psx_gpu: 2x sprite fixes and integration

4 files changed, 64 insertions, 41 deletions

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.c b/plugins/gpu_neon/psx_gpu/psx_gpu.c
index 9b5a64d..2d552aa 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c
@@ -3574,7 +3574,7 @@ void setup_sprite_16bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
   u32 left_offset = u & 0x7;
   u32 width_rounded = width + left_offset + 7;
 
-  u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + (s32)(x - left_offset);
+  u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + (s32)(x - left_offset);
   u32 right_width = width_rounded & 0x7;
   u32 block_width = width_rounded / 8;
   u32 fb_ptr_pitch = (1024 + 8) - (block_width * 8);
@@ -3696,7 +3696,7 @@ void setup_sprite_untextured(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
 {
   u32 right_width = ((width - 1) & 0x7) + 1;
   u32 right_mask_bits = (0xFF << right_width);
-  u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + x;
+  u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + x;
   u32 block_width = (width + 7) / 8;
   u32 fb_ptr_pitch = 1024 - ((block_width - 1) * 8);
   u32 blocks_remaining;
@@ -4736,3 +4736,4 @@ void triangle_benchmark(psx_gpu_struct *psx_gpu)
 
 #endif
 
+#include "psx_gpu_4x.c"
diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c b/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
index 5e87bea..f8afcf1 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
@@ -50,7 +50,7 @@
     tbl_16(texels_high, texels, clut_high);                                    \

     zip_8x16b(pixels, texels_low, texels_high);                                \

                                                                                \

-    zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);		 \

+    zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);    \

     block->texels = pixels_wide;                                               \

     block->draw_mask_bits = left_mask_bits_a;                                  \

     block->fb_ptr = fb_ptr;                                                    \

@@ -77,7 +77,7 @@
     tbl_16(texels_high, texels, clut_high);                                    \

     zip_8x16b(pixels, texels_low, texels_high);                                \

                                                                                \

-    zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);	   \

+    zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);    \

     block->texels = pixels_wide;                                               \

     block->draw_mask_bits = right_mask_bits_a;                                 \

     block->fb_ptr = fb_ptr + 16;                                               \

@@ -111,7 +111,7 @@
 {                                                                              \

   vec_8x8u texels_low, texels_high;                                            \

   vec_8x16u pixels, pixels_wide;                                               \

-  setup_sprite_tile_add_blocks(sub_tile_height);                               \

+  setup_sprite_tile_add_blocks_4x(sub_tile_height);                            \

   u32 edge##_mask_bits_a = edge##_mask_bits & 0xFF;                            \

   u32 edge##_mask_bits_b = edge##_mask_bits >> 8;                              \

                                                                                \

@@ -122,7 +122,7 @@
     tbl_16(texels_high, texels, clut_high);                                    \

     zip_8x16b(pixels, texels_low, texels_high);                                \

                                                                                \

-    zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);		 \

+    zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);    \

     block->texels = pixels_wide;                                               \

     block->draw_mask_bits = edge##_mask_bits_a;                                \

     block->fb_ptr = fb_ptr;                                                    \

@@ -195,12 +195,12 @@
                                                                                \

     block->r = texels_wide.low;                                                \

     block->draw_mask_bits = right_mask_bits_a;                                 \

-    block->fb_ptr = fb_ptr + 1024;                                             \

+    block->fb_ptr = fb_ptr + 1024 + 16;                                        \

     block++;                                                                   \

                                                                                \

     block->r = texels_wide.high;                                               \

     block->draw_mask_bits = right_mask_bits_b;                                 \

-    block->fb_ptr = fb_ptr + 24 + 1024;                                        \

+    block->fb_ptr = fb_ptr + 24;                                               \

     block++;                                                                   \

                                                                                \

     block->r = texels_wide.high;                                               \

@@ -218,7 +218,7 @@
 

 #define setup_sprite_tile_half_8bpp_4x(edge)                                   \

 {                                                                              \

-  setup_sprite_tile_add_blocks_4x(sub_tile_height * 2);                        \

