plugins/gpu_neon/psx_gpu/psx_gpu.c

   1 /*
   2  * Copyright (C) 2011 Gilead Kutnick "Exophase" <exophase@gmail.com>
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public License as
   6  * published by the Free Software Foundation; either version 2 of
   7  * the License, or (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  12  * General Public License for more details.
  13  */
  14
  15 #include <stdio.h>
  16 #include <stdlib.h>
  17 #include <string.h>
  18
  19 #include "common.h"
  20
  21 u32 span_pixels = 0;
  22 u32 span_pixel_blocks = 0;
  23 u32 span_pixel_blocks_unaligned = 0;
  24 u32 spans = 0;
  25 u32 triangles = 0;
  26 u32 sprites = 0;
  27 u32 sprites_4bpp = 0;
  28 u32 sprites_8bpp = 0;
  29 u32 sprites_16bpp = 0;
  30 u32 sprite_blocks = 0;
  31 u32 sprites_untextured = 0;
  32 u32 lines = 0;
  33 u32 trivial_rejects = 0;
  34 u32 texels_4bpp = 0;
  35 u32 texels_8bpp = 0;
  36 u32 texels_16bpp = 0;
  37 u32 texel_blocks_4bpp = 0;
  38 u32 texel_blocks_8bpp = 0;
  39 u32 texel_blocks_16bpp = 0;
  40 u32 texel_blocks_untextured = 0;
  41 u32 blend_blocks = 0;
  42 u32 untextured_pixels = 0;
  43 u32 blend_pixels = 0;
  44 u32 transparent_pixels = 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_triangles = 0;
  53 u32 false_modulated_sprites = 0;
  54
  55 u32 reciprocal_table[512];
  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   psx_gpu->dirty_textures_4bpp_mask |= mask;
 245   psx_gpu->dirty_textures_8bpp_mask |= mask;
 246   psx_gpu->dirty_textures_8bpp_alternate_mask |= mask;
 247
 248   return mask;
 249 }
 250
 251
 252 void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
 253  u32 texture_page);
 254
 255 #ifndef NEON_BUILD
 256
 257 void update_texture_4bpp_cache(psx_gpu_struct *psx_gpu)
 258 {
 259   u32 current_texture_page = psx_gpu->current_texture_page;
 260   u8 *texture_page_ptr = psx_gpu->texture_page_ptr;
 261   u16 *vram_ptr = psx_gpu->vram_ptr;
 262
 263   u32 texel_block;
 264   u32 tile_x, tile_y;
 265   u32 sub_x, sub_y;
 266
 267   vram_ptr += (current_texture_page >> 4) * 256 * 1024;
 268   vram_ptr += (current_texture_page & 0xF) * 64;
 269
 270   texture_cache_loads++;
 271
 272   tile_y = 16;
 273   tile_x = 16;
 274   sub_x = 4;
 275   sub_y = 16;
 276
 277   psx_gpu->dirty_textures_4bpp_mask &= ~(psx_gpu->current_texture_mask);
 278
 279   while(tile_y)
 280   {
 281     while(tile_x)
 282     {
 283       while(sub_y)
 284       {
 285         while(sub_x)
 286         {
 287           texel_block = *vram_ptr;
 288           texture_page_ptr[0] = texel_block & 0xF;
 289           texture_page_ptr[1] = (texel_block >> 4) & 0xF;
 290           texture_page_ptr[2] = (texel_block >> 8) & 0xF;
 291           texture_page_ptr[3] = texel_block >> 12;
 292
 293           vram_ptr++;
 294           texture_page_ptr += 4;
 295
 296           sub_x--;
 297         }
 298
 299         vram_ptr -= 4;
 300         sub_x = 4;
 301
 302         sub_y--;
 303         vram_ptr += 1024;
 304       }
 305
 306       sub_y = 16;
 307
 308       vram_ptr -= (1024 * 16) - 4;
 309       tile_x--;
 310     }
 311
 312     tile_x = 16;
 313
 314     vram_ptr += (16 * 1024) - (4 * 16);
 315     tile_y--;
 316   }
 317 }
 318
 319 void update_texture_8bpp_cache_slice(psx_gpu_struct *psx_gpu,
 320  u32 texture_page)
 321 {
 322   u16 *texture_page_ptr = psx_gpu->texture_page_ptr;
 323   u16 *vram_ptr = psx_gpu->vram_ptr;
 324
 325   u32 tile_x, tile_y;
 326   u32 sub_y;
 327
 328   vec_8x16u texels;
 329
 330   texture_cache_loads++;
 331
 332   vram_ptr += (texture_page >> 4) * 256 * 1024;
 333   vram_ptr += (texture_page & 0xF) * 64;
 334
 335   if((texture_page ^ psx_gpu->current_texture_page) & 0x1)
 336     texture_page_ptr += (8 * 16) * 8;
 337
 338   tile_x = 8;
 339   tile_y = 16;
 340
 341   sub_y = 16;
 342
 343   while(tile_y)
 344   {
 345     while(tile_x)
 346     {
 347       while(sub_y)
 348       {
 349         load_128b(texels, vram_ptr);
 350         store_128b(texels, texture_page_ptr);
 351
 352         texture_page_ptr += 8;
 353         vram_ptr += 1024;
 354
 355         sub_y--;
 356       }
 357
 358       sub_y = 16;
 359
 360       vram_ptr -= (1024 * 16);
 361       vram_ptr += 8;
 362
 363       tile_x--;
 364     }
 365
 366     tile_x = 8;
 367
 368     vram_ptr -= (8 * 8);
 369     vram_ptr += (16 * 1024);
 370
 371     texture_page_ptr += (8 * 16) * 8;
 372     tile_y--;
 373   }
 374 }
 375
 376 #endif
 377
 378
 379 void update_texture_8bpp_cache(psx_gpu_struct *psx_gpu)
 380 {
 381   u32 current_texture_page = psx_gpu->current_texture_page;
 382   u32 update_textures =
 383    psx_gpu->dirty_textures_8bpp_mask & psx_gpu->current_texture_mask;
 384
 385   psx_gpu->dirty_textures_8bpp_mask &= ~update_textures;
 386
 387   if(update_textures & (1 << current_texture_page))
 388   {
 389     update_texture_8bpp_cache_slice(psx_gpu, current_texture_page);
 390     update_textures &= ~(1 << current_texture_page);
 391   }
 392
 393   if(update_textures)
 394   {
 395     u32 adjacent_texture_page = ((current_texture_page + 1) & 0xF) |
 396      (current_texture_page & 0x10);
 397
 398     update_texture_8bpp_cache_slice(psx_gpu, adjacent_texture_page);
 399   }
 400 }
 401
 402 void setup_blocks_shaded_untextured_undithered_unswizzled_indirect(
 403  psx_gpu_struct *psx_gpu);
 404
 405 void flush_render_block_buffer(psx_gpu_struct *psx_gpu)
 406 {
 407   if(psx_gpu->num_blocks)
 408   {
 409     render_block_handler_struct *render_block_handler =
 410      psx_gpu->render_block_handler;
 411
 412     render_block_handler->texture_blocks(psx_gpu);
 413     render_block_handler->shade_blocks(psx_gpu);
 414     render_block_handler->blend_blocks(psx_gpu);
 415
 416     span_pixel_blocks += psx_gpu->num_blocks;
 417     render_buffer_flushes++;
 418
 419     psx_gpu->num_blocks = 0;
 420   }
 421 }
 422
 423
 424 void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
 425  vertex_struct *b, vertex_struct *c);
 426
 427 #ifndef NEON_BUILD
 428
 429 #define setup_gradient_calculation_input(set, vertex)                          \
 430   /* First type is:  uvrg bxxx xxxx                                          */\
 431   /* Second type is: yyyy ybyy uvrg                                          */\
 432   /* Since x_a and y_c are the same the same variable is used for both.      */\
 433   x##set##_a_y##set##_c.e[0] = vertex->u;                                      \
 434   x##set##_a_y##set##_c.e[1] = vertex->v;                                      \
 435   x##set##_a_y##set##_c.e[2] = vertex->r;                                      \
 436   x##set##_a_y##set##_c.e[3] = vertex->g;                                      \
 437   dup_4x16b(x##set##_b, vertex->x);                                            \
 438   dup_4x16b(x##set##_c, vertex->x);                                            \
 439   dup_4x16b(y##set##_a, vertex->y);                                            \
 440   dup_4x16b(y##set##_b, vertex->y);                                            \
 441   x##set##_b.e[0] = vertex->b;                                                 \
 442   y##set##_b.e[1] = vertex->b                                                  \
 443
 444
 445 void compute_all_gradients(psx_gpu_struct *psx_gpu, vertex_struct *a,
 446  vertex_struct *b, vertex_struct *c)
 447 {
 448   u32 triangle_area = psx_gpu->triangle_area;
 449   u32 winding_mask_scalar;
 450
 451   u32 triangle_area_shift;
 452   u64 triangle_area_reciprocal =
 453    fixed_reciprocal(triangle_area, &triangle_area_shift);
 454   triangle_area_shift = -(triangle_area_shift - FIXED_BITS);
 455
 456   // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) =
 457   // ( d0       *  d1      ) - ( d2       *  d3      ) =
 458   // ( m0                  ) - ( m1                  ) = gradient
 459
 460   // This is split to do 12 elements at a time over three sets: a, b, and c.
