plugins/gpu_neon/psx_gpu/psx_gpu.c

   1 /*
   2  * Copyright (C) 2011 Gilead Kutnick "Exophase" <exophase@gmail.com>
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public License as
   6  * published by the Free Software Foundation; either version 2 of
   7  * the License, or (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * General Public License for more details.
  13  */
  14
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18
  19 #include "common.h"
  20
  21 u32 span_pixels = 0;
  22 u32 span_pixel_blocks = 0;
  23 u32 spans = 0;
  24 u32 triangles = 0;
  25 u32 sprites = 0;
  26 u32 sprites_4bpp = 0;
  27 u32 sprites_8bpp = 0;
  28 u32 sprites_16bpp = 0;
  29 u32 sprite_blocks = 0;
  30 u32 sprites_untextured = 0;
  31 u32 lines = 0;
  32 u32 trivial_rejects = 0;
  33 u32 texels_4bpp = 0;
  34 u32 texels_8bpp = 0;
  35 u32 texels_16bpp = 0;
  36 u32 texel_blocks_4bpp = 0;
  37 u32 texel_blocks_8bpp = 0;
  38 u32 texel_blocks_16bpp = 0;
  39 u32 texel_blocks_untextured = 0;
  40 u32 blend_blocks = 0;
  41 u32 render_buffer_flushes = 0;
  42 u32 state_changes = 0;
  43 u32 left_split_triangles = 0;
  44 u32 flat_triangles = 0;
  45 u32 clipped_triangles = 0;
  46 u32 zero_block_spans = 0;
  47 u32 texture_cache_loads = 0;
  48 u32 false_modulated_blocks = 0;
  49
  50 /* double size for enhancement */
  51 u32 reciprocal_table[512 * 2];
  52
  53
  54 typedef s32 fixed_type;
  55
  56 #define EDGE_STEP_BITS 32
  57 #define FIXED_BITS     12
  58
  59 #define fixed_center(value)                                                    \
  60   ((((fixed_type)(value)) << FIXED_BITS) + (1 << (FIXED_BITS - 1)))            \
  61
  62 #define int_to_fixed(value)                                                    \
  63   (((fixed_type)(value)) << FIXED_BITS)                                        \
  64
  65 #define fixed_to_int(value)                                                    \
  66   ((value) >> FIXED_BITS)                                                      \
  67
  68 #define fixed_to_double(value)                                                 \
  69   ((value) / (double)(1 << FIXED_BITS))                                        \
  70
  71 #define double_to_fixed(value)                                                 \
  72   (fixed_type)(((value) * (double)(1 << FIXED_BITS)))                          \
  73
  74 typedef void (setup_blocks_function_type)(psx_gpu_struct *psx_gpu);
  75 typedef void (texture_blocks_function_type)(psx_gpu_struct *psx_gpu);
  76 typedef void (shade_blocks_function_type)(psx_gpu_struct *psx_gpu);
  77 typedef void (blend_blocks_function_type)(psx_gpu_struct *psx_gpu);
  78
  79 typedef void (setup_sprite_function_type)(psx_gpu_struct *psx_gpu, s32 x,
  80  s32 y, s32 u, s32 v, s32 width, s32 height, u32 color);
  81
  82 struct render_block_handler_struct
  83 {
  84   void *setup_blocks;
  85   texture_blocks_function_type *texture_blocks;
  86   shade_blocks_function_type *shade_blocks;
  87   blend_blocks_function_type *blend_blocks;
  88 };
  89
  90 #ifndef NEON_BUILD
  91
  92 u32 fixed_reciprocal(u32 denominator, u32 *_shift)
  93 {
  94   u32 shift = __builtin_clz(denominator);
  95   u32 denominator_normalized = denominator << shift;
  96
  97   double numerator = (1ULL << 62) + denominator_normalized;
  98   double numerator_b;
  99
 100   double denominator_normalized_dp_b;
 101   u64 denominator_normalized_dp_u64;
 102
 103   u32 reciprocal;
 104   double reciprocal_dp;
 105
 106   u64 numerator_u64 = (denominator_normalized >> 10) |
 107    ((u64)(62 + 1023) << 52);
 108   *((u64 *)(&numerator_b)) = numerator_u64;
 109
 110   denominator_normalized_dp_u64 =
 111    (u64)(denominator_normalized << 21) |
 112    ((u64)((denominator_normalized >> 11) + ((1022 + 31) << 20)) << 32);
 113   *((u64 *)(&denominator_normalized_dp_b)) = denominator_normalized_dp_u64;
 114
 115   // Implement with a DP divide
 116   reciprocal_dp = numerator / denominator_normalized_dp_b;
 117   reciprocal = reciprocal_dp;
 118
 119   if(reciprocal == 0x80000001)
 120     reciprocal = 0x80000000;
 121
 122   *_shift = 62 - shift;
 123   return reciprocal;
 124 }
 125
 126 double reciprocal_estimate(double a)
 127 {
 128   int q, s;
 129   double r;
 130
 131   q = (int)(a * 512.0);
 132   /* a in units of 1/512 rounded down */
 133   r = 1.0 / (((double)q + 0.5) / 512.0); /* reciprocal r */
 134   s = (int)(256.0 * r + 0.5);
 135
 136   /* r in units of 1/256 rounded to nearest */
 137
 138   return (double)s / 256.0;
 139 }
 140
 141 u32 reciprocal_estimate_u32(u32 value)
 142 {
 143   u64 dp_value_u64;
 144   volatile double dp_value;
 145   volatile u64 *dp_value_ptr = (volatile u64 *)&dp_value;
 146
 147   if((value >> 31) == 0)
 148     return 0xFFFFFFFF;
 149
 150   dp_value_u64 = (0x3FEULL << (31 + 21)) | ((u64)(value & 0x7FFFFFFF) << 21);
 151
 152   *dp_value_ptr = dp_value_u64;
 153
 154   dp_value = reciprocal_estimate(dp_value);
 155   dp_value_u64 = *dp_value_ptr;
 156
 157   return (0x80000000 | ((dp_value_u64 >> 21) & 0x7FFFFFFF));
 158 }
 159
 160 u32 fixed_reciprocal_nr(u32 value, u32 *_shift)
 161 {
 162   u32 shift = __builtin_clz(value);
 163   u32 value_normalized = value << shift;
 164
 165   *_shift = 62 - shift;
 166
 167   value_normalized -= 2;
 168
 169   u32 reciprocal_normalized = reciprocal_estimate_u32(value_normalized) >> 1;
 170
 171   u32 temp = -(((u64)value_normalized * (u32)reciprocal_normalized) >> 31);
 172   reciprocal_normalized = (((u64)reciprocal_normalized * temp) >> 31);
 173   temp = -(((u64)value_normalized * (u32)reciprocal_normalized) >> 31);
 174   reciprocal_normalized = (((u64)reciprocal_normalized * temp) >> 31);
 175   temp = -(((u64)value_normalized * (u32)reciprocal_normalized) >> 31);
 176   reciprocal_normalized = (((u64)reciprocal_normalized * temp) >> 31);
 177
 178   return reciprocal_normalized;
 179 }
 180
 181 #endif
 182
 183
 184 s32 triangle_signed_area_x2(s32 x0, s32 y0, s32 x1, s32 y1, s32 x2, s32 y2)
 185 {
 186         return ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0));
 187 }
 188
 189 u32 texture_region_mask(s32 x1, s32 y1, s32 x2, s32 y2)
 190 {
 191   s32 coverage_x, coverage_y;
 192
 193   u32 mask_up_left;
 194   u32 mask_down_right;
 195
 196   coverage_x = x2 >> 6;
 197   coverage_y = y2 >> 8;
 198
 199   if(coverage_x < 0)
 200     coverage_x = 0;
 201
 202   if(coverage_x > 31)
 203     coverage_x = 31;
 204
 205   mask_down_right = ~(0xFFFFFFFF << (coverage_x + 1)) & 0xFFFF;
 206
 207   if(coverage_y >= 1)
 208     mask_down_right |= mask_down_right << 16;
 209
 210   coverage_x = x1 >> 6;
 211
 212   mask_up_left = 0xFFFF0000 << coverage_x;
 213   if(coverage_x < 0)
 214     mask_up_left = 0xFFFF0000;
 215
 216   coverage_y = y1 >> 8;
 217   if(coverage_y <= 0)
 218     mask_up_left |= mask_up_left >> 16;
 219
 220   return mask_up_left & mask_down_right;
 221 }
 222
 223 u32 invalidate_texture_cache_region(psx_gpu_struct *psx_gpu, u32 x1, u32 y1,
 224  u32 x2, u32 y2)
 225 {
 226   u32 mask = texture_region_mask(x1, y1, x2, y2);
 227
 228   psx_gpu->dirty_textures_4bpp_mask |= mask;
 229   psx_gpu->dirty_textures_8bpp_mask |= mask;
 230   psx_gpu->dirty_textures_8bpp_alternate_mask |= mask;
 231
 232   return mask;
 233 }
 234
 235 u32 invalidate_texture_cache_region_viewport(psx_gpu_struct *psx_gpu, u32 x1,
 236  u32 y1, u32 x2, u32 y2)
 237 {
 238   u32 mask = texture_region_mask(x1, y1, x2, y2) &
 239    psx_gpu->viewport_mask;
 240
 241   psx_gpu->dirty_textures_4bpp_mask |= mask;
 242   psx_gpu->dirty_textures_8bpp_mask |= mask;
 243   psx_gpu->dirty_textures_8bpp_alternate_mask |= mask;
 244
 245   return mask;
 246 }
 247
 248 void update_texture_cache_region(psx_gpu_struct *psx_gpu, u32 x1, u32 y1,
 249  u32 x2, u32 y2)
 250 {
 251   u32 mask = texture_region_mask(x1, y1, x2, y2);
 252   u32 texture_page;
 253   u8 *texture_page_ptr;
 254   u16 *vram_ptr;
 255   u32 texel_block;
 256   u32 sub_x, sub_y;
 257
 258   psx_gpu->dirty_textures_8bpp_mask |= mask;
 259   psx_gpu->dirty_textures_8bpp_alternate_mask |= mask;
 260
 261   if ((psx_gpu->dirty_textures_4bpp_mask & mask) == 0 &&
 262       (x1 & 3) == 0 && (y1 & 15) == 0 && x2 - x1 < 4 && y2 - y1 < 16)
 263   {
 264     texture_page = ((x1 / 64) & 15) + (y1 / 256) * 16;
 265     texture_page_ptr = psx_gpu->texture_4bpp_cache[texture_page];
 266     texture_page_ptr += (x1 / 4 & 15) * 16*16 + (y1 / 16 & 15) * 16*16*16;
 267     vram_ptr = psx_gpu->vram_ptr + x1 + y1 * 1024;
 268     sub_x = 4;
 269     sub_y = 16;
 270
 271     while(sub_y)
 272     {
 273       while(sub_x)
 274       {
 275         texel_block = *vram_ptr;
 276
 277         texture_page_ptr[0] = texel_block & 0xF;
 278         texture_page_ptr[1] = (texel_block >> 4) & 0xF;
 279         texture_page_ptr[2] = (texel_block >> 8) & 0xF;
 280         texture_page_ptr[3] = texel_block >> 12;
 281
 282         vram_ptr++;
 283         texture_page_ptr += 4;
 284
 285         sub_x--;
 286       }
 287
 288       vram_ptr -= 4;
 289       sub_x = 4;
 290
 291       sub_y--;
 292       vram_ptr += 1024;
 293     }
 294   }
 295   else
 296   {
 297     psx_gpu->dirty_textures_4bpp_mask |= mask;
 298   }
 299 }
 300
 301 void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
 302  u32 texture_page);
 303
 304 #ifndef NEON_BUILD
 305
 306 void update_texture_4bpp_cache(psx_gpu_struct *psx_gpu)
 307 {
 308   u32 current_texture_page = psx_gpu->current_texture_page;
 309   u8 *texture_page_ptr = psx_gpu->texture_page_base;
 310   u16 *vram_ptr = psx_gpu->vram_ptr;
 311
 312   u32 texel_block;
 313   u32 tile_x, tile_y;
 314   u32 sub_x, sub_y;
 315
 316   vram_ptr += (current_texture_page >> 4) * 256 * 1024;
 317   vram_ptr += (current_texture_page & 0xF) * 64;
 318
 319   texture_cache_loads++;
 320
 321   tile_y = 16;
 322   tile_x = 16;
 323   sub_x = 4;
 324   sub_y = 16;
 325
 326   psx_gpu->dirty_textures_4bpp_mask &= ~(psx_gpu->current_texture_mask);
 327
 328   while(tile_y)
 329   {
 330     while(tile_x)
 331     {
 332       while(sub_y)
 333       {
 334         while(sub_x)
 335         {
 336           texel_block = *vram_ptr;
 337
 338           texture_page_ptr[0] = texel_block & 0xF;
 339           texture_page_ptr[1] = (texel_block >> 4) & 0xF;
 340           texture_page_ptr[2] = (texel_block >> 8) & 0xF;
 341           texture_page_ptr[3] = texel_block >> 12;
 342
 343           vram_ptr++;
 344           texture_page_ptr += 4;
 345
 346           sub_x--;
 347         }
 348
 349         vram_ptr -= 4;
 350         sub_x = 4;
 351
 352         sub_y--;
 353         vram_ptr += 1024;
 354       }
 355
 356       sub_y = 16;
 357
 358       vram_ptr -= (1024 * 16) - 4;
 359       tile_x--;
 360     }
 361
 362     tile_x = 16;
 363
 364     vram_ptr += (16 * 1024) - (4 * 16);
 365     tile_y--;
 366   }
 367 }
 368
 369 void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
 370  u32 texture_page)
 371 {
 372   u16 *texture_page_ptr = psx_gpu->texture_page_base;
 373   u16 *vram_ptr = psx_gpu->vram_ptr;
 374
 375   u32 tile_x, tile_y;
 376   u32 sub_y;
 377
 378   vec_8x16u texels;
 379
 380   texture_cache_loads++;
 381
 382   vram_ptr += (texture_page >> 4) * 256 * 1024;
 383   vram_ptr += (texture_page & 0xF) * 64;
 384
 385   if((texture_page ^ psx_gpu->current_texture_page) & 0x1)
 386     texture_page_ptr += (8 * 16) * 8;
 387
 388   tile_x = 8;
 389   tile_y = 16;
 390
 391   sub_y = 16;
 392
 393   while(tile_y)
 394   {
 395     while(tile_x)
 396     {
 397       while(sub_y)
 398       {
 399         load_128b(texels, vram_ptr);
 400         store_128b(texels, texture_page_ptr);
 401
 402         texture_page_ptr += 8;
 403         vram_ptr += 1024;
 404
 405         sub_y--;
 406       }
 407
 408       sub_y = 16;
 409
 410       vram_ptr -= (1024 * 16);
 411       vram_ptr += 8;
 412
 413       tile_x--;
 414     }
 415
 416     tile_x = 8;
 417
 418     vram_ptr -= (8 * 8);
 419     vram_ptr += (16 * 1024);
 420
 421     texture_page_ptr += (8 * 16) * 8;
 422     tile_y--;
 423   }
 424 }
 425
 426 #endif
 427
 428
 429 void update_texture_8bpp_cache(psx_gpu_struct *psx_gpu)
 430 {
 431   u32 current_texture_page = psx_gpu->current_texture_page;
 432   u32 update_textures =
 433    psx_gpu->dirty_textures_8bpp_mask & psx_gpu->current_texture_mask;
 434
 435   psx_gpu->dirty_textures_8bpp_mask &= ~update_textures;
 436
 437   if(update_textures & (1 << current_texture_page))
 438   {
 439     update_texture_8bpp_cache_slice(psx_gpu, current_texture_page);
 440     update_textures &= ~(1 << current_texture_page);
 441   }
 442
 443   if(update_textures)
 444   {
 445     u32 adjacent_texture_page = ((current_texture_page + 1) & 0xF) |
 446      (current_texture_page & 0x10);
 447
 448     update_texture_8bpp_cache_slice(psx_gpu, adjacent_texture_page);
 449   }
 450 }
 451
 452 void setup_blocks_shaded_untextured_undithered_unswizzled_indirect(
 453  psx_gpu_struct *psx_gpu);
 454
 455 void flush_render_block_buffer(psx_gpu_struct *psx_gpu)
 456 {
 457   if((psx_gpu->render_mode & RENDER_INTERLACE_ENABLED) &&
 458    (psx_gpu->primitive_type == PRIMITIVE_TYPE_SPRITE))
 459   {
 460     u32 num_blocks_dest = 0;
 461     block_struct *block_src = psx_gpu->blocks;
 462     block_struct *block_dest = psx_gpu->blocks;
 463
 464     u16 *vram_ptr = psx_gpu->vram_ptr;
 465     u32 i;
 466
 467     if(psx_gpu->render_mode & RENDER_INTERLACE_ODD)
 468     {
 469       for(i = 0; i < psx_gpu->num_blocks; i++)
 470       {
 471         u32 fb_offset = (u32)((u8 *)block_src->fb_ptr - (u8 *)vram_ptr);
 472         if(fb_offset & (1 << 11))
 473         {
 474           *block_dest = *block_src;
 475           num_blocks_dest++;
 476           block_dest++;
 477         }
 478         block_src++;
 479       }
 480     }
 481     else
 482     {
 483       for(i = 0; i < psx_gpu->num_blocks; i++)
 484       {
 485         u32 fb_offset = (u32)((u8 *)block_src->fb_ptr - (u8 *)vram_ptr);
 486         if((fb_offset & (1 << 11)) == 0)
 487         {
 488           *block_dest = *block_src;
 489           num_blocks_dest++;
 490           block_dest++;
 491         }
 492         block_src++;
 493       }
 494     }
 495
 496     psx_gpu->num_blocks = num_blocks_dest;
 497   }
 498
 499   if(psx_gpu->num_blocks)
 500   {
 501     render_block_handler_struct *render_block_handler =
 502      psx_gpu->render_block_handler;
 503
 504     render_block_handler->texture_blocks(psx_gpu);
 505     render_block_handler->shade_blocks(psx_gpu);
 506     render_block_handler->blend_blocks(psx_gpu);
 507
 508 #ifdef PROFILE
 509     span_pixel_blocks += psx_gpu->num_blocks;
 510     render_buffer_flushes++;
 511 #endif
 512
 513     psx_gpu->num_blocks = 0;
 514   }
 515 }
 516
 517
 518 void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
 519  vertex_struct *b, vertex_struct *c);
 520
 521 #ifndef NEON_BUILD
 522
 523 #define setup_gradient_calculation_input(set, vertex)                          \
 524   /* First type is:  uvrg bxxx xxxx                                          */\
 525   /* Second type is: yyyy ybyy uvrg                                          */\
 526   /* Since x_a and y_c are the same the same variable is used for both.      */\
 527   x##set##_a_y##set##_c.e[0] = vertex->u;                                      \
 528   x##set##_a_y##set##_c.e[1] = vertex->v;                                      \
 529   x##set##_a_y##set##_c.e[2] = vertex->r;                                      \
 530   x##set##_a_y##set##_c.e[3] = vertex->g;                                      \
 531   dup_4x16b(x##set##_b, vertex->x);                                            \
 532   dup_4x16b(x##set##_c, vertex->x);                                            \
 533   dup_4x16b(y##set##_a, vertex->y);                                            \
 534   dup_4x16b(y##set##_b, vertex->y);                                            \
 535   x##set##_b.e[0] = vertex->b;                                                 \
 536   y##set##_b.e[1] = vertex->b                                                  \
 537
 538
 539 void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
 540  vertex_struct *b, vertex_struct *c)
 541 {
 542   u32 triangle_area = psx_gpu->triangle_area;
 543   u32 winding_mask_scalar;
 544
 545   u32 triangle_area_shift;
 546   u64 triangle_area_reciprocal =
 547    fixed_reciprocal(triangle_area, &triangle_area_shift);
 548   triangle_area_shift = -(triangle_area_shift - FIXED_BITS);
 549
 550   // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) =
 551   // ( d0       *  d1      ) - ( d2       *  d3      ) =
 552   // ( m0                  ) - ( m1                  ) = gradient
 553
 554   // This is split to do 12 elements at a time over three sets: a, b, and c.
 555   // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so
 556   // two of the slots are unused.
 557
 558   // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as
 559   // is g.
