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