 461   // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so
 462   // two of the slots are unused.
 463
 464   // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as
 465   // is g.
 466
 467   vec_4x16s x0_a_y0_c, x0_b, x0_c;
 468   vec_4x16s y0_a, y0_b;
 469   vec_4x16s x1_a_y1_c, x1_b, x1_c;
 470   vec_4x16s y1_a, y1_b;
 471   vec_4x16s x2_a_y2_c, x2_b, x2_c;
 472   vec_4x16s y2_a, y2_b;
 473
 474   vec_4x32u uvrg_base;
 475   vec_4x32u b_base;
 476   vec_4x32u const_0x8000;
 477
 478   vec_4x16s d0_a_d3_c, d0_b, d0_c;
 479   vec_4x16s d1_a, d1_b, d1_c_d2_a;
 480   vec_4x16s d2_b, d2_c;
 481   vec_4x16s d3_a, d3_b;
 482
 483   vec_4x32s m0_a, m0_b, m0_c;
 484   vec_4x32s m1_a, m1_b, m1_c;
 485
 486   vec_4x32u gradient_area_a, gradient_area_c;
 487   vec_2x32u gradient_area_b;
 488
 489   vec_4x32u gradient_area_sign_a, gradient_area_sign_c;
 490   vec_2x32u gradient_area_sign_b;
 491   vec_4x32u winding_mask;
 492
 493   vec_2x64u gradient_wide_a0, gradient_wide_a1;
 494   vec_2x64u gradient_wide_c0, gradient_wide_c1;
 495   vec_2x64u gradient_wide_b;
 496
 497   vec_4x32u gradient_a, gradient_c;
 498   vec_2x32u gradient_b;
 499   vec_16x8s gradient_shift;
 500
 501   setup_gradient_calculation_input(0, a);
 502   setup_gradient_calculation_input(1, b);
 503   setup_gradient_calculation_input(2, c);
 504
 505   dup_4x32b(const_0x8000, 0x8000);
 506   shl_long_4x16b(uvrg_base, x0_a_y0_c, 16);
 507   shl_long_4x16b(b_base, x0_b, 16);
 508
 509   add_4x32b(uvrg_base, uvrg_base, const_0x8000);
 510   add_4x32b(b_base, b_base, const_0x8000);
 511
 512   // Can probably pair these, but it'll require careful register allocation
 513   sub_4x16b(d0_a_d3_c, x1_a_y1_c, x0_a_y0_c);
 514   sub_4x16b(d1_c_d2_a, x2_a_y2_c, x1_a_y1_c);
 515
 516   sub_4x16b(d0_b, x1_b, x0_b);
 517   sub_4x16b(d0_c, x1_c, x0_c);
 518
 519   sub_4x16b(d1_a, y2_a, y1_a);
 520   sub_4x16b(d1_b, y2_b, y1_b);
 521
 522   sub_4x16b(d2_b, x2_b, x1_b);
 523   sub_4x16b(d2_c, x2_c, x1_c);
 524
 525   sub_4x16b(d3_a, y1_a, y0_a);
 526   sub_4x16b(d3_b, y1_b, y0_b);
 527
 528   mul_long_4x16b(m0_a, d0_a_d3_c, d1_a);
 529   mul_long_4x16b(m0_b, d0_b, d1_b);
 530   mul_long_4x16b(m0_c, d0_c, d1_c_d2_a);
 531
 532   mul_long_4x16b(m1_a, d1_c_d2_a, d3_a);
 533   mul_long_4x16b(m1_b, d2_b, d3_b);
 534   mul_long_4x16b(m1_c, d2_c, d0_a_d3_c);
 535
 536   sub_4x32b(gradient_area_a, m0_a, m1_a);
 537   sub_2x32b(gradient_area_b, m0_b.low, m1_b.low);
 538   sub_4x32b(gradient_area_c, m0_c, m1_c);
 539
 540   cmpltz_4x32b(gradient_area_sign_a, gradient_area_a);
 541   cmpltz_2x32b(gradient_area_sign_b, gradient_area_b);
 542   cmpltz_4x32b(gradient_area_sign_c, gradient_area_c);
 543
 544   abs_4x32b(gradient_area_a, gradient_area_a);
 545   abs_2x32b(gradient_area_b, gradient_area_b);
 546   abs_4x32b(gradient_area_c, gradient_area_c);
 547
 548   winding_mask_scalar = -psx_gpu->triangle_winding;
 549
 550   dup_4x32b(winding_mask, winding_mask_scalar);
 551   eor_4x32b(gradient_area_sign_a, gradient_area_sign_a, winding_mask);
 552   eor_2x32b(gradient_area_sign_b, gradient_area_sign_b, winding_mask);
 553   eor_4x32b(gradient_area_sign_c, gradient_area_sign_c, winding_mask);
 554
 555   mul_scalar_long_2x32b(gradient_wide_a0,
 556    vector_cast(vec_2x32s, gradient_area_a.low),
 557    (s64)triangle_area_reciprocal);
 558   mul_scalar_long_2x32b(gradient_wide_a1,
 559    vector_cast(vec_2x32s, gradient_area_a.high),
 560    (s64)triangle_area_reciprocal);
 561   mul_scalar_long_2x32b(gradient_wide_b,
 562    vector_cast(vec_2x32s, gradient_area_b),
 563    (s64)triangle_area_reciprocal);
 564   mul_scalar_long_2x32b(gradient_wide_c0,
 565    vector_cast(vec_2x32s, gradient_area_c.low),
 566    (s64)triangle_area_reciprocal);
 567   mul_scalar_long_2x32b(gradient_wide_c1,
 568    vector_cast(vec_2x32s, gradient_area_c.high),
 569    (s64)triangle_area_reciprocal);
 570
 571   dup_16x8b(gradient_shift, triangle_area_shift);
 572   shl_reg_2x64b(gradient_wide_a0, gradient_wide_a0,
 573    vector_cast(vec_2x64u, gradient_shift));
 574   shl_reg_2x64b(gradient_wide_a1, gradient_wide_a1,
 575    vector_cast(vec_2x64u, gradient_shift));
 576   shl_reg_2x64b(gradient_wide_b, gradient_wide_b,
 577    vector_cast(vec_2x64u, gradient_shift));
 578   shl_reg_2x64b(gradient_wide_c0, gradient_wide_c0,
 579    vector_cast(vec_2x64u, gradient_shift));
 580   shl_reg_2x64b(gradient_wide_c1, gradient_wide_c1,
 581    vector_cast(vec_2x64u, gradient_shift));
 582
 583   mov_narrow_2x64b(gradient_a.low, gradient_wide_a0);
 584   mov_narrow_2x64b(gradient_a.high, gradient_wide_a1);
 585   mov_narrow_2x64b(gradient_b, gradient_wide_b);
 586   mov_narrow_2x64b(gradient_c.low, gradient_wide_c0);
 587   mov_narrow_2x64b(gradient_c.high, gradient_wide_c1);
 588
 589   shl_4x32b(gradient_a, gradient_a, 4);
 590   shl_2x32b(gradient_b, gradient_b, 4);
 591   shl_4x32b(gradient_c, gradient_c, 4);
 592
 593   eor_4x32b(gradient_a, gradient_a, gradient_area_sign_a);
 594   eor_2x32b(gradient_b, gradient_b, gradient_area_sign_b);
 595   eor_4x32b(gradient_c, gradient_c, gradient_area_sign_c);
 596
 597   sub_4x32b(gradient_a, gradient_a, gradient_area_sign_a);
 598   sub_2x32b(gradient_b, gradient_b, gradient_area_sign_b);
 599   sub_4x32b(gradient_c, gradient_c, gradient_area_sign_c);
 600
 601   u32 left_adjust = a->x;
 602   mls_scalar_4x32b(uvrg_base, gradient_a, left_adjust);
 603   mls_scalar_2x32b(b_base.low, gradient_b, left_adjust);
 604
 605   vec_4x32u uvrg_dx2;
 606   vec_2x32u b_dx2;
 607
 608   vec_4x32u uvrg_dx3;
 609   vec_2x32u b_dx3;
 610
 611   vec_4x32u zero;
 612
 613   eor_4x32b(zero, zero, zero);
 614   add_4x32b(uvrg_dx2, gradient_a, gradient_a);
 615   add_2x32b(b_dx2, gradient_b, gradient_b);
 616   add_4x32b(uvrg_dx3, gradient_a, uvrg_dx2);
 617   add_2x32b(b_dx3, gradient_b, b_dx2);
 618
 619   // Can be done with vst4, assuming that the zero, dx, dx2, and dx3 are
 620   // lined up properly
 621   psx_gpu->u_block_span.e[0] = zero.e[0];
 622   psx_gpu->u_block_span.e[1] = gradient_a.e[0];
 623   psx_gpu->u_block_span.e[2] = uvrg_dx2.e[0];
 624   psx_gpu->u_block_span.e[3] = uvrg_dx3.e[0];
 625
 626   psx_gpu->v_block_span.e[0] = zero.e[1];
 627   psx_gpu->v_block_span.e[1] = gradient_a.e[1];
 628   psx_gpu->v_block_span.e[2] = uvrg_dx2.e[1];
 629   psx_gpu->v_block_span.e[3] = uvrg_dx3.e[1];
 630
 631   psx_gpu->r_block_span.e[0] = zero.e[2];
 632   psx_gpu->r_block_span.e[1] = gradient_a.e[2];
 633   psx_gpu->r_block_span.e[2] = uvrg_dx2.e[2];
 634   psx_gpu->r_block_span.e[3] = uvrg_dx3.e[2];
 635
 636   psx_gpu->g_block_span.e[0] = zero.e[3];
 637   psx_gpu->g_block_span.e[1] = gradient_a.e[3];
 638   psx_gpu->g_block_span.e[2] = uvrg_dx2.e[3];
 639   psx_gpu->g_block_span.e[3] = uvrg_dx3.e[3];
 640
 641   psx_gpu->b_block_span.e[0] = zero.e[0];
 642   psx_gpu->b_block_span.e[1] = gradient_b.e[0];
 643   psx_gpu->b_block_span.e[2] = b_dx2.e[0];
 644   psx_gpu->b_block_span.e[3] = b_dx3.e[0];
 645
 646   psx_gpu->uvrg = uvrg_base;
 647   psx_gpu->b = b_base.e[0];
 648
 649   psx_gpu->uvrg_dx = gradient_a;
 650   psx_gpu->uvrg_dy = gradient_c;
 651   psx_gpu->b_dy = gradient_b.e[1];
 652 }
 653 #endif
 654
 655 #define vector_check(_a, _b)                                                   \
 656   if(memcmp(&_a, &_b, sizeof(_b)))                                             \
 657   {                                                                            \
 658     if(sizeof(_b) == 8)                                                        \
 659     {                                                                          \
 660       printf("mismatch on %s vs %s: (%x %x) vs (%x %x)\n",                     \
 661        #_a, #_b, _a.e[0], _a.e[1], _b.e[0], _b.e[1]);                          \
 662     }                                                                          \
 663     else                                                                       \