 560
 561   vec_4x16s x0_a_y0_c, x0_b, x0_c;
 562   vec_4x16s y0_a, y0_b;
 563   vec_4x16s x1_a_y1_c, x1_b, x1_c;
 564   vec_4x16s y1_a, y1_b;
 565   vec_4x16s x2_a_y2_c, x2_b, x2_c;
 566   vec_4x16s y2_a, y2_b;
 567
 568   vec_4x32u uvrg_base;
 569   vec_4x32u b_base;
 570   vec_4x32u uvrgb_phase;
 571
 572   vec_4x16s d0_a_d3_c, d0_b, d0_c;
 573   vec_4x16s d1_a, d1_b, d1_c_d2_a;
 574   vec_4x16s d2_b, d2_c;
 575   vec_4x16s d3_a, d3_b;
 576
 577   vec_4x32s m0_a, m0_b, m0_c;
 578   vec_4x32s m1_a, m1_b, m1_c;
 579
 580   vec_4x32u gradient_area_a, gradient_area_c;
 581   vec_2x32u gradient_area_b;
 582
 583   vec_4x32u gradient_area_sign_a, gradient_area_sign_c;
 584   vec_2x32u gradient_area_sign_b;
 585   vec_4x32u winding_mask;
 586
 587   vec_2x64u gradient_wide_a0, gradient_wide_a1;
 588   vec_2x64u gradient_wide_c0, gradient_wide_c1;
 589   vec_2x64u gradient_wide_b;
 590
 591   vec_4x32u gradient_a, gradient_c;
 592   vec_2x32u gradient_b;
 593   vec_16x8s gradient_shift;
 594
 595   setup_gradient_calculation_input(0, a);
 596   setup_gradient_calculation_input(1, b);
 597   setup_gradient_calculation_input(2, c);
 598
 599   dup_4x32b(uvrgb_phase, psx_gpu->uvrgb_phase);
 600   shl_long_4x16b(uvrg_base, x0_a_y0_c, 16);
 601   shl_long_4x16b(b_base, x0_b, 16);
 602
 603   add_4x32b(uvrg_base, uvrg_base, uvrgb_phase);
 604   add_4x32b(b_base, b_base, uvrgb_phase);
 605
 606   // Can probably pair these, but it'll require careful register allocation
 607   sub_4x16b(d0_a_d3_c, x1_a_y1_c, x0_a_y0_c);
 608   sub_4x16b(d1_c_d2_a, x2_a_y2_c, x1_a_y1_c);
 609
 610   sub_4x16b(d0_b, x1_b, x0_b);
 611   sub_4x16b(d0_c, x1_c, x0_c);
 612
 613   sub_4x16b(d1_a, y2_a, y1_a);
 614   sub_4x16b(d1_b, y2_b, y1_b);
 615
 616   sub_4x16b(d2_b, x2_b, x1_b);
 617   sub_4x16b(d2_c, x2_c, x1_c);
 618
 619   sub_4x16b(d3_a, y1_a, y0_a);
 620   sub_4x16b(d3_b, y1_b, y0_b);
 621
 622   mul_long_4x16b(m0_a, d0_a_d3_c, d1_a);
 623   mul_long_4x16b(m0_b, d0_b, d1_b);
 624   mul_long_4x16b(m0_c, d0_c, d1_c_d2_a);
 625
 626   mul_long_4x16b(m1_a, d1_c_d2_a, d3_a);
 627   mul_long_4x16b(m1_b, d2_b, d3_b);
 628   mul_long_4x16b(m1_c, d2_c, d0_a_d3_c);
 629
 630   sub_4x32b(gradient_area_a, m0_a, m1_a);
 631   sub_2x32b(gradient_area_b, m0_b.low, m1_b.low);
 632   sub_4x32b(gradient_area_c, m0_c, m1_c);
 633
 634   cmpltz_4x32b(gradient_area_sign_a, gradient_area_a);
 635   cmpltz_2x32b(gradient_area_sign_b, gradient_area_b);
 636   cmpltz_4x32b(gradient_area_sign_c, gradient_area_c);
 637
 638   abs_4x32b(gradient_area_a, gradient_area_a);
 639   abs_2x32b(gradient_area_b, gradient_area_b);
 640   abs_4x32b(gradient_area_c, gradient_area_c);
 641
 642   winding_mask_scalar = -psx_gpu->triangle_winding;
 643
 644   dup_4x32b(winding_mask, winding_mask_scalar);
 645   eor_4x32b(gradient_area_sign_a, gradient_area_sign_a, winding_mask);
 646   eor_2x32b(gradient_area_sign_b, gradient_area_sign_b, winding_mask);
 647   eor_4x32b(gradient_area_sign_c, gradient_area_sign_c, winding_mask);
 648
 649   mul_scalar_long_2x32b(gradient_wide_a0,
 650    vector_cast(vec_2x32s, gradient_area_a.low),
 651    (s64)triangle_area_reciprocal);
 652   mul_scalar_long_2x32b(gradient_wide_a1,
 653    vector_cast(vec_2x32s, gradient_area_a.high),
 654    (s64)triangle_area_reciprocal);
 655   mul_scalar_long_2x32b(gradient_wide_b,
 656    vector_cast(vec_2x32s, gradient_area_b),
 657    (s64)triangle_area_reciprocal);
 658   mul_scalar_long_2x32b(gradient_wide_c0,
 659    vector_cast(vec_2x32s, gradient_area_c.low),
 660    (s64)triangle_area_reciprocal);
 661   mul_scalar_long_2x32b(gradient_wide_c1,
 662    vector_cast(vec_2x32s, gradient_area_c.high),
 663    (s64)triangle_area_reciprocal);
 664
 665   dup_16x8b(gradient_shift, triangle_area_shift);
 666   shl_reg_2x64b(gradient_wide_a0, gradient_wide_a0,
 667    vector_cast(vec_2x64u, gradient_shift));
 668   shl_reg_2x64b(gradient_wide_a1, gradient_wide_a1,
 669    vector_cast(vec_2x64u, gradient_shift));
 670   shl_reg_2x64b(gradient_wide_b, gradient_wide_b,
 671    vector_cast(vec_2x64u, gradient_shift));
 672   shl_reg_2x64b(gradient_wide_c0, gradient_wide_c0,
 673    vector_cast(vec_2x64u, gradient_shift));
 674   shl_reg_2x64b(gradient_wide_c1, gradient_wide_c1,
 675    vector_cast(vec_2x64u, gradient_shift));
 676
 677   mov_narrow_2x64b(gradient_a.low, gradient_wide_a0);
 678   mov_narrow_2x64b(gradient_a.high, gradient_wide_a1);
 679   mov_narrow_2x64b(gradient_b, gradient_wide_b);
 680   mov_narrow_2x64b(gradient_c.low, gradient_wide_c0);
 681   mov_narrow_2x64b(gradient_c.high, gradient_wide_c1);
 682
 683   shl_4x32b(gradient_a, gradient_a, 4);
 684   shl_2x32b(gradient_b, gradient_b, 4);
 685   shl_4x32b(gradient_c, gradient_c, 4);
 686
 687   eor_4x32b(gradient_a, gradient_a, gradient_area_sign_a);
 688   eor_2x32b(gradient_b, gradient_b, gradient_area_sign_b);
 689   eor_4x32b(gradient_c, gradient_c, gradient_area_sign_c);
 690
 691   sub_4x32b(gradient_a, gradient_a, gradient_area_sign_a);
 692   sub_2x32b(gradient_b, gradient_b, gradient_area_sign_b);
 693   sub_4x32b(gradient_c, gradient_c, gradient_area_sign_c);
 694
 695   u32 left_adjust = a->x;
 696   mls_scalar_4x32b(uvrg_base, gradient_a, left_adjust);
 697   mls_scalar_2x32b(b_base.low, gradient_b, left_adjust);
 698
 699   vec_4x32u uvrg_dx2;
 700   vec_2x32u b_dx2;
 701
 702   vec_4x32u uvrg_dx3;
 703   vec_2x32u b_dx3;
 704
 705   vec_4x32u zero;
 706
 707   eor_4x32b(zero, zero, zero);
 708   add_4x32b(uvrg_dx2, gradient_a, gradient_a);
 709   add_2x32b(b_dx2, gradient_b, gradient_b);
 710   add_4x32b(uvrg_dx3, gradient_a, uvrg_dx2);
 711   add_2x32b(b_dx3, gradient_b, b_dx2);
 712
 713   // Can be done with vst4, assuming that the zero, dx, dx2, and dx3 are
 714   // lined up properly
 715   psx_gpu->u_block_span.e[0] = zero.e[0];
 716   psx_gpu->u_block_span.e[1] = gradient_a.e[0];
 717   psx_gpu->u_block_span.e[2] = uvrg_dx2.e[0];
 718   psx_gpu->u_block_span.e[3] = uvrg_dx3.e[0];
 719
 720   psx_gpu->v_block_span.e[0] = zero.e[1];
 721   psx_gpu->v_block_span.e[1] = gradient_a.e[1];
 722   psx_gpu->v_block_span.e[2] = uvrg_dx2.e[1];
 723   psx_gpu->v_block_span.e[3] = uvrg_dx3.e[1];
 724
 725   psx_gpu->r_block_span.e[0] = zero.e[2];
 726   psx_gpu->r_block_span.e[1] = gradient_a.e[2];
 727   psx_gpu->r_block_span.e[2] = uvrg_dx2.e[2];
 728   psx_gpu->r_block_span.e[3] = uvrg_dx3.e[2];
 729
 730   psx_gpu->g_block_span.e[0] = zero.e[3];
 731   psx_gpu->g_block_span.e[1] = gradient_a.e[3];
 732   psx_gpu->g_block_span.e[2] = uvrg_dx2.e[3];
 733   psx_gpu->g_block_span.e[3] = uvrg_dx3.e[3];
 734
 735   psx_gpu->b_block_span.e[0] = zero.e[0];
 736   psx_gpu->b_block_span.e[1] = gradient_b.e[0];
 737   psx_gpu->b_block_span.e[2] = b_dx2.e[0];
 738   psx_gpu->b_block_span.e[3] = b_dx3.e[0];
 739
 740   psx_gpu->uvrg = uvrg_base;
 741   psx_gpu->b = b_base.e[0];
 742
 743   psx_gpu->uvrg_dx = gradient_a;
 744   psx_gpu->uvrg_dy = gradient_c;
 745   psx_gpu->b_dy = gradient_b.e[1];
 746 }
 747 #endif
 748
 749 #define vector_check(_a, _b)                                                   \
 750   if(memcmp(&_a, &_b, sizeof(_b)))                                             \
 751   {                                                                            \
 752     if(sizeof(_b) == 8)                                                        \
 753     {                                                                          \
 754       printf("mismatch on %s vs %s: (%x %x) vs (%x %x)\n",                     \
 755        #_a, #_b, _a.e[0], _a.e[1], _b.e[0], _b.e[1]);                          \
 756     }                                                                          \
 757     else                                                                       \
 758     {                                                                          \
 759       printf("mismatch on %s vs %s: (%x %x %x %x) vs (%x %x %x %x)\n",         \
 760        #_a, #_b, _a.e[0], _a.e[1], _a.e[2], _a.e[3], _b.e[0], _b.e[1],         \
 761        _b.e[2], _b.e[3]);                                                      \
 762     }                                                                          \
 763   }                                                                            \
 764
 765 #define scalar_check(_a, _b)                                                   \
 766   if(_a != _b)                                                                 \
 767     printf("mismatch on %s %s: %x vs %x\n", #_a, #_b, _a, _b)                  \
 768
 769
 770 #ifndef NDEBUG
 771 #define setup_spans_debug_check(span_edge_data_element)                        \
 772 {                                                                              \
 773   u32 _num_spans = &span_edge_data_element - psx_gpu->span_edge_data;          \
 774   if (_num_spans > MAX_SPANS)                                                  \
 775     *(int *)0 = 1;                                                             \
 776   if (_num_spans < psx_gpu->num_spans)                                         \
 777   {                                                                            \
 778     if(span_edge_data_element.num_blocks > MAX_BLOCKS_PER_ROW)                 \
 779       *(int *)0 = 1;                                                           \
 780     if(span_edge_data_element.y > 2048)                                        \
 781       *(int *)0 = 1;                                                           \
 782   }                                                                            \
 783 }                                                                              \
 784
 785 #else
 786 #define setup_spans_debug_check(span_edge_data_element)                        \
 787
 788 #endif
 789
 790 #define setup_spans_prologue_alternate_yes()                                   \
 791   vec_2x64s alternate_x;                                                       \
 792   vec_2x64s alternate_dx_dy;                                                   \
 793   vec_4x32s alternate_x_32;                                                    \
 794   vec_2x32s alternate_x_16;                                                    \
 795                                                                                \
 796   vec_4x16u alternate_select;                                                  \
 797   vec_4x16s y_mid_point;                                                       \
 798                                                                                \
 799   s32 y_b = v_b->y;                                                            \
 800   s64 edge_alt;                                                                \
 801   s32 edge_dx_dy_alt;                                                          \
 802   u32 edge_shift_alt                                                           \
 803
 804 #define setup_spans_prologue_alternate_no()                                    \
 805
 806 #define setup_spans_prologue(alternate_active)                                 \
 807   edge_data_struct *span_edge_data;                                            \
 808   vec_4x32u *span_uvrg_offset;                                                 \
 809   u32 *span_b_offset;                                                          \
 810                                                                                \
 811   s32 clip;                                                                    \
 812                                                                                \
 813   vec_2x64s edges_xy;                                                          \
 814   vec_2x32s edges_dx_dy;                                                       \
 815   vec_2x32u edge_shifts;                                                       \
 816                                                                                \
 817   vec_2x64s left_x, right_x;                                                   \
 818   vec_2x64s left_dx_dy, right_dx_dy;                                           \
 819   vec_4x32s left_x_32, right_x_32;                                             \
 820   vec_8x16s left_right_x_16;                                                   \
 821   vec_4x16s y_x4;                                                              \
 822   vec_8x16s left_edge;                                                         \
 823   vec_8x16s right_edge;                                                        \
 824   vec_4x16u span_shift;                                                        \
 825                                                                                \
 826   vec_2x32u c_0x01;                                                            \
 827   vec_4x16u c_0x04;                                                            \
 828   vec_4x16u c_0xFFFE;                                                          \
 829   vec_4x16u c_0x07;                                                            \
 830                                                                                \
 831   vec_2x32s x_starts;                                                          \
 832   vec_2x32s x_ends;                                                            \
 833                                                                                \
 834   s32 x_a = v_a->x;                                                            \
 835   s32 x_b = v_b->x;                                                            \
 836   s32 x_c = v_c->x;                                                            \
 837   s32 y_a = v_a->y;                                                            \
 838   s32 y_c = v_c->y;                                                            \
 839                                                                                \
 840   vec_4x32u uvrg = psx_gpu->uvrg;                                              \
 841   vec_4x32u uvrg_dy = psx_gpu->uvrg_dy;                                        \
 842   u32 b = psx_gpu->b;                                                          \
 843   u32 b_dy = psx_gpu->b_dy;                                                    \
 844                                                                                \
 845   dup_2x32b(c_0x01, 0x01);                                                     \
 846   setup_spans_prologue_alternate_##alternate_active()                          \
 847
 848 #define setup_spans_prologue_b()                                               \
 849   span_edge_data = psx_gpu->span_edge_data;                                    \
 850   span_uvrg_offset = psx_gpu->span_uvrg_offset;                                \
 851   span_b_offset = psx_gpu->span_b_offset;                                      \
 852                                                                                \
 853   vec_8x16u c_0x0001;                                                          \
 854                                                                                \
 855   dup_8x16b(c_0x0001, 0x0001);                                                 \
 856   dup_8x16b(left_edge, psx_gpu->viewport_start_x);                             \
 857   dup_8x16b(right_edge, psx_gpu->viewport_end_x);                              \
 858   add_8x16b(right_edge, right_edge, c_0x0001);                                 \
 859   dup_4x16b(c_0x04, 0x04);                                                     \
 860   dup_4x16b(c_0x07, 0x07);                                                     \
 861   dup_4x16b(c_0xFFFE, 0xFFFE);                                                 \
 862
 863
 864 #define compute_edge_delta_x2()                                                \
 865 {                                                                              \
 866   vec_2x32s heights;                                                           \
 867   vec_2x32s height_reciprocals;                                                \
 868   vec_2x32s heights_b;                                                         \
 869   vec_4x32u widths;                                                            \
 870                                                                                \
 871   u32 edge_shift = reciprocal_table[height];                                   \
 872                                                                                \
 873   dup_2x32b(heights, height);                                                  \
 874   sub_2x32b(widths, x_ends, x_starts);                                         \
 875                                                                                \
 876   dup_2x32b(edge_shifts, edge_shift);                                          \
 877   sub_2x32b(heights_b, heights, c_0x01);                                       \
 878   shr_2x32b(height_reciprocals, edge_shifts, 10);                              \
 879                                                                                \
 880   mla_2x32b(heights_b, x_starts, heights);                                     \
 881   bic_immediate_4x16b(vector_cast(vec_4x16u, edge_shifts), 0xE0);              \
 882   mul_2x32b(edges_dx_dy, widths, height_reciprocals);                          \
 883   mul_long_2x32b(edges_xy, heights_b, height_reciprocals);                     \
 884 }                                                                              \
 885
 886 #define compute_edge_delta_x3(start_c, height_a, height_b)                     \
 887 {                                                                              \
 888   vec_2x32s heights;                                                           \
 889   vec_2x32s height_reciprocals;                                                \
 890   vec_2x32s heights_b;                                                         \
 891   vec_2x32u widths;                                                            \
 892                                                                                \
 893   u32 width_alt;                                                               \
 894   s32 height_b_alt;                                                            \
 895   u32 height_reciprocal_alt;                                                   \
 896                                                                                \
 897   heights.e[0] = height_a;                                                     \
 898   heights.e[1] = height_b;                                                     \
 899                                                                                \
 900   edge_shifts.e[0] = reciprocal_table[height_a];                               \
 901   edge_shifts.e[1] = reciprocal_table[height_b];                               \
 902   edge_shift_alt = reciprocal_table[height_minor_b];                           \
 903                                                                                \
 904   sub_2x32b(widths, x_ends, x_starts);                                         \
 905   width_alt = x_c - start_c;                                                   \
 906                                                                                \
 907   shr_2x32b(height_reciprocals, edge_shifts, 10);                              \
 908   height_reciprocal_alt = edge_shift_alt >> 10;                                \
 909                                                                                \
 910   bic_immediate_4x16b(vector_cast(vec_4x16u, edge_shifts), 0xE0);              \
 911   edge_shift_alt &= 0x1F;                                                      \
 912                                                                                \
 913   sub_2x32b(heights_b, heights, c_0x01);                                       \
 914   height_b_alt = height_minor_b - 1;                                           \
 915                                                                                \
 916   mla_2x32b(heights_b, x_starts, heights);                                     \
 917   height_b_alt += height_minor_b * start_c;                                    \
 918                                                                                \
 919   mul_long_2x32b(edges_xy, heights_b, height_reciprocals);                     \
 920   edge_alt = (s64)height_b_alt * height_reciprocal_alt;                        \
 921                                                                                \
 922   mul_2x32b(edges_dx_dy, widths, height_reciprocals);                          \
 923   edge_dx_dy_alt = width_alt * height_reciprocal_alt;                          \
 924 }                                                                              \
 925
 926
 927 #define setup_spans_adjust_y_up()                                              \
 928   sub_4x32b(y_x4, y_x4, c_0x04)                                                \
 929
 930 #define setup_spans_adjust_y_down()                                            \
 931   add_4x32b(y_x4, y_x4, c_0x04)                                                \
 932
 933 #define setup_spans_adjust_interpolants_up()                                   \
 934   sub_4x32b(uvrg, uvrg, uvrg_dy);                                              \
 935   b -= b_dy                                                                    \
 936
 937 #define setup_spans_adjust_interpolants_down()                                 \
 938   add_4x32b(uvrg, uvrg, uvrg_dy);                                              \
 939   b += b_dy                                                                    \
 940
 941
 942 #define setup_spans_clip_interpolants_increment()                              \
 943   mla_scalar_4x32b(uvrg, uvrg_dy, clip);                                       \
 944   b += b_dy * clip                                                             \
 945
 946 #define setup_spans_clip_interpolants_decrement()                              \
 947   mls_scalar_4x32b(uvrg, uvrg_dy, clip);                                       \
 948   b -= b_dy * clip                                                             \
 949
 950 #define setup_spans_clip_alternate_yes()                                       \
 951   edge_alt += edge_dx_dy_alt * (s64)(clip)                                     \
 952
 953 #define setup_spans_clip_alternate_no()                                        \
 954
 955 #define setup_spans_clip(direction, alternate_active)                          \
 956 {                                                                              \
 957   clipped_triangles++;                                                         \
 958   mla_scalar_long_2x32b(edges_xy, edges_dx_dy, (s64)clip);                     \
 959   setup_spans_clip_alternate_##alternate_active();                             \
 960   setup_spans_clip_interpolants_##direction();                                 \
 961 }                                                                              \
 962
 963
 964 #define setup_spans_adjust_edges_alternate_no(left_index, right_index)         \
 965 {                                                                              \
 966   vec_2x64u edge_shifts_64;                                                    \
 967   vec_2x64s edges_dx_dy_64;                                                    \
 968                                                                                \
 969   mov_wide_2x32b(edge_shifts_64, edge_shifts);                                 \
 970   shl_variable_2x64b(edges_xy, edges_xy, edge_shifts_64);                      \
 971                                                                                \
 972   mov_wide_2x32b(edges_dx_dy_64, edges_dx_dy);                                 \
 973   shl_variable_2x64b(edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64);          \
 974                                                                                \
 975   left_x.e[0] = edges_xy.e[left_index];                                        \
 976   right_x.e[0] = edges_xy.e[right_index];                                      \
 977                                                                                \
 978   left_dx_dy.e[0] = edges_dx_dy_64.e[left_index];                              \
 979   left_dx_dy.e[1] = edges_dx_dy_64.e[left_index];                              \
 980   right_dx_dy.e[0] = edges_dx_dy_64.e[right_index];                            \
 981   right_dx_dy.e[1] = edges_dx_dy_64.e[right_index];                            \
 982                                                                                \
 983   add_1x64b(left_x.high, left_x.low, left_dx_dy.low);                          \
 984   add_1x64b(right_x.high, right_x.low, right_dx_dy.low);                       \
 985                                                                                \
 986   add_2x64b(left_dx_dy, left_dx_dy, left_dx_dy);                               \
 987   add_2x64b(right_dx_dy, right_dx_dy, right_dx_dy);                            \
 988 }                                                                              \
 989
 990 #define setup_spans_adjust_edges_alternate_yes(left_index, right_index)        \
 991 {                                                                              \
 992   setup_spans_adjust_edges_alternate_no(left_index, right_index);              \
 993   s64 edge_dx_dy_alt_64;                                                       \
 994                                                                                \
 995   dup_4x16b(y_mid_point, y_b);                                                 \
 996                                                                                \
 997   edge_alt <<= edge_shift_alt;                                                 \
 998   edge_dx_dy_alt_64 = (s64)edge_dx_dy_alt << edge_shift_alt;                   \
 999                                                                                \
1000   alternate_x.e[0] = edge_alt;                                                 \
1001   alternate_dx_dy.e[0] = edge_dx_dy_alt_64;                                    \
1002   alternate_dx_dy.e[1] = edge_dx_dy_alt_64;                                    \
1003                                                                                \
1004   add_1x64b(alternate_x.high, alternate_x.low, alternate_dx_dy.low);           \
1005   add_2x64b(alternate_dx_dy, alternate_dx_dy, alternate_dx_dy);                \
1006 }                                                                              \
1007
1008
1009 #define setup_spans_y_select_up()                                              \
1010   cmplt_4x16b(alternate_select, y_x4, y_mid_point)                             \
1011
1012 #define setup_spans_y_select_down()                                            \
1013   cmpgt_4x16b(alternate_select, y_x4, y_mid_point)                             \
1014
1015 #define setup_spans_y_select_alternate_yes(direction)                          \
1016   setup_spans_y_select_##direction()                                           \
1017
1018 #define setup_spans_y_select_alternate_no(direction)                           \
1019
1020 #define setup_spans_alternate_select_left()                                    \
1021   bit_4x16b(left_right_x_16.low, alternate_x_16, alternate_select)             \
1022
1023 #define setup_spans_alternate_select_right()                                   \
1024   bit_4x16b(left_right_x_16.high, alternate_x_16, alternate_select)            \
1025
1026 #define setup_spans_alternate_select_none()                                    \
1027
1028 #define setup_spans_increment_alternate_yes()                                  \
1029   shr_narrow_2x64b(alternate_x_32.low, alternate_x, 32);                       \
1030   add_2x64b(alternate_x, alternate_x, alternate_dx_dy);                        \
1031   shr_narrow_2x64b(alternate_x_32.high, alternate_x, 32);                      \
1032   add_2x64b(alternate_x, alternate_x, alternate_dx_dy);                        \
1033   mov_narrow_4x32b(alternate_x_16, alternate_x_32)                             \
1034
1035 #define setup_spans_increment_alternate_no()                                   \
1036
1037 #define setup_spans_set_x4(alternate, direction, alternate_active)             \
1038 {                                                                              \
1039   span_uvrg_offset[0] = uvrg;                                                  \
1040   span_b_offset[0] = b;                                                        \
1041   setup_spans_adjust_interpolants_##direction();                               \
1042                                                                                \
1043   span_uvrg_offset[1] = uvrg;                                                  \
1044   span_b_offset[1] = b;                                                        \
1045   setup_spans_adjust_interpolants_##direction();                               \
1046                                                                                \
1047   span_uvrg_offset[2] = uvrg;                                                  \
1048   span_b_offset[2] = b;                                                        \
1049   setup_spans_adjust_interpolants_##direction();                               \
1050                                                                                \
1051   span_uvrg_offset[3] = uvrg;                                                  \
1052   span_b_offset[3] = b;                                                        \
1053   setup_spans_adjust_interpolants_##direction();                               \
1054                                                                                \
1055   span_uvrg_offset += 4;                                                       \
1056   span_b_offset += 4;                                                          \
1057                                                                                \
1058   shr_narrow_2x64b(left_x_32.low, left_x, 32);                                 \
1059   shr_narrow_2x64b(right_x_32.low, right_x, 32);                               \
1060                                                                                \
1061   add_2x64b(left_x, left_x, left_dx_dy);                                       \
1062   add_2x64b(right_x, right_x, right_dx_dy);                                    \
1063                                                                                \
1064   shr_narrow_2x64b(left_x_32.high, left_x, 32);                                \
1065   shr_narrow_2x64b(right_x_32.high, right_x, 32);                              \
1066                                                                                \
1067   add_2x64b(left_x, left_x, left_dx_dy);                                       \
1068   add_2x64b(right_x, right_x, right_dx_dy);                                    \
1069                                                                                \
1070   mov_narrow_4x32b(left_right_x_16.low, left_x_32);                            \
1071   mov_narrow_4x32b(left_right_x_16.high, right_x_32);                          \
1072                                                                                \
1073   setup_spans_increment_alternate_##alternate_active();                        \
1074   setup_spans_y_select_alternate_##alternate_active(direction);                \
1075   setup_spans_alternate_select_##alternate();                                  \
1076                                                                                \
1077   max_8x16b(left_right_x_16, left_right_x_16, left_edge);                      \
1078   min_8x16b(left_right_x_16, left_right_x_16, right_edge);                     \
1079                                                                                \
1080   sub_4x16b(left_right_x_16.high, left_right_x_16.high, left_right_x_16.low);  \
1081   add_4x16b(left_right_x_16.high, left_right_x_16.high, c_0x07);               \
1082   and_4x16b(span_shift, left_right_x_16.high, c_0x07);                         \
1083   shl_variable_4x16b(span_shift, c_0xFFFE, span_shift);                        \
1084   shr_4x16b(left_right_x_16.high, left_right_x_16.high, 3);                    \
1085                                                                                \
1086   u32 i;                                                                       \
1087   for(i = 0; i < 4; i++)                                                       \
1088   {                                                                            \
1089     span_edge_data[i].left_x = left_right_x_16.low.e[i];                       \
1090     span_edge_data[i].num_blocks = left_right_x_16.high.e[i];                  \
1091     span_edge_data[i].right_mask = span_shift.e[i];                            \
1092     span_edge_data[i].y = y_x4.e[i];                                           \
1093     setup_spans_debug_check(span_edge_data[i]);                                \
1094   }                                                                            \
1095                                                                                \
1096   span_edge_data += 4;                                                         \
1097                                                                                \
1098   setup_spans_adjust_y_##direction();                                          \
1099 }                                                                              \
1100
1101
1102 #define setup_spans_alternate_adjust_yes()                                     \
1103   edge_alt -= edge_dx_dy_alt * (s64)height_minor_a                             \
1104
1105 #define setup_spans_alternate_adjust_no()                                      \
1106
1107
1108 #define setup_spans_down(left_index, right_index, alternate, alternate_active) \
1109   setup_spans_alternate_adjust_##alternate_active();                           \
1110   if(y_c > psx_gpu->viewport_end_y)                                            \
1111     height -= y_c - psx_gpu->viewport_end_y - 1;                               \
1112                                                                                \
1113   clip = psx_gpu->viewport_start_y - y_a;                                      \
1114   if(clip > 0)                                                                 \
1115   {                                                                            \
1116     height -= clip;                                                            \
1117     y_a += clip;                                                               \
1118     setup_spans_clip(increment, alternate_active);                             \
1119   }                                                                            \
1120                                                                                \
1121   setup_spans_prologue_b();                                                    \
1122                                                                                \
1123   if(height > 0)                                                               \
1124   {                                                                            \
1125     y_x4.e[0] = y_a;                                                           \
1126     y_x4.e[1] = y_a + 1;                                                       \
1127     y_x4.e[2] = y_a + 2;                                                       \
1128     y_x4.e[3] = y_a + 3;                                                       \
1129     setup_spans_adjust_edges_alternate_##alternate_active(left_index,          \
1130      right_index);                                                             \
1131                                                                                \
1132     psx_gpu->num_spans = height;                                               \
1133     do                                                                         \
1134     {                                                                          \
1135       setup_spans_set_x4(alternate, down, alternate_active);                   \
1136       height -= 4;                                                             \
1137     } while(height > 0);                                                       \
1138   }                                                                            \
1139
1140
1141 #define setup_spans_alternate_pre_increment_yes()                              \
