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