 664     {                                                                          \
 665       printf("mismatch on %s vs %s: (%x %x %x %x) vs (%x %x %x %x)\n",         \
 666        #_a, #_b, _a.e[0], _a.e[1], _a.e[2], _a.e[3], _b.e[0], _b.e[1],         \
 667        _b.e[2], _b.e[3]);                                                      \
 668     }                                                                          \
 669   }                                                                            \
 670
 671 #define scalar_check(_a, _b)                                                   \
 672   if(_a != _b)                                                                 \
 673     printf("mismatch on %s %s: %x vs %x\n", #_a, #_b, _a, _b)                  \
 674
 675
 676 #define setup_spans_prologue_alternate_yes()                                   \
 677   vec_2x64s alternate_x;                                                       \
 678   vec_2x64s alternate_dx_dy;                                                   \
 679   vec_4x32s alternate_x_32;                                                    \
 680   vec_2x32s alternate_x_16;                                                    \
 681                                                                                \
 682   vec_4x16u alternate_select;                                                  \
 683   vec_4x16s y_mid_point;                                                       \
 684                                                                                \
 685   s32 y_b = v_b->y;                                                            \
 686   s64 edge_alt;                                                                \
 687   s32 edge_dx_dy_alt;                                                          \
 688   u32 edge_shift_alt                                                           \
 689
 690 #define setup_spans_prologue_alternate_no()                                    \
 691
 692 #define setup_spans_prologue(alternate_active)                                 \
 693   edge_data_struct *span_edge_data;                                            \
 694   vec_4x32u *span_uvrg_offset;                                                 \
 695   u32 *span_b_offset;                                                          \
 696                                                                                \
 697   s32 clip;                                                                    \
 698                                                                                \
 699   vec_2x64s edges_xy;                                                          \
 700   vec_2x32s edges_dx_dy;                                                       \
 701   vec_2x32u edge_shifts;                                                       \
 702                                                                                \
 703   vec_2x64s left_x, right_x;                                                   \
 704   vec_2x64s left_dx_dy, right_dx_dy;                                           \
 705   vec_4x32s left_x_32, right_x_32;                                             \
 706   vec_8x16s left_right_x_16;                                                   \
 707   vec_4x16s y_x4;                                                              \
 708   vec_8x16s left_edge;                                                         \
 709   vec_8x16s right_edge;                                                        \
 710   vec_4x16u span_shift;                                                        \
 711                                                                                \
 712   vec_2x32u c_0x01;                                                            \
 713   vec_4x16u c_0x04;                                                            \
 714   vec_4x16u c_0xFFFE;                                                          \
 715   vec_4x16u c_0x07;                                                            \
 716                                                                                \
 717   vec_2x32s x_starts;                                                          \
 718   vec_2x32s x_ends;                                                            \
 719                                                                                \
 720   s32 x_a = v_a->x;                                                            \
 721   s32 x_b = v_b->x;                                                            \
 722   s32 x_c = v_c->x;                                                            \
 723   s32 y_a = v_a->y;                                                            \
 724   s32 y_c = v_c->y;                                                            \
 725                                                                                \
 726   vec_4x32u uvrg = psx_gpu->uvrg;                                              \
 727   vec_4x32u uvrg_dy = psx_gpu->uvrg_dy;                                        \
 728   u32 b = psx_gpu->b;                                                          \
 729   u32 b_dy = psx_gpu->b_dy;                                                    \
 730                                                                                \
 731   dup_2x32b(c_0x01, 0x01);                                                     \
 732   setup_spans_prologue_alternate_##alternate_active()                          \
 733
 734 #define setup_spans_prologue_b()                                               \
 735   span_edge_data = psx_gpu->span_edge_data;                                    \
 736   span_uvrg_offset = psx_gpu->span_uvrg_offset;                                \
 737   span_b_offset = psx_gpu->span_b_offset;                                      \
 738                                                                                \
 739   vec_8x16u c_0x0001;                                                          \
 740                                                                                \
 741   dup_8x16b(c_0x0001, 0x0001);                                                 \
 742   dup_8x16b(left_edge, psx_gpu->viewport_start_x);                             \
 743   dup_8x16b(right_edge, psx_gpu->viewport_end_x);                              \
 744   add_8x16b(right_edge, right_edge, c_0x0001);                                 \
 745   dup_4x16b(c_0x04, 0x04);                                                     \
 746   dup_4x16b(c_0x07, 0x07);                                                     \
 747   dup_4x16b(c_0xFFFE, 0xFFFE);                                                 \
 748
 749
 750 #define compute_edge_delta_x2()                                                \
 751 {                                                                              \
 752   vec_2x32s heights;                                                           \
 753   vec_2x32s height_reciprocals;                                                \
 754   vec_2x32s heights_b;                                                         \
 755   vec_4x32u widths;                                                            \
 756                                                                                \
 757   u32 edge_shift = reciprocal_table[height];                                   \
 758                                                                                \
 759   dup_2x32b(heights, height);                                                  \
 760   sub_2x32b(widths, x_ends, x_starts);                                         \
 761                                                                                \
 762   dup_2x32b(edge_shifts, edge_shift);                                          \
 763   sub_2x32b(heights_b, heights, c_0x01);                                       \