1142   edge_alt += edge_dx_dy_alt                                                   \
1143
1144 #define setup_spans_alternate_pre_increment_no()                               \
1145
1146 #define setup_spans_up_decrement_height_yes()                                  \
1147   height--                                                                     \
1148
1149 #define setup_spans_up_decrement_height_no()                                   \
1150   {}                                                                           \
1151
1152 #define setup_spans_up(left_index, right_index, alternate, alternate_active)   \
1153   setup_spans_alternate_adjust_##alternate_active();                           \
1154   y_a--;                                                                       \
1155                                                                                \
1156   if(y_c < psx_gpu->viewport_start_y)                                          \
1157     height -= psx_gpu->viewport_start_y - y_c;                                 \
1158   else                                                                         \
1159     setup_spans_up_decrement_height_##alternate_active();                      \
1160                                                                                \
1161   clip = y_a - psx_gpu->viewport_end_y;                                        \
1162   if(clip > 0)                                                                 \
1163   {                                                                            \
1164     height -= clip;                                                            \
1165     y_a -= clip;                                                               \
1166     setup_spans_clip(decrement, alternate_active);                             \
1167   }                                                                            \
1168                                                                                \
1169   setup_spans_prologue_b();                                                    \
1170                                                                                \
1171   if(height > 0)                                                               \
1172   {                                                                            \
1173     y_x4.e[0] = y_a;                                                           \
1174     y_x4.e[1] = y_a - 1;                                                       \
1175     y_x4.e[2] = y_a - 2;                                                       \
1176     y_x4.e[3] = y_a - 3;                                                       \
1177     add_wide_2x32b(edges_xy, edges_xy, edges_dx_dy);                           \
1178     setup_spans_alternate_pre_increment_##alternate_active();                  \
1179     setup_spans_adjust_edges_alternate_##alternate_active(left_index,          \
1180      right_index);                                                             \
1181     setup_spans_adjust_interpolants_up();                                      \
1182                                                                                \
1183     psx_gpu->num_spans = height;                                               \
1184     while(height > 0)                                                          \
1185     {                                                                          \
1186       setup_spans_set_x4(alternate, up, alternate_active);                     \
1187       height -= 4;                                                             \
1188     }                                                                          \
1189   }                                                                            \
1190
1191 #define index_left  0
1192 #define index_right 1
1193
1194 #define setup_spans_up_up(minor, major)                                        \
1195   setup_spans_prologue(yes);                                                   \
1196   s32 height_minor_a = y_a - y_b;                                              \
1197   s32 height_minor_b = y_b - y_c;                                              \
1198   s32 height = y_a - y_c;                                                      \
1199                                                                                \
1200   dup_2x32b(x_starts, x_a);                                                    \
1201   x_ends.e[0] = x_c;                                                           \
1202   x_ends.e[1] = x_b;                                                           \
1203                                                                                \
1204   compute_edge_delta_x3(x_b, height, height_minor_a);                          \
1205   setup_spans_up(index_##major, index_##minor, minor, yes)                     \
1206
1207
1208 void setup_spans_up_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1209  vertex_struct *v_b, vertex_struct *v_c);
1210 void setup_spans_up_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1211  vertex_struct *v_b, vertex_struct *v_c);
1212 void setup_spans_down_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1213  vertex_struct *v_b, vertex_struct *v_c);
1214 void setup_spans_down_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1215  vertex_struct *v_b, vertex_struct *v_c);
1216 void setup_spans_up_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1217  vertex_struct *v_b, vertex_struct *v_c);
1218 void setup_spans_up_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1219  vertex_struct *v_b, vertex_struct *v_c);
1220 void setup_spans_down_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1221  vertex_struct *v_b, vertex_struct *v_c);
1222 void setup_spans_down_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1223  vertex_struct *v_b, vertex_struct *v_c);
1224 void setup_spans_up_down(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1225  vertex_struct *v_b, vertex_struct *v_c);
1226
1227
1228 #ifndef NEON_BUILD
1229
1230 void setup_spans_up_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1231  vertex_struct *v_b, vertex_struct *v_c)
1232 {
1233   setup_spans_up_up(left, right);
1234 }
1235
1236 void setup_spans_up_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1237  vertex_struct *v_b, vertex_struct *v_c)
1238 {
1239   setup_spans_up_up(right, left);
1240 }
1241
1242 #define setup_spans_down_down(minor, major)                                    \
1243   setup_spans_prologue(yes);                                                   \
1244   s32 height_minor_a = y_b - y_a;                                              \
1245   s32 height_minor_b = y_c - y_b;                                              \
1246   s32 height = y_c - y_a;                                                      \
1247                                                                                \
1248   dup_2x32b(x_starts, x_a);                                                    \
1249   x_ends.e[0] = x_c;                                                           \
1250   x_ends.e[1] = x_b;                                                           \
1251                                                                                \
1252   compute_edge_delta_x3(x_b, height, height_minor_a);                          \
1253   setup_spans_down(index_##major, index_##minor, minor, yes)                   \
1254
1255 void setup_spans_down_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1256  vertex_struct *v_b, vertex_struct *v_c)
1257 {
1258   setup_spans_down_down(left, right);
1259 }
1260
1261 void setup_spans_down_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1262  vertex_struct *v_b, vertex_struct *v_c)
1263 {
1264   setup_spans_down_down(right, left);
1265 }
1266
1267 #define setup_spans_up_flat()                                                  \
1268   s32 height = y_a - y_c;                                                      \
1269                                                                                \
1270   flat_triangles++;                                                            \
1271   compute_edge_delta_x2();                                                     \
1272   setup_spans_up(index_left, index_right, none, no)                            \
1273
1274 void setup_spans_up_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1275  vertex_struct *v_b, vertex_struct *v_c)
1276 {
1277   setup_spans_prologue(no);
1278   x_starts.e[0] = x_a;
1279   x_starts.e[1] = x_b;
1280   dup_2x32b(x_ends, x_c);
1281
1282   setup_spans_up_flat();
1283 }
1284
1285 void setup_spans_up_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1286  vertex_struct *v_b, vertex_struct *v_c)
1287 {
1288   setup_spans_prologue(no);
1289   dup_2x32b(x_starts, x_a);
1290   x_ends.e[0] = x_b;
1291   x_ends.e[1] = x_c;
1292
1293   setup_spans_up_flat();
1294 }
1295
1296 #define setup_spans_down_flat()                                                \
1297   s32 height = y_c - y_a;                                                      \
1298                                                                                \
1299   flat_triangles++;                                                            \
1300   compute_edge_delta_x2();                                                     \
1301   setup_spans_down(index_left, index_right, none, no)                          \
1302
1303 void setup_spans_down_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1304  vertex_struct *v_b, vertex_struct *v_c)
1305 {
1306   setup_spans_prologue(no);
1307   x_starts.e[0] = x_a;
1308   x_starts.e[1] = x_b;
1309   dup_2x32b(x_ends, x_c);
1310
1311   setup_spans_down_flat();
1312 }
1313
1314 void setup_spans_down_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1315  vertex_struct *v_b, vertex_struct *v_c)
1316 {
1317   setup_spans_prologue(no);
1318   dup_2x32b(x_starts, x_a);
1319   x_ends.e[0] = x_b;
1320   x_ends.e[1] = x_c;
1321
1322   setup_spans_down_flat();
1323 }
1324
1325 void setup_spans_up_down(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1326  vertex_struct *v_b, vertex_struct *v_c)
1327 {
1328   setup_spans_prologue(no);
1329
1330   s32 y_b = v_b->y;
1331   s64 edge_alt;
1332   s32 edge_dx_dy_alt;
1333   u32 edge_shift_alt;
1334
1335   s32 middle_y = y_a;
1336   s32 height_minor_a = y_a - y_b;
1337   s32 height_minor_b = y_c - y_a;
1338   s32 height_major = y_c - y_b;
1339
1340   vec_2x64s edges_xy_b;
1341   vec_2x32s edges_dx_dy_b;
1342   vec_2x32u edge_shifts_b;
1343
1344   vec_2x32s height_increment;
1345
1346   x_starts.e[0] = x_a;
1347   x_starts.e[1] = x_c;
1348   dup_2x32b(x_ends, x_b);
1349
1350   compute_edge_delta_x3(x_a, height_minor_a, height_major);
1351
1352   height_increment.e[0] = 0;
1353   height_increment.e[1] = height_minor_b;
1354
1355   mla_long_2x32b(edges_xy, edges_dx_dy, height_increment);
1356
1357   edges_xy_b.e[0] = edge_alt;
1358   edges_xy_b.e[1] = edges_xy.e[1];
1359
1360   edge_shifts_b = edge_shifts;
1361   edge_shifts_b.e[0] = edge_shift_alt;
1362
1363   neg_2x32b(edges_dx_dy_b, edges_dx_dy);
1364   edges_dx_dy_b.e[0] = edge_dx_dy_alt;
1365
1366   y_a--;
1367
1368   if(y_b < psx_gpu->viewport_start_y)
1369     height_minor_a -= psx_gpu->viewport_start_y - y_b;
1370
1371   clip = y_a - psx_gpu->viewport_end_y;
1372   if(clip > 0)
1373   {
1374     height_minor_a -= clip;
1375     y_a -= clip;
1376     setup_spans_clip(decrement, no);
1377   }
1378
1379   setup_spans_prologue_b();
1380
1381   if(height_minor_a > 0)
1382   {
1383     y_x4.e[0] = y_a;
1384     y_x4.e[1] = y_a - 1;
1385     y_x4.e[2] = y_a - 2;
1386     y_x4.e[3] = y_a - 3;
1387     add_wide_2x32b(edges_xy, edges_xy, edges_dx_dy);
1388     setup_spans_adjust_edges_alternate_no(index_left, index_right);
1389     setup_spans_adjust_interpolants_up();
1390
1391     psx_gpu->num_spans = height_minor_a;
1392     while(height_minor_a > 0)
1393     {
1394       setup_spans_set_x4(none, up, no);
1395       height_minor_a -= 4;
1396     }
1397
1398     span_edge_data += height_minor_a;
1399     span_uvrg_offset += height_minor_a;
1400     span_b_offset += height_minor_a;
1401   }
1402
1403   edges_xy = edges_xy_b;
1404   edges_dx_dy = edges_dx_dy_b;
1405   edge_shifts = edge_shifts_b;
1406
1407   uvrg = psx_gpu->uvrg;
1408   b = psx_gpu->b;
1409
1410   y_a = middle_y;
1411
1412   if(y_c > psx_gpu->viewport_end_y)
1413     height_minor_b -= y_c - psx_gpu->viewport_end_y - 1;
1414
1415   clip = psx_gpu->viewport_start_y - y_a;
1416   if(clip > 0)
1417   {
1418     height_minor_b -= clip;
1419     y_a += clip;
1420     setup_spans_clip(increment, no);
1421   }
1422
1423   if(height_minor_b > 0)
1424   {
1425     y_x4.e[0] = y_a;
1426     y_x4.e[1] = y_a + 1;
1427     y_x4.e[2] = y_a + 2;
1428     y_x4.e[3] = y_a + 3;
1429     setup_spans_adjust_edges_alternate_no(index_left, index_right);
1430
1431     // FIXME: overflow corner case
1432     if(psx_gpu->num_spans + height_minor_b == MAX_SPANS)
1433       height_minor_b &= ~3;
1434
1435     psx_gpu->num_spans += height_minor_b;
1436     while(height_minor_b > 0)
1437     {
1438       setup_spans_set_x4(none, down, no);
1439       height_minor_b -= 4;
1440     }
1441   }
1442
1443   left_split_triangles++;
1444 }
1445
1446 #endif
1447
1448
1449 #define dither_table_entry_normal(value)                                       \
1450   (value)                                                                      \
1451
1452
1453 #define setup_blocks_load_msb_mask_indirect()                                  \
1454
1455 #define setup_blocks_load_msb_mask_direct()                                    \
1456   vec_8x16u msb_mask;                                                          \
1457   dup_8x16b(msb_mask, psx_gpu->mask_msb);                                      \
1458
1459
1460 #define setup_blocks_variables_shaded_textured(target)                         \
1461   vec_4x32u u_block;                                                           \
1462   vec_4x32u v_block;                                                           \
1463   vec_4x32u r_block;                                                           \
1464   vec_4x32u g_block;                                                           \
1465   vec_4x32u b_block;                                                           \
1466   vec_4x32u uvrg_dx = psx_gpu->uvrg_dx;                                        \
1467   vec_4x32u uvrg_dx4;                                                          \
1468   vec_4x32u uvrg_dx8;                                                          \
1469   vec_4x32u uvrg;                                                              \
1470   u32 b_dx = psx_gpu->b_block_span.e[1];                                       \
1471   u32 b_dx4 = b_dx << 2;                                                       \
1472   u32 b_dx8 = b_dx << 3;                                                       \
1473   u32 b;                                                                       \
1474                                                                                \
1475   vec_16x8u texture_mask;                                                      \
1476   shl_4x32b(uvrg_dx4, uvrg_dx, 2);                                             \
1477   shl_4x32b(uvrg_dx8, uvrg_dx, 3);                                             \
1478   dup_8x8b(texture_mask.low, psx_gpu->texture_mask_width);                     \
1479   dup_8x8b(texture_mask.high, psx_gpu->texture_mask_height)                    \
1480
1481 #define setup_blocks_variables_shaded_untextured(target)                       \
1482   vec_4x32u r_block;                                                           \
1483   vec_4x32u g_block;                                                           \
1484   vec_4x32u b_block;                                                           \
1485   vec_4x32u rgb_dx;                                                            \
1486   vec_4x32u rgb_dx4;                                                           \
1487   vec_4x32u rgb_dx8;                                                           \
1488   vec_4x32u rgb;                                                               \
1489                                                                                \
1490   vec_8x8u d64_0x07;                                                           \
1491   vec_8x8u d64_1;                                                              \
1492   vec_8x8u d64_4;                                                              \
1493   vec_8x8u d64_128;                                                            \
1494                                                                                \
1495   dup_8x8b(d64_0x07, 0x07);                                                    \
1496   dup_8x8b(d64_1, 1);                                                          \
1497   dup_8x8b(d64_4, 4);                                                          \
1498   dup_8x8b(d64_128, 128);                                                      \
1499                                                                                \
1500   rgb_dx.low = psx_gpu->uvrg_dx.high;                                          \
1501   rgb_dx.e[2] = psx_gpu->b_block_span.e[1];                                    \
1502   shl_4x32b(rgb_dx4, rgb_dx, 2);                                               \
1503   shl_4x32b(rgb_dx8, rgb_dx, 3)                                                \
1504
1505 #define setup_blocks_variables_unshaded_textured(target)                       \
1506   vec_4x32u u_block;                                                           \
1507   vec_4x32u v_block;                                                           \
1508   vec_2x32u uv_dx = psx_gpu->uvrg_dx.low;                                      \
1509   vec_2x32u uv_dx4;                                                            \
1510   vec_2x32u uv_dx8;                                                            \
1511   vec_2x32u uv = psx_gpu->uvrg.low;                                            \
1512                                                                                \
1513   vec_16x8u texture_mask;                                                      \
1514   shl_2x32b(uv_dx4, uv_dx, 2);                                                 \
1515   shl_2x32b(uv_dx8, uv_dx, 3);                                                 \
1516   dup_8x8b(texture_mask.low, psx_gpu->texture_mask_width);                     \
1517   dup_8x8b(texture_mask.high, psx_gpu->texture_mask_height)                    \
1518
1519
1520 #define setup_blocks_variables_unshaded_untextured_direct()                    \
1521   or_8x16b(colors, colors, msb_mask)                                           \
1522
1523 #define setup_blocks_variables_unshaded_untextured_indirect()                  \
1524
1525 #define setup_blocks_variables_unshaded_untextured(target)                     \
1526   u32 color = psx_gpu->triangle_color;                                         \
1527   vec_8x16u colors;                                                            \
1528                                                                                \
1529   u32 color_r = color & 0xFF;                                                  \
1530   u32 color_g = (color >> 8) & 0xFF;                                           \
1531   u32 color_b = (color >> 16) & 0xFF;                                          \
1532                                                                                \
1533   color = (color_r >> 3) | ((color_g >> 3) << 5) |                             \
1534    ((color_b >> 3) << 10);                                                     \
1535   dup_8x16b(colors, color);                                                    \
1536   setup_blocks_variables_unshaded_untextured_##target()                        \
1537
1538 #define setup_blocks_span_initialize_dithered_textured()                       \
1539   vec_8x16u dither_offsets;                                                    \
1540   shl_long_8x8b(dither_offsets, dither_offsets_short, 4)                       \
1541
1542 #define setup_blocks_span_initialize_dithered_untextured()                     \
1543   vec_8x8u dither_offsets;                                                     \
1544   add_8x8b(dither_offsets, dither_offsets_short, d64_4)                        \
1545
1546 #define setup_blocks_span_initialize_dithered(texturing)                       \
1547   u32 dither_row = psx_gpu->dither_table[y & 0x3];                             \
1548   u32 dither_shift = (span_edge_data->left_x & 0x3) * 8;                       \
1549   vec_8x8s dither_offsets_short;                                               \
1550                                                                                \
1551   dither_row =                                                                 \
1552    (dither_row >> dither_shift) | (dither_row << (32 - dither_shift));         \
1553   dup_2x32b(vector_cast(vec_2x32u, dither_offsets_short), dither_row);         \
1554   setup_blocks_span_initialize_dithered_##texturing()                          \
1555
1556 #define setup_blocks_span_initialize_undithered(texturing)                     \
1557
1558
1559 #define setup_blocks_span_initialize_shaded_textured()                         \
1560 {                                                                              \
1561   vec_4x32u block_span;                                                        \
1562   u32 offset = span_edge_data->left_x;                                         \
1563                                                                                \
1564   uvrg = *span_uvrg_offset;                                                    \
1565   mla_scalar_4x32b(uvrg, uvrg_dx, offset);                                     \
1566   b = *span_b_offset;                                                          \
1567   b += b_dx * offset;                                                          \
1568                                                                                \
1569   dup_4x32b(u_block, uvrg.e[0]);                                               \
1570   dup_4x32b(v_block, uvrg.e[1]);                                               \
1571   dup_4x32b(r_block, uvrg.e[2]);                                               \
1572   dup_4x32b(g_block, uvrg.e[3]);                                               \
1573   dup_4x32b(b_block, b);                                                       \
1574                                                                                \
1575   block_span = psx_gpu->u_block_span;                                          \
1576   add_4x32b(u_block, u_block, block_span);                                     \
1577   block_span = psx_gpu->v_block_span;                                          \
1578   add_4x32b(v_block, v_block, block_span);                                     \
1579   block_span = psx_gpu->r_block_span;                                          \
1580   add_4x32b(r_block, r_block, block_span);                                     \
1581   block_span = psx_gpu->g_block_span;                                          \
1582   add_4x32b(g_block, g_block, block_span);                                     \
1583   block_span = psx_gpu->b_block_span;                                          \
1584   add_4x32b(b_block, b_block, block_span);                                     \
1585 }
1586
1587 #define setup_blocks_span_initialize_shaded_untextured()                       \
1588 {                                                                              \
1589   vec_4x32u block_span;                                                        \
1590   u32 offset = span_edge_data->left_x;                                         \
1591                                                                                \
1592   rgb.low = span_uvrg_offset->high;                                            \
1593   rgb.high.e[0] = *span_b_offset;                                              \
1594   mla_scalar_4x32b(rgb, rgb_dx, offset);                                       \
1595                                                                                \
1596   dup_4x32b(r_block, rgb.e[0]);                                                \
1597   dup_4x32b(g_block, rgb.e[1]);                                                \
1598   dup_4x32b(b_block, rgb.e[2]);                                                \
1599                                                                                \
1600   block_span = psx_gpu->r_block_span;                                          \
1601   add_4x32b(r_block, r_block, block_span);                                     \
1602   block_span = psx_gpu->g_block_span;                                          \
1603   add_4x32b(g_block, g_block, block_span);                                     \
1604   block_span = psx_gpu->b_block_span;                                          \
1605   add_4x32b(b_block, b_block, block_span);                                     \
1606 }                                                                              \
1607
1608 #define setup_blocks_span_initialize_unshaded_textured()                       \
1609 {                                                                              \
1610   vec_4x32u block_span;                                                        \
1611   u32 offset = span_edge_data->left_x;                                         \
1612                                                                                \
1613   uv = span_uvrg_offset->low;                                                  \
1614   mla_scalar_2x32b(uv, uv_dx, offset);                                         \
1615                                                                                \
1616   dup_4x32b(u_block, uv.e[0]);                                                 \
1617   dup_4x32b(v_block, uv.e[1]);                                                 \
1618                                                                                \
1619   block_span = psx_gpu->u_block_span;                                          \
1620   add_4x32b(u_block, u_block, block_span);                                     \
1621   block_span = psx_gpu->v_block_span;                                          \
1622   add_4x32b(v_block, v_block, block_span);                                     \
1623 }                                                                              \
1624
1625 #define setup_blocks_span_initialize_unshaded_untextured()                     \
1626
1627
1628 #define setup_blocks_texture_swizzled()                                        \
1629 {                                                                              \
1630   vec_8x8u u_saved = u;                                                        \
1631   sli_8x8b(u, v, 4);                                                           \
1632   sri_8x8b(v, u_saved, 4);                                                     \
1633 }                                                                              \
1634
1635 #define setup_blocks_texture_unswizzled()                                      \
1636
1637 #define setup_blocks_store_shaded_textured(swizzling, dithering, target,       \
1638  edge_type)                                                                    \
1639 {                                                                              \
1640   vec_8x16u u_whole;                                                           \
1641   vec_8x16u v_whole;                                                           \
1642   vec_8x16u r_whole;                                                           \
1643   vec_8x16u g_whole;                                                           \
1644   vec_8x16u b_whole;                                                           \
1645                                                                                \
1646   vec_8x8u u;                                                                  \
1647   vec_8x8u v;                                                                  \
1648   vec_8x8u r;                                                                  \
1649   vec_8x8u g;                                                                  \
1650   vec_8x8u b;                                                                  \
1651   vec_8x16u uv;                                                                \
1652                                                                                \
1653   vec_4x32u dx4;                                                               \
1654   vec_4x32u dx8;                                                               \
1655                                                                                \
1656   shr_narrow_4x32b(u_whole.low, u_block, 16);                                  \
1657   shr_narrow_4x32b(v_whole.low, v_block, 16);                                  \
1658   shr_narrow_4x32b(r_whole.low, r_block, 16);                                  \
1659   shr_narrow_4x32b(g_whole.low, g_block, 16);                                  \
1660   shr_narrow_4x32b(b_whole.low, b_block, 16);                                  \
1661                                                                                \
1662   dup_4x32b(dx4, uvrg_dx4.e[0]);                                               \
1663   add_high_narrow_4x32b(u_whole.high, u_block, dx4);                           \
1664   dup_4x32b(dx4, uvrg_dx4.e[1]);                                               \
1665   add_high_narrow_4x32b(v_whole.high, v_block, dx4);                           \
1666   dup_4x32b(dx4, uvrg_dx4.e[2]);                                               \
1667   add_high_narrow_4x32b(r_whole.high, r_block, dx4);                           \
1668   dup_4x32b(dx4, uvrg_dx4.e[3]);                                               \
1669   add_high_narrow_4x32b(g_whole.high, g_block, dx4);                           \
1670   dup_4x32b(dx4, b_dx4);                                                       \
1671   add_high_narrow_4x32b(b_whole.high, b_block, dx4);                           \
1672                                                                                \
1673   mov_narrow_8x16b(u, u_whole);                                                \
1674   mov_narrow_8x16b(v, v_whole);                                                \
1675   mov_narrow_8x16b(r, r_whole);                                                \
1676   mov_narrow_8x16b(g, g_whole);                                                \
1677   mov_narrow_8x16b(b, b_whole);                                                \
1678                                                                                \
1679   dup_4x32b(dx8, uvrg_dx8.e[0]);                                               \
1680   add_4x32b(u_block, u_block, dx8);                                            \
1681   dup_4x32b(dx8, uvrg_dx8.e[1]);                                               \
1682   add_4x32b(v_block, v_block, dx8);                                            \
1683   dup_4x32b(dx8, uvrg_dx8.e[2]);                                               \
1684   add_4x32b(r_block, r_block, dx8);                                            \
1685   dup_4x32b(dx8, uvrg_dx8.e[3]);                                               \
1686   add_4x32b(g_block, g_block, dx8);                                            \
1687   dup_4x32b(dx8, b_dx8);                                                       \
1688   add_4x32b(b_block, b_block, dx8);                                            \
1689                                                                                \
1690   and_8x8b(u, u, texture_mask.low);                                            \
1691   and_8x8b(v, v, texture_mask.high);                                           \
1692   setup_blocks_texture_##swizzling();                                          \
1693                                                                                \
1694   zip_8x16b(uv, u, v);                                                         \
1695   block->uv = uv;                                                              \
1696   block->r = r;                                                                \
1697   block->g = g;                                                                \
1698   block->b = b;                                                                \
1699   block->dither_offsets = vector_cast(vec_8x16u, dither_offsets);              \
1700   block->fb_ptr = fb_ptr;                                                      \
1701 }                                                                              \
1702
1703 #define setup_blocks_store_unshaded_textured(swizzling, dithering, target,     \
1704  edge_type)                                                                    \
1705 {                                                                              \
1706   vec_8x16u u_whole;                                                           \
1707   vec_8x16u v_whole;                                                           \
1708                                                                                \
1709   vec_8x8u u;                                                                  \
1710   vec_8x8u v;                                                                  \
1711   vec_8x16u uv;                                                                \
1712                                                                                \
1713   vec_4x32u dx4;                                                               \
1714   vec_4x32u dx8;                                                               \
1715                                                                                \
1716   shr_narrow_4x32b(u_whole.low, u_block, 16);                                  \
1717   shr_narrow_4x32b(v_whole.low, v_block, 16);                                  \
1718                                                                                \
1719   dup_4x32b(dx4, uv_dx4.e[0]);                                                 \
1720   add_high_narrow_4x32b(u_whole.high, u_block, dx4);                           \
1721   dup_4x32b(dx4, uv_dx4.e[1]);                                                 \
1722   add_high_narrow_4x32b(v_whole.high, v_block, dx4);                           \
1723                                                                                \
1724   mov_narrow_8x16b(u, u_whole);                                                \
1725   mov_narrow_8x16b(v, v_whole);                                                \
1726                                                                                \
1727   dup_4x32b(dx8, uv_dx8.e[0]);                                                 \
1728   add_4x32b(u_block, u_block, dx8);                                            \
1729   dup_4x32b(dx8, uv_dx8.e[1]);                                                 \
1730   add_4x32b(v_block, v_block, dx8);                                            \
1731                                                                                \
1732   and_8x8b(u, u, texture_mask.low);                                            \
1733   and_8x8b(v, v, texture_mask.high);                                           \
1734   setup_blocks_texture_##swizzling();                                          \
1735                                                                                \
1736   zip_8x16b(uv, u, v);                                                         \
1737   block->uv = uv;                                                              \
1738   block->dither_offsets = vector_cast(vec_8x16u, dither_offsets);              \
1739   block->fb_ptr = fb_ptr;                                                      \
1740 }                                                                              \
1741
1742 #define setup_blocks_store_shaded_untextured_dithered()                        \
1743   addq_8x8b(r, r, dither_offsets);                                             \
1744   addq_8x8b(g, g, dither_offsets);                                             \
1745   addq_8x8b(b, b, dither_offsets);                                             \
1746                                                                                \
1747   subq_8x8b(r, r, d64_4);                                                      \
1748   subq_8x8b(g, g, d64_4);                                                      \
1749   subq_8x8b(b, b, d64_4)                                                       \
1750
