X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=plugins%2Fgpu_neon%2Fpsx_gpu%2Fpsx_gpu_4x.c;h=d7ec340959439937f86232a152c1b9bac1207405;hb=0b4038f8edd327a3a9a2fbdefbc25ece921bc2ab;hp=83c6680f4fa97f30e125ebe9b7acb15d333050fa;hpb=7956599fa5f666016f71870d9889748c97839041;p=pcsx_rearmed.git

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c b/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
index 83c6680f..d7ec3409 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu_4x.c
@@ -1,8 +1,26 @@
+#define select_enhancement_buf_index(psx_gpu, x) \
+  ((psx_gpu)->enhancement_buf_by_x16[(u32)(x) / \
+    (1024u / sizeof((psx_gpu)->enhancement_buf_by_x16))])
+
 #define select_enhancement_buf_ptr(psx_gpu, x) \
   ((psx_gpu)->enhancement_buf_ptr + \
-   ((psx_gpu)->enhancement_buf_by_x16[(x) / 16] << 20))
+    (select_enhancement_buf_index(psx_gpu, x) << 20))
+
+#if !defined(NEON_BUILD) || defined(SIMD_BUILD)
+
+#ifndef zip_4x32b
+
+#define vector_cast(vec_to, source) source
+
+#define zip_4x32b(dest, source_a, source_b) {                                  \
+  u32 _i; for(_i = 0; _i < 4; _i++) {                                          \
+    (dest).e[_i * 2 + 0] = (source_a).e[_i];                                   \
+    (dest).e[_i * 2 + 1] = (source_b).e[_i];                                   \
+  }                                                                            \
+}
+
+#endif
 
-#ifndef NEON_BUILD
 void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
  s32 v, s32 width, s32 height, u32 color)
 {
@@ -56,7 +74,8 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       texture_block_ptr =
        texture_page_ptr + (texture_offset_base & texture_mask);
 
-      load_128b(texels, texture_block_ptr);
+      //load_128b(texels, texture_block_ptr);
+      texels = *(vec_8x16u *)texture_block_ptr;
       
       zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);
       block->texels = texels_wide;
@@ -117,7 +136,8 @@ 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);
+      //load_128b(texels, texture_block_ptr);
+      texels = *(vec_8x16u *)texture_block_ptr;
 
       zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);
       block->texels = texels_wide;
@@ -147,7 +167,8 @@ void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
       while(blocks_remaining)
       {
         texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);
-        load_128b(texels, texture_block_ptr);
+        //load_128b(texels, texture_block_ptr);
+        texels = *(vec_8x16u *)texture_block_ptr;
 
         zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);
         block->texels = texels_wide;
@@ -178,7 +199,8 @@ 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);
+      //load_128b(texels, texture_block_ptr);
+      texels = *(vec_8x16u *)texture_block_ptr;
       
       zip_4x32b(vector_cast(vec_4x32u, texels_wide), texels.low, texels.low);
       block->texels = texels_wide;
@@ -291,7 +313,7 @@ static void setup_sprite_untextured_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y,
   render_sprite_blocks_switch_block_texture_mode_4x(4bpp),                     \
   render_sprite_blocks_switch_block_texture_mode_4x(8bpp),                     \
   render_sprite_blocks_switch_block_texture_mode_4x(16bpp),                    \
-  render_sprite_blocks_switch_block_texture_mode_4x(4bpp)                      \
+  render_sprite_blocks_switch_block_texture_mode_4x(16bpp)                     \
 
 
 render_block_handler_struct render_sprite_block_handlers_4x[] =