 764   shr_2x32b(height_reciprocals, edge_shifts, 12);                              \
 765                                                                                \
 766   mla_2x32b(heights_b, x_starts, heights);                                     \
 767   bic_immediate_4x16b(vector_cast(vec_4x16u, edge_shifts), 0xE0);              \
 768   mul_2x32b(edges_dx_dy, widths, height_reciprocals);                          \
 769   mul_long_2x32b(edges_xy, heights_b, height_reciprocals);                     \
 770 }                                                                              \
 771
 772 #define compute_edge_delta_x3(start_c, height_a, height_b)                     \
 773 {                                                                              \
 774   vec_2x32s heights;                                                           \
 775   vec_2x32s height_reciprocals;                                                \
 776   vec_2x32s heights_b;                                                         \
 777   vec_2x32u widths;                                                            \
 778                                                                                \
 779   u32 width_alt;                                                               \
 780   s32 height_b_alt;                                                            \
 781   u32 height_reciprocal_alt;                                                   \
 782                                                                                \
 783   heights.e[0] = height_a;                                                     \
 784   heights.e[1] = height_b;                                                     \
 785                                                                                \
 786   edge_shifts.e[0] = reciprocal_table[height_a];                               \
 787   edge_shifts.e[1] = reciprocal_table[height_b];                               \
 788   edge_shift_alt = reciprocal_table[height_minor_b];                           \
 789                                                                                \
 790   sub_2x32b(widths, x_ends, x_starts);                                         \
 791   width_alt = x_c - start_c;                                                   \
 792                                                                                \
 793   shr_2x32b(height_reciprocals, edge_shifts, 12);                              \
 794   height_reciprocal_alt = edge_shift_alt >> 12;                                \
 795                                                                                \
 796   bic_immediate_4x16b(vector_cast(vec_4x16u, edge_shifts), 0xE0);              \
 797   edge_shift_alt &= 0x1F;                                                      \
 798                                                                                \
 799   sub_2x32b(heights_b, heights, c_0x01);                                       \
 800   height_b_alt = height_minor_b - 1;                                           \
 801                                                                                \
 802   mla_2x32b(heights_b, x_starts, heights);                                     \
 803   height_b_alt += height_minor_b * start_c;                                    \
 804                                                                                \
 805   mul_long_2x32b(edges_xy, heights_b, height_reciprocals);                     \
 806   edge_alt = (s64)height_b_alt * height_reciprocal_alt;                        \
 807                                                                                \
 808   mul_2x32b(edges_dx_dy, widths, height_reciprocals);                          \
 809   edge_dx_dy_alt = width_alt * height_reciprocal_alt;                          \
 810 }                                                                              \
 811
 812
 813 #define setup_spans_adjust_y_up()                                              \
 814   sub_4x32b(y_x4, y_x4, c_0x04)                                                \
 815
 816 #define setup_spans_adjust_y_down()                                            \
 817   add_4x32b(y_x4, y_x4, c_0x04)                                                \
 818
 819 #define setup_spans_adjust_interpolants_up()                                   \
 820   sub_4x32b(uvrg, uvrg, uvrg_dy);                                              \
 821   b -= b_dy                                                                    \
 822
 823 #define setup_spans_adjust_interpolants_down()                                 \
 824   add_4x32b(uvrg, uvrg, uvrg_dy);                                              \
 825   b += b_dy                                                                    \
 826
 827
 828 #define setup_spans_clip_interpolants_increment()                              \
 829   mla_scalar_4x32b(uvrg, uvrg_dy, clip);                                       \
 830   b += b_dy * clip                                                             \
 831
 832 #define setup_spans_clip_interpolants_decrement()                              \
 833   mls_scalar_4x32b(uvrg, uvrg_dy, clip);                                       \
 834   b -= b_dy * clip                                                             \
 835
 836 #define setup_spans_clip_alternate_yes()                                       \
 837   edge_alt += edge_dx_dy_alt * (s64)(clip)                                     \
 838
 839 #define setup_spans_clip_alternate_no()                                        \
 840
 841 #define setup_spans_clip(direction, alternate_active)                          \
 842 {                                                                              \
 843   clipped_triangles++;                                                         \
 844   mla_scalar_long_2x32b(edges_xy, edges_dx_dy, (s64)clip);                     \
 845   setup_spans_clip_alternate_##alternate_active();                             \
 846   setup_spans_clip_interpolants_##direction();                                 \
 847 }                                                                              \
 848
 849
 850 #define setup_spans_adjust_edges_alternate_no(left_index, right_index)         \
 851 {                                                                              \
 852   vec_2x64u edge_shifts_64;                                                    \
 853   vec_2x64s edges_dx_dy_64;                                                    \
 854                                                                                \
 855   mov_wide_2x32b(edge_shifts_64, edge_shifts);                                 \
 856   shl_variable_2x64b(edges_xy, edges_xy, edge_shifts_64);                      \
 857                                                                                \
 858   mov_wide_2x32b(edges_dx_dy_64, edges_dx_dy);                                 \
 859   shl_variable_2x64b(edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64);          \
 860                                                                                \
 861   left_x.e[0] = edges_xy.e[left_index];                                        \
 862   right_x.e[0] = edges_xy.e[right_index];                                      \
 863                                                                                \
 864   left_dx_dy.e[0] = edges_dx_dy_64.e[left_index];                              \
 865   left_dx_dy.e[1] = edges_dx_dy_64.e[left_index];                              \
 866   right_dx_dy.e[0] = edges_dx_dy_64.e[right_index];                            \
 867   right_dx_dy.e[1] = edges_dx_dy_64.e[right_index];                            \
 868                                                                                \
 869   add_1x64b(left_x.high, left_x.low, left_dx_dy.low);                          \
