X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=plugins%2Fgpu_neon%2Fpsx_gpu%2Fpsx_gpu.c;h=a79254da9f17bd496c993364d1157e3892e1abd6;hb=37725e8cc9157e2e7819538ee5c98279b8dbefff;hp=ce72af55f54b2db7e71a74ea95334a88d164d27a;hpb=c111e8f8fb8a0d3bd7b05c743a48d942e107cc79;p=pcsx_rearmed.git

diff --git a/plugins/gpu_neon/psx_gpu/psx_gpu.c b/plugins/gpu_neon/psx_gpu/psx_gpu.c
index ce72af55..a79254da 100644
--- a/plugins/gpu_neon/psx_gpu/psx_gpu.c
+++ b/plugins/gpu_neon/psx_gpu/psx_gpu.c
@@ -17,6 +17,10 @@
 #include <string.h>
 
 #include "common.h"
+#ifndef NEON_BUILD
+#include "vector_ops.h"
+#endif
+#include "psx_gpu_simd.h"
 
 u32 span_pixels = 0;
 u32 span_pixel_blocks = 0;
@@ -298,9 +302,6 @@ void update_texture_cache_region(psx_gpu_struct *psx_gpu, u32 x1, u32 y1,
   }
 }
 
-void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
- u32 texture_page);
-
 #ifndef NEON_BUILD
 
 void update_texture_4bpp_cache(psx_gpu_struct *psx_gpu)
@@ -449,9 +450,6 @@ void update_texture_8bpp_cache(psx_gpu_struct *psx_gpu)
   }
 }
 
-void setup_blocks_shaded_untextured_undithered_unswizzled_indirect(
- psx_gpu_struct *psx_gpu);
-
 void flush_render_block_buffer(psx_gpu_struct *psx_gpu)
 {
   if((psx_gpu->render_mode & RENDER_INTERLACE_ENABLED) &&
@@ -515,9 +513,6 @@ void flush_render_block_buffer(psx_gpu_struct *psx_gpu)
 }
 
 
-void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
- vertex_struct *b, vertex_struct *c);
-
 #ifndef NEON_BUILD
 
 #define setup_gradient_calculation_input(set, vertex)                          \
@@ -769,13 +764,18 @@ void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
 
 #ifndef NDEBUG
 #define setup_spans_debug_check(span_edge_data_element)                        \
-  if (&span_edge_data_element - psx_gpu->span_edge_data < psx_gpu->num_spans)  \
+{                                                                              \
+  u32 _num_spans = &span_edge_data_element - psx_gpu->span_edge_data;          \
+  if (_num_spans > MAX_SPANS)                                                  \
+    *(int *)0 = 1;                                                             \
+  if (_num_spans < psx_gpu->num_spans)                                         \
   {                                                                            \
     if(span_edge_data_element.num_blocks > MAX_BLOCKS_PER_ROW)                 \
       *(int *)0 = 1;                                                           \
     if(span_edge_data_element.y > 2048)                                        \
       *(int *)0 = 1;                                                           \
   }                                                                            \
+}                                                                              \
 
 #else
 #define setup_spans_debug_check(span_edge_data_element)                        \
@@ -1200,26 +1200,6 @@ void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
   setup_spans_up(index_##major, index_##minor, minor, yes)                     \
 
 
-void setup_spans_up_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_up_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_down_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_down_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_up_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_up_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_down_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_down_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-void setup_spans_up_down(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
- vertex_struct *v_b, vertex_struct *v_c);
-
-
 #ifndef NEON_BUILD
 
 void setup_spans_up_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
@@ -1423,12 +1403,16 @@ void setup_spans_up_down(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
     y_x4.e[3] = y_a + 3;
     setup_spans_adjust_edges_alternate_no(index_left, index_right);
 