+  setup_sprite_tile_add_blocks_4x(sub_tile_height);                            \

   vec_16x8u texels_wide;                                                       \

   u32 edge##_mask_bits_a = edge##_mask_bits & 0xFF;                            \

   u32 edge##_mask_bits_b = edge##_mask_bits >> 8;                              \

@@ -387,7 +387,7 @@ void setup_sprite_##texture_mode##_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y,   \
   u32 offset_u_right = width_rounded & 0xF;                                    \

                                                                                \

   u32 left_block_mask = ~(0xFFFFFFFF << (offset_u * 2));                       \

-  u32 right_block_mask = 0xFFFFFFFE << (offset_u_right * 2);                   \

+  u32 right_block_mask = 0xFFFFFFFC << (offset_u_right * 2);                   \

                                                                                \

   u32 left_mask_bits;                                                          \

   u32 right_mask_bits;                                                         \

@@ -404,7 +404,7 @@ void setup_sprite_##texture_mode##_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y,   \
   u32 texture_offset_base = texture_offset;                                    \

   u32 control_mask;                                                            \

                                                                                \

-  u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + (x - offset_u);               \

+  u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + (x - offset_u * 2);       \

   u32 num_blocks = psx_gpu->num_blocks;                                        \

   block_struct *block = psx_gpu->blocks + num_blocks;                          \

                                                                                \

@@ -416,7 +416,7 @@ void setup_sprite_##texture_mode##_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y,   \
   control_mask = tile_width == 1;                                              \

   control_mask |= (tile_height == 1) << 1;                                     \

   control_mask |= ((left_block_mask & 0xFFFF) == 0xFFFF) << 2;                 \

-  control_mask |= (((right_block_mask >> 8) & 0xFFFF) == 0xFFFF) << 3;         \

+  control_mask |= (((right_block_mask >> 16) & 0xFFFF) == 0xFFFF) << 3;        \

                                                                                \

   sprites_##texture_mode++;                                                    \

                                                                                \

@@ -514,13 +514,13 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
   u32 left_offset = u & 0x7;

   u32 width_rounded = width + left_offset + 7;

 

-  u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + (s32)(x - left_offset);

+  u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + (s32)(x - left_offset * 2);

   u32 right_width = width_rounded & 0x7;

   u32 block_width = width_rounded / 8;

-  u32 fb_ptr_pitch = (1024 + 8) - (block_width * 8);

+  u32 fb_ptr_pitch = (2048 + 16) - (block_width * 16);

 

   u32 left_mask_bits = ~(0xFFFF << (left_offset * 2));

-  u32 right_mask_bits = 0xFE << (right_width * 2);

+  u32 right_mask_bits = 0xFFFC << (right_width * 2);

 

   u32 texture_offset_base = u + (v * 1024);

   u32 texture_mask =

@@ -563,7 +563,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
 

       load_128b(texels, texture_block_ptr);

       

-      zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

+      zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

       block->texels = texels_wide;

       block->draw_mask_bits = mask_bits_a;

       block->fb_ptr = fb_ptr;          

@@ -574,7 +574,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       block->fb_ptr = fb_ptr + 1024;          

       block++;

       

-      zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

+      zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

       block->texels = texels_wide;

       block->draw_mask_bits = mask_bits_b;

       block->fb_ptr = fb_ptr + 8;

@@ -613,7 +613,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       if(num_blocks > MAX_BLOCKS)

       {

         flush_render_block_buffer(psx_gpu);

-        num_blocks = block_width;

+        num_blocks = block_width * 4;

         block = psx_gpu->blocks;

       }

 

@@ -624,7 +624,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       

       load_128b(texels, texture_block_ptr);

 

-      zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

+      zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

       block->texels = texels_wide;

       block->draw_mask_bits = left_mask_bits_a;

       block->fb_ptr = fb_ptr;

@@ -635,7 +635,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       block->fb_ptr = fb_ptr + 1024;

       block++;      

 

-      zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

+      zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

       block->texels = texels_wide;

       block->draw_mask_bits = left_mask_bits_b;

       block->fb_ptr = fb_ptr + 8;

@@ -654,7 +654,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
         texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);

         load_128b(texels, texture_block_ptr);

 

-        zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

+        zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

         block->texels = texels_wide;

         block->draw_mask_bits = 0;

         block->fb_ptr = fb_ptr;

@@ -665,7 +665,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
         block->fb_ptr = fb_ptr + 1024;

         block++;      

 