 870   add_1x64b(right_x.high, right_x.low, right_dx_dy.low);                       \
 871                                                                                \
 872   add_2x64b(left_dx_dy, left_dx_dy, left_dx_dy);                               \
 873   add_2x64b(right_dx_dy, right_dx_dy, right_dx_dy);                            \
 874 }                                                                              \
 875
 876 #define setup_spans_adjust_edges_alternate_yes(left_index, right_index)        \
 877 {                                                                              \
 878   setup_spans_adjust_edges_alternate_no(left_index, right_index);              \
 879   s64 edge_dx_dy_alt_64;                                                       \
 880                                                                                \
 881   dup_4x16b(y_mid_point, y_b);                                                 \
 882                                                                                \
 883   edge_alt <<= edge_shift_alt;                                                 \
 884   edge_dx_dy_alt_64 = (s64)edge_dx_dy_alt << edge_shift_alt;                   \
 885                                                                                \
 886   alternate_x.e[0] = edge_alt;                                                 \
 887   alternate_dx_dy.e[0] = edge_dx_dy_alt_64;                                    \
 888   alternate_dx_dy.e[1] = edge_dx_dy_alt_64;                                    \
 889                                                                                \
 890   add_1x64b(alternate_x.high, alternate_x.low, alternate_dx_dy.low);           \
 891   add_2x64b(alternate_dx_dy, alternate_dx_dy, alternate_dx_dy);                \
 892 }                                                                              \
 893
 894
 895 #define setup_spans_y_select_up()                                              \
 896   cmplt_4x16b(alternate_select, y_x4, y_mid_point)                             \
 897
 898 #define setup_spans_y_select_down()                                            \
 899   cmpgt_4x16b(alternate_select, y_x4, y_mid_point)                             \
 900
 901 #define setup_spans_y_select_alternate_yes(direction)                          \
 902   setup_spans_y_select_##direction()                                           \
 903
 904 #define setup_spans_y_select_alternate_no(direction)                           \
 905
 906 #define setup_spans_alternate_select_left()                                    \
 907   bit_4x16b(left_right_x_16.low, alternate_x_16, alternate_select)             \
 908
 909 #define setup_spans_alternate_select_right()                                   \
 910   bit_4x16b(left_right_x_16.high, alternate_x_16, alternate_select)            \
 911
 912 #define setup_spans_alternate_select_none()                                    \
 913
 914 #define setup_spans_increment_alternate_yes()                                  \
 915   shr_narrow_2x64b(alternate_x_32.low, alternate_x, 32);                       \
 916   add_2x64b(alternate_x, alternate_x, alternate_dx_dy);                        \
 917   shr_narrow_2x64b(alternate_x_32.high, alternate_x, 32);                      \
 918   add_2x64b(alternate_x, alternate_x, alternate_dx_dy);                        \
 919   mov_narrow_4x32b(alternate_x_16, alternate_x_32)                             \
 920
 921 #define setup_spans_increment_alternate_no()                                   \
 922
 923 #define setup_spans_set_x4(alternate, direction, alternate_active)             \
 924 {                                                                              \
 925   span_uvrg_offset[0] = uvrg;                                                  \
 926   span_b_offset[0] = b;                                                        \
 927   setup_spans_adjust_interpolants_##direction();                               \
 928                                                                                \
 929   span_uvrg_offset[1] = uvrg;                                                  \
 930   span_b_offset[1] = b;                                                        \
 931   setup_spans_adjust_interpolants_##direction();                               \
 932                                                                                \
 933   span_uvrg_offset[2] = uvrg;                                                  \
 934   span_b_offset[2] = b;                                                        \
 935   setup_spans_adjust_interpolants_##direction();                               \
 936                                                                                \
 937   span_uvrg_offset[3] = uvrg;                                                  \
 938   span_b_offset[3] = b;                                                        \
 939   setup_spans_adjust_interpolants_##direction();                               \
 940                                                                                \
 941   span_uvrg_offset += 4;                                                       \
 942   span_b_offset += 4;                                                          \
 943                                                                                \
 944   shr_narrow_2x64b(left_x_32.low, left_x, 32);                                 \
 945   shr_narrow_2x64b(right_x_32.low, right_x, 32);                               \
 946                                                                                \
 947   add_2x64b(left_x, left_x, left_dx_dy);                                       \
 948   add_2x64b(right_x, right_x, right_dx_dy);                                    \
 949                                                                                \
 950   shr_narrow_2x64b(left_x_32.high, left_x, 32);                                \
 951   shr_narrow_2x64b(right_x_32.high, right_x, 32);                              \
 952                                                                                \
 953   add_2x64b(left_x, left_x, left_dx_dy);                                       \
 954   add_2x64b(right_x, right_x, right_dx_dy);                                    \
 955                                                                                \
 956   mov_narrow_4x32b(left_right_x_16.low, left_x_32);                            \
 957   mov_narrow_4x32b(left_right_x_16.high, right_x_32);                          \
 958                                                                                \
 959   setup_spans_increment_alternate_##alternate_active();                        \
 960   setup_spans_y_select_alternate_##alternate_active(direction);                \
 961   setup_spans_alternate_select_##alternate();                                  \
 962                                                                                \
 963   max_8x16b(left_right_x_16, left_right_x_16, left_edge);                      \
 964   min_8x16b(left_right_x_16, left_right_x_16, right_edge);                     \
 965                                                                                \
 966   sub_4x16b(left_right_x_16.high, left_right_x_16.high, left_right_x_16.low);  \
 967   add_4x16b(left_right_x_16.high, left_right_x_16.high, c_0x07);               \
 968   and_4x16b(span_shift, left_right_x_16.high, c_0x07);                         \
 969   shl_variable_4x16b(span_shift, c_0xFFFE, span_shift);                        \
 970   shr_4x16b(left_right_x_16.high, left_right_x_16.high, 3);                    \
 971                                                                                \
 972   u32 i;                                                                       \
 973   for(i = 0; i < 4; i++)                                                       \