+    // FIXME: overflow corner case
+    if(psx_gpu->num_spans + height_minor_b == MAX_SPANS)
+      height_minor_b &= ~3;
+
     psx_gpu->num_spans += height_minor_b;
-    do
+    while(height_minor_b > 0)
     {
       setup_spans_set_x4(none, down, no);
       height_minor_b -= 4;
-    } while(height_minor_b > 0);
+    }
   }
 
   left_split_triangles++;
@@ -1932,30 +1916,6 @@ void setup_blocks_##shading##_##texturing##_##dithering##_##sw##_##target(     \
   psx_gpu->num_blocks = num_blocks;                                            \
 }                                                                              \
 
-void setup_blocks_shaded_textured_dithered_unswizzled_indirect(psx_gpu_struct
- *psx_gpu);
-
-void setup_blocks_shaded_untextured_dithered_unswizzled_indirect(psx_gpu_struct
- *psx_gpu);
-void setup_blocks_shaded_untextured_undithered_unswizzled_indirect(
- psx_gpu_struct *psx_gpu);
-void setup_blocks_shaded_untextured_dithered_unswizzled_direct(psx_gpu_struct
- *psx_gpu);
-void setup_blocks_shaded_untextured_undithered_unswizzled_direct(
- psx_gpu_struct *psx_gpu);
-
-void setup_blocks_unshaded_textured_dithered_unswizzled_indirect(psx_gpu_struct
- *psx_gpu);
-void setup_blocks_unshaded_untextured_undithered_unswizzled_indirect(
- psx_gpu_struct *psx_gpu);
-void setup_blocks_unshaded_untextured_undithered_unswizzled_direct(
- psx_gpu_struct *psx_gpu);
-
-void setup_blocks_shaded_textured_dithered_swizzled_indirect(psx_gpu_struct
- *psx_gpu);
-void setup_blocks_unshaded_textured_dithered_swizzled_indirect(psx_gpu_struct
- *psx_gpu);
-
 
 //setup_blocks_builder(unshaded, untextured, undithered, unswizzled, direct);
 
@@ -1975,15 +1935,6 @@ setup_blocks_builder(shaded, untextured, dithered, unswizzled, direct);
 setup_blocks_builder(unshaded, untextured, undithered, unswizzled, indirect);
 setup_blocks_builder(unshaded, untextured, undithered, unswizzled, direct);
 
-#endif
-
-void texture_blocks_untextured(psx_gpu_struct *psx_gpu);
-void texture_blocks_4bpp(psx_gpu_struct *psx_gpu);
-void texture_blocks_8bpp(psx_gpu_struct *psx_gpu);
-void texture_blocks_16bpp(psx_gpu_struct *psx_gpu);
-
-#ifndef NEON_BUILD
-
 void texture_blocks_untextured(psx_gpu_struct *psx_gpu)
 {
   if(psx_gpu->primitive_type != PRIMITIVE_TYPE_SPRITE)
@@ -2268,27 +2219,6 @@ void shade_blocks_##shading##_textured_modulated_##dithering##_##target(       \
   }                                                                            \
 }                                                                              \
 
-void shade_blocks_shaded_textured_modulated_dithered_direct(psx_gpu_struct
- *psx_gpu);
-void shade_blocks_shaded_textured_modulated_undithered_direct(psx_gpu_struct
- *psx_gpu);
-void shade_blocks_unshaded_textured_modulated_dithered_direct(psx_gpu_struct
- *psx_gpu);
-void shade_blocks_unshaded_textured_modulated_undithered_direct(psx_gpu_struct
- *psx_gpu);
-
-void shade_blocks_shaded_textured_modulated_dithered_indirect(psx_gpu_struct
- *psx_gpu);
-void shade_blocks_shaded_textured_modulated_undithered_indirect(psx_gpu_struct
- *psx_gpu);
-void shade_blocks_unshaded_textured_modulated_dithered_indirect(psx_gpu_struct
- *psx_gpu);
-void shade_blocks_unshaded_textured_modulated_undithered_indirect(psx_gpu_struct
- *psx_gpu);
-
-void shade_blocks_textured_unmodulated_indirect(psx_gpu_struct *psx_gpu);
-void shade_blocks_textured_unmodulated_direct(psx_gpu_struct *psx_gpu);
-
 #ifndef NEON_BUILD
 
 shade_blocks_textured_modulated_builder(shaded, dithered, direct);
@@ -2374,14 +2304,6 @@ void shade_blocks_textured_unmodulated_dithered_##target(psx_gpu_struct        \
 shade_blocks_textured_unmodulated_builder(indirect)
 shade_blocks_textured_unmodulated_builder(direct)
 