1751 #define setup_blocks_store_shaded_untextured_undithered()                      \
1752
1753
1754 #define setup_blocks_store_untextured_pixels_indirect_full(_pixels)            \
1755   block->pixels = _pixels;                                                     \
1756   block->fb_ptr = fb_ptr                                                       \
1757
1758 #define setup_blocks_store_untextured_pixels_indirect_edge(_pixels)            \
1759   block->pixels = _pixels;                                                     \
1760   block->fb_ptr = fb_ptr                                                       \
1761
1762 #define setup_blocks_store_shaded_untextured_seed_pixels_indirect()            \
1763   mul_long_8x8b(pixels, r, d64_1)                                              \
1764
1765
1766 #define setup_blocks_store_untextured_pixels_direct_full(_pixels)              \
1767   store_8x16b(_pixels, fb_ptr)                                                 \
1768
1769 #define setup_blocks_store_untextured_pixels_direct_edge(_pixels)              \
1770 {                                                                              \
1771   vec_8x16u fb_pixels;                                                         \
1772   vec_8x16u draw_mask;                                                         \
1773   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
1774                                                                                \
1775   load_8x16b(fb_pixels, fb_ptr);                                               \
1776   dup_8x16b(draw_mask, span_edge_data->right_mask);                            \
1777   tst_8x16b(draw_mask, draw_mask, test_mask);                                  \
1778   bif_8x16b(fb_pixels, _pixels, draw_mask);                                    \
1779   store_8x16b(fb_pixels, fb_ptr);                                              \
1780 }                                                                              \
1781
1782 #define setup_blocks_store_shaded_untextured_seed_pixels_direct()              \
1783   pixels = msb_mask;                                                           \
1784   mla_long_8x8b(pixels, r, d64_1)                                              \
1785
1786
1787 #define setup_blocks_store_shaded_untextured(swizzling, dithering, target,     \
1788  edge_type)                                                                    \
1789 {                                                                              \
1790   vec_8x16u r_whole;                                                           \
1791   vec_8x16u g_whole;                                                           \
1792   vec_8x16u b_whole;                                                           \
1793                                                                                \
1794   vec_8x8u r;                                                                  \
1795   vec_8x8u g;                                                                  \
1796   vec_8x8u b;                                                                  \
1797                                                                                \
1798   vec_4x32u dx4;                                                               \
1799   vec_4x32u dx8;                                                               \
1800                                                                                \
1801   vec_8x16u pixels;                                                            \
1802                                                                                \
1803   shr_narrow_4x32b(r_whole.low, r_block, 16);                                  \
1804   shr_narrow_4x32b(g_whole.low, g_block, 16);                                  \
1805   shr_narrow_4x32b(b_whole.low, b_block, 16);                                  \
1806                                                                                \
1807   dup_4x32b(dx4, rgb_dx4.e[0]);                                                \
1808   add_high_narrow_4x32b(r_whole.high, r_block, dx4);                           \
1809   dup_4x32b(dx4, rgb_dx4.e[1]);                                                \
1810   add_high_narrow_4x32b(g_whole.high, g_block, dx4);                           \
1811   dup_4x32b(dx4, rgb_dx4.e[2]);                                                \
1812   add_high_narrow_4x32b(b_whole.high, b_block, dx4);                           \
1813                                                                                \
1814   mov_narrow_8x16b(r, r_whole);                                                \
1815   mov_narrow_8x16b(g, g_whole);                                                \
1816   mov_narrow_8x16b(b, b_whole);                                                \
1817                                                                                \
1818   dup_4x32b(dx8, rgb_dx8.e[0]);                                                \
1819   add_4x32b(r_block, r_block, dx8);                                            \
1820   dup_4x32b(dx8, rgb_dx8.e[1]);                                                \
1821   add_4x32b(g_block, g_block, dx8);                                            \
1822   dup_4x32b(dx8, rgb_dx8.e[2]);                                                \
1823   add_4x32b(b_block, b_block, dx8);                                            \
1824                                                                                \
1825   setup_blocks_store_shaded_untextured_##dithering();                          \
1826                                                                                \
1827   shr_8x8b(r, r, 3);                                                           \
1828   bic_8x8b(g, g, d64_0x07);                                                    \
1829   bic_8x8b(b, b, d64_0x07);                                                    \
1830                                                                                \
1831   setup_blocks_store_shaded_untextured_seed_pixels_##target();                 \
1832   mla_long_8x8b(pixels, g, d64_4);                                             \
1833   mla_long_8x8b(pixels, b, d64_128)                                            \
1834                                                                                \
1835   setup_blocks_store_untextured_pixels_##target##_##edge_type(pixels);         \
1836 }                                                                              \
1837
1838 #define setup_blocks_store_unshaded_untextured(swizzling, dithering, target,   \
1839  edge_type)                                                                    \
1840   setup_blocks_store_untextured_pixels_##target##_##edge_type(colors)          \
1841
1842
1843 #define setup_blocks_store_draw_mask_textured_indirect(_block, bits)           \
1844   (_block)->draw_mask_bits = bits                                              \
1845
1846 #define setup_blocks_store_draw_mask_untextured_indirect(_block, bits)         \
1847 {                                                                              \
1848   vec_8x16u bits_mask;                                                         \
1849   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
1850   dup_8x16b(bits_mask, bits);                                                  \
1851   tst_8x16b(bits_mask, bits_mask, test_mask);                                  \
1852   (_block)->draw_mask = bits_mask;                                             \
1853 }                                                                              \
1854
1855 #define setup_blocks_store_draw_mask_untextured_direct(_block, bits)           \
1856
1857
1858 #define setup_blocks_add_blocks_indirect()                                     \
1859   num_blocks += span_num_blocks;                                               \
1860                                                                                \
1861   if(num_blocks > MAX_BLOCKS)                                                  \
1862   {                                                                            \
1863     psx_gpu->num_blocks = num_blocks - span_num_blocks;                        \
1864     flush_render_block_buffer(psx_gpu);                                        \
1865     num_blocks = span_num_blocks;                                              \
1866     block = psx_gpu->blocks;                                                   \
1867   }                                                                            \
1868
1869 #define setup_blocks_add_blocks_direct()                                       \
1870   texel_blocks_untextured += span_num_blocks;                                  \
1871   span_pixel_blocks += span_num_blocks                                         \
1872
1873
1874 #define setup_blocks_builder(shading, texturing, dithering, sw, target)        \
1875 void setup_blocks_##shading##_##texturing##_##dithering##_##sw##_##target(     \
1876  psx_gpu_struct *psx_gpu)                                                      \
1877 {                                                                              \
1878   setup_blocks_load_msb_mask_##target();                                       \
1879   setup_blocks_variables_##shading##_##texturing(target);                      \
1880                                                                                \
1881   edge_data_struct *span_edge_data = psx_gpu->span_edge_data;                  \
1882   vec_4x32u *span_uvrg_offset = psx_gpu->span_uvrg_offset;                     \
1883   u32 *span_b_offset = psx_gpu->span_b_offset;                                 \
1884                                                                                \
1885   block_struct *block = psx_gpu->blocks + psx_gpu->num_blocks;                 \
1886                                                                                \
1887   u32 num_spans = psx_gpu->num_spans;                                          \
1888                                                                                \
1889   u16 *fb_ptr;                                                                 \
1890   u32 y;                                                                       \
1891                                                                                \
1892   u32 num_blocks = psx_gpu->num_blocks;                                        \
1893   u32 span_num_blocks;                                                         \
1894                                                                                \
1895   while(num_spans)                                                             \
1896   {                                                                            \
1897     span_num_blocks = span_edge_data->num_blocks;                              \
1898     if(span_num_blocks)                                                        \
1899     {                                                                          \
1900       y = span_edge_data->y;                                                   \
1901       fb_ptr = psx_gpu->vram_out_ptr + span_edge_data->left_x + (y * 1024);    \
1902                                                                                \
1903       setup_blocks_span_initialize_##shading##_##texturing();                  \
1904       setup_blocks_span_initialize_##dithering(texturing);                     \
1905                                                                                \
1906       setup_blocks_add_blocks_##target();                                      \
1907                                                                                \
1908       s32 pixel_span = span_num_blocks * 8;                                    \
1909       pixel_span -= __builtin_popcount(span_edge_data->right_mask & 0xFF);     \
1910       span_pixels += pixel_span;                                               \
1911                                                                                \
1912       span_num_blocks--;                                                       \
1913       while(span_num_blocks)                                                   \
1914       {                                                                        \
1915         setup_blocks_store_##shading##_##texturing(sw, dithering, target,      \
1916          full);                                                                \
1917         setup_blocks_store_draw_mask_##texturing##_##target(block, 0x00);      \
1918                                                                                \
1919         fb_ptr += 8;                                                           \
1920         block++;                                                               \
1921         span_num_blocks--;                                                     \
1922       }                                                                        \
1923                                                                                \
1924       setup_blocks_store_##shading##_##texturing(sw, dithering, target, edge); \
1925       setup_blocks_store_draw_mask_##texturing##_##target(block,               \
1926        span_edge_data->right_mask);                                            \
1927                                                                                \
1928       block++;                                                                 \
1929     }                                                                          \
1930     else                                                                       \
1931     {                                                                          \
1932       zero_block_spans++;                                                      \
1933     }                                                                          \
1934                                                                                \
1935     num_spans--;                                                               \
1936     span_edge_data++;                                                          \
1937     span_uvrg_offset++;                                                        \
1938     span_b_offset++;                                                           \
1939   }                                                                            \
1940                                                                                \
1941   psx_gpu->num_blocks = num_blocks;                                            \
1942 }                                                                              \
1943
1944 void setup_blocks_shaded_textured_dithered_unswizzled_indirect(psx_gpu_struct
1945  *psx_gpu);
1946
1947 void setup_blocks_shaded_untextured_dithered_unswizzled_indirect(psx_gpu_struct
1948  *psx_gpu);
1949 void setup_blocks_shaded_untextured_undithered_unswizzled_indirect(
1950  psx_gpu_struct *psx_gpu);
1951 void setup_blocks_shaded_untextured_dithered_unswizzled_direct(psx_gpu_struct
1952  *psx_gpu);
1953 void setup_blocks_shaded_untextured_undithered_unswizzled_direct(
1954  psx_gpu_struct *psx_gpu);
1955
1956 void setup_blocks_unshaded_textured_dithered_unswizzled_indirect(psx_gpu_struct
1957  *psx_gpu);
1958 void setup_blocks_unshaded_untextured_undithered_unswizzled_indirect(
1959  psx_gpu_struct *psx_gpu);
1960 void setup_blocks_unshaded_untextured_undithered_unswizzled_direct(
1961  psx_gpu_struct *psx_gpu);
1962
1963 void setup_blocks_shaded_textured_dithered_swizzled_indirect(psx_gpu_struct
1964  *psx_gpu);
1965 void setup_blocks_unshaded_textured_dithered_swizzled_indirect(psx_gpu_struct
1966  *psx_gpu);
1967
1968
1969 //setup_blocks_builder(unshaded, untextured, undithered, unswizzled, direct);
1970
1971 #ifndef NEON_BUILD
1972
1973 setup_blocks_builder(shaded, textured, dithered, swizzled, indirect);
1974 setup_blocks_builder(shaded, textured, dithered, unswizzled, indirect);
1975
1976 setup_blocks_builder(unshaded, textured, dithered, unswizzled, indirect);
1977 setup_blocks_builder(unshaded, textured, dithered, swizzled, indirect);
1978
1979 setup_blocks_builder(shaded, untextured, undithered, unswizzled, indirect);
1980 setup_blocks_builder(shaded, untextured, dithered, unswizzled, indirect);
1981 setup_blocks_builder(shaded, untextured, undithered, unswizzled, direct);
1982 setup_blocks_builder(shaded, untextured, dithered, unswizzled, direct);
1983
1984 setup_blocks_builder(unshaded, untextured, undithered, unswizzled, indirect);
1985 setup_blocks_builder(unshaded, untextured, undithered, unswizzled, direct);
1986
1987 #endif
1988
1989 void texture_blocks_untextured(psx_gpu_struct *psx_gpu);
1990 void texture_blocks_4bpp(psx_gpu_struct *psx_gpu);
1991 void texture_blocks_8bpp(psx_gpu_struct *psx_gpu);
1992 void texture_blocks_16bpp(psx_gpu_struct *psx_gpu);
1993
1994 #ifndef NEON_BUILD
1995
1996 void texture_blocks_untextured(psx_gpu_struct *psx_gpu)
1997 {
1998   if(psx_gpu->primitive_type != PRIMITIVE_TYPE_SPRITE)
1999     texel_blocks_untextured += psx_gpu->num_blocks;
2000 }
2001
2002 void texture_blocks_4bpp(psx_gpu_struct *psx_gpu)
2003 {
2004   block_struct *block = psx_gpu->blocks;
2005   u32 num_blocks = psx_gpu->num_blocks;
2006   texel_blocks_4bpp += num_blocks;
2007
2008   vec_8x8u texels_low;
2009   vec_8x8u texels_high;
2010   vec_8x8u texels;
2011   vec_8x16u pixels;
2012
2013   vec_8x16u clut_a;
2014   vec_8x16u clut_b;
2015   vec_16x8u clut_low;
2016   vec_16x8u clut_high;
2017
2018   u8 *texture_ptr_8bpp = psx_gpu->texture_page_ptr;
2019   u16 *clut_ptr = psx_gpu->clut_ptr;
2020
2021   // Can be done with one deinterleaving load on NEON
2022   load_8x16b(clut_a, clut_ptr);
2023   load_8x16b(clut_b, clut_ptr + 8);
2024   unzip_16x8b(clut_low, clut_high, clut_a, clut_b);
2025
2026   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_4bpp_mask)
2027     update_texture_4bpp_cache(psx_gpu);
2028
2029   while(num_blocks)
2030   {
2031     texels.e[0] = texture_ptr_8bpp[block->uv.e[0]];
2032     texels.e[1] = texture_ptr_8bpp[block->uv.e[1]];
2033     texels.e[2] = texture_ptr_8bpp[block->uv.e[2]];
2034     texels.e[3] = texture_ptr_8bpp[block->uv.e[3]];
2035     texels.e[4] = texture_ptr_8bpp[block->uv.e[4]];
2036     texels.e[5] = texture_ptr_8bpp[block->uv.e[5]];
2037     texels.e[6] = texture_ptr_8bpp[block->uv.e[6]];
2038     texels.e[7] = texture_ptr_8bpp[block->uv.e[7]];
2039
2040     tbl_16(texels_low, texels, clut_low);
2041     tbl_16(texels_high, texels, clut_high);
2042
2043     // Can be done with an interleaving store on NEON
2044     zip_8x16b(pixels, texels_low, texels_high);
2045
2046     block->texels = pixels;
2047
2048     num_blocks--;
2049     block++;
2050   }
2051 }
2052
2053 void texture_blocks_8bpp(psx_gpu_struct *psx_gpu)
2054 {
2055   block_struct *block = psx_gpu->blocks;
2056   u32 num_blocks = psx_gpu->num_blocks;
2057
2058   texel_blocks_8bpp += num_blocks;
2059
2060   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_8bpp_mask)
2061     update_texture_8bpp_cache(psx_gpu);
2062
2063   vec_8x16u texels;
2064   u8 *texture_ptr_8bpp = psx_gpu->texture_page_ptr;
2065
2066   u32 texel;
2067   u32 offset;
2068   u32 i;
2069
2070   while(num_blocks)
2071   {
2072     for(i = 0; i < 8; i++)
2073     {
2074       offset = block->uv.e[i];
2075
2076       texel = texture_ptr_8bpp[offset];
2077       texels.e[i] = psx_gpu->clut_ptr[texel];
2078     }
2079
2080     block->texels = texels;
2081
2082     num_blocks--;
2083     block++;
2084   }
2085 }
2086
2087 void texture_blocks_16bpp(psx_gpu_struct *psx_gpu)
2088 {
2089   block_struct *block = psx_gpu->blocks;
2090   u32 num_blocks = psx_gpu->num_blocks;
2091
2092   texel_blocks_16bpp += num_blocks;
2093
2094   vec_8x16u texels;
2095
2096   u16 *texture_ptr_16bpp = psx_gpu->texture_page_ptr;
2097   u32 offset;
2098   u32 i;
2099
2100   while(num_blocks)
2101   {
2102     for(i = 0; i < 8; i++)
2103     {
2104       offset = block->uv.e[i];
2105       offset += ((offset & 0xFF00) * 3);
2106
2107       texels.e[i] = texture_ptr_16bpp[offset];
2108     }
2109
2110     block->texels = texels;
2111
2112     num_blocks--;
2113     block++;
2114   }
2115 }
2116
2117 #endif
2118
2119
2120 #define shade_blocks_load_msb_mask_indirect()                                  \
2121
2122 #define shade_blocks_load_msb_mask_direct()                                    \
2123   vec_8x16u msb_mask;                                                          \
2124   dup_8x16b(msb_mask, psx_gpu->mask_msb);                                      \
2125
2126 #define shade_blocks_store_indirect(_draw_mask, _pixels)                       \
2127   block->draw_mask = _draw_mask;                                               \
2128   block->pixels = _pixels                                                      \
2129
2130 #define shade_blocks_store_direct(_draw_mask, _pixels)                         \
2131 {                                                                              \
2132   vec_8x16u fb_pixels;                                                         \
2133   or_8x16b(_pixels, _pixels, msb_mask);                                        \
2134   load_8x16b(fb_pixels, block->fb_ptr);                                        \
2135   bif_8x16b(fb_pixels, _pixels, _draw_mask);                                   \
2136   store_8x16b(fb_pixels, block->fb_ptr);                                       \
2137 }                                                                              \
2138
2139
2140 #define shade_blocks_textured_false_modulated_check_dithered(target)           \
2141   if(psx_gpu->triangle_color == 0x808080)                                      \
2142   {                                                                            \
2143     false_modulated_blocks += num_blocks;                                      \
2144   }                                                                            \
2145
2146 #define shade_blocks_textured_false_modulated_check_undithered(target)         \
2147   if(psx_gpu->triangle_color == 0x808080)                                      \
2148   {                                                                            \
2149                                                                                \
2150     shade_blocks_textured_unmodulated_##target(psx_gpu);                       \
2151     false_modulated_blocks += num_blocks;                                      \
2152     return;                                                                    \
2153   }                                                                            \
2154
2155
2156 #define shade_blocks_textured_modulated_shaded_primitive_load(dithering,       \
2157  target)                                                                       \
2158
2159 #define shade_blocks_textured_modulated_unshaded_primitive_load(dithering,     \
2160  target)                                                                       \
2161 {                                                                              \
2162   u32 color = psx_gpu->triangle_color;                                         \
2163   dup_8x8b(colors_r, color);                                                   \
2164   dup_8x8b(colors_g, color >> 8);                                              \
2165   dup_8x8b(colors_b, color >> 16);                                             \
2166   shade_blocks_textured_false_modulated_check_##dithering(target);             \
2167 }                                                                              \
2168
2169 #define shade_blocks_textured_modulated_shaded_block_load()                    \
2170   colors_r = block->r;                                                         \
2171   colors_g = block->g;                                                         \
2172   colors_b = block->b                                                          \
2173
2174 #define shade_blocks_textured_modulated_unshaded_block_load()                  \
2175
2176 #define shade_blocks_textured_modulate_dithered(component)                     \
2177   pixels_##component = block->dither_offsets;                                  \
2178   mla_long_8x8b(pixels_##component, texels_##component, colors_##component)    \
2179
2180 #define shade_blocks_textured_modulate_undithered(component)                   \
2181   mul_long_8x8b(pixels_##component, texels_##component, colors_##component)    \
2182
2183 #define shade_blocks_textured_modulated_builder(shading, dithering, target)    \
2184 void shade_blocks_##shading##_textured_modulated_##dithering##_##target(       \
2185  psx_gpu_struct *psx_gpu)                                                      \
2186 {                                                                              \
2187   block_struct *block = psx_gpu->blocks;                                       \
2188   u32 num_blocks = psx_gpu->num_blocks;                                        \
2189   vec_8x16u texels;                                                            \
2190                                                                                \
2191   vec_8x8u texels_r;                                                           \
2192   vec_8x8u texels_g;                                                           \
2193   vec_8x8u texels_b;                                                           \
2194                                                                                \
2195   vec_8x8u colors_r;                                                           \
2196   vec_8x8u colors_g;                                                           \
2197   vec_8x8u colors_b;                                                           \
2198                                                                                \
2199   vec_8x8u pixels_r_low;                                                       \
2200   vec_8x8u pixels_g_low;                                                       \
2201   vec_8x8u pixels_b_low;                                                       \
2202   vec_8x16u pixels;                                                            \
2203                                                                                \
2204   vec_8x16u pixels_r;                                                          \
2205   vec_8x16u pixels_g;                                                          \
2206   vec_8x16u pixels_b;                                                          \
2207                                                                                \
2208   vec_8x16u draw_mask;                                                         \
2209   vec_8x16u zero_mask;                                                         \
2210                                                                                \
2211   vec_8x8u d64_0x07;                                                           \
2212   vec_8x8u d64_0x1F;                                                           \
2213   vec_8x8u d64_1;                                                              \
2214   vec_8x8u d64_4;                                                              \
2215   vec_8x8u d64_128;                                                            \
2216                                                                                \
2217   vec_8x16u d128_0x8000;                                                       \
2218                                                                                \
2219   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
2220   u32 draw_mask_bits;                                                          \
2221   shade_blocks_load_msb_mask_##target();                                       \
2222                                                                                \
2223   dup_8x8b(d64_0x07, 0x07);                                                    \
2224   dup_8x8b(d64_0x1F, 0x1F);                                                    \
2225   dup_8x8b(d64_1, 1);                                                          \
2226   dup_8x8b(d64_4, 4);                                                          \
2227   dup_8x8b(d64_128, 128);                                                      \
2228                                                                                \
2229   dup_8x16b(d128_0x8000, 0x8000);                                              \
2230                                                                                \
2231   shade_blocks_textured_modulated_##shading##_primitive_load(dithering,        \
2232    target);                                                                    \
2233                                                                                \
2234   while(num_blocks)                                                            \
2235   {                                                                            \
2236     draw_mask_bits = block->draw_mask_bits;                                    \
2237     dup_8x16b(draw_mask, draw_mask_bits);                                      \
2238     tst_8x16b(draw_mask, draw_mask, test_mask);                                \
2239                                                                                \
2240     shade_blocks_textured_modulated_##shading##_block_load();                  \
2241                                                                                \
2242     texels = block->texels;                                                    \
2243                                                                                \
2244     mov_narrow_8x16b(texels_r, texels);                                        \
2245     shr_narrow_8x16b(texels_g, texels, 5);                                     \
2246     shr_narrow_8x16b(texels_b, texels, 7);                                     \
2247                                                                                \
2248     and_8x8b(texels_r, texels_r, d64_0x1F);                                    \
2249     and_8x8b(texels_g, texels_g, d64_0x1F);                                    \
2250     shr_8x8b(texels_b, texels_b, 3);                                           \
2251                                                                                \
2252     shade_blocks_textured_modulate_##dithering(r);                             \
2253     shade_blocks_textured_modulate_##dithering(g);                             \
2254     shade_blocks_textured_modulate_##dithering(b);                             \
2255                                                                                \
2256     cmpeqz_8x16b(zero_mask, texels);                                           \
2257     and_8x16b(pixels, texels, d128_0x8000);                                    \
2258                                                                                \
2259     shrq_narrow_signed_8x16b(pixels_r_low, pixels_r, 4);                       \
2260     shrq_narrow_signed_8x16b(pixels_g_low, pixels_g, 4);                       \
2261     shrq_narrow_signed_8x16b(pixels_b_low, pixels_b, 4);                       \
2262                                                                                \
2263     or_8x16b(zero_mask, draw_mask, zero_mask);                                 \
2264                                                                                \
2265     shr_8x8b(pixels_r_low, pixels_r_low, 3);                                   \
2266     bic_8x8b(pixels_g_low, pixels_g_low, d64_0x07);                            \
2267     bic_8x8b(pixels_b_low, pixels_b_low, d64_0x07);                            \
2268                                                                                \
2269     mla_long_8x8b(pixels, pixels_r_low, d64_1);                                \
2270     mla_long_8x8b(pixels, pixels_g_low, d64_4);                                \
2271     mla_long_8x8b(pixels, pixels_b_low, d64_128);                              \
2272                                                                                \
2273     shade_blocks_store_##target(zero_mask, pixels);                            \
2274                                                                                \
2275     num_blocks--;                                                              \
2276     block++;                                                                   \
2277   }                                                                            \
2278 }                                                                              \
2279
2280 void shade_blocks_shaded_textured_modulated_dithered_direct(psx_gpu_struct
2281  *psx_gpu);
2282 void shade_blocks_shaded_textured_modulated_undithered_direct(psx_gpu_struct
2283  *psx_gpu);
2284 void shade_blocks_unshaded_textured_modulated_dithered_direct(psx_gpu_struct
2285  *psx_gpu);
2286 void shade_blocks_unshaded_textured_modulated_undithered_direct(psx_gpu_struct
2287  *psx_gpu);
2288
2289 void shade_blocks_shaded_textured_modulated_dithered_indirect(psx_gpu_struct
2290  *psx_gpu);
2291 void shade_blocks_shaded_textured_modulated_undithered_indirect(psx_gpu_struct
2292  *psx_gpu);
2293 void shade_blocks_unshaded_textured_modulated_dithered_indirect(psx_gpu_struct
2294  *psx_gpu);
2295 void shade_blocks_unshaded_textured_modulated_undithered_indirect(psx_gpu_struct
2296  *psx_gpu);
2297
2298 void shade_blocks_textured_unmodulated_indirect(psx_gpu_struct *psx_gpu);
2299 void shade_blocks_textured_unmodulated_direct(psx_gpu_struct *psx_gpu);
2300
2301 #ifndef NEON_BUILD
2302
2303 shade_blocks_textured_modulated_builder(shaded, dithered, direct);
2304 shade_blocks_textured_modulated_builder(shaded, undithered, direct);
2305 shade_blocks_textured_modulated_builder(unshaded, dithered, direct);
2306 shade_blocks_textured_modulated_builder(unshaded, undithered, direct);
2307
2308 shade_blocks_textured_modulated_builder(shaded, dithered, indirect);
2309 shade_blocks_textured_modulated_builder(shaded, undithered, indirect);
2310 shade_blocks_textured_modulated_builder(unshaded, dithered, indirect);
2311 shade_blocks_textured_modulated_builder(unshaded, undithered, indirect);
2312
2313 #endif
2314
2315
2316 #define shade_blocks_textured_unmodulated_builder(target)                      \
2317 void shade_blocks_textured_unmodulated_##target(psx_gpu_struct *psx_gpu)       \
2318 {                                                                              \
2319   block_struct *block = psx_gpu->blocks;                                       \
2320   u32 num_blocks = psx_gpu->num_blocks;                                        \
2321   vec_8x16u draw_mask;                                                         \
2322   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
2323   u32 draw_mask_bits;                                                          \
2324                                                                                \
2325   vec_8x16u pixels;                                                            \
2326   shade_blocks_load_msb_mask_##target();                                       \
2327                                                                                \
2328   while(num_blocks)                                                            \
2329   {                                                                            \
2330     vec_8x16u zero_mask;                                                       \
2331                                                                                \
2332     draw_mask_bits = block->draw_mask_bits;                                    \
2333     dup_8x16b(draw_mask, draw_mask_bits);                                      \
2334     tst_8x16b(draw_mask, draw_mask, test_mask);                                \
2335                                                                                \
2336     pixels = block->texels;                                                    \
2337                                                                                \
2338     cmpeqz_8x16b(zero_mask, pixels);                                           \
2339     or_8x16b(zero_mask, draw_mask, zero_mask);                                 \
2340                                                                                \
2341     shade_blocks_store_##target(zero_mask, pixels);                            \
2342                                                                                \
2343     num_blocks--;                                                              \
2344     block++;                                                                   \
2345   }                                                                            \
