psx_gpu: increase reciprocal accuracy

[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 9e84ace..98aacc3 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);
@@ -802,7 +854,7 @@ void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
                                                                                \
   dup_2x32b(edge_shifts, edge_shift);                                          \
   sub_2x32b(heights_b, heights, c_0x01);                                       \
-  shr_2x32b(height_reciprocals, edge_shifts, 12);                              \
+  shr_2x32b(height_reciprocals, edge_shifts, 10);                              \
                                                                                \
   mla_2x32b(heights_b, x_starts, heights);                                     \
   bic_immediate_4x16b(vector_cast(vec_4x16u, edge_shifts), 0xE0);              \
@@ -831,8 +883,8 @@ void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
   sub_2x32b(widths, x_ends, x_starts);                                         \
   width_alt = x_c - start_c;                                                   \
                                                                                \
-  shr_2x32b(height_reciprocals, edge_shifts, 12);                              \
-  height_reciprocal_alt = edge_shift_alt >> 12;                                \
+  shr_2x32b(height_reciprocals, edge_shifts, 10);                              \
+  height_reciprocal_alt = edge_shift_alt >> 10;                                \
                                                                                \
   bic_immediate_4x16b(vector_cast(vec_4x16u, edge_shifts), 0xE0);              \
   edge_shift_alt &= 0x1F;                                                      \
@@ -2918,7 +2970,7 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
       vertex_swap(a, b);
   }
 
-  if((c->x - a->x) >= 1024)
+  if((c->x - psx_gpu->offset_x) >= 1024 || (c->x - a->x) >= 1024)
   {
 #ifdef PROFILE
     trivial_rejects++;
@@ -4474,12 +4526,12 @@ void initialize_reciprocal_table(void)
   {
     shift = __builtin_clz(height);
     height_normalized = height << shift;
-    height_reciprocal = ((1ULL << 50) + (height_normalized - 1)) /
+    height_reciprocal = ((1ULL << 52) + (height_normalized - 1)) /
      height_normalized;
 
-    shift = 32 - (50 - shift);
+    shift = 32 - (52 - shift);
 
-    reciprocal_table[height] = (height_reciprocal << 12) | shift;
+    reciprocal_table[height] = (height_reciprocal << 10) | shift;
   }
 }