-#endif
-
-
-void shade_blocks_unshaded_untextured_indirect(psx_gpu_struct *psx_gpu);
-void shade_blocks_unshaded_untextured_direct(psx_gpu_struct *psx_gpu);
-
-#ifndef NEON_BUILD
-                                                                               
 void shade_blocks_unshaded_untextured_indirect(psx_gpu_struct *psx_gpu)
 {
 }
@@ -2593,27 +2515,6 @@ void                                                                           \
   }                                                                            \
 }                                                                              \
 
-void blend_blocks_textured_average_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_average_on(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_add_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_add_on(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_subtract_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_subtract_on(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_add_fourth_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_add_fourth_on(psx_gpu_struct *psx_gpu);
-
-void blend_blocks_untextured_average_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_untextured_average_on(psx_gpu_struct *psx_gpu);
-void blend_blocks_untextured_add_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_untextured_add_on(psx_gpu_struct *psx_gpu);
-void blend_blocks_untextured_subtract_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_untextured_subtract_on(psx_gpu_struct *psx_gpu);
-void blend_blocks_untextured_add_fourth_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_untextured_add_fourth_on(psx_gpu_struct *psx_gpu);
-
-void blend_blocks_textured_unblended_off(psx_gpu_struct *psx_gpu);
-void blend_blocks_textured_unblended_on(psx_gpu_struct *psx_gpu);
-
 #ifndef NEON_BUILD
 
 void blend_blocks_textured_unblended_off(psx_gpu_struct *psx_gpu)
@@ -2876,7 +2777,7 @@ blend_blocks_builder(textured, unblended, on);
   render_blocks_switch_block_texture_mode(4bpp),                               \
   render_blocks_switch_block_texture_mode(8bpp),                               \
   render_blocks_switch_block_texture_mode(16bpp),                              \
-  render_blocks_switch_block_texture_mode(4bpp)                                \
+  render_blocks_switch_block_texture_mode(16bpp)                               \
 
 
 render_block_handler_struct render_triangle_block_handlers[] =
@@ -3166,9 +3067,6 @@ void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
     render_triangle_p(psx_gpu, vertex_ptrs, flags);
 }
 
-
-void texture_sprite_blocks_8bpp(psx_gpu_struct *psx_gpu);
-
 #ifndef NEON_BUILD
 
 void texture_sprite_blocks_8bpp(psx_gpu_struct *psx_gpu)
@@ -3871,20 +3769,6 @@ void setup_sprite_##texture_mode##x4mode(psx_gpu_struct *psx_gpu, s32 x, s32 y,\
   }                                                                            \
 }                                                                              \
 
-void setup_sprite_4bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
- s32 width, s32 height, u32 color);
-void setup_sprite_8bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
- s32 width, s32 height, u32 color);
-void setup_sprite_16bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
- s32 width, s32 height, u32 color);
-
-void setup_sprite_4bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
- s32 width, s32 height, u32 color);
-void setup_sprite_8bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
- s32 width, s32 height, u32 color);
-void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
- s32 width, s32 height, u32 color);
-
 #ifndef NEON_BUILD
 setup_sprite_tiled_builder(4bpp,);
 setup_sprite_tiled_builder(8bpp,);
@@ -4013,11 +3897,17 @@ void setup_sprite_16bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
   }
 }
 
