {
+    u32 num_blocks_dest = 0;
+    block_struct *block_src = psx_gpu->blocks; 
+    block_struct *block_dest = psx_gpu->blocks;
+
+    u16 *vram_ptr = psx_gpu->vram_ptr;
+    u32 i;
+
+    if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+    {
+      for(i = 0; i < psx_gpu->num_blocks; i++)
+      {
+        u32 fb_offset = (u32)((u8 *)block_src->fb_ptr - (u8 *)vram_ptr);
+        if(fb_offset & (1 << 11))
+        {
+          *block_dest = *block_src;
+          num_blocks_dest++;
+          block_dest++;
+        }
+        block_src++;
+      }
+    }
+    else
+    {
+      for(i = 0; i < psx_gpu->num_blocks; i++)
+      {
+        u32 fb_offset = (u32)((u8 *)block_src->fb_ptr - (u8 *)vram_ptr);
+        if((fb_offset & (1 << 11)) == 0)
+        {
+          *block_dest = *block_src;
+          num_blocks_dest++;
+          block_dest++;
+        }
+        block_src++;
+      }
+    }
+
+    psx_gpu->num_blocks = num_blocks_dest;
+  }
+
   if(psx_gpu->num_blocks)
   {
     render_block_handler_struct *render_block_handler =
@@ -2987,6 +3081,28 @@ 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)
+  {
+    u32 i;
+
+    if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+    {
+      for(i = 0; i < psx_gpu->num_spans; i++)
+      {
+        if((psx_gpu->span_edge_data[i].y & 1) == 0)
+          psx_gpu->span_edge_data[i].num_blocks = 0;
+      }
+    }
+    else
+    {
+      for(i = 0; i < psx_gpu->num_spans; i++)
+      {
+        if(psx_gpu->span_edge_data[i].y & 1)
+          psx_gpu->span_edge_data[i].num_blocks = 0;
+      }
+    }
+  }
+
   u32 render_state = flags &
    (RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND | 
    RENDER_FLAGS_TEXTURE_MAP | RENDER_FLAGS_SHADE);
@@ -4322,30 +4438,48 @@ void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
   flush_render_block_buffer(psx_gpu);
   invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
 
-#ifndef NEON_BUILD
   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);
+  u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
+   psx_gpu->mask_msb;
+  u32 color_32bpp = color_16bpp | (color_16bpp << 16);
 
-  u16 *vram_ptr = psx_gpu->vram_ptr + x + (y * 1024);
-  u32 draw_x, draw_y;
+  u32 *vram_ptr = (u32 *)(psx_gpu->vram_ptr + x + (y * 1024));
 
-  for(draw_y = 0; draw_y < height; draw_y++)
+  u32 pitch = 512 - (width / 2);
+  u32 num_width;
+
+  if(psx_gpu->interlace_mode & RENDER_INTERLACE_ENABLED)
   {
-    for(draw_x = 0; draw_x < width; draw_x++)
+    pitch += 512;
+    height /= 2;
+
+    if(psx_gpu->interlace_mode & RENDER_INTERLACE_ODD)
+      vram_ptr += 512; 
+  }
+
+  while(height)
+  {
+    num_width = width;
+    while(num_width)
     {
-      vram_ptr[draw_x] = color_16bpp;
+      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 += 1024;
+    vram_ptr += pitch;
+    height--;
   }
-#else
-  void render_block_fill_body(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
-   u32 width, u32 height);
-
-  render_block_fill_body(psx_gpu, color, x, y, width, height);
-#endif
 }
 
 void render_block_copy(psx_gpu_struct *psx_gpu, u16 *source, u32 x, u32 y,
@@ -4439,6 +4573,8 @@ 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;
+
   memset(psx_gpu->vram_ptr, 0, sizeof(u16) * 1024 * 512);
 
   initialize_reciprocal_table();