 974   {                                                                            \
 975     span_edge_data[i].left_x = left_right_x_16.low.e[i];                       \
 976     span_edge_data[i].num_blocks = left_right_x_16.high.e[i];                  \
 977     span_edge_data[i].right_mask = span_shift.e[i];                            \
 978     span_edge_data[i].y = y_x4.e[i];                                           \
 979   }                                                                            \
 980                                                                                \
 981   span_edge_data += 4;                                                         \
 982                                                                                \
 983   setup_spans_adjust_y_##direction();                                          \
 984 }                                                                              \
 985
 986
 987 #define setup_spans_alternate_adjust_yes()                                     \
 988   edge_alt -= edge_dx_dy_alt * (s64)height_minor_a                             \
 989
 990 #define setup_spans_alternate_adjust_no()                                      \
 991
 992
 993 #define setup_spans_down(left_index, right_index, alternate, alternate_active) \
 994   setup_spans_alternate_adjust_##alternate_active();                           \
 995   if(y_c > psx_gpu->viewport_end_y)                                            \
 996     height -= y_c - psx_gpu->viewport_end_y - 1;                               \
 997                                                                                \
 998   clip = psx_gpu->viewport_start_y - y_a;                                      \
 999   if(clip > 0)                                                                 \
1000   {                                                                            \
1001     height -= clip;                                                            \
1002     y_a += clip;                                                               \
1003     setup_spans_clip(increment, alternate_active);                             \
1004   }                                                                            \
1005                                                                                \
1006   setup_spans_prologue_b();                                                    \
1007                                                                                \
1008   if(height > 0)                                                               \
1009   {                                                                            \
1010     y_x4.e[0] = y_a;                                                           \
1011     y_x4.e[1] = y_a + 1;                                                       \
1012     y_x4.e[2] = y_a + 2;                                                       \
1013     y_x4.e[3] = y_a + 3;                                                       \
1014     setup_spans_adjust_edges_alternate_##alternate_active(left_index,          \
1015      right_index);                                                             \
1016                                                                                \
1017     psx_gpu->num_spans = height;                                               \
1018     do                                                                         \
1019     {                                                                          \
1020       setup_spans_set_x4(alternate, down, alternate_active);                   \
1021       height -= 4;                                                             \
1022     } while(height > 0);                                                       \
1023   }                                                                            \
1024
1025
1026 #define setup_spans_alternate_pre_increment_yes()                              \
1027   edge_alt += edge_dx_dy_alt                                                   \
1028
1029 #define setup_spans_alternate_pre_increment_no()                               \
1030
1031 #define setup_spans_up_decrement_height_yes()                                  \
1032   height--                                                                     \
1033
1034 #define setup_spans_up_decrement_height_no()                                   \
1035   {}                                                                           \
1036
1037 #define setup_spans_up(left_index, right_index, alternate, alternate_active)   \
1038   setup_spans_alternate_adjust_##alternate_active();                           \
1039   y_a--;                                                                       \
1040                                                                                \
1041   if(y_c < psx_gpu->viewport_start_y)                                          \
1042     height -= psx_gpu->viewport_start_y - y_c;                                 \
1043   else                                                                         \
1044     setup_spans_up_decrement_height_##alternate_active();                      \
1045                                                                                \
1046   clip = y_a - psx_gpu->viewport_end_y;                                        \
1047   if(clip > 0)                                                                 \
1048   {                                                                            \
1049     height -= clip;                                                            \
1050     y_a -= clip;                                                               \
1051     setup_spans_clip(decrement, alternate_active);                             \
1052   }                                                                            \
1053                                                                                \
1054   setup_spans_prologue_b();                                                    \
1055                                                                                \
1056   if(height > 0)                                                               \
1057   {                                                                            \
1058     y_x4.e[0] = y_a;                                                           \
1059     y_x4.e[1] = y_a - 1;                                                       \
1060     y_x4.e[2] = y_a - 2;                                                       \
1061     y_x4.e[3] = y_a - 3;                                                       \
1062     add_wide_2x32b(edges_xy, edges_xy, edges_dx_dy);                           \
1063     setup_spans_alternate_pre_increment_##alternate_active();                  \
1064     setup_spans_adjust_edges_alternate_##alternate_active(left_index,          \
1065      right_index);                                                             \
1066     setup_spans_adjust_interpolants_up();                                      \
1067                                                                                \
1068     psx_gpu->num_spans = height;                                               \
1069     while(height > 0)                                                          \
1070     {                                                                          \
1071       setup_spans_set_x4(alternate, up, alternate_active);                     \
1072       height -= 4;                                                             \
1073     }                                                                          \
1074   }                                                                            \
1075
1076 #define index_left  0
1077 #define index_right 1
1078
1079 #define setup_spans_up_up(minor, major)                                        \
1080   setup_spans_prologue(yes);                                                   \