2346 }                                                                              \
2347
2348 #define shade_blocks_textured_unmodulated_dithered_builder(target)             \
2349 void shade_blocks_textured_unmodulated_dithered_##target(psx_gpu_struct        \
2350  *psx_gpu)                                                                     \
2351 {                                                                              \
2352   block_struct *block = psx_gpu->blocks;                                       \
2353   u32 num_blocks = psx_gpu->num_blocks;                                        \
2354   vec_8x16u draw_mask;                                                         \
2355   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
2356   u32 draw_mask_bits;                                                          \
2357                                                                                \
2358   vec_8x16u pixels;                                                            \
2359   shade_blocks_load_msb_mask_##target();                                       \
2360                                                                                \
2361   while(num_blocks)                                                            \
2362   {                                                                            \
2363     vec_8x16u zero_mask;                                                       \
2364                                                                                \
2365     draw_mask_bits = block->draw_mask_bits;                                    \
2366     dup_8x16b(draw_mask, draw_mask_bits);                                      \
2367     tst_8x16b(draw_mask, draw_mask, test_mask);                                \
2368                                                                                \
2369     pixels = block->texels;                                                    \
2370                                                                                \
2371     cmpeqz_8x16b(zero_mask, pixels);                                           \
2372     or_8x16b(zero_mask, draw_mask, zero_mask);                                 \
2373                                                                                \
2374     shade_blocks_store_##target(zero_mask, pixels);                            \
2375                                                                                \
2376     num_blocks--;                                                              \
2377     block++;                                                                   \
2378   }                                                                            \
2379 }                                                                              \
2380
2381 #ifndef NEON_BUILD
2382
2383 shade_blocks_textured_unmodulated_builder(indirect)
2384 shade_blocks_textured_unmodulated_builder(direct)
2385
2386 #endif
2387
2388
2389 void shade_blocks_unshaded_untextured_indirect(psx_gpu_struct *psx_gpu);
2390 void shade_blocks_unshaded_untextured_direct(psx_gpu_struct *psx_gpu);
2391
2392 #ifndef NEON_BUILD
2393
2394 void shade_blocks_unshaded_untextured_indirect(psx_gpu_struct *psx_gpu)
2395 {
2396 }
2397
2398 void shade_blocks_unshaded_untextured_direct(psx_gpu_struct *psx_gpu)
2399 {
2400   block_struct *block = psx_gpu->blocks;
2401   u32 num_blocks = psx_gpu->num_blocks;
2402
2403   vec_8x16u pixels = block->pixels;
2404   shade_blocks_load_msb_mask_direct();
2405
2406   while(num_blocks)
2407   {
2408     shade_blocks_store_direct(block->draw_mask, pixels);
2409
2410     num_blocks--;
2411     block++;
2412   }
2413 }
2414
2415 #endif
2416
2417 void shade_blocks_shaded_untextured(psx_gpu_struct *psx_gpu)
2418 {
2419 }
2420
2421
2422 #define blend_blocks_mask_evaluate_on()                                        \
2423   vec_8x16u mask_pixels;                                                       \
2424   cmpltz_8x16b(mask_pixels, framebuffer_pixels);                               \
2425   or_8x16b(draw_mask, draw_mask, mask_pixels)                                  \
2426
2427 #define blend_blocks_mask_evaluate_off()                                       \
2428
2429 #define blend_blocks_average()                                                 \
2430 {                                                                              \
2431   vec_8x16u pixels_no_msb;                                                     \
2432   vec_8x16u fb_pixels_no_msb;                                                  \
2433                                                                                \
2434   vec_8x16u d128_0x0421;                                                       \
2435   vec_8x16u d128_0x8000;                                                       \
2436                                                                                \
2437   dup_8x16b(d128_0x0421, 0x0421);                                              \
2438   dup_8x16b(d128_0x8000, 0x8000);                                              \
2439                                                                                \
2440   eor_8x16b(blend_pixels, pixels, framebuffer_pixels);                         \
2441   bic_8x16b(pixels_no_msb, pixels, d128_0x8000);                               \
2442   and_8x16b(blend_pixels, blend_pixels, d128_0x0421);                          \
2443   sub_8x16b(blend_pixels, pixels_no_msb, blend_pixels);                        \
2444   bic_8x16b(fb_pixels_no_msb, framebuffer_pixels, d128_0x8000);                \
2445   average_8x16b(blend_pixels, fb_pixels_no_msb, blend_pixels);                 \
2446 }                                                                              \
2447
2448 #define blend_blocks_add()                                                     \
2449 {                                                                              \
2450   vec_8x16u pixels_rb, pixels_g;                                               \
2451   vec_8x16u fb_rb, fb_g;                                                       \
2452                                                                                \
2453   vec_8x16u d128_0x7C1F;                                                       \
2454   vec_8x16u d128_0x03E0;                                                       \
2455                                                                                \
2456   dup_8x16b(d128_0x7C1F, 0x7C1F);                                              \
2457   dup_8x16b(d128_0x03E0, 0x03E0);                                              \
2458                                                                                \
2459   and_8x16b(pixels_rb, pixels, d128_0x7C1F);                                   \
2460   and_8x16b(pixels_g, pixels, d128_0x03E0);                                    \
2461                                                                                \
2462   and_8x16b(fb_rb, framebuffer_pixels, d128_0x7C1F);                           \
2463   and_8x16b(fb_g, framebuffer_pixels, d128_0x03E0);                            \
2464                                                                                \
2465   add_8x16b(fb_rb, fb_rb, pixels_rb);                                          \
2466   add_8x16b(fb_g, fb_g, pixels_g);                                             \
2467                                                                                \
2468   min_16x8b(vector_cast(vec_16x8u, fb_rb), vector_cast(vec_16x8u, fb_rb),      \
2469    vector_cast(vec_16x8u, d128_0x7C1F));                                       \
2470   min_8x16b(fb_g, fb_g, d128_0x03E0);                                          \
2471                                                                                \
2472   or_8x16b(blend_pixels, fb_rb, fb_g);                                         \
2473 }                                                                              \
2474
2475 #define blend_blocks_subtract()                                                \
2476 {                                                                              \
2477   vec_8x16u pixels_rb, pixels_g;                                               \
2478   vec_8x16u fb_rb, fb_g;                                                       \
2479                                                                                \
2480   vec_8x16u d128_0x7C1F;                                                       \
2481   vec_8x16u d128_0x03E0;                                                       \
2482                                                                                \
2483   dup_8x16b(d128_0x7C1F, 0x7C1F);                                              \
2484   dup_8x16b(d128_0x03E0, 0x03E0);                                              \
2485                                                                                \
2486   and_8x16b(pixels_rb, pixels, d128_0x7C1F);                                   \
2487   and_8x16b(pixels_g, pixels, d128_0x03E0);                                    \
2488                                                                                \
2489   and_8x16b(fb_rb, framebuffer_pixels, d128_0x7C1F);                           \
2490   and_8x16b(fb_g, framebuffer_pixels, d128_0x03E0);                            \
2491                                                                                \
2492   subs_16x8b(vector_cast(vec_16x8u, fb_rb),                                    \
2493    vector_cast(vec_16x8u, fb_rb), vector_cast(vec_16x8u, pixels_rb));          \
2494   subs_8x16b(fb_g, fb_g, pixels_g);                                            \
2495                                                                                \
2496   or_8x16b(blend_pixels, fb_rb, fb_g);                                         \
2497 }                                                                              \
2498
2499 #define blend_blocks_add_fourth()                                              \
2500 {                                                                              \
2501   vec_8x16u pixels_rb, pixels_g;                                               \
2502   vec_8x16u pixels_fourth;                                                     \
2503   vec_8x16u fb_rb, fb_g;                                                       \
2504                                                                                \
2505   vec_8x16u d128_0x7C1F;                                                       \
2506   vec_8x16u d128_0x1C07;                                                       \
2507   vec_8x16u d128_0x03E0;                                                       \
2508   vec_8x16u d128_0x00E0;                                                       \
2509                                                                                \
2510   dup_8x16b(d128_0x7C1F, 0x7C1F);                                              \
2511   dup_8x16b(d128_0x1C07, 0x1C07);                                              \
2512   dup_8x16b(d128_0x03E0, 0x03E0);                                              \
2513   dup_8x16b(d128_0x00E0, 0x00E0);                                              \
2514                                                                                \
2515   shr_8x16b(pixels_fourth, vector_cast(vec_8x16s, pixels), 2);                 \
2516                                                                                \
2517   and_8x16b(fb_rb, framebuffer_pixels, d128_0x7C1F);                           \
2518   and_8x16b(fb_g, framebuffer_pixels, d128_0x03E0);                            \
2519                                                                                \
2520   and_8x16b(pixels_rb, pixels_fourth, d128_0x1C07);                            \
2521   and_8x16b(pixels_g, pixels_fourth, d128_0x00E0);                             \
2522                                                                                \
2523   add_8x16b(fb_rb, fb_rb, pixels_rb);                                          \
2524   add_8x16b(fb_g, fb_g, pixels_g);                                             \
2525                                                                                \
2526   min_16x8b(vector_cast(vec_16x8u, fb_rb), vector_cast(vec_16x8u, fb_rb),      \
2527    vector_cast(vec_16x8u, d128_0x7C1F));                                       \
2528   min_8x16b(fb_g, fb_g, d128_0x03E0);                                          \
2529                                                                                \
2530   or_8x16b(blend_pixels, fb_rb, fb_g);                                         \
2531 }                                                                              \
2532
2533 #define blend_blocks_blended_combine_textured()                                \
2534 {                                                                              \
2535   vec_8x16u blend_mask;                                                        \
2536   cmpltz_8x16b(blend_mask, pixels);                                            \
2537                                                                                \
2538   or_immediate_8x16b(blend_pixels, blend_pixels, 0x8000);                      \
2539   bif_8x16b(blend_pixels, pixels, blend_mask);                                 \
2540 }                                                                              \
2541
2542 #define blend_blocks_blended_combine_untextured()                              \
2543
2544
2545 #define blend_blocks_body_blend(blend_mode, texturing)                         \
2546 {                                                                              \
2547   blend_blocks_##blend_mode();                                                 \
2548   blend_blocks_blended_combine_##texturing();                                  \
2549 }                                                                              \
2550
2551 #define blend_blocks_body_average(texturing)                                   \
2552   blend_blocks_body_blend(average, texturing)                                  \
2553
2554 #define blend_blocks_body_add(texturing)                                       \
2555   blend_blocks_body_blend(add, texturing)                                      \
2556
2557 #define blend_blocks_body_subtract(texturing)                                  \
2558   blend_blocks_body_blend(subtract, texturing)                                 \
2559
2560 #define blend_blocks_body_add_fourth(texturing)                                \
2561   blend_blocks_body_blend(add_fourth, texturing)                               \
2562
2563 #define blend_blocks_body_unblended(texturing)                                 \
2564   blend_pixels = pixels                                                        \
2565
2566
2567 #define blend_blocks_builder(texturing, blend_mode, mask_evaluate)             \
2568 void                                                                           \
2569  blend_blocks_##texturing##_##blend_mode##_##mask_evaluate(psx_gpu_struct      \
2570  *psx_gpu)                                                                     \
2571 {                                                                              \
2572   block_struct *block = psx_gpu->blocks;                                       \
2573   u32 num_blocks = psx_gpu->num_blocks;                                        \
2574   vec_8x16u draw_mask;                                                         \
2575   vec_8x16u pixels;                                                            \
2576   vec_8x16u blend_pixels;                                                      \
2577   vec_8x16u framebuffer_pixels;                                                \
2578   vec_8x16u msb_mask;                                                          \
2579                                                                                \
2580   u16 *fb_ptr;                                                                 \
2581                                                                                \
2582   dup_8x16b(msb_mask, psx_gpu->mask_msb);                                      \
2583                                                                                \
2584   while(num_blocks)                                                            \
2585   {                                                                            \
2586     pixels = block->pixels;                                                    \
2587     draw_mask = block->draw_mask;                                              \
2588     fb_ptr = block->fb_ptr;                                                    \
2589                                                                                \
2590     load_8x16b(framebuffer_pixels, fb_ptr);                                    \
2591                                                                                \
2592     blend_blocks_mask_evaluate_##mask_evaluate();                              \
2593     blend_blocks_body_##blend_mode(texturing);                                 \
2594                                                                                \
2595     or_8x16b(blend_pixels, blend_pixels, msb_mask);                            \
2596     bif_8x16b(framebuffer_pixels, blend_pixels, draw_mask);                    \
2597     store_8x16b(framebuffer_pixels, fb_ptr);                                   \
2598                                                                                \
2599     blend_blocks++;                                                            \
2600     num_blocks--;                                                              \
2601     block++;                                                                   \
2602   }                                                                            \
2603 }                                                                              \
2604
2605 void blend_blocks_textured_average_off(psx_gpu_struct *psx_gpu);
2606 void blend_blocks_textured_average_on(psx_gpu_struct *psx_gpu);
2607 void blend_blocks_textured_add_off(psx_gpu_struct *psx_gpu);
2608 void blend_blocks_textured_add_on(psx_gpu_struct *psx_gpu);
2609 void blend_blocks_textured_subtract_off(psx_gpu_struct *psx_gpu);
2610 void blend_blocks_textured_subtract_on(psx_gpu_struct *psx_gpu);
2611 void blend_blocks_textured_add_fourth_off(psx_gpu_struct *psx_gpu);
2612 void blend_blocks_textured_add_fourth_on(psx_gpu_struct *psx_gpu);
2613
2614 void blend_blocks_untextured_average_off(psx_gpu_struct *psx_gpu);
2615 void blend_blocks_untextured_average_on(psx_gpu_struct *psx_gpu);
2616 void blend_blocks_untextured_add_off(psx_gpu_struct *psx_gpu);
2617 void blend_blocks_untextured_add_on(psx_gpu_struct *psx_gpu);
2618 void blend_blocks_untextured_subtract_off(psx_gpu_struct *psx_gpu);
2619 void blend_blocks_untextured_subtract_on(psx_gpu_struct *psx_gpu);
2620 void blend_blocks_untextured_add_fourth_off(psx_gpu_struct *psx_gpu);
2621 void blend_blocks_untextured_add_fourth_on(psx_gpu_struct *psx_gpu);
2622
2623 void blend_blocks_textured_unblended_off(psx_gpu_struct *psx_gpu);
2624 void blend_blocks_textured_unblended_on(psx_gpu_struct *psx_gpu);
2625
2626 #ifndef NEON_BUILD
2627
2628 void blend_blocks_textured_unblended_off(psx_gpu_struct *psx_gpu)
2629 {
2630 }
2631
2632 blend_blocks_builder(textured, average, off);
2633 blend_blocks_builder(textured, average, on);
2634 blend_blocks_builder(textured, add, off);
2635 blend_blocks_builder(textured, add, on);
2636 blend_blocks_builder(textured, subtract, off);
2637 blend_blocks_builder(textured, subtract, on);
2638 blend_blocks_builder(textured, add_fourth, off);
2639 blend_blocks_builder(textured, add_fourth, on);
2640
2641 blend_blocks_builder(untextured, average, off);
2642 blend_blocks_builder(untextured, average, on);
2643 blend_blocks_builder(untextured, add, off);
2644 blend_blocks_builder(untextured, add, on);
2645 blend_blocks_builder(untextured, subtract, off);
2646 blend_blocks_builder(untextured, subtract, on);
2647 blend_blocks_builder(untextured, add_fourth, off);
2648 blend_blocks_builder(untextured, add_fourth, on);
2649
2650 blend_blocks_builder(textured, unblended, on);
2651
2652 #endif
2653
2654
2655 #define vertex_swap(_a, _b)                                                    \
2656 {                                                                              \
2657   vertex_struct *temp_vertex = _a;                                             \
2658   _a = _b;                                                                     \
2659   _b = temp_vertex;                                                            \
2660   triangle_winding ^= 1;                                                       \
2661 }                                                                              \
2662
2663
2664 // Setup blocks parametric-variables:
2665 // SHADE  TEXTURE_MAP SWIZZLING
2666 // 0      0           x
2667 // 0      1           0
2668 // 0      1           1
2669 // 1      0           x
2670 // 1      1           0
2671 // 1      1           1
2672 // 8 inputs, 6 combinations
2673
2674 #define setup_blocks_switch_untextured_unshaded(dithering, target)             \
2675   setup_blocks_unshaded_untextured_undithered_unswizzled_##target              \
2676
2677 #define setup_blocks_switch_untextured_shaded(dithering, target)               \
2678   setup_blocks_shaded_untextured_##dithering##_unswizzled_##target             \
2679
2680 #define setup_blocks_switch_untextured(shading, texture_mode, dithering,       \
2681  target)                                                                       \
2682   setup_blocks_switch_untextured_##shading(dithering, target)                  \
2683
2684 #define setup_blocks_switch_texture_mode_4bpp(shading)                         \
2685   setup_blocks_##shading##_textured_dithered_swizzled_indirect                 \
2686
2687 #define setup_blocks_switch_texture_mode_8bpp(shading)                         \
2688   setup_blocks_##shading##_textured_dithered_swizzled_indirect                 \
2689
2690 #define setup_blocks_switch_texture_mode_16bpp(shading)                        \
2691   setup_blocks_##shading##_textured_dithered_unswizzled_indirect               \
2692
2693 #define setup_blocks_switch_textured(shading, texture_mode, dithering, target) \
2694   setup_blocks_switch_texture_mode_##texture_mode(shading)                     \
2695
2696 #define setup_blocks_switch_blended(shading, texturing, texture_mode,          \
2697  dithering, mask_evaluate)                                                     \
2698   setup_blocks_switch_##texturing(shading, texture_mode, dithering, indirect)  \
2699
2700 #define setup_blocks_switch_unblended_on(shading, texturing, texture_mode,     \
2701  dithering)                                                                    \
2702   setup_blocks_switch_##texturing(shading, texture_mode, dithering, indirect)  \
2703
2704 #define setup_blocks_switch_unblended_off(shading, texturing, texture_mode,    \
2705  dithering)                                                                    \
2706   setup_blocks_switch_##texturing(shading, texture_mode, dithering, direct)    \
2707
2708 #define setup_blocks_switch_unblended(shading, texturing, texture_mode,        \
2709  dithering, mask_evaluate)                                                     \
2710   setup_blocks_switch_unblended_##mask_evaluate(shading, texturing,            \
2711    texture_mode, dithering)                                                    \
2712
2713 #define setup_blocks_switch(shading, texturing, texture_mode, dithering,       \
2714  blending, mask_evaluate)                                                      \
2715   setup_blocks_switch_##blending(shading, texturing, texture_mode,             \
2716    dithering, mask_evaluate)                                                   \
2717
2718
2719 // Texture blocks:
2720
2721 #define texture_blocks_switch_untextured(texture_mode)                         \
2722   texture_blocks_untextured                                                    \
2723
2724 #define texture_blocks_switch_textured(texture_mode)                           \
2725   texture_blocks_##texture_mode                                                \
2726
2727 #define texture_blocks_switch(texturing, texture_mode)                         \
2728   texture_blocks_switch_##texturing(texture_mode)                              \
2729
2730
2731 // Shade blocks parametric-variables:
2732 // SHADE  TEXTURE_MAP  MODULATE_TEXELS  dither_mode
2733 // 0      0            x                x
2734 // 0      1            0                0
2735 // 0      1            0                1
2736 // x      1            1                x
2737 // 1      0            x                0
2738 // 1      0            x                1
2739 // 1      1            0                0
2740 // 1      1            0                1
2741 // 16 inputs, 8 combinations
2742
2743 #define shade_blocks_switch_unshaded_untextured(modulation, dithering, target) \
2744   shade_blocks_unshaded_untextured_##target                                    \
2745
2746 #define shade_blocks_switch_unshaded_textured_unmodulated(dithering, target)   \
2747   shade_blocks_textured_unmodulated_##target                                   \
2748
2749 #define shade_blocks_switch_unshaded_textured_modulated(dithering, target)     \
2750   shade_blocks_unshaded_textured_modulated_##dithering##_##target              \
2751
2752 #define shade_blocks_switch_unshaded_textured(modulation, dithering, target)   \
2753   shade_blocks_switch_unshaded_textured_##modulation(dithering, target)        \
2754
2755 #define shade_blocks_switch_unshaded(texturing, modulation, dithering, target) \
2756   shade_blocks_switch_unshaded_##texturing(modulation, dithering, target)      \
2757
2758 #define shade_blocks_switch_shaded_untextured(modulation, dithering, target)   \
2759   shade_blocks_shaded_untextured                                               \
2760
2761 #define shade_blocks_switch_shaded_textured_unmodulated(dithering, target)     \
2762   shade_blocks_textured_unmodulated_##target                                   \
2763
2764 #define shade_blocks_switch_shaded_textured_modulated(dithering, target)       \
2765   shade_blocks_shaded_textured_modulated_##dithering##_##target                \
2766
2767 #define shade_blocks_switch_shaded_textured(modulation, dithering, target)     \
2768   shade_blocks_switch_shaded_textured_##modulation(dithering, target)          \
2769
2770 #define shade_blocks_switch_shaded(texturing, modulation, dithering, target)   \
2771   shade_blocks_switch_shaded_##texturing(modulation, dithering, target)        \
2772
2773 #define shade_blocks_switch_mask_off(shading, texturing, modulation,           \
2774  dithering)                                                                    \
2775   shade_blocks_switch_##shading(texturing, modulation, dithering, direct)      \
2776
2777 #define shade_blocks_switch_mask_on(shading, texturing, modulation,            \
2778  dithering)                                                                    \
2779   shade_blocks_switch_##shading(texturing, modulation, dithering, indirect)    \
2780
2781 #define shade_blocks_switch_blended(shading, texturing, modulation, dithering, \
2782  mask_evaluate)                                                                \
2783   shade_blocks_switch_##shading(texturing, modulation, dithering, indirect)    \
2784
2785 #define shade_blocks_switch_unblended(shading, texturing, modulation,          \
2786  dithering, mask_evaluate)                                                     \
2787   shade_blocks_switch_mask_##mask_evaluate(shading, texturing, modulation,     \
2788    dithering)                                                                  \
2789
2790 #define shade_blocks_switch(shading, texturing, modulation, dithering,         \
2791  blending, mask_evaluate)                                                      \
2792   shade_blocks_switch_##blending(shading, texturing, modulation, dithering,    \
2793    mask_evaluate)                                                              \
2794
2795
2796 // Blend blocks parametric-variables:
2797 // TEXTURE_MAP BLEND  BM_A BM_B mask_evaluate
2798 // x           0      x    x    0
2799 // x           0      x    x    1
2800 // 0           1      0    0    0
2801 // 0           1      0    0    1
2802 // 0           1      0    1    0
2803 // 0           1      0    1    1
2804 // 0           1      1    0    0
2805 // 0           1      1    0    1
2806 // 0           1      1    1    0
2807 // 0           1      1    1    1
2808 // 1           1      0    0    0
2809 // 1           1      0    0    1
2810 // 1           1      0    1    0
2811 // 1           1      0    1    1
2812 // 1           1      1    0    0
2813 // 1           1      1    0    1
2814 // 1           1      1    1    0
2815 // 1           1      1    1    1
2816 // 32 inputs, 18 combinations
2817
2818 #define blend_blocks_switch_unblended(texturing, blend_mode, mask_evaluate)    \
2819   blend_blocks_textured_unblended_##mask_evaluate                              \
2820
2821 #define blend_blocks_switch_blended(texturing, blend_mode, mask_evaluate)      \
2822   blend_blocks_##texturing##_##blend_mode##_##mask_evaluate                    \
2823
2824 #define blend_blocks_switch(texturing, blending, blend_mode, mask_evaluate)    \
2825   blend_blocks_switch_##blending(texturing, blend_mode, mask_evaluate)         \
2826
2827
2828 #define render_blocks_switch_block_modulation(texture_mode, blend_mode,        \
2829  mask_evaluate, shading, dithering, texturing, blending, modulation)           \
2830 {                                                                              \
2831   setup_blocks_switch(shading, texturing, texture_mode, dithering, blending,   \
2832    mask_evaluate),                                                             \
2833   texture_blocks_switch(texturing, texture_mode),                              \
2834   shade_blocks_switch(shading, texturing, modulation, dithering, blending,     \
2835    mask_evaluate),                                                             \
2836   blend_blocks_switch(texturing, blending, blend_mode, mask_evaluate)          \
2837 }                                                                              \
2838
2839 #define render_blocks_switch_block_blending(texture_mode, blend_mode,          \
2840  mask_evaluate, shading, dithering, texturing, blending)                       \
2841   render_blocks_switch_block_modulation(texture_mode, blend_mode,              \
2842    mask_evaluate, shading, dithering, texturing, blending, modulated),         \
2843   render_blocks_switch_block_modulation(texture_mode, blend_mode,              \
2844    mask_evaluate, shading, dithering, texturing, blending, unmodulated)        \
2845
2846 #define render_blocks_switch_block_texturing(texture_mode, blend_mode,         \
2847  mask_evaluate, shading, dithering, texturing)                                 \
2848   render_blocks_switch_block_blending(texture_mode, blend_mode,                \
2849    mask_evaluate, shading, dithering, texturing, unblended),                   \
2850   render_blocks_switch_block_blending(texture_mode, blend_mode,                \
2851    mask_evaluate, shading, dithering, texturing, blended)                      \
2852
2853 #define render_blocks_switch_block_dithering(texture_mode, blend_mode,         \
2854  mask_evaluate, shading, dithering)                                            \
2855   render_blocks_switch_block_texturing(texture_mode, blend_mode,               \
2856    mask_evaluate, shading, dithering, untextured),                             \
2857   render_blocks_switch_block_texturing(texture_mode, blend_mode,               \
2858    mask_evaluate, shading, dithering, textured)                                \
2859
2860 #define render_blocks_switch_block_shading(texture_mode, blend_mode,           \
2861  mask_evaluate, shading)                                                       \
2862   render_blocks_switch_block_dithering(texture_mode, blend_mode,               \
2863    mask_evaluate, shading, undithered),                                        \
2864   render_blocks_switch_block_dithering(texture_mode, blend_mode,               \
2865    mask_evaluate, shading, dithered)                                           \
2866
2867 #define render_blocks_switch_block_mask_evaluate(texture_mode, blend_mode,     \
2868  mask_evaluate)                                                                \
2869   render_blocks_switch_block_shading(texture_mode, blend_mode, mask_evaluate,  \
2870    unshaded),                                                                  \
2871   render_blocks_switch_block_shading(texture_mode, blend_mode, mask_evaluate,  \
2872    shaded)                                                                     \
2873
2874 #define render_blocks_switch_block_blend_mode(texture_mode, blend_mode)        \
2875   render_blocks_switch_block_mask_evaluate(texture_mode, blend_mode, off),     \
2876   render_blocks_switch_block_mask_evaluate(texture_mode, blend_mode, on)       \
2877
2878 #define render_blocks_switch_block_texture_mode(texture_mode)                  \
2879   render_blocks_switch_block_blend_mode(texture_mode, average),                \
2880   render_blocks_switch_block_blend_mode(texture_mode, add),                    \
2881   render_blocks_switch_block_blend_mode(texture_mode, subtract),               \
2882   render_blocks_switch_block_blend_mode(texture_mode, add_fourth)              \
2883
2884 #define render_blocks_switch_block()                                           \
2885   render_blocks_switch_block_texture_mode(4bpp),                               \
2886   render_blocks_switch_block_texture_mode(8bpp),                               \
2887   render_blocks_switch_block_texture_mode(16bpp),                              \
2888   render_blocks_switch_block_texture_mode(4bpp)                                \
2889
2890
2891 render_block_handler_struct render_triangle_block_handlers[] =
2892 {
2893   render_blocks_switch_block()
2894 };
2895
2896 #undef render_blocks_switch_block_modulation
2897
2898 #define render_blocks_switch_block_modulation(texture_mode, blend_mode,        \
2899  mask_evaluate, shading, dithering, texturing, blending, modulation)           \
2900   "render flags:\n"                                                            \
2901   "texture mode:     " #texture_mode "\n"                                      \
2902   "blend mode:       " #blend_mode "\n"                                        \
2903   "mask evaluation:  " #mask_evaluate "\n"                                     \
2904   #shading "\n"                                                                \
2905   #dithering "\n"                                                              \
2906   #texturing "\n"                                                              \
2907   #blending "\n"                                                               \
2908   #modulation "\n"                                                             \
2909
2910 char *render_block_flag_strings[] =
2911 {
2912   render_blocks_switch_block()
2913 };
2914
2915
2916 #define triangle_y_direction_up   1
2917 #define triangle_y_direction_flat 2
2918 #define triangle_y_direction_down 0
2919
2920 #define triangle_winding_positive 0
2921 #define triangle_winding_negative 1
2922
2923 #define triangle_set_direction(direction_variable, value)                      \
2924   u32 direction_variable = (u32)(value) >> 31;                                 \
2925   if(value == 0)                                                               \
2926     direction_variable = 2                                                     \
2927
2928 #define triangle_case(direction_a, direction_b, direction_c, winding)          \
2929   case (triangle_y_direction_##direction_a |                                   \
2930    (triangle_y_direction_##direction_b << 2) |                                 \
2931    (triangle_y_direction_##direction_c << 4) |                                 \
2932    (triangle_winding_##winding << 6))                                          \
2933
2934 static int prepare_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
2935  vertex_struct *vertexes_out[3])
2936 {
2937   s32 y_top, y_bottom;
2938   s32 triangle_area;
2939   u32 triangle_winding = 0;
2940
2941   vertex_struct *a = &(vertexes[0]);
2942   vertex_struct *b = &(vertexes[1]);