-        zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

+        zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

         block->texels = texels_wide;

         block->draw_mask_bits = 0;

         block->fb_ptr = fb_ptr + 8;

@@ -677,7 +677,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
         block++;

         

         texture_offset += 8;

-        fb_ptr += 8;

+        fb_ptr += 16;

 

         blocks_remaining--;

       }

@@ -685,7 +685,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);

       load_128b(texels, texture_block_ptr);

       

-      zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

+      zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);

       block->texels = texels_wide;

       block->draw_mask_bits = right_mask_bits_a;

       block->fb_ptr = fb_ptr;

@@ -696,7 +696,7 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       block->fb_ptr = fb_ptr + 1024;

       block++;      

 

-      zip4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

+      zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.high, texels.high);

       block->texels = texels_wide;

       block->draw_mask_bits = right_mask_bits_b;

       block->fb_ptr = fb_ptr + 8;

@@ -717,11 +717,17 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
 

 #endif

 

+static void setup_sprite_untextured_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y,

+ s32 u, s32 v, s32 width, s32 height, u32 color)

+{

+  setup_sprite_untextured(psx_gpu, x, y, u, v, width * 2, height * 2, color);

+}

+

 #define setup_sprite_blocks_switch_textured_4x(texture_mode)                   \

   setup_sprite_##texture_mode##_4x                                             \

 

 #define setup_sprite_blocks_switch_untextured_4x(texture_mode)                 \

-  setup_sprite_untextured                                                      \

+  setup_sprite_untextured_4x                                                   \

 

 #define setup_sprite_blocks_switch_4x(texturing, texture_mode)                 \

   setup_sprite_blocks_switch_##texturing##_4x(texture_mode)                    \

@@ -802,9 +808,6 @@ render_block_handler_struct render_sprite_block_handlers_4x[] =
 void render_sprite_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, u32 u, u32 v,

  s32 width, s32 height, u32 flags, u32 color)

 {

-  x *= 2;

-  y *= 2;

-

   s32 x_right = x + width - 1;

   s32 y_bottom = y + height - 1;

 

@@ -837,6 +840,9 @@ void render_sprite_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, u32 u, u32 v,
   if((width <= 0) || (height <= 0))

     return;

 

+  x *= 2;

+  y *= 2;

+

 #ifdef PROFILE

   span_pixels += width * height;

   spans += height;

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c b/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
index f920c73..4260bc7 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_parse.c
@@ -760,6 +760,8 @@ breakloop:
   return list - list_start;
 }
 
+#ifdef PCSX
+
 static void *select_enhancement_buf_ptr(psx_gpu_struct *psx_gpu, u32 x)
 {
   u32 b;
@@ -864,6 +866,8 @@ static void do_quad_enhanced(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
   }
 }
 
+#if 0
+
 #define fill_vertex(i, x_, y_, u_, v_, rgb_) \
   vertexes[i].x = x_; \
   vertexes[i].y = y_; \
@@ -873,7 +877,7 @@ static void do_quad_enhanced(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
   vertexes[i].g = (rgb_) >> 8; \
   vertexes[i].b = (rgb_) >> 16
 
-static void do_esprite_in_triangles(psx_gpu_struct *psx_gpu, int x, int y,
+static void do_sprite_enhanced(psx_gpu_struct *psx_gpu, int x, int y,
  u32 u, u32 v, u32 w, u32 h, u32 cmd_rgb)
 {
   vertex_struct *vertex_ptrs[3];
@@ -928,6 +932,17 @@ static void do_esprite_in_triangles(psx_gpu_struct *psx_gpu, int x, int y,
 
   psx_gpu->render_state_base = render_state_base_saved;
 }
+#else
+static void do_sprite_enhanced(psx_gpu_struct *psx_gpu, int x, int y,
+ u32 u, u32 v, u32 w, u32 h, u32 cmd_rgb)
+{
+  u32 flags = (cmd_rgb >> 24);
+  u32 color = cmd_rgb & 0xffffff;
+
+  psx_gpu->vram_out_ptr = psx_gpu->enhancement_current_buf_ptr;
+  render_sprite_4x(psx_gpu, x, y, u, v, w, h, flags, color);
+}
+#endif
 
 u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_command)
 {
@@ -1216,7 +1231,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
         u32 height = list_s16[5] & 0x1FF;
 
         render_sprite(psx_gpu, x, y, 0, 0, width, height, current_command, list[0]);
-        do_esprite_in_triangles(psx_gpu, x, y, 0, 0, width, height, list[0]);
+        do_sprite_enhanced(psx_gpu, x, y, 0, 0, width, height, list[0]);
         break;
       }
   