1081   s32 height_minor_a = y_a - y_b;                                              \
1082   s32 height_minor_b = y_b - y_c;                                              \
1083   s32 height = y_a - y_c;                                                      \
1084                                                                                \
1085   dup_2x32b(x_starts, x_a);                                                    \
1086   x_ends.e[0] = x_c;                                                           \
1087   x_ends.e[1] = x_b;                                                           \
1088                                                                                \
1089   compute_edge_delta_x3(x_b, height, height_minor_a);                          \
1090   setup_spans_up(index_##major, index_##minor, minor, yes)                     \
1091
1092
1093 void setup_spans_up_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1094  vertex_struct *v_b, vertex_struct *v_c);
1095 void setup_spans_up_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1096  vertex_struct *v_b, vertex_struct *v_c);
1097 void setup_spans_down_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1098  vertex_struct *v_b, vertex_struct *v_c);
1099 void setup_spans_down_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1100  vertex_struct *v_b, vertex_struct *v_c);
1101 void setup_spans_up_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1102  vertex_struct *v_b, vertex_struct *v_c);
1103 void setup_spans_up_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1104  vertex_struct *v_b, vertex_struct *v_c);
1105 void setup_spans_down_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1106  vertex_struct *v_b, vertex_struct *v_c);
1107 void setup_spans_down_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1108  vertex_struct *v_b, vertex_struct *v_c);
1109 void setup_spans_up_down(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1110  vertex_struct *v_b, vertex_struct *v_c);
1111
1112
1113 #ifndef NEON_BUILD
1114
1115 void setup_spans_up_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1116  vertex_struct *v_b, vertex_struct *v_c)
1117 {
1118   setup_spans_up_up(left, right);
1119 }
1120
1121 void setup_spans_up_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1122  vertex_struct *v_b, vertex_struct *v_c)
1123 {
1124   setup_spans_up_up(right, left);
1125 }
1126
1127 #define setup_spans_down_down(minor, major)                                    \
1128   setup_spans_prologue(yes);                                                   \
1129   s32 height_minor_a = y_b - y_a;                                              \
1130   s32 height_minor_b = y_c - y_b;                                              \
1131   s32 height = y_c - y_a;                                                      \
1132                                                                                \
1133   dup_2x32b(x_starts, x_a);                                                    \
1134   x_ends.e[0] = x_c;                                                           \
1135   x_ends.e[1] = x_b;                                                           \
1136                                                                                \
1137   compute_edge_delta_x3(x_b, height, height_minor_a);                          \
1138   setup_spans_down(index_##major, index_##minor, minor, yes)                   \
1139
1140 void setup_spans_down_left(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1141  vertex_struct *v_b, vertex_struct *v_c)
1142 {
1143   setup_spans_down_down(left, right);
1144 }
1145
1146 void setup_spans_down_right(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1147  vertex_struct *v_b, vertex_struct *v_c)
1148 {
1149   setup_spans_down_down(right, left);
1150 }
1151
1152 #define setup_spans_up_flat()                                                  \
1153   s32 height = y_a - y_c;                                                      \
1154                                                                                \
1155   flat_triangles++;                                                            \
1156   compute_edge_delta_x2();                                                     \
1157   setup_spans_up(index_left, index_right, none, no)                            \
1158
1159 void setup_spans_up_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1160  vertex_struct *v_b, vertex_struct *v_c)
1161 {
1162   setup_spans_prologue(no);
1163   x_starts.e[0] = x_a;
1164   x_starts.e[1] = x_b;
1165   dup_2x32b(x_ends, x_c);
1166
1167   setup_spans_up_flat();
1168 }
1169
1170 void setup_spans_up_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1171  vertex_struct *v_b, vertex_struct *v_c)
1172 {
1173   setup_spans_prologue(no);
1174   dup_2x32b(x_starts, x_a);
1175   x_ends.e[0] = x_b;
1176   x_ends.e[1] = x_c;
1177
1178   setup_spans_up_flat();
1179 }
1180
1181 #define setup_spans_down_flat()                                                \
1182   s32 height = y_c - y_a;                                                      \
1183                                                                                \
1184   flat_triangles++;                                                            \
1185   compute_edge_delta_x2();                                                     \
1186   setup_spans_down(index_left, index_right, none, no)                          \
1187
1188 void setup_spans_down_a(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1189  vertex_struct *v_b, vertex_struct *v_c)
1190 {
1191   setup_spans_prologue(no);
1192   x_starts.e[0] = x_a;
1193   x_starts.e[1] = x_b;
1194   dup_2x32b(x_ends, x_c);
1195
1196   setup_spans_down_flat();
1197 }
1198
1199 void setup_spans_down_b(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1200  vertex_struct *v_b, vertex_struct *v_c)
1201 {
1202   setup_spans_prologue(no);
1203   dup_2x32b(x_starts, x_a);
1204   x_ends.e[0] = x_b;
1205   x_ends.e[1] = x_c;
1206
1207   setup_spans_down_flat();
1208 }
1209
1210 void setup_spans_up_down(psx_gpu_struct *psx_gpu, vertex_struct *v_a,
1211  vertex_struct *v_b, vertex_struct *v_c)
1212 {
1213   setup_spans_prologue(no);
1214
1215   s32 y_b = v_b->y;
1216   s64 edge_alt;
1217   s32 edge_dx_dy_alt;
1218   u32 edge_shift_alt;
1219
1220   s32 middle_y = y_a;
1221   s32 height_minor_a = y_a - y_b;
1222   s32 height_minor_b = y_c - y_a;
1223   s32 height_major = y_c - y_b;
1224
1225   vec_2x64s edges_xy_b;
1226   vec_2x32s edges_dx_dy_b;
1227   vec_2x32u edge_shifts_b;
1228
1229   vec_2x32s height_increment;
1230
1231   x_starts.e[0] = x_a;
1232   x_starts.e[1] = x_c;
1233   dup_2x32b(x_ends, x_b);
1234
1235   compute_edge_delta_x3(x_a, height_minor_a, height_major);
1236
1237   height_increment.e[0] = 0;
1238   height_increment.e[1] = height_minor_b;
1239
1240   mla_long_2x32b(edges_xy, edges_dx_dy, height_increment);
1241
1242   edges_xy_b.e[0] = edge_alt;
1243   edges_xy_b.e[1] = edges_xy.e[1];
1244
1245   edge_shifts_b = edge_shifts;
1246   edge_shifts_b.e[0] = edge_shift_alt;
1247
1248   neg_2x32b(edges_dx_dy_b, edges_dx_dy);
1249   edges_dx_dy_b.e[0] = edge_dx_dy_alt;
1250
1251   y_a--;
1252
1253   if(y_b < psx_gpu->viewport_start_y)
1254     height_minor_a -= psx_gpu->viewport_start_y - y_b;
1255
1256   clip = y_a - psx_gpu->viewport_end_y;
1257   if(clip > 0)
1258   {
1259     height_minor_a -= clip;
1260     y_a -= clip;
1261     setup_spans_clip(decrement, no);
1262   }
1263
1264   setup_spans_prologue_b();
1265
1266   if(height_minor_a > 0)
1267   {
1268     y_x4.e[0] = y_a;
1269     y_x4.e[1] = y_a - 1;
1270     y_x4.e[2] = y_a - 2;
1271     y_x4.e[3] = y_a - 3;
1272     add_wide_2x32b(edges_xy, edges_xy, edges_dx_dy);
1273     setup_spans_adjust_edges_alternate_no(index_left, index_right);
1274     setup_spans_adjust_interpolants_up();
1275
1276     psx_gpu->num_spans = height_minor_a;