2943   vertex_struct *c = &(vertexes[2]);
2944
2945   triangle_area = triangle_signed_area_x2(a->x, a->y, b->x, b->y, c->x, c->y);
2946
2947 #ifdef PROFILE
2948   triangles++;
2949 #endif
2950
2951   if(triangle_area == 0)
2952   {
2953 #ifdef PROFILE
2954     trivial_rejects++;
2955 #endif
2956     return 0;
2957   }
2958
2959   if(b->y < a->y)
2960     vertex_swap(a, b);
2961
2962   if(c->y < b->y)
2963   {
2964     vertex_swap(b, c);
2965
2966     if(b->y < a->y)
2967       vertex_swap(a, b);
2968   }
2969
2970   y_bottom = c->y;
2971   y_top = a->y;
2972
2973   if((y_bottom - y_top) >= 512)
2974   {
2975 #ifdef PROFILE
2976     trivial_rejects++;
2977 #endif
2978     return 0;
2979   }
2980
2981   if(triangle_area < 0)
2982   {
2983     triangle_area = -triangle_area;
2984     triangle_winding ^= 1;
2985     vertex_swap(a, c);
2986   }
2987
2988   if(b->x < a->x)
2989     vertex_swap(a, b);
2990
2991   if(c->x < b->x)
2992   {
2993     vertex_swap(b, c);
2994
2995     if(b->x < a->x)
2996       vertex_swap(a, b);
2997   }
2998
2999   if((c->x - psx_gpu->offset_x) >= 1024 || (c->x - a->x) >= 1024)
3000   {
3001 #ifdef PROFILE
3002     trivial_rejects++;
3003 #endif
3004     return 0;
3005   }
3006
3007   if(invalidate_texture_cache_region_viewport(psx_gpu, a->x, y_top, c->x,
3008    y_bottom) == 0)
3009   {
3010 #ifdef PROFILE
3011     trivial_rejects++;
3012 #endif
3013     return 0;
3014   }
3015
3016   psx_gpu->triangle_area = triangle_area;
3017   psx_gpu->triangle_winding = triangle_winding;
3018
3019   vertexes_out[0] = a;
3020   vertexes_out[1] = b;
3021   vertexes_out[2] = c;
3022
3023   return 1;
3024 }
3025
3026 static void render_triangle_p(psx_gpu_struct *psx_gpu,
3027  vertex_struct *vertex_ptrs[3], u32 flags)
3028 {
3029   psx_gpu->num_spans = 0;
3030
3031   vertex_struct *a = vertex_ptrs[0];
3032   vertex_struct *b = vertex_ptrs[1];
3033   vertex_struct *c = vertex_ptrs[2];
3034
3035   s32 y_delta_a = b->y - a->y;
3036   s32 y_delta_b = c->y - b->y;
3037   s32 y_delta_c = c->y - a->y;
3038
3039   triangle_set_direction(y_direction_a, y_delta_a);
3040   triangle_set_direction(y_direction_b, y_delta_b);
3041   triangle_set_direction(y_direction_c, y_delta_c);
3042
3043   compute_all_gradients(psx_gpu, a, b, c);
3044
3045   switch(y_direction_a | (y_direction_b << 2) | (y_direction_c << 4) |
3046    (psx_gpu->triangle_winding << 6))
3047   {
3048     triangle_case(up, up, up, negative):
3049     triangle_case(up, up, flat, negative):
3050     triangle_case(up, up, down, negative):
3051       setup_spans_up_right(psx_gpu, a, b, c);
3052       break;
3053
3054     triangle_case(flat, up, up, negative):
3055     triangle_case(flat, up, flat, negative):
3056     triangle_case(flat, up, down, negative):
3057       setup_spans_up_a(psx_gpu, a, b, c);
3058       break;
3059
3060     triangle_case(down, up, up, negative):
3061       setup_spans_up_down(psx_gpu, a, c, b);
3062       break;
3063
3064     triangle_case(down, up, flat, negative):
3065       setup_spans_down_a(psx_gpu, a, c, b);
3066       break;
3067
3068     triangle_case(down, up, down, negative):
3069       setup_spans_down_right(psx_gpu, a, c, b);
3070       break;
3071
3072     triangle_case(down, flat, up, negative):
3073     triangle_case(down, flat, flat, negative):
3074     triangle_case(down, flat, down, negative):
3075       setup_spans_down_b(psx_gpu, a, b, c);
3076       break;
3077
3078     triangle_case(down, down, up, negative):
3079     triangle_case(down, down, flat, negative):
3080     triangle_case(down, down, down, negative):
3081       setup_spans_down_left(psx_gpu, a, b, c);
3082       break;
3083
3084     triangle_case(up, up, up, positive):
3085     triangle_case(up, up, flat, positive):
3086     triangle_case(up, up, down, positive):
3087       setup_spans_up_left(psx_gpu, a, b, c);
3088       break;
3089
3090     triangle_case(up, flat, up, positive):
3091     triangle_case(up, flat, flat, positive):
3092     triangle_case(up, flat, down, positive):
3093       setup_spans_up_b(psx_gpu, a, b, c);
3094       break;
3095
3096     triangle_case(up, down, up, positive):
3097       setup_spans_up_right(psx_gpu, a, c, b);
3098       break;
3099
3100     triangle_case(up, down, flat, positive):
3101       setup_spans_up_a(psx_gpu, a, c, b);
3102       break;
3103
3104     triangle_case(up, down, down, positive):
3105       setup_spans_up_down(psx_gpu, a, b, c);
3106       break;
3107
3108     triangle_case(flat, down, up, positive):
3109     triangle_case(flat, down, flat, positive):
3110     triangle_case(flat, down, down, positive):
3111       setup_spans_down_a(psx_gpu, a, b, c);
3112       break;
3113
3114     triangle_case(down, down, up, positive):
3115     triangle_case(down, down, flat, positive):
3116     triangle_case(down, down, down, positive):
3117       setup_spans_down_right(psx_gpu, a, b, c);
3118       break;
3119   }
3120
3121 #ifdef PROFILE
3122   spans += psx_gpu->num_spans;
3123 #endif
3124
3125   if(unlikely(psx_gpu->render_mode & RENDER_INTERLACE_ENABLED))
3126   {
3127     u32 i;
3128
3129     if(psx_gpu->render_mode & RENDER_INTERLACE_ODD)
3130     {
3131       for(i = 0; i < psx_gpu->num_spans; i++)
3132       {
3133         if((psx_gpu->span_edge_data[i].y & 1) == 0)
3134           psx_gpu->span_edge_data[i].num_blocks = 0;
3135       }
3136     }
3137     else
3138     {
3139       for(i = 0; i < psx_gpu->num_spans; i++)
3140       {
3141         if(psx_gpu->span_edge_data[i].y & 1)
3142           psx_gpu->span_edge_data[i].num_blocks = 0;
3143       }
3144     }
3145   }
3146
3147   u32 render_state = flags &
3148    (RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND |
3149    RENDER_FLAGS_TEXTURE_MAP | RENDER_FLAGS_SHADE);
3150   render_state |= psx_gpu->render_state_base;
3151
3152   if((psx_gpu->render_state != render_state) ||
3153    (psx_gpu->primitive_type != PRIMITIVE_TYPE_TRIANGLE))
3154   {
3155     psx_gpu->render_state = render_state;
3156     flush_render_block_buffer(psx_gpu);
3157 #ifdef PROFILE
3158     state_changes++;
3159 #endif
3160   }
3161
3162   psx_gpu->primitive_type = PRIMITIVE_TYPE_TRIANGLE;
3163
3164   psx_gpu->render_block_handler =
3165    &(render_triangle_block_handlers[render_state]);
3166   ((setup_blocks_function_type *)psx_gpu->render_block_handler->setup_blocks)
3167    (psx_gpu);
3168 }
3169
3170 void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
3171  u32 flags)
3172 {
3173   vertex_struct *vertex_ptrs[3];
3174   if (prepare_triangle(psx_gpu, vertexes, vertex_ptrs))
3175     render_triangle_p(psx_gpu, vertex_ptrs, flags);
3176 }
3177
3178
3179 void texture_sprite_blocks_8bpp(psx_gpu_struct *psx_gpu);
3180
3181 #ifndef NEON_BUILD
3182
3183 void texture_sprite_blocks_8bpp(psx_gpu_struct *psx_gpu)
3184 {
3185   block_struct *block = psx_gpu->blocks;
3186   u32 num_blocks = psx_gpu->num_blocks;
3187
3188   vec_8x16u texels;
3189   vec_8x8u texel_indexes;
3190
3191   u16 *clut_ptr = psx_gpu->clut_ptr;
3192   u32 i;
3193
3194   while(num_blocks)
3195   {
3196     texel_indexes = block->r;
3197
3198     for(i = 0; i < 8; i++)
3199     {
3200       texels.e[i] = clut_ptr[texel_indexes.e[i]];
3201     }
3202
3203     block->texels = texels;
3204
3205     num_blocks--;
3206     block++;
3207   }
3208 }
3209
3210 #endif
3211
3212
3213 #define setup_sprite_tiled_initialize_4bpp_clut()                              \
3214   u16 *clut_ptr = psx_gpu->clut_ptr;                                           \
3215   vec_8x16u clut_a, clut_b;                                                    \
3216   vec_16x8u clut_low, clut_high;                                               \
3217                                                                                \
3218   load_8x16b(clut_a, clut_ptr);                                                \
3219   load_8x16b(clut_b, clut_ptr + 8);                                            \
3220   unzip_16x8b(clut_low, clut_high, clut_a, clut_b)                             \
3221
3222 #define setup_sprite_tiled_initialize_4bpp()                                   \
3223   setup_sprite_tiled_initialize_4bpp_clut();                                   \
3224                                                                                \
3225   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_4bpp_mask)        \
3226     update_texture_4bpp_cache(psx_gpu)                                         \
3227
3228 #define setup_sprite_tiled_initialize_8bpp()                                   \
3229   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_8bpp_mask)        \
3230     update_texture_8bpp_cache(psx_gpu)                                         \
3231
3232
3233 #define setup_sprite_tile_fetch_texel_block_8bpp(offset)                       \
3234   texture_block_ptr = psx_gpu->texture_page_ptr +                              \
3235    ((texture_offset + offset) & texture_mask);                                 \
3236                                                                                \
3237   load_64b(texels, texture_block_ptr)                                          \
3238
3239
3240 #define setup_sprite_tile_add_blocks(tile_num_blocks)                          \
3241   num_blocks += tile_num_blocks;                                               \
3242   sprite_blocks += tile_num_blocks;                                            \
3243                                                                                \
3244   if(num_blocks > MAX_BLOCKS)                                                  \
3245   {                                                                            \
3246     flush_render_block_buffer(psx_gpu);                                        \
3247     num_blocks = tile_num_blocks;                                              \
3248     block = psx_gpu->blocks;                                                   \
3249   }                                                                            \
3250
3251 #define setup_sprite_tile_full_4bpp(edge)                                      \
3252 {                                                                              \
3253   vec_8x8u texels_low, texels_high;                                            \
3254   vec_8x16u pixels;                                                            \
3255   setup_sprite_tile_add_blocks(sub_tile_height * 2);                           \
3256                                                                                \
3257   while(sub_tile_height)                                                       \
3258   {                                                                            \
3259     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3260     tbl_16(texels_low, texels, clut_low);                                      \
3261     tbl_16(texels_high, texels, clut_high);                                    \
3262     zip_8x16b(pixels, texels_low, texels_high);                                \
3263                                                                                \
3264     block->texels = pixels;                                                    \
3265     block->draw_mask_bits = left_mask_bits;                                    \
3266     block->fb_ptr = fb_ptr;                                                    \
3267     block++;                                                                   \
3268                                                                                \
3269     setup_sprite_tile_fetch_texel_block_8bpp(8);                               \
3270     tbl_16(texels_low, texels, clut_low);                                      \
3271     tbl_16(texels_high, texels, clut_high);                                    \
3272     zip_8x16b(pixels, texels_low, texels_high);                                \
3273                                                                                \
3274     block->texels = pixels;                                                    \
3275     block->draw_mask_bits = right_mask_bits;                                   \
3276     block->fb_ptr = fb_ptr + 8;                                                \
3277     block++;                                                                   \
3278                                                                                \
3279     fb_ptr += 1024;                                                            \
3280     texture_offset += 0x10;                                                    \
3281     sub_tile_height--;                                                         \
3282   }                                                                            \
3283   texture_offset += 0xF00;                                                     \
3284   psx_gpu->num_blocks = num_blocks;                                            \
3285 }                                                                              \
3286
3287 #define setup_sprite_tile_half_4bpp(edge)                                      \
3288 {                                                                              \
3289   vec_8x8u texels_low, texels_high;                                            \
3290   vec_8x16u pixels;                                                            \
3291   setup_sprite_tile_add_blocks(sub_tile_height);                               \
3292                                                                                \
3293   while(sub_tile_height)                                                       \
3294   {                                                                            \
3295     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3296     tbl_16(texels_low, texels, clut_low);                                      \
3297     tbl_16(texels_high, texels, clut_high);                                    \
3298     zip_8x16b(pixels, texels_low, texels_high);                                \
3299                                                                                \
3300     block->texels = pixels;                                                    \
3301     block->draw_mask_bits = edge##_mask_bits;                                  \
3302     block->fb_ptr = fb_ptr;                                                    \
3303     block++;                                                                   \
3304                                                                                \
3305     fb_ptr += 1024;                                                            \
3306     texture_offset += 0x10;                                                    \
3307     sub_tile_height--;                                                         \
3308   }                                                                            \
3309   texture_offset += 0xF00;                                                     \
3310   psx_gpu->num_blocks = num_blocks;                                            \
3311 }                                                                              \
3312
3313
3314 #define setup_sprite_tile_full_8bpp(edge)                                      \
3315 {                                                                              \
3316   setup_sprite_tile_add_blocks(sub_tile_height * 2);                           \
3317                                                                                \
3318   while(sub_tile_height)                                                       \
3319   {                                                                            \
3320     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3321     block->r = texels;                                                         \
3322     block->draw_mask_bits = left_mask_bits;                                    \
3323     block->fb_ptr = fb_ptr;                                                    \
3324     block++;                                                                   \
3325                                                                                \
3326     setup_sprite_tile_fetch_texel_block_8bpp(8);                               \
3327     block->r = texels;                                                         \
3328     block->draw_mask_bits = right_mask_bits;                                   \
3329     block->fb_ptr = fb_ptr + 8;                                                \
3330     block++;                                                                   \
3331                                                                                \
3332     fb_ptr += 1024;                                                            \
3333     texture_offset += 0x10;                                                    \
3334     sub_tile_height--;                                                         \
3335   }                                                                            \
3336   texture_offset += 0xF00;                                                     \
3337   psx_gpu->num_blocks = num_blocks;                                            \
3338 }                                                                              \
3339
3340 #define setup_sprite_tile_half_8bpp(edge)                                      \
3341 {                                                                              \
3342   setup_sprite_tile_add_blocks(sub_tile_height * 2);                           \
3343                                                                                \
3344   while(sub_tile_height)                                                       \
3345   {                                                                            \
3346     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3347     block->r = texels;                                                         \
3348     block->draw_mask_bits = edge##_mask_bits;                                  \
3349     block->fb_ptr = fb_ptr;                                                    \
3350     block++;                                                                   \
3351                                                                                \
3352     fb_ptr += 1024;                                                            \
3353     texture_offset += 0x10;                                                    \
3354     sub_tile_height--;                                                         \
3355   }                                                                            \
3356   texture_offset += 0xF00;                                                     \
3357   psx_gpu->num_blocks = num_blocks;                                            \
3358 }                                                                              \
3359
3360
3361 #define setup_sprite_tile_column_edge_pre_adjust_half_right()                  \
3362   texture_offset = texture_offset_base + 8;                                    \
3363   fb_ptr += 8                                                                  \
3364
3365 #define setup_sprite_tile_column_edge_pre_adjust_half_left()                   \
3366   texture_offset = texture_offset_base                                         \
3367
3368 #define setup_sprite_tile_column_edge_pre_adjust_half(edge)                    \
3369   setup_sprite_tile_column_edge_pre_adjust_half_##edge()                       \
3370
3371 #define setup_sprite_tile_column_edge_pre_adjust_full(edge)                    \
3372   texture_offset = texture_offset_base                                         \
3373
3374 #define setup_sprite_tile_column_edge_post_adjust_half_right()                 \
3375   fb_ptr -= 8                                                                  \
3376
3377 #define setup_sprite_tile_column_edge_post_adjust_half_left()                  \
3378
3379 #define setup_sprite_tile_column_edge_post_adjust_half(edge)                   \
3380   setup_sprite_tile_column_edge_post_adjust_half_##edge()                      \
3381
3382 #define setup_sprite_tile_column_edge_post_adjust_full(edge)                   \
3383
3384
3385 #define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode,  \
3386  x4mode)                                                                       \
3387 do                                                                             \
3388 {                                                                              \
3389   sub_tile_height = column_data;                                               \
3390   setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge);          \
3391   setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge);                \
3392   setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge);         \
3393 } while(0)                                                                     \
3394
3395 #define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode,   \
3396  x4mode)                                                                       \
3397 do                                                                             \
3398 {                                                                              \
3399   u32 tiles_remaining = column_data >> 16;                                     \
3400   sub_tile_height = column_data & 0xFF;                                        \
3401   setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge);          \
3402   setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge);                \
3403   tiles_remaining -= 1;                                                        \
3404                                                                                \
3405   while(tiles_remaining)                                                       \
3406   {                                                                            \
3407     sub_tile_height = 16;                                                      \
3408     setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge);              \
3409     tiles_remaining--;                                                         \
3410   }                                                                            \
3411                                                                                \
3412   sub_tile_height = (column_data >> 8) & 0xFF;                                 \
3413   setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge);                \
3414   setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge);         \
3415 } while(0)                                                                     \
3416
3417
3418 #define setup_sprite_column_data_single()                                      \
3419   column_data = height                                                         \
3420
3421 #define setup_sprite_column_data_multi()                                       \
3422   column_data = 16 - offset_v;                                                 \
3423   column_data |= ((height_rounded & 0xF) + 1) << 8;                            \
3424   column_data |= (tile_height - 1) << 16                                       \
3425
3426
3427 #define RIGHT_MASK_BIT_SHIFT 8
3428 #define RIGHT_MASK_BIT_SHIFT_4x 16
3429
3430 #define setup_sprite_tile_column_width_single(texture_mode, multi_height,      \
3431  edge_mode, edge, x4mode)                                                      \
3432 {                                                                              \
3433   setup_sprite_column_data_##multi_height();                                   \
3434   left_mask_bits = left_block_mask | right_block_mask;                         \
3435   right_mask_bits = left_mask_bits >> RIGHT_MASK_BIT_SHIFT##x4mode;            \
3436                                                                                \
3437   setup_sprite_tile_column_height_##multi_height(edge_mode, edge,              \
3438    texture_mode, x4mode);                                                      \
3439 }                                                                              \
3440
3441 #define setup_sprite_tiled_advance_column()                                    \
3442   texture_offset_base += 0x100;                                                \
3443   if((texture_offset_base & 0xF00) == 0)                                       \
3444     texture_offset_base -= (0x100 + 0xF00)                                     \
3445
3446 #define FB_PTR_MULTIPLIER 1
3447 #define FB_PTR_MULTIPLIER_4x 2
3448
3449 #define setup_sprite_tile_column_width_multi(texture_mode, multi_height,       \
3450  left_mode, right_mode, x4mode)                                                \
3451 {                                                                              \
3452   setup_sprite_column_data_##multi_height();                                   \
3453   s32 fb_ptr_advance_column = (16 - (1024 * height))                           \
3454     * FB_PTR_MULTIPLIER##x4mode;                                               \
3455                                                                                \
3456   tile_width -= 2;                                                             \
3457   left_mask_bits = left_block_mask;                                            \
3458   right_mask_bits = left_mask_bits >> RIGHT_MASK_BIT_SHIFT##x4mode;            \
3459                                                                                \
3460   setup_sprite_tile_column_height_##multi_height(left_mode, right,             \
3461    texture_mode, x4mode);                                                      \
3462   fb_ptr += fb_ptr_advance_column;                                             \
3463                                                                                \
3464   left_mask_bits = 0x00;                                                       \
3465   right_mask_bits = 0x00;                                                      \
3466                                                                                \
3467   while(tile_width)                                                            \
3468   {                                                                            \
3469     setup_sprite_tiled_advance_column();                                       \
3470     setup_sprite_tile_column_height_##multi_height(full, none,                 \
3471      texture_mode, x4mode);                                                    \
3472     fb_ptr += fb_ptr_advance_column;                                           \
3473     tile_width--;                                                              \
3474   }                                                                            \
3475                                                                                \
3476   left_mask_bits = right_block_mask;                                           \
3477   right_mask_bits = left_mask_bits >> RIGHT_MASK_BIT_SHIFT##x4mode;            \
3478                                                                                \
3479   setup_sprite_tiled_advance_column();                                         \
3480   setup_sprite_tile_column_height_##multi_height(right_mode, left,             \
3481    texture_mode, x4mode);                                                      \
3482 }                                                                              \
3483
3484
3485 /* 4x stuff */
3486 #define setup_sprite_tiled_initialize_4bpp_4x()                                \
3487   setup_sprite_tiled_initialize_4bpp_clut()                                    \
3488
3489 #define setup_sprite_tiled_initialize_8bpp_4x()                                \
3490
3491
3492 #define setup_sprite_tile_full_4bpp_4x(edge)                                   \
3493 {                                                                              \
3494   vec_8x8u texels_low, texels_high;                                            \
3495   vec_8x16u pixels, pixels_wide;                                               \
3496   setup_sprite_tile_add_blocks(sub_tile_height * 2 * 4);                       \
3497   u32 left_mask_bits_a = left_mask_bits & 0xFF;                                \
3498   u32 left_mask_bits_b = left_mask_bits >> 8;                                  \
3499   u32 right_mask_bits_a = right_mask_bits & 0xFF;                              \
3500   u32 right_mask_bits_b = right_mask_bits >> 8;                                \
3501                                                                                \
3502   while(sub_tile_height)                                                       \
3503   {                                                                            \
3504     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3505     tbl_16(texels_low, texels, clut_low);                                      \
3506     tbl_16(texels_high, texels, clut_high);                                    \
3507     zip_8x16b(pixels, texels_low, texels_high);                                \
3508                                                                                \
3509     zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);    \
3510     block->texels = pixels_wide;                                               \
3511     block->draw_mask_bits = left_mask_bits_a;                                  \
3512     block->fb_ptr = fb_ptr;                                                    \
3513     block++;                                                                   \
3514                                                                                \
3515     block->texels = pixels_wide;                                               \
3516     block->draw_mask_bits = left_mask_bits_a;                                  \
3517     block->fb_ptr = fb_ptr + 1024;                                             \
3518     block++;                                                                   \
3519                                                                                \
3520     zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.high, pixels.high);  \
3521     block->texels = pixels_wide;                                               \
3522     block->draw_mask_bits = left_mask_bits_b;                                  \
3523     block->fb_ptr = fb_ptr + 8;                                                \
3524     block++;                                                                   \
3525                                                                                \
3526     block->texels = pixels_wide;                                               \
3527     block->draw_mask_bits = left_mask_bits_b;                                  \
3528     block->fb_ptr = fb_ptr + 1024 + 8;                                         \
3529     block++;                                                                   \
3530                                                                                \
3531     setup_sprite_tile_fetch_texel_block_8bpp(8);                               \
3532     tbl_16(texels_low, texels, clut_low);                                      \
3533     tbl_16(texels_high, texels, clut_high);                                    \
3534     zip_8x16b(pixels, texels_low, texels_high);                                \
3535                                                                                \
3536     zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);    \
3537     block->texels = pixels_wide;                                               \
3538     block->draw_mask_bits = right_mask_bits_a;                                 \
3539     block->fb_ptr = fb_ptr + 16;                                               \
3540     block++;                                                                   \
3541                                                                                \
3542     block->texels = pixels_wide;                                               \
3543     block->draw_mask_bits = right_mask_bits_a;                                 \
3544     block->fb_ptr = fb_ptr + 1024 + 16;                                        \
3545     block++;                                                                   \
3546                                                                                \
3547     zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.high, pixels.high);  \
3548     block->texels = pixels_wide;                                               \
3549     block->draw_mask_bits = right_mask_bits_b;                                 \
3550     block->fb_ptr = fb_ptr + 24;                                               \
3551     block++;                                                                   \
3552                                                                                \
3553     block->texels = pixels_wide;                                               \
3554     block->draw_mask_bits = right_mask_bits_b;                                 \
3555     block->fb_ptr = fb_ptr + 1024 + 24;                                        \
3556     block++;                                                                   \
3557                                                                                \
3558     fb_ptr += 2048;                                                            \
3559     texture_offset += 0x10;                                                    \
3560     sub_tile_height--;                                                         \
3561   }                                                                            \
3562   texture_offset += 0xF00;                                                     \
3563   psx_gpu->num_blocks = num_blocks;                                            \
3564 }                                                                              \
3565
3566 #define setup_sprite_tile_half_4bpp_4x(edge)                                   \
3567 {                                                                              \
3568   vec_8x8u texels_low, texels_high;                                            \
3569   vec_8x16u pixels, pixels_wide;                                               \
3570   setup_sprite_tile_add_blocks(sub_tile_height * 4);                           \
3571   u32 edge##_mask_bits_a = edge##_mask_bits & 0xFF;                            \
3572   u32 edge##_mask_bits_b = edge##_mask_bits >> 8;                              \
3573                                                                                \
3574   while(sub_tile_height)                                                       \
3575   {                                                                            \
3576     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3577     tbl_16(texels_low, texels, clut_low);                                      \
3578     tbl_16(texels_high, texels, clut_high);                                    \
3579     zip_8x16b(pixels, texels_low, texels_high);                                \
3580                                                                                \
3581     zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.low, pixels.low);    \
3582     block->texels = pixels_wide;                                               \
3583     block->draw_mask_bits = edge##_mask_bits_a;                                \
3584     block->fb_ptr = fb_ptr;                                                    \
3585     block++;                                                                   \
3586                                                                                \
3587     block->texels = pixels_wide;                                               \
3588     block->draw_mask_bits = edge##_mask_bits_a;                                \
3589     block->fb_ptr = fb_ptr + 1024;                                             \
3590     block++;                                                                   \
3591                                                                                \
3592     zip_4x32b(vector_cast(vec_4x32u, pixels_wide), pixels.high, pixels.high);  \
3593     block->texels = pixels_wide;                                               \
3594     block->draw_mask_bits = edge##_mask_bits_b;                                \
3595     block->fb_ptr = fb_ptr + 8;                                                \
3596     block++;                                                                   \
3597                                                                                \
3598     block->texels = pixels_wide;                                               \
3599     block->draw_mask_bits = edge##_mask_bits_b;                                \
3600     block->fb_ptr = fb_ptr + 1024 + 8;                                         \