@@ -1233,7 +1248,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
 
         render_sprite(psx_gpu, x, y, u, v, width, height,
          current_command, list[0]);
-        do_esprite_in_triangles(psx_gpu, x, y, u, v, width, height, list[0]);
+        do_sprite_enhanced(psx_gpu, x, y, u, v, width, height, list[0]);
         break;
       }
   
@@ -1246,7 +1261,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
         s32 y = sign_extend_11bit(list_s16[3] + psx_gpu->offset_y);
 
         render_sprite(psx_gpu, x, y, 0, 0, 1, 1, current_command, list[0]);
-        do_esprite_in_triangles(psx_gpu, x, y, 0, 0, 1, 1, list[0]);
+        do_sprite_enhanced(psx_gpu, x, y, 0, 0, 1, 1, list[0]);
         break;
       }
   
@@ -1259,7 +1274,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
         s32 y = sign_extend_11bit(list_s16[3] + psx_gpu->offset_y);
 
         render_sprite(psx_gpu, x, y, 0, 0, 8, 8, current_command, list[0]);
-        do_esprite_in_triangles(psx_gpu, x, y, 0, 0, 8, 8, list[0]);
+        do_sprite_enhanced(psx_gpu, x, y, 0, 0, 8, 8, list[0]);
         break;
       }
   
@@ -1277,7 +1292,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
 
         render_sprite(psx_gpu, x, y, u, v, 8, 8,
          current_command, list[0]);
-        do_esprite_in_triangles(psx_gpu, x, y, u, v, 8, 8, list[0]);
+        do_sprite_enhanced(psx_gpu, x, y, u, v, 8, 8, list[0]);
         break;
       }
   
@@ -1290,7 +1305,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
         s32 y = sign_extend_11bit(list_s16[3] + psx_gpu->offset_y);
 
         render_sprite(psx_gpu, x, y, 0, 0, 16, 16, current_command, list[0]);
-        do_esprite_in_triangles(psx_gpu, x, y, 0, 0, 16, 16, list[0]);
+        do_sprite_enhanced(psx_gpu, x, y, 0, 0, 16, 16, list[0]);
         break;
       }
   
@@ -1307,7 +1322,7 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
         set_clut(psx_gpu, list_s16[5]);
 
         render_sprite(psx_gpu, x, y, u, v, 16, 16, current_command, list[0]);
-        do_esprite_in_triangles(psx_gpu, x, y, u, v, 16, 16, list[0]);
+        do_sprite_enhanced(psx_gpu, x, y, u, v, 16, 16, list[0]);
         break;
       }
   
@@ -1454,13 +1469,14 @@ u32 gpu_parse_enhanced(psx_gpu_struct *psx_gpu, u32 *list, u32 size, u32 *last_c
     }
   }
 
-#ifdef PCSX
+  enhancement_disable();
+
 breakloop:
-#endif
-enhancement_disable();
   if (last_command != NULL)
     *last_command = current_command;
   return list - list_start;
 }
 
+#endif /* PCSX */
+
 // vim:shiftwidth=2:expandtab
diff --git a/plugins/gpu_neon/psx_gpu/vector_ops.h b/plugins/gpu_neon/psx_gpu/vector_ops.h
index b58b190..c91e7d9 100644
--- a/plugins/gpu_neon/psx_gpu/vector_ops.h
+++ b/plugins/gpu_neon/psx_gpu/vector_ops.h
@@ -396,7 +396,7 @@ build_vector_types(s);
 
 #define zip_4x32b(dest, source_a, source_b)                                    \
   foreach_element(4, (dest).e[_i] =                                            \
-   (u8)(source_a).e[_i] | ((u8)(source_b).e[_i] << 16))                        \
+   (u16)(source_a).e[_i] | ((u16)(source_b).e[_i] << 16))                      \
 
 #define zip_2x64b(dest, source_a, source_b)                                    \
   foreach_element(2, (dest).e[_i] =                                            \