-#endif
-
 void setup_sprite_untextured(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
  s32 v, s32 width, s32 height, u32 color)
 {
+  if((psx_gpu->render_state & (RENDER_STATE_MASK_EVALUATE |
+   RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND)) == 0 &&
+   (psx_gpu->render_mode & RENDER_INTERLACE_ENABLED) == 0)
+  {
+    setup_sprite_untextured_simple(psx_gpu, x, y, u, v, width, height, color);
+    return;
+  }
+
   u32 right_width = ((width - 1) & 0x7) + 1;
   u32 right_mask_bits = (0xFF << right_width);
   u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + x;
@@ -4083,6 +3973,66 @@ void setup_sprite_untextured(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
   }
 }
 
+#endif
+
+void setup_sprite_untextured_simple(psx_gpu_struct *psx_gpu, s32 x, s32 y,
+ s32 u, s32 v, s32 width, s32 height, u32 color)
+{
+  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);
+
+  u16 *vram_ptr16 = psx_gpu->vram_out_ptr + x + (y * 1024);
+  u32 *vram_ptr;
+
+  u32 num_width;
+
+  if(psx_gpu->num_blocks > MAX_BLOCKS)
+  {
+    flush_render_block_buffer(psx_gpu);
+  }
+
+  while(height)
+  {
+    num_width = width;
+
+    vram_ptr = (void *)vram_ptr16;
+    if((long)vram_ptr16 & 2)
+    {
+      *vram_ptr16 = color_32bpp;
+      vram_ptr = (void *)(vram_ptr16 + 1);
+      num_width--;
+    }
+
+    while(num_width >= 4 * 2)
+    {
+      vram_ptr[0] = color_32bpp;
+      vram_ptr[1] = color_32bpp;
+      vram_ptr[2] = color_32bpp;
+      vram_ptr[3] = color_32bpp;
+
+      vram_ptr += 4;
+      num_width -= 4 * 2;
+    }
+
+    while(num_width >= 2)
+    {
+      *vram_ptr++ = color_32bpp;
+      num_width -= 2;
+    }
+
+    if(num_width > 0)
+    {
+      *(u16 *)vram_ptr = color_32bpp;
+    }
+
+    vram_ptr16 += 1024;
+    height--;
+  }
+}
 
 
 #define setup_sprite_blocks_switch_textured(texture_mode)                      \
@@ -4171,7 +4121,7 @@ void setup_sprite_untextured(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
   render_sprite_blocks_switch_block_texture_mode(4bpp),                        \
   render_sprite_blocks_switch_block_texture_mode(8bpp),                        \
   render_sprite_blocks_switch_block_texture_mode(16bpp),                       \
-  render_sprite_blocks_switch_block_texture_mode(4bpp)                         \
+  render_sprite_blocks_switch_block_texture_mode(16bpp)                        \
 
 
 render_block_handler_struct render_sprite_block_handlers[] =
@@ -4264,7 +4214,7 @@ void render_sprite(psx_gpu_struct *psx_gpu, s32 x, s32 y, u32 u, u32 v,
 }
 
 #define draw_pixel_line_mask_evaluate_yes()                                    \
-  if(*vram_ptr & 0x8000)                                                       \
+  if((*vram_ptr & 0x8000) == 0)                                                \
 
 #define draw_pixel_line_mask_evaluate_no()                                     \
     
@@ -4976,6 +4926,7 @@ void initialize_psx_gpu(psx_gpu_struct *psx_gpu, u16 *vram)
   memset(psx_gpu->vram_ptr, 0, sizeof(u16) * 1024 * 512);
 
   initialize_reciprocal_table();
+  psx_gpu->reciprocal_table_ptr = reciprocal_table;
 
   //    00 01 10 11
   // 00  0  4  1  5
@@ -5006,7 +4957,7 @@ u64 get_us(void)
   return (tv.tv_sec * 1000000ULL) + tv.tv_usec;
 }
 
-#ifdef NEON_BUILD
+#if 0 //def NEON_BUILD
 
 u32 get_counter()
 {