3601     block++;                                                                   \
3602                                                                                \
3603     fb_ptr += 2048;                                                            \
3604     texture_offset += 0x10;                                                    \
3605     sub_tile_height--;                                                         \
3606   }                                                                            \
3607   texture_offset += 0xF00;                                                     \
3608   psx_gpu->num_blocks = num_blocks;                                            \
3609 }                                                                              \
3610
3611
3612 #define setup_sprite_tile_full_8bpp_4x(edge)                                   \
3613 {                                                                              \
3614   setup_sprite_tile_add_blocks(sub_tile_height * 2 * 4);                       \
3615   vec_16x8u texels_wide;                                                       \
3616   u32 left_mask_bits_a = left_mask_bits & 0xFF;                                \
3617   u32 left_mask_bits_b = left_mask_bits >> 8;                                  \
3618   u32 right_mask_bits_a = right_mask_bits & 0xFF;                              \
3619   u32 right_mask_bits_b = right_mask_bits >> 8;                                \
3620                                                                                \
3621   while(sub_tile_height)                                                       \
3622   {                                                                            \
3623     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3624     zip_8x16b(vector_cast(vec_8x16u, texels_wide), texels, texels);            \
3625     block->r = texels_wide.low;                                                \
3626     block->draw_mask_bits = left_mask_bits_a;                                  \
3627     block->fb_ptr = fb_ptr;                                                    \
3628     block++;                                                                   \
3629                                                                                \
3630     block->r = texels_wide.low;                                                \
3631     block->draw_mask_bits = left_mask_bits_a;                                  \
3632     block->fb_ptr = fb_ptr + 1024;                                             \
3633     block++;                                                                   \
3634                                                                                \
3635     block->r = texels_wide.high;                                               \
3636     block->draw_mask_bits = left_mask_bits_b;                                  \
3637     block->fb_ptr = fb_ptr + 8;                                                \
3638     block++;                                                                   \
3639                                                                                \
3640     block->r = texels_wide.high;                                               \
3641     block->draw_mask_bits = left_mask_bits_b;                                  \
3642     block->fb_ptr = fb_ptr + 1024 + 8;                                         \
3643     block++;                                                                   \
3644                                                                                \
3645     setup_sprite_tile_fetch_texel_block_8bpp(8);                               \
3646     zip_8x16b(vector_cast(vec_8x16u, texels_wide), texels, texels);            \
3647     block->r = texels_wide.low;                                                \
3648     block->draw_mask_bits = right_mask_bits_a;                                 \
3649     block->fb_ptr = fb_ptr + 16;                                               \
3650     block++;                                                                   \
3651                                                                                \
3652     block->r = texels_wide.low;                                                \
3653     block->draw_mask_bits = right_mask_bits_a;                                 \
3654     block->fb_ptr = fb_ptr + 1024 + 16;                                        \
3655     block++;                                                                   \
3656                                                                                \
3657     block->r = texels_wide.high;                                               \
3658     block->draw_mask_bits = right_mask_bits_b;                                 \
3659     block->fb_ptr = fb_ptr + 24;                                               \
3660     block++;                                                                   \
3661                                                                                \
3662     block->r = texels_wide.high;                                               \
3663     block->draw_mask_bits = right_mask_bits_b;                                 \
3664     block->fb_ptr = fb_ptr + 24 + 1024;                                        \
3665     block++;                                                                   \
3666                                                                                \
3667     fb_ptr += 2048;                                                            \
3668     texture_offset += 0x10;                                                    \
3669     sub_tile_height--;                                                         \
3670   }                                                                            \
3671   texture_offset += 0xF00;                                                     \
3672   psx_gpu->num_blocks = num_blocks;                                            \
3673 }                                                                              \
3674
3675 #define setup_sprite_tile_half_8bpp_4x(edge)                                   \
3676 {                                                                              \
3677   setup_sprite_tile_add_blocks(sub_tile_height * 4);                           \
3678   vec_16x8u texels_wide;                                                       \
3679   u32 edge##_mask_bits_a = edge##_mask_bits & 0xFF;                            \
3680   u32 edge##_mask_bits_b = edge##_mask_bits >> 8;                              \
3681                                                                                \
3682   while(sub_tile_height)                                                       \
3683   {                                                                            \
3684     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3685     zip_8x16b(vector_cast(vec_8x16u, texels_wide), texels, texels);            \
3686     block->r = texels_wide.low;                                                \
3687     block->draw_mask_bits = edge##_mask_bits_a;                                \
3688     block->fb_ptr = fb_ptr;                                                    \
3689     block++;                                                                   \
3690                                                                                \
3691     block->r = texels_wide.low;                                                \
3692     block->draw_mask_bits = edge##_mask_bits_a;                                \
3693     block->fb_ptr = fb_ptr + 1024;                                             \
3694     block++;                                                                   \
3695                                                                                \
3696     block->r = texels_wide.high;                                               \
3697     block->draw_mask_bits = edge##_mask_bits_b;                                \
3698     block->fb_ptr = fb_ptr + 8;                                                \
3699     block++;                                                                   \
3700                                                                                \
3701     block->r = texels_wide.high;                                               \
3702     block->draw_mask_bits = edge##_mask_bits_b;                                \
3703     block->fb_ptr = fb_ptr + 8 + 1024;                                         \
3704     block++;                                                                   \
3705                                                                                \
3706     fb_ptr += 2048;                                                            \
3707     texture_offset += 0x10;                                                    \
3708     sub_tile_height--;                                                         \
3709   }                                                                            \
3710   texture_offset += 0xF00;                                                     \
3711   psx_gpu->num_blocks = num_blocks;                                            \
3712 }                                                                              \
3713
3714
3715 #define setup_sprite_tile_column_edge_pre_adjust_half_right_4x()               \
3716   texture_offset = texture_offset_base + 8;                                    \
3717   fb_ptr += 16                                                                 \
3718
3719 #define setup_sprite_tile_column_edge_pre_adjust_half_left_4x()                \
3720   texture_offset = texture_offset_base                                         \
3721
3722 #define setup_sprite_tile_column_edge_pre_adjust_half_4x(edge)                 \
3723   setup_sprite_tile_column_edge_pre_adjust_half_##edge##_4x()                  \
3724
3725 #define setup_sprite_tile_column_edge_pre_adjust_full_4x(edge)                 \
3726   texture_offset = texture_offset_base                                         \
3727
3728 #define setup_sprite_tile_column_edge_post_adjust_half_right_4x()              \
3729   fb_ptr -= 16                                                                 \
3730
3731 #define setup_sprite_tile_column_edge_post_adjust_half_left_4x()               \
3732
3733 #define setup_sprite_tile_column_edge_post_adjust_half_4x(edge)                \
3734   setup_sprite_tile_column_edge_post_adjust_half_##edge##_4x()                 \
3735
3736 #define setup_sprite_tile_column_edge_post_adjust_full_4x(edge)                \
3737
3738
3739 #define setup_sprite_offset_u_adjust()                                         \
3740
3741 #define setup_sprite_comapre_left_block_mask()                                 \
3742   ((left_block_mask & 0xFF) == 0xFF)                                           \
3743
3744 #define setup_sprite_comapre_right_block_mask()                                \
3745   (((right_block_mask >> 8) & 0xFF) == 0xFF)                                   \
3746
3747
3748 #define setup_sprite_offset_u_adjust_4x()                                      \
3749   offset_u *= 2;                                                               \
3750   offset_u_right = offset_u_right * 2 + 1                                      \
3751
3752 #define setup_sprite_comapre_left_block_mask_4x()                              \
3753   ((left_block_mask & 0xFFFF) == 0xFFFF)                                       \
3754
3755 #define setup_sprite_comapre_right_block_mask_4x()                             \
3756   (((right_block_mask >> 16) & 0xFFFF) == 0xFFFF)                              \
3757
3758
3759 #define setup_sprite_tiled_builder(texture_mode, x4mode)                       \
3760 void setup_sprite_##texture_mode##x4mode(psx_gpu_struct *psx_gpu, s32 x, s32 y,\
3761  s32 u, s32 v, s32 width, s32 height, u32 color)                               \
3762 {                                                                              \
3763   s32 offset_u = u & 0xF;                                                      \
3764   s32 offset_v = v & 0xF;                                                      \
3765                                                                                \
3766   s32 width_rounded = offset_u + width + 15;                                   \
3767   s32 height_rounded = offset_v + height + 15;                                 \
3768   s32 tile_height = height_rounded / 16;                                       \
3769   s32 tile_width = width_rounded / 16;                                         \
3770   u32 offset_u_right = width_rounded & 0xF;                                    \
3771                                                                                \
3772   setup_sprite_offset_u_adjust##x4mode();                                      \
3773                                                                                \
3774   u32 left_block_mask = ~(0xFFFFFFFF << offset_u);                             \
3775   u32 right_block_mask = 0xFFFFFFFE << offset_u_right;                         \
3776                                                                                \
3777   u32 left_mask_bits;                                                          \
3778   u32 right_mask_bits;                                                         \
3779                                                                                \
3780   u32 sub_tile_height;                                                         \
3781   u32 column_data;                                                             \
3782                                                                                \
3783   u32 texture_mask = (psx_gpu->texture_mask_width & 0xF) |                     \
3784    ((psx_gpu->texture_mask_height & 0xF) << 4) |                               \
3785    ((psx_gpu->texture_mask_width >> 4) << 8) |                                 \
3786    ((psx_gpu->texture_mask_height >> 4) << 12);                                \
3787   u32 texture_offset = ((v & 0xF) << 4) | ((u & 0xF0) << 4) |                  \
3788    ((v & 0xF0) << 8);                                                          \
3789   u32 texture_offset_base = texture_offset;                                    \
3790   u32 control_mask;                                                            \
3791                                                                                \
3792   u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + (x - offset_u);           \
3793   u32 num_blocks = psx_gpu->num_blocks;                                        \
3794   block_struct *block = psx_gpu->blocks + num_blocks;                          \
3795                                                                                \
3796   u16 *texture_block_ptr;                                                      \
3797   vec_8x8u texels;                                                             \
3798                                                                                \
3799   setup_sprite_tiled_initialize_##texture_mode##x4mode();                      \
3800                                                                                \
3801   control_mask = tile_width == 1;                                              \
3802   control_mask |= (tile_height == 1) << 1;                                     \
3803   control_mask |= setup_sprite_comapre_left_block_mask##x4mode() << 2;         \
3804   control_mask |= setup_sprite_comapre_right_block_mask##x4mode() << 3;        \
3805                                                                                \
3806   sprites_##texture_mode++;                                                    \
3807                                                                                \
3808   switch(control_mask)                                                         \
3809   {                                                                            \
3810     default:                                                                   \
3811     case 0x0:                                                                  \
3812       setup_sprite_tile_column_width_multi(texture_mode, multi, full, full,    \
3813        x4mode);                                                                \
3814       break;                                                                   \
3815                                                                                \
3816     case 0x1:                                                                  \
3817       setup_sprite_tile_column_width_single(texture_mode, multi, full, none,   \
3818        x4mode);                                                                \
3819       break;                                                                   \
3820                                                                                \
3821     case 0x2:                                                                  \
3822       setup_sprite_tile_column_width_multi(texture_mode, single, full, full,   \
3823        x4mode);                                                                \
3824       break;                                                                   \
3825                                                                                \
3826     case 0x3:                                                                  \
3827       setup_sprite_tile_column_width_single(texture_mode, single, full, none,  \
3828        x4mode);                                                                \
3829       break;                                                                   \
3830                                                                                \
3831     case 0x4:                                                                  \
3832       setup_sprite_tile_column_width_multi(texture_mode, multi, half, full,    \
3833        x4mode);                                                                \
3834       break;                                                                   \
3835                                                                                \
3836     case 0x5:                                                                  \
3837       setup_sprite_tile_column_width_single(texture_mode, multi, half, right,  \
3838        x4mode);                                                                \
3839       break;                                                                   \
3840                                                                                \
3841     case 0x6:                                                                  \
3842       setup_sprite_tile_column_width_multi(texture_mode, single, half, full,   \
3843        x4mode);                                                                \
3844       break;                                                                   \
3845                                                                                \
3846     case 0x7:                                                                  \
3847       setup_sprite_tile_column_width_single(texture_mode, single, half, right, \
3848        x4mode);                                                                \
3849       break;                                                                   \
3850                                                                                \
3851     case 0x8:                                                                  \
3852       setup_sprite_tile_column_width_multi(texture_mode, multi, full, half,    \
3853        x4mode);                                                                \
3854       break;                                                                   \
3855                                                                                \
3856     case 0x9:                                                                  \
3857       setup_sprite_tile_column_width_single(texture_mode, multi, half, left,   \
3858        x4mode);                                                                \
3859       break;                                                                   \
3860                                                                                \
3861     case 0xA:                                                                  \
3862       setup_sprite_tile_column_width_multi(texture_mode, single, full, half,   \
3863        x4mode);                                                                \
3864       break;                                                                   \
3865                                                                                \
3866     case 0xB:                                                                  \
3867       setup_sprite_tile_column_width_single(texture_mode, single, half, left,  \
3868        x4mode);                                                                \
3869       break;                                                                   \
3870                                                                                \
3871     case 0xC:                                                                  \
3872       setup_sprite_tile_column_width_multi(texture_mode, multi, half, half,    \
3873        x4mode);                                                                \
3874       break;                                                                   \
3875                                                                                \
3876     case 0xE:                                                                  \
3877       setup_sprite_tile_column_width_multi(texture_mode, single, half, half,   \
3878        x4mode);                                                                \
3879       break;                                                                   \
3880   }                                                                            \
3881 }                                                                              \
3882
3883 void setup_sprite_4bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3884  s32 width, s32 height, u32 color);
3885 void setup_sprite_8bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3886  s32 width, s32 height, u32 color);
3887 void setup_sprite_16bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3888  s32 width, s32 height, u32 color);
3889
3890 void setup_sprite_4bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3891  s32 width, s32 height, u32 color);
3892 void setup_sprite_8bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3893  s32 width, s32 height, u32 color);
3894 void setup_sprite_16bpp_4x(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3895  s32 width, s32 height, u32 color);
3896
3897 void setup_sprite_untextured(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
3898  s32 v, s32 width, s32 height, u32 color);
3899 void setup_sprite_untextured_simple(psx_gpu_struct *psx_gpu, s32 x, s32 y,
3900  s32 u, s32 v, s32 width, s32 height, u32 color);
3901
3902 #ifndef NEON_BUILD
3903 setup_sprite_tiled_builder(4bpp,);
3904 setup_sprite_tiled_builder(8bpp,);
3905
3906 setup_sprite_tiled_builder(4bpp,_4x);
3907 setup_sprite_tiled_builder(8bpp,_4x);
3908
3909 void setup_sprite_16bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
3910  s32 v, s32 width, s32 height, u32 color)
3911 {
3912   u32 left_offset = u & 0x7;
3913   u32 width_rounded = width + left_offset + 7;
3914
3915   u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + (s32)(x - left_offset);
3916   u32 right_width = width_rounded & 0x7;
3917   u32 block_width = width_rounded / 8;
3918   u32 fb_ptr_pitch = (1024 + 8) - (block_width * 8);
3919
3920   u32 left_mask_bits = ~(0xFF << left_offset);
3921   u32 right_mask_bits = 0xFE << right_width;
3922
3923   u32 texture_offset_base = u + (v * 1024);
3924   u32 texture_mask =
3925    psx_gpu->texture_mask_width | (psx_gpu->texture_mask_height * 1024);
3926
3927   u32 blocks_remaining;
3928   u32 num_blocks = psx_gpu->num_blocks;
3929   block_struct *block = psx_gpu->blocks + num_blocks;
3930
3931   u16 *texture_page_ptr = psx_gpu->texture_page_ptr;
3932   u16 *texture_block_ptr;
3933
3934   texture_offset_base &= ~0x7;
3935
3936   sprites_16bpp++;
3937
3938   if(block_width == 1)
3939   {
3940     u32 mask_bits = left_mask_bits | right_mask_bits;
3941
3942     while(height)
3943     {
3944       num_blocks++;
3945       sprite_blocks++;
3946
3947       if(num_blocks > MAX_BLOCKS)
3948       {
3949         flush_render_block_buffer(psx_gpu);
3950         num_blocks = 1;
3951         block = psx_gpu->blocks;
3952       }
3953
3954       texture_block_ptr =
3955        texture_page_ptr + (texture_offset_base & texture_mask);
3956
3957       load_128b(block->texels, texture_block_ptr);
3958       block->draw_mask_bits = mask_bits;
3959       block->fb_ptr = fb_ptr;
3960
3961       block++;
3962
3963       texture_offset_base += 1024;
3964       fb_ptr += 1024;
3965
3966       height--;
3967       psx_gpu->num_blocks = num_blocks;
3968     }
3969   }
3970   else
3971   {
3972     u32 texture_offset;
3973
3974     while(height)
3975     {
3976       blocks_remaining = block_width - 2;
3977       num_blocks += block_width;
3978       sprite_blocks += block_width;
3979
3980       if(num_blocks > MAX_BLOCKS)
3981       {
3982         flush_render_block_buffer(psx_gpu);
3983         num_blocks = block_width;
3984         block = psx_gpu->blocks;
3985       }
3986
3987       texture_offset = texture_offset_base;
3988       texture_offset_base += 1024;
3989
3990       texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);
3991       load_128b(block->texels, texture_block_ptr);
3992
3993       block->draw_mask_bits = left_mask_bits;
3994       block->fb_ptr = fb_ptr;
3995
3996       texture_offset += 8;
3997       fb_ptr += 8;
3998       block++;
3999
4000       while(blocks_remaining)
4001       {
4002         texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);
4003         load_128b(block->texels, texture_block_ptr);
4004
4005         block->draw_mask_bits = 0;
4006         block->fb_ptr = fb_ptr;
4007
4008         texture_offset += 8;
4009         fb_ptr += 8;
4010         block++;
4011
4012         blocks_remaining--;
4013       }
4014
4015       texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);
4016       load_128b(block->texels, texture_block_ptr);
4017
4018       block->draw_mask_bits = right_mask_bits;
4019       block->fb_ptr = fb_ptr;
4020
4021       fb_ptr += fb_ptr_pitch;
4022       block++;
4023
4024       height--;
4025       psx_gpu->num_blocks = num_blocks;
4026     }
4027   }
4028 }
4029
4030 void setup_sprite_untextured(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
4031  s32 v, s32 width, s32 height, u32 color)
4032 {
4033   if((psx_gpu->render_state & (RENDER_STATE_MASK_EVALUATE |
4034    RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND)) == 0)
4035   {
4036     setup_sprite_untextured_simple(psx_gpu, x, y, u, v, width, height, color);
4037     return;
4038   }
4039
4040   u32 right_width = ((width - 1) & 0x7) + 1;
4041   u32 right_mask_bits = (0xFF << right_width);
4042   u16 *fb_ptr = psx_gpu->vram_out_ptr + (y * 1024) + x;
4043   u32 block_width = (width + 7) / 8;
4044   u32 fb_ptr_pitch = 1024 - ((block_width - 1) * 8);
4045   u32 blocks_remaining;
4046   u32 num_blocks = psx_gpu->num_blocks;
4047   block_struct *block = psx_gpu->blocks + num_blocks;
4048
4049   u32 color_r = color & 0xFF;
4050   u32 color_g = (color >> 8) & 0xFF;
4051   u32 color_b = (color >> 16) & 0xFF;
4052   vec_8x16u colors;
4053   vec_8x16u right_mask;
4054   vec_8x16u test_mask = psx_gpu->test_mask;
4055   vec_8x16u zero_mask;
4056
4057   sprites_untextured++;
4058
4059   color = (color_r >> 3) | ((color_g >> 3) << 5) | ((color_b >> 3) << 10);
4060
4061   dup_8x16b(colors, color);
4062   dup_8x16b(zero_mask, 0x00);
4063   dup_8x16b(right_mask, right_mask_bits);
4064   tst_8x16b(right_mask, right_mask, test_mask);
4065
4066   while(height)
4067   {
4068     blocks_remaining = block_width - 1;
4069     num_blocks += block_width;
4070
4071 #ifdef PROFILE
4072     sprite_blocks += block_width;
4073 #endif
4074
4075     if(num_blocks > MAX_BLOCKS)
4076     {
4077       flush_render_block_buffer(psx_gpu);
4078       num_blocks = block_width;
4079       block = psx_gpu->blocks;
4080     }
4081
4082     while(blocks_remaining)
4083     {
4084       block->pixels = colors;
4085       block->draw_mask = zero_mask;
4086       block->fb_ptr = fb_ptr;
4087
4088       fb_ptr += 8;
4089       block++;
4090       blocks_remaining--;
4091     }
4092
4093     block->pixels = colors;
4094     block->draw_mask = right_mask;
4095     block->fb_ptr = fb_ptr;
4096
4097     block++;
4098     fb_ptr += fb_ptr_pitch;
4099
4100     height--;
4101     psx_gpu->num_blocks = num_blocks;
4102   }
4103 }
4104
4105 #endif
4106
4107 void setup_sprite_untextured_simple(psx_gpu_struct *psx_gpu, s32 x, s32 y,
4108  s32 u, s32 v, s32 width, s32 height, u32 color)
4109 {
4110   u32 r = color & 0xFF;
4111   u32 g = (color >> 8) & 0xFF;
4112   u32 b = (color >> 16) & 0xFF;
4113   u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
4114    psx_gpu->mask_msb;
4115   u32 color_32bpp = color_16bpp | (color_16bpp << 16);
4116
4117   u16 *vram_ptr16 = psx_gpu->vram_out_ptr + x + (y * 1024);
4118   u32 *vram_ptr;
4119
4120   u32 num_width;
4121
4122   if(psx_gpu->num_blocks > MAX_BLOCKS)
4123   {
4124     flush_render_block_buffer(psx_gpu);
4125   }
4126
4127   while(height)
4128   {
4129     num_width = width;
4130
4131     vram_ptr = (void *)vram_ptr16;
4132     if((long)vram_ptr16 & 2)
4133     {
4134       *vram_ptr16 = color_32bpp;
4135       vram_ptr = (void *)(vram_ptr16 + 1);
4136       num_width--;
4137     }
4138
4139     while(num_width >= 4 * 2)
4140     {
4141       vram_ptr[0] = color_32bpp;
4142       vram_ptr[1] = color_32bpp;
4143       vram_ptr[2] = color_32bpp;
4144       vram_ptr[3] = color_32bpp;
4145
4146       vram_ptr += 4;
4147       num_width -= 4 * 2;
4148     }
4149
4150     while(num_width >= 2)
4151     {
4152       *vram_ptr++ = color_32bpp;
4153       num_width -= 2;
4154     }
4155
4156     if(num_width > 0)
4157     {
4158       *(u16 *)vram_ptr = color_32bpp;
4159     }
4160
4161     vram_ptr16 += 1024;
4162     height--;
4163   }
4164 }
4165
4166
4167 #define setup_sprite_blocks_switch_textured(texture_mode)                      \
4168   setup_sprite_##texture_mode                                                  \
4169
4170 #define setup_sprite_blocks_switch_untextured(texture_mode)                    \
4171   setup_sprite_untextured                                                      \
4172
4173 #define setup_sprite_blocks_switch(texturing, texture_mode)                    \
4174   setup_sprite_blocks_switch_##texturing(texture_mode)                         \
4175
4176
4177 #define texture_sprite_blocks_switch_4bpp()                                    \
4178   texture_blocks_untextured                                                    \
4179
4180 #define texture_sprite_blocks_switch_8bpp()                                    \
4181   texture_sprite_blocks_8bpp                                                   \
4182
4183 #define texture_sprite_blocks_switch_16bpp()                                   \
4184   texture_blocks_untextured                                                    \
4185
4186 #define texture_sprite_blocks_switch_untextured(texture_mode)                  \
4187   texture_blocks_untextured                                                    \
4188
4189 #define texture_sprite_blocks_switch_textured(texture_mode)                    \
4190   texture_sprite_blocks_switch_##texture_mode()                                \
4191
4192 #define render_sprite_blocks_switch_block_modulation(texture_mode, blend_mode, \
4193  mask_evaluate, shading, dithering, texturing, blending, modulation)           \
4194 {                                                                              \
4195   setup_sprite_blocks_switch(texturing, texture_mode),                         \
4196   texture_sprite_blocks_switch_##texturing(texture_mode),                      \
4197   shade_blocks_switch(unshaded, texturing, modulation, undithered, blending,   \
4198    mask_evaluate),                                                             \
4199   blend_blocks_switch(texturing, blending, blend_mode, mask_evaluate)          \
4200 }                                                                              \
4201
4202 #define render_sprite_blocks_switch_block_blending(texture_mode, blend_mode,   \
4203  mask_evaluate, shading, dithering, texturing, blending)                       \
4204   render_sprite_blocks_switch_block_modulation(texture_mode, blend_mode,       \
4205    mask_evaluate, shading, dithering, texturing, blending, modulated),         \
4206   render_sprite_blocks_switch_block_modulation(texture_mode, blend_mode,       \
4207    mask_evaluate, shading, dithering, texturing, blending, unmodulated)        \
4208
4209 #define render_sprite_blocks_switch_block_texturing(texture_mode, blend_mode,  \
4210  mask_evaluate, shading, dithering, texturing)                                 \
4211   render_sprite_blocks_switch_block_blending(texture_mode, blend_mode,         \
4212    mask_evaluate, shading, dithering, texturing, unblended),                   \
4213   render_sprite_blocks_switch_block_blending(texture_mode, blend_mode,         \
4214    mask_evaluate, shading, dithering, texturing, blended)                      \
4215
4216 #define render_sprite_blocks_switch_block_dithering(texture_mode, blend_mode,  \
4217  mask_evaluate, shading, dithering)                                            \
4218   render_sprite_blocks_switch_block_texturing(texture_mode, blend_mode,        \
4219    mask_evaluate, shading, dithering, untextured),                             \
4220   render_sprite_blocks_switch_block_texturing(texture_mode, blend_mode,        \
4221    mask_evaluate, shading, dithering, textured)                                \
4222
4223 #define render_sprite_blocks_switch_block_shading(texture_mode, blend_mode,    \
4224  mask_evaluate, shading)                                                       \
4225   render_sprite_blocks_switch_block_dithering(texture_mode, blend_mode,        \
4226    mask_evaluate, shading, undithered),                                        \
4227   render_sprite_blocks_switch_block_dithering(texture_mode, blend_mode,        \
4228    mask_evaluate, shading, dithered)                                           \
4229
4230 #define render_sprite_blocks_switch_block_mask_evaluate(texture_mode,          \
4231  blend_mode, mask_evaluate)                                                    \
4232   render_sprite_blocks_switch_block_shading(texture_mode, blend_mode,          \
4233    mask_evaluate, unshaded),                                                   \
4234   render_sprite_blocks_switch_block_shading(texture_mode, blend_mode,          \
4235    mask_evaluate, shaded)                                                      \
4236
4237 #define render_sprite_blocks_switch_block_blend_mode(texture_mode, blend_mode) \
4238   render_sprite_blocks_switch_block_mask_evaluate(texture_mode, blend_mode,    \
4239    off),                                                                       \
4240   render_sprite_blocks_switch_block_mask_evaluate(texture_mode, blend_mode,    \
4241    on)                                                                         \
4242
4243 #define render_sprite_blocks_switch_block_texture_mode(texture_mode)           \
4244   render_sprite_blocks_switch_block_blend_mode(texture_mode, average),         \
4245   render_sprite_blocks_switch_block_blend_mode(texture_mode, add),             \
4246   render_sprite_blocks_switch_block_blend_mode(texture_mode, subtract),        \
4247   render_sprite_blocks_switch_block_blend_mode(texture_mode, add_fourth)       \
4248
4249 #define render_sprite_blocks_switch_block()                                    \
4250   render_sprite_blocks_switch_block_texture_mode(4bpp),                        \
4251   render_sprite_blocks_switch_block_texture_mode(8bpp),                        \
4252   render_sprite_blocks_switch_block_texture_mode(16bpp),                       \
4253   render_sprite_blocks_switch_block_texture_mode(4bpp)                         \
4254
4255
4256 render_block_handler_struct render_sprite_block_handlers[] =
4257 {
4258   render_sprite_blocks_switch_block()
4259 };
4260
4261
4262 void render_sprite(psx_gpu_struct *psx_gpu, s32 x, s32 y, u32 u, u32 v,
4263  s32 width, s32 height, u32 flags, u32 color)
4264 {
4265   s32 x_right = x + width - 1;
4266   s32 y_bottom = y + height - 1;
4267
4268 #ifdef PROFILE
4269   sprites++;
4270 #endif
4271
4272   if(invalidate_texture_cache_region_viewport(psx_gpu, x, y, x_right,
4273    y_bottom) == 0)
4274   {
4275     return;
4276   }
4277
4278   if(x < psx_gpu->viewport_start_x)
4279   {
4280     u32 clip = psx_gpu->viewport_start_x - x;
4281     x += clip;
4282     u += clip;
4283     width -= clip;
4284   }
4285
4286   if(y < psx_gpu->viewport_start_y)
4287   {
4288     s32 clip = psx_gpu->viewport_start_y - y;
4289     y += clip;
4290     v += clip;
4291     height -= clip;
4292   }
4293
4294   if(x_right > psx_gpu->viewport_end_x)
4295     width -= x_right - psx_gpu->viewport_end_x;
4296
4297   if(y_bottom > psx_gpu->viewport_end_y)
4298     height -= y_bottom - psx_gpu->viewport_end_y;
4299
4300   if((width <= 0) || (height <= 0))
4301     return;
4302
4303 #ifdef PROFILE
4304   span_pixels += width * height;
4305   spans += height;
4306 #endif
4307
4308   u32 render_state = flags &
4309    (RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND |
4310    RENDER_FLAGS_TEXTURE_MAP);
4311   render_state |=
4312    (psx_gpu->render_state_base & ~RENDER_STATE_DITHER);
4313
4314   if((psx_gpu->render_state != render_state) ||
4315    (psx_gpu->primitive_type != PRIMITIVE_TYPE_SPRITE))
4316   {
4317     psx_gpu->render_state = render_state;
4318     flush_render_block_buffer(psx_gpu);
4319 #ifdef PROFILE
4320     state_changes++;
4321 #endif
4322   }
4323
4324   psx_gpu->primitive_type = PRIMITIVE_TYPE_SPRITE;
4325
4326   color &= 0xFFFFFF;
4327
4328   if(psx_gpu->triangle_color != color)
4329   {
4330     flush_render_block_buffer(psx_gpu);
4331     psx_gpu->triangle_color = color;
4332   }
4333
4334   if(color == 0x808080)
4335     render_state |= RENDER_FLAGS_MODULATE_TEXELS;
4336
4337   render_block_handler_struct *render_block_handler =
4338    &(render_sprite_block_handlers[render_state]);
4339   psx_gpu->render_block_handler = render_block_handler;
4340
4341   ((setup_sprite_function_type *)render_block_handler->setup_blocks)
4342    (psx_gpu, x, y, u, v, width, height, color);
4343 }
4344
4345 #define draw_pixel_line_mask_evaluate_yes()                                    \
4346   if(*vram_ptr & 0x8000)                                                       \
4347
4348 #define draw_pixel_line_mask_evaluate_no()                                     \
4349
4350
4351 #define draw_pixel_line_shaded()                                               \
4352 {                                                                              \
4353   color_r = fixed_to_int(current_r);                                           \
4354   color_g = fixed_to_int(current_g);                                           \
4355   color_b = fixed_to_int(current_b);                                           \
4356                                                                                \
4357   current_r += gradient_r;                                                     \
4358   current_g += gradient_g;                                                     \
4359   current_b += gradient_b;                                                     \
4360 }                                                                              \
4361
4362 #define draw_pixel_line_unshaded()                                             \
4363 {                                                                              \
4364   color_r = color & 0xFF;                                                      \
4365   color_g = (color >> 8) & 0xFF;                                               \
4366   color_b = (color >> 16) & 0xFF;                                              \
4367 }                                                                              \
4368
4369
4370 #define draw_pixel_line_dithered(_x, _y)                                       \
4371 {                                                                              \
4372   u32 dither_xor = _x ^ _y;                                                    \
4373   s32 dither_offset = (dither_xor >> 1) & 0x1;                                 \
4374   dither_offset |= (_y & 0x1) << 1;                                            \
4375   dither_offset |= (dither_xor & 0x1) << 2;                                    \
4376   dither_offset -= 4;                                                          \
4377                                                                                \
4378   color_r += dither_offset;                                                    \
4379   color_g += dither_offset;                                                    \
4380   color_b += dither_offset;                                                    \
4381                                                                                \
4382   if(color_r < 0)                                                              \
4383     color_r = 0;                                                               \
4384                                                                                \
4385   if(color_g < 0)                                                              \
4386     color_g = 0;                                                               \
4387                                                                                \
4388   if(color_b < 0)                                                              \
4389     color_b = 0;                                                               \
4390                                                                                \
4391   if(color_r > 255)                                                            \
4392     color_r = 255;                                                             \
4393                                                                                \
4394   if(color_g > 255)                                                            \
4395     color_g = 255;                                                             \
4396                                                                                \
4397   if(color_b > 255)                                                            \
4398     color_b = 255;                                                             \
4399 }                                                                              \
4400
4401 #define draw_pixel_line_undithered(_x, _y)                                     \
4402
4403
4404 #define draw_pixel_line_average()                                              \
4405   color_r = (color_r + fb_r) / 2;                                              \
4406   color_g = (color_g + fb_g) / 2;                                              \
4407   color_b = (color_b + fb_b) / 2                                               \
4408
4409 #define draw_pixel_line_add()                                                  \
4410   color_r += fb_r;                                                             \
4411   color_g += fb_g;                                                             \
4412   color_b += fb_b;                                                             \
4413                                                                                \
4414   if(color_r > 31)                                                             \
4415     color_r = 31;                                                              \
4416                                                                                \
4417   if(color_g > 31)                                                             \
4418     color_g = 31;                                                              \
4419                                                                                \
4420   if(color_b > 31)                                                             \
4421     color_b = 31                                                               \
4422                                                                                \
4423
4424 #define draw_pixel_line_subtract()                                             \
4425   color_r = fb_r - color_r;                                                    \
4426   color_g = fb_g - color_g;                                                    \
4427   color_b = fb_b - color_b;                                                    \
4428                                                                                \
4429   if(color_r < 0)                                                              \
4430     color_r = 0;                                                               \
4431                                                                                \
4432   if(color_g < 0)                                                              \
4433     color_g = 0;                                                               \
4434                                                                                \
4435   if(color_b < 0)                                                              \
4436     color_b = 0                                                                \
4437
4438 #define draw_pixel_line_add_fourth()                                           \
4439   color_r = fb_r + (color_r / 4);                                              \
4440   color_g = fb_g + (color_g / 4);                                              \
4441   color_b = fb_b + (color_b / 4);                                              \
4442                                                                                \
4443   if(color_r > 31)                                                             \
4444     color_r = 31;                                                              \
4445                                                                                \
4446   if(color_g > 31)                                                             \
4447     color_g = 31;                                                              \
4448                                                                                \
4449   if(color_b > 31)                                                             \
4450     color_b = 31                                                               \
4451
4452
4453 #define draw_pixel_line_blended(blend_mode)                                    \
4454   s32 fb_pixel = *vram_ptr;                                                    \
4455   s32 fb_r = fb_pixel & 0x1F;                                                  \
4456   s32 fb_g = (fb_pixel >> 5) & 0x1F;                                           \
4457   s32 fb_b = (fb_pixel >> 10) & 0x1F;                                          \
4458                                                                                \
4459   draw_pixel_line_##blend_mode()                                               \
4460
4461 #define draw_pixel_line_unblended(blend_mode)                                  \
4462
4463
4464 #define draw_pixel_line(_x, _y, shading, blending, dithering, mask_evaluate,   \
4465  blend_mode)                                                                   \
4466   if((_x >= psx_gpu->viewport_start_x) && (_y >= psx_gpu->viewport_start_y) && \
4467    (_x <= psx_gpu->viewport_end_x) && (_y <= psx_gpu->viewport_end_y))         \
4468   {                                                                            \
4469     draw_pixel_line_mask_evaluate_##mask_evaluate()                            \
4470     {                                                                          \
4471       draw_pixel_line_##shading();                                             \
4472       draw_pixel_line_##dithering(_x, _y);                                     \
4473                                                                                \
4474       color_r >>= 3;                                                           \
4475       color_g >>= 3;                                                           \
4476       color_b >>= 3;                                                           \
4477                                                                                \
4478       draw_pixel_line_##blending(blend_mode);                                  \
4479                                                                                \
4480       *vram_ptr = color_r | (color_g << 5) | (color_b << 10) |                 \
4481        psx_gpu->mask_msb;                                                      \
4482     }                                                                          \
4483   }                                                                            \
4484
4485 #define update_increment(value)                                                \
4486   value++                                                                      \
4487
4488 #define update_decrement(value)                                                \
4489   value--                                                                      \
4490
4491 #define update_vram_row_increment(value)                                       \
4492   vram_ptr += 1024                                                             \
4493
4494 #define update_vram_row_decrement(value)                                       \
4495   vram_ptr -= 1024                                                             \
4496
4497 #define compare_increment(a, b)                                                \
4498   (a <= b)                                                                     \
4499
4500 #define compare_decrement(a, b)                                                \
4501   (a >= b)                                                                     \
4502
4503 #define set_line_gradients(minor)                                              \
4504 {                                                                              \
4505   s32 gradient_divisor = delta_##minor;                                        \
4506   if(gradient_divisor != 0)                                                    \
4507   {                                                                            \
4508     gradient_r = int_to_fixed(vertex_b->r - vertex_a->r) / gradient_divisor;   \
4509     gradient_g = int_to_fixed(vertex_b->g - vertex_a->g) / gradient_divisor;   \
4510     gradient_b = int_to_fixed(vertex_b->b - vertex_a->b) / gradient_divisor;   \
4511   }                                                                            \
4512   else                                                                         \
4513   {                                                                            \
4514     gradient_r = 0;                                                            \
4515     gradient_g = 0;                                                            \
4516     gradient_b = 0;                                                            \
4517   }                                                                            \
4518   current_r = fixed_center(vertex_a->r);                                       \
4519   current_g = fixed_center(vertex_a->g);                                       \
4520   current_b = fixed_center(vertex_a->b);                                       \
4521 }
4522
4523 #define draw_line_span_horizontal(direction, shading, blending, dithering,     \
4524  mask_evaluate, blend_mode)                                                    \
4525 do                                                                             \
4526 {                                                                              \
4527   error_step = delta_y * 2;                                                    \
4528   error_wrap = delta_x * 2;                                                    \
4529   error = delta_x;                                                             \
4530                                                                                \
4531   current_y = y_a;                                                             \
4532   set_line_gradients(x);                                                       \
4533                                                                                \
4534   for(current_x = x_a; current_x <= x_b; current_x++)                          \
4535   {                                                                            \
4536     draw_pixel_line(current_x, current_y, shading, blending, dithering,        \
4537      mask_evaluate, blend_mode);                                               \
4538     error += error_step;                                                       \
4539     vram_ptr++;                                                                \
4540                                                                                \
4541     if(error >= error_wrap)                                                    \
4542     {                                                                          \
4543       update_##direction(current_y);                                           \
4544       update_vram_row_##direction();                                           \
4545       error -= error_wrap;                                                     \
4546     }                                                                          \
4547   }                                                                            \
4548 } while(0)                                                                     \
4549
4550 #define draw_line_span_vertical(direction, shading, blending, dithering,       \
4551  mask_evaluate, blend_mode)                                                    \
4552 do                                                                             \
4553 {                                                                              \
4554   error_step = delta_x * 2;                                                    \
4555   error_wrap = delta_y * 2;                                                    \
4556   error = delta_y;                                                             \
4557                                                                                \
4558   current_x = x_a;                                                             \
4559   set_line_gradients(y);                                                       \
4560                                                                                \
4561   for(current_y = y_a; compare_##direction(current_y, y_b);                    \
4562    update_##direction(current_y))                                              \
4563   {                                                                            \
4564     draw_pixel_line(current_x, current_y, shading, blending, dithering,        \
4565      mask_evaluate, blend_mode);                                               \
4566     error += error_step;                                                       \
4567     update_vram_row_##direction();                                             \
4568                                                                                \
4569     if(error > error_wrap)                                                     \
4570     {                                                                          \
4571       vram_ptr++;                                                              \
4572       current_x++;                                                             \
4573       error -= error_wrap;                                                     \
4574     }                                                                          \
4575   }                                                                            \
4576 } while(0)                                                                     \
4577
4578
4579 #define render_line_body(shading, blending, dithering, mask_evaluate,          \
4580  blend_mode)                                                                   \
4581   if(delta_y < 0)                                                              \
4582   {                                                                            \
4583     delta_y *= -1;                                                             \
4584                                                                                \
4585     if(delta_x > delta_y)                                                      \
4586     {                                                                          \
4587       draw_line_span_horizontal(decrement, shading, blending, dithering,       \
4588        mask_evaluate, blend_mode);                                             \
4589     }                                                                          \
4590     else                                                                       \
4591     {                                                                          \
4592       draw_line_span_vertical(decrement, shading, blending, dithering,         \
4593        mask_evaluate, blend_mode);                                             \
4594     }                                                                          \
4595   }                                                                            \
4596   else                                                                         \
4597   {                                                                            \
4598     if(delta_x > delta_y)                                                      \
4599     {                                                                          \
4600       draw_line_span_horizontal(increment, shading, blending, dithering,       \
4601        mask_evaluate, blend_mode);                                             \
4602     }                                                                          \
4603     else                                                                       \
4604     {                                                                          \
4605       draw_line_span_vertical(increment, shading, blending, dithering,         \
4606        mask_evaluate, blend_mode);                                             \
4607     }                                                                          \
4608   }                                                                            \
4609
4610
4611 void render_line(psx_gpu_struct *psx_gpu, vertex_struct *vertexes, u32 flags,
4612  u32 color, int double_resolution)
4613 {
4614   s32 color_r, color_g, color_b;
4615   u32 triangle_winding = 0;
4616
4617   fixed_type gradient_r = 0;
4618   fixed_type gradient_g = 0;
4619   fixed_type gradient_b = 0;
4620   fixed_type current_r = 0;
4621   fixed_type current_g = 0;
4622   fixed_type current_b = 0;
4623
4624   s32 y_a, y_b;
4625   s32 x_a, x_b;
4626
4627   s32 delta_x, delta_y;
4628
4629   s32 current_x;
4630   s32 current_y;
4631
4632   u32 error_step;
4633   u32 error;
4634   u32 error_wrap;
4635
4636   u16 *vram_ptr;
4637
4638   flush_render_block_buffer(psx_gpu);
4639   psx_gpu->primitive_type = PRIMITIVE_TYPE_LINE;
4640
4641   vertex_struct *vertex_a = &(vertexes[0]);
4642   vertex_struct *vertex_b = &(vertexes[1]);
4643
4644   u32 control_mask;
4645
4646 #ifdef PROFILE
4647   lines++;
4648 #endif
4649
4650   if(vertex_a->x >= vertex_b->x)
4651   {
4652     vertex_swap(vertex_a, vertex_b);
4653   }
4654
4655   x_a = vertex_a->x;
4656   x_b = vertex_b->x;
4657
4658   y_a = vertex_a->y;
4659   y_b = vertex_b->y;
4660
4661   delta_x = x_b - x_a;
4662   delta_y = y_b - y_a;
4663
4664   if(delta_x >= 1024 || delta_y >= 512 || delta_y <= -512)
4665     return;
4666
4667   if(double_resolution)
4668   {
4669     x_a *= 2;
4670     x_b *= 2;
4671     y_a *= 2;
4672     y_b *= 2;
4673     delta_x *= 2;
4674     delta_y *= 2;
4675   }
4676
4677   flags &= ~RENDER_FLAGS_TEXTURE_MAP;
4678
4679   vram_ptr = psx_gpu->vram_out_ptr + (y_a * 1024) + x_a;
4680
4681   control_mask = 0x0;
4682
4683   if(flags & RENDER_FLAGS_SHADE)
4684     control_mask |= 0x1;
4685
4686   if(flags & RENDER_FLAGS_BLEND)
4687   {
4688     control_mask |= 0x2;
4689     control_mask |= ((psx_gpu->render_state_base >> 6) & 0x3) << 4;
4690   }
4691
4692   if(psx_gpu->render_state_base & RENDER_STATE_DITHER)
4693     control_mask |= 0x4;
4694
4695   if(psx_gpu->render_state_base & RENDER_STATE_MASK_EVALUATE)
4696     control_mask |= 0x8;
4697
4698   switch(control_mask)
4699   {
4700     case 0x0:
4701       render_line_body(unshaded, unblended, undithered, no, none);
4702       break;
4703
4704     case 0x1:
4705       render_line_body(shaded, unblended, undithered, no, none);
4706       break;
4707
4708     case 0x2:
4709       render_line_body(unshaded, blended, undithered, no, average);
4710       break;
4711
4712     case 0x3:
4713       render_line_body(shaded, blended, undithered, no, average);
4714       break;
4715
4716     case 0x4:
4717       render_line_body(unshaded, unblended, dithered, no, none);
4718       break;
4719
4720     case 0x5:
4721       render_line_body(shaded, unblended, dithered, no, none);
4722       break;
4723
4724     case 0x6:
4725       render_line_body(unshaded, blended, dithered, no, average);
4726       break;
4727
4728     case 0x7:
4729       render_line_body(shaded, blended, dithered, no, average);
4730       break;
4731
4732     case 0x8:
4733       render_line_body(unshaded, unblended, undithered, yes, none);
4734       break;
4735
4736     case 0x9:
4737       render_line_body(shaded, unblended, undithered, yes, none);
4738       break;
4739
4740     case 0xA:
4741       render_line_body(unshaded, blended, undithered, yes, average);
4742       break;
4743
4744     case 0xB:
4745       render_line_body(shaded, blended, undithered, yes, average);
4746       break;
4747
4748     case 0xC:
4749       render_line_body(unshaded, unblended, dithered, yes, none);
4750       break;
4751
4752     case 0xD:
4753       render_line_body(shaded, unblended, dithered, yes, none);
4754       break;
4755
4756     case 0xE:
4757       render_line_body(unshaded, blended, dithered, yes, average);
4758       break;
4759
4760     case 0xF:
4761       render_line_body(shaded, blended, dithered, yes, average);
4762       break;
4763
4764     case 0x12:
4765       render_line_body(unshaded, blended, undithered, no, add);
4766       break;
4767
4768     case 0x13:
4769       render_line_body(shaded, blended, undithered, no, add);
4770       break;
4771
4772     case 0x16:
4773       render_line_body(unshaded, blended, dithered, no, add);
4774       break;
4775
4776     case 0x17:
4777       render_line_body(shaded, blended, dithered, no, add);
4778       break;
4779
4780     case 0x1A:
4781       render_line_body(unshaded, blended, undithered, yes, add);
4782       break;
4783
4784     case 0x1B:
4785       render_line_body(shaded, blended, undithered, yes, add);
4786       break;
4787
4788     case 0x1E:
4789       render_line_body(unshaded, blended, dithered, yes, add);
4790       break;
4791
4792     case 0x1F:
4793       render_line_body(shaded, blended, dithered, yes, add);
4794       break;
4795
4796     case 0x22:
4797       render_line_body(unshaded, blended, undithered, no, subtract);
4798       break;
4799
4800     case 0x23:
4801       render_line_body(shaded, blended, undithered, no, subtract);
4802       break;
4803
4804     case 0x26:
4805       render_line_body(unshaded, blended, dithered, no, subtract);
4806       break;
4807
4808     case 0x27:
4809       render_line_body(shaded, blended, dithered, no, subtract);
4810       break;
4811
4812     case 0x2A:
4813       render_line_body(unshaded, blended, undithered, yes, subtract);
4814       break;
4815
4816     case 0x2B:
4817       render_line_body(shaded, blended, undithered, yes, subtract);
4818       break;
4819
4820     case 0x2E:
4821       render_line_body(unshaded, blended, dithered, yes, subtract);
4822       break;
4823
4824     case 0x2F:
4825       render_line_body(shaded, blended, dithered, yes, subtract);
4826       break;
4827
4828     case 0x32:
4829       render_line_body(unshaded, blended, undithered, no, add_fourth);
4830       break;
4831
4832     case 0x33:
4833       render_line_body(shaded, blended, undithered, no, add_fourth);
4834       break;
4835
4836     case 0x36:
4837       render_line_body(unshaded, blended, dithered, no, add_fourth);
4838       break;
4839
4840     case 0x37:
4841       render_line_body(shaded, blended, dithered, no, add_fourth);
4842       break;
4843
4844     case 0x3A:
4845       render_line_body(unshaded, blended, undithered, yes, add_fourth);
4846       break;
4847
4848     case 0x3B:
4849       render_line_body(shaded, blended, undithered, yes, add_fourth);
4850       break;
4851
4852     case 0x3E:
4853       render_line_body(unshaded, blended, dithered, yes, add_fourth);
4854       break;
4855
4856     case 0x3F:
4857       render_line_body(shaded, blended, dithered, yes, add_fourth);
4858       break;
4859   }
4860 }
4861
4862
4863 void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
4864  u32 width, u32 height)
4865 {
4866   if((width == 0) || (height == 0))
4867     return;
4868
4869   invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
4870
4871   u32 r = color & 0xFF;
4872   u32 g = (color >> 8) & 0xFF;
4873   u32 b = (color >> 16) & 0xFF;
4874   u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
4875    psx_gpu->mask_msb;
4876   u32 color_32bpp = color_16bpp | (color_16bpp << 16);
4877
4878   u32 *vram_ptr = (u32 *)(psx_gpu->vram_out_ptr + x + (y * 1024));
4879
4880   u32 pitch = 512 - (width / 2);
4881   u32 num_width;
4882
4883   if(psx_gpu->render_mode & RENDER_INTERLACE_ENABLED)
4884   {
4885     pitch += 512;
4886     height /= 2;
4887
4888     if(psx_gpu->render_mode & RENDER_INTERLACE_ODD)
4889       vram_ptr += 512;
4890   }
4891
4892   while(height)
4893   {
4894     num_width = width;
4895     while(num_width)
4896     {
4897       vram_ptr[0] = color_32bpp;
4898       vram_ptr[1] = color_32bpp;
4899       vram_ptr[2] = color_32bpp;
4900       vram_ptr[3] = color_32bpp;
4901       vram_ptr[4] = color_32bpp;
4902       vram_ptr[5] = color_32bpp;
4903       vram_ptr[6] = color_32bpp;
4904       vram_ptr[7] = color_32bpp;
4905
4906       vram_ptr += 8;
4907       num_width -= 16;
4908     }
4909
4910     vram_ptr += pitch;
4911     height--;
4912   }
4913 }
4914
4915 void render_block_fill_enh(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
4916  u32 width, u32 height)
4917 {
4918   if((width == 0) || (height == 0))
4919     return;
4920
4921   if(width > 1024)
4922     width = 1024;
4923
4924   u32 r = color & 0xFF;
4925   u32 g = (color >> 8) & 0xFF;
4926   u32 b = (color >> 16) & 0xFF;
4927   u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10) |
4928    psx_gpu->mask_msb;
4929   u32 color_32bpp = color_16bpp | (color_16bpp << 16);
4930
4931   u32 *vram_ptr = (u32 *)(psx_gpu->vram_out_ptr + x + (y * 1024));
4932
4933   u32 pitch = 1024 / 2 - (width / 2);
4934   u32 num_width;
4935
4936   while(height)
4937   {
4938     num_width = width;
4939     while(num_width)
4940     {
4941       vram_ptr[0] = color_32bpp;
4942       vram_ptr[1] = color_32bpp;
4943       vram_ptr[2] = color_32bpp;
4944       vram_ptr[3] = color_32bpp;
4945       vram_ptr[4] = color_32bpp;
4946       vram_ptr[5] = color_32bpp;
4947       vram_ptr[6] = color_32bpp;
4948       vram_ptr[7] = color_32bpp;
4949
4950       vram_ptr += 8;
4951       num_width -= 16;
4952     }
4953
4954     vram_ptr += pitch;
4955     height--;
4956   }
4957 }
4958
4959 void render_block_copy(psx_gpu_struct *psx_gpu, u16 *source, u32 x, u32 y,
4960  u32 width, u32 height, u32 pitch)
4961 {
4962   u16 *vram_ptr = psx_gpu->vram_ptr + x + (y * 1024);
4963   u32 draw_x, draw_y;
4964   u32 mask_msb = psx_gpu->mask_msb;
4965
4966   if((width == 0) || (height == 0))
4967     return;
4968
4969   flush_render_block_buffer(psx_gpu);
4970   invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
4971
4972   for(draw_y = 0; draw_y < height; draw_y++)
4973   {
4974     for(draw_x = 0; draw_x < width; draw_x++)
4975     {
4976       vram_ptr[draw_x] = source[draw_x] | mask_msb;
4977     }
4978
4979     source += pitch;
4980     vram_ptr += 1024;
4981   }
4982 }
4983
4984 void render_block_move(psx_gpu_struct *psx_gpu, u32 source_x, u32 source_y,
4985  u32 dest_x, u32 dest_y, u32 width, u32 height)
4986 {
4987   render_block_copy(psx_gpu, psx_gpu->vram_ptr + source_x + (source_y * 1024),
4988    dest_x, dest_y, width, height, 1024);
4989 }
4990
4991
4992 void initialize_reciprocal_table(void)
4993 {
4994   u32 height;
4995   u32 height_normalized;
4996   u32 height_reciprocal;
4997   s32 shift;
4998
4999   for(height = 1; height < sizeof(reciprocal_table)
5000        / sizeof(reciprocal_table[0]); height++)
5001   {
5002     shift = __builtin_clz(height);
5003     height_normalized = height << shift;
5004     height_reciprocal = ((1ULL << 51) + (height_normalized - 1)) /
5005      height_normalized;
5006
5007     shift = 32 - (51 - shift);
5008
5009     reciprocal_table[height] = (height_reciprocal << 10) | shift;
5010   }
5011 }
5012
5013
5014 #define dither_table_row(a, b, c, d)                                           \
5015  ((a & 0xFF) | ((b & 0xFF) << 8) | ((c & 0xFF) << 16) | ((d & 0xFF) << 24))    \
5016
5017 void initialize_psx_gpu(psx_gpu_struct *psx_gpu, u16 *vram)
5018 {
5019   vec_8x16u test_mask =
5020    { { { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80 } } };
5021
5022   psx_gpu->test_mask = test_mask;
5023
5024   psx_gpu->dirty_textures_4bpp_mask = 0xFFFFFFFF;
5025   psx_gpu->dirty_textures_8bpp_mask = 0xFFFFFFFF;
5026   psx_gpu->dirty_textures_8bpp_alternate_mask = 0xFFFFFFFF;
5027   psx_gpu->viewport_mask = 0;
5028   psx_gpu->current_texture_page = 0;
5029   psx_gpu->current_texture_mask = 0;
5030   psx_gpu->last_8bpp_texture_page = 0;
5031
5032   psx_gpu->clut_settings = 0;
5033   psx_gpu->texture_settings = 0;
5034   psx_gpu->render_state = 0;
5035   psx_gpu->render_state_base = 0;
5036   psx_gpu->num_blocks = 0;
5037   psx_gpu->uvrgb_phase = 0x8000;
5038
5039   psx_gpu->vram_ptr = vram;
5040   psx_gpu->vram_out_ptr = vram;
5041
5042   psx_gpu->texture_page_base = psx_gpu->vram_ptr;
5043   psx_gpu->texture_page_ptr = psx_gpu->vram_ptr;
5044   psx_gpu->clut_ptr = psx_gpu->vram_ptr;
5045
5046   psx_gpu->mask_msb = 0;
5047
5048   psx_gpu->texture_window_x = 0;
5049   psx_gpu->texture_window_y = 0;
5050   psx_gpu->texture_mask_width = 0xFF;
5051   psx_gpu->texture_mask_height = 0xFF;
5052
5053   psx_gpu->render_mode = 0;
5054
5055   memset(psx_gpu->vram_ptr, 0, sizeof(u16) * 1024 * 512);
5056
5057   initialize_reciprocal_table();
5058
5059   //    00 01 10 11
5060   // 00  0  4  1  5
5061   // 01  6  2  7  3
5062   // 10  1  5  0  4
5063   // 11  7  3  6  2
5064   // (minus ones(4) * 4)
5065
5066   // d0: (1 3 5 7): x1 ^ y1
5067   // d1: (2 3 6 7): y0
5068   // d2: (4 5 6 7): x0 ^ y0
5069
5070   psx_gpu->dither_table[0] = dither_table_row(-4, 0, -3, 1);
5071   psx_gpu->dither_table[1] = dither_table_row(2, -2, 3, -1);
5072   psx_gpu->dither_table[2] = dither_table_row(-3, 1, -4, 0);
5073   psx_gpu->dither_table[3] = dither_table_row(3, -1, 2, -2);
5074
5075   psx_gpu->primitive_type = PRIMITIVE_TYPE_UNKNOWN;
5076
5077   psx_gpu->enhancement_x_threshold = 256;
5078 }
5079
5080 u64 get_us(void)
5081 {
5082   struct timeval tv;
5083   gettimeofday(&tv, NULL);
5084
5085   return (tv.tv_sec * 1000000ULL) + tv.tv_usec;
5086 }
5087
5088 #ifdef NEON_BUILD
5089
5090 u32 get_counter()
5091 {
5092   u32 counter;
5093   __asm__ volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(counter));
5094
5095   return counter;
5096 }
5097
5098 void init_counter(void)
5099 {
5100   u32 value;
5101   asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(value));
5102   value |= 5; // master enable, ccnt reset
5103   value &= ~8; // ccnt divider 0
5104   asm volatile("mcr p15, 0, %0, c9, c12, 0" :: "r"(value));
5105   // enable cycle counter
5106   asm volatile("mcr p15, 0, %0, c9, c12, 1" :: "r"(1 << 31));
5107 }
5108
5109 void triangle_benchmark(psx_gpu_struct *psx_gpu)
5110 {
5111   u32 i;
5112
5113   u32 ticks;
5114   u32 ticks_elapsed;
5115
5116   const u32 iterations = 500000;
5117
5118   psx_gpu->num_blocks = 64;
5119   psx_gpu->clut_ptr = psx_gpu->vram_ptr;
5120
5121   for(i = 0; i < 64; i++)
5122   {
5123     memset(&(psx_gpu->blocks[i].r), 0, 16);
5124   }
5125
5126   init_counter();
5127
5128   ticks = get_counter();
5129
5130   for(i = 0; i < iterations; i++)
5131   {
5132     texture_sprite_blocks_8bpp(psx_gpu);
5133   }
5134
5135   ticks_elapsed = get_counter() - ticks;
5136
5137   printf("benchmark: %lf cycles\n", (double)ticks_elapsed / (iterations * 64));
5138 }
5139
5140 #endif
5141
5142 #include "psx_gpu_4x.c"