1277     while(height_minor_a > 0)
1278     {
1279       setup_spans_set_x4(none, up, no);
1280       height_minor_a -= 4;
1281     }
1282
1283     span_edge_data += height_minor_a;
1284     span_uvrg_offset += height_minor_a;
1285     span_b_offset += height_minor_a;
1286   }
1287
1288   edges_xy = edges_xy_b;
1289   edges_dx_dy = edges_dx_dy_b;
1290   edge_shifts = edge_shifts_b;
1291
1292   uvrg = psx_gpu->uvrg;
1293   b = psx_gpu->b;
1294
1295   y_a = middle_y;
1296
1297   if(y_c > psx_gpu->viewport_end_y)
1298     height_minor_b -= y_c - psx_gpu->viewport_end_y - 1;
1299
1300   clip = psx_gpu->viewport_start_y - y_a;
1301   if(clip > 0)
1302   {
1303     height_minor_b -= clip;
1304     y_a += clip;
1305     setup_spans_clip(increment, no);
1306   }
1307
1308   if(height_minor_b > 0)
1309   {
1310     y_x4.e[0] = y_a;
1311     y_x4.e[1] = y_a + 1;
1312     y_x4.e[2] = y_a + 2;
1313     y_x4.e[3] = y_a + 3;
1314     setup_spans_adjust_edges_alternate_no(index_left, index_right);
1315
1316     psx_gpu->num_spans += height_minor_b;
1317     do
1318     {
1319       setup_spans_set_x4(none, down, no);
1320       height_minor_b -= 4;
1321     } while(height_minor_b > 0);
1322   }
1323
1324   left_split_triangles++;
1325 }
1326
1327 #endif
1328
1329
1330 #define dither_table_entry_normal(value)                                       \
1331   (value)                                                                      \
1332
1333
1334 #define setup_blocks_load_msb_mask_indirect()                                  \
1335
1336 #define setup_blocks_load_msb_mask_direct()                                    \
1337   vec_8x16u msb_mask;                                                          \
1338   dup_8x16b(msb_mask, psx_gpu->mask_msb);                                      \
1339
1340
1341 #define setup_blocks_variables_shaded_textured(target)                         \
1342   vec_4x32u u_block;                                                           \
1343   vec_4x32u v_block;                                                           \
1344   vec_4x32u r_block;                                                           \
1345   vec_4x32u g_block;                                                           \
1346   vec_4x32u b_block;                                                           \
1347   vec_4x32u uvrg_dx = psx_gpu->uvrg_dx;                                        \
1348   vec_4x32u uvrg_dx4;                                                          \
1349   vec_4x32u uvrg_dx8;                                                          \
1350   vec_4x32u uvrg;                                                              \
1351   u32 b_dx = psx_gpu->b_block_span.e[1];                                       \
1352   u32 b_dx4 = b_dx << 2;                                                       \
1353   u32 b_dx8 = b_dx << 3;                                                       \
1354   u32 b;                                                                       \
1355                                                                                \
1356   vec_16x8u texture_mask;                                                      \
1357   shl_4x32b(uvrg_dx4, uvrg_dx, 2);                                             \
1358   shl_4x32b(uvrg_dx8, uvrg_dx, 3);                                             \
1359   dup_8x8b(texture_mask.low, psx_gpu->texture_mask_width);                     \
1360   dup_8x8b(texture_mask.high, psx_gpu->texture_mask_height)                    \
1361
1362 #define setup_blocks_variables_shaded_untextured(target)                       \
1363   vec_4x32u r_block;                                                           \
1364   vec_4x32u g_block;                                                           \
1365   vec_4x32u b_block;                                                           \
1366   vec_4x32u rgb_dx;                                                            \
1367   vec_4x32u rgb_dx4;                                                           \
1368   vec_4x32u rgb_dx8;                                                           \
1369   vec_4x32u rgb;                                                               \
1370                                                                                \
1371   vec_8x8u d64_0x07;                                                           \
1372   vec_8x8u d64_1;                                                              \
1373   vec_8x8u d64_4;                                                              \
1374   vec_8x8u d64_128;                                                            \
1375                                                                                \
1376   dup_8x8b(d64_0x07, 0x07);                                                    \
1377   dup_8x8b(d64_1, 1);                                                          \
1378   dup_8x8b(d64_4, 4);                                                          \
1379   dup_8x8b(d64_128, 128);                                                      \
1380                                                                                \
1381   rgb_dx.low = psx_gpu->uvrg_dx.high;                                          \
1382   rgb_dx.e[2] = psx_gpu->b_block_span.e[1];                                    \
1383   shl_4x32b(rgb_dx4, rgb_dx, 2);                                               \
1384   shl_4x32b(rgb_dx8, rgb_dx, 3)                                                \
1385
1386 #define setup_blocks_variables_unshaded_textured(target)                       \
1387   vec_4x32u u_block;                                                           \
1388   vec_4x32u v_block;                                                           \
1389   vec_2x32u uv_dx = psx_gpu->uvrg_dx.low;                                      \
1390   vec_2x32u uv_dx4;                                                            \
1391   vec_2x32u uv_dx8;                                                            \
1392   vec_2x32u uv = psx_gpu->uvrg.low;                                            \
1393                                                                                \
1394   vec_16x8u texture_mask;                                                      \
1395   shl_2x32b(uv_dx4, uv_dx, 2);                                                 \
1396   shl_2x32b(uv_dx8, uv_dx, 3);                                                 \
1397   dup_8x8b(texture_mask.low, psx_gpu->texture_mask_width);                     \
1398   dup_8x8b(texture_mask.high, psx_gpu->texture_mask_height)                    \
1399
1400
1401 #define setup_blocks_variables_unshaded_untextured_direct()                    \
1402   or_8x16b(colors, colors, msb_mask)                                           \
1403
1404 #define setup_blocks_variables_unshaded_untextured_indirect()                  \
1405
1406 #define setup_blocks_variables_unshaded_untextured(target)                     \
1407   u32 color = psx_gpu->triangle_color;                                         \
1408   vec_8x16u colors;                                                            \
1409                                                                                \
1410   u32 color_r = color & 0xFF;                                                  \
1411   u32 color_g = (color >> 8) & 0xFF;                                           \
1412   u32 color_b = (color >> 16) & 0xFF;                                          \
1413                                                                                \
1414   color = (color_r >> 3) | ((color_g >> 3) << 5) |                             \
1415    ((color_b >> 3) << 10);                                                     \
1416   dup_8x16b(colors, color);                                                    \
1417   setup_blocks_variables_unshaded_untextured_##target()                        \
1418
1419 #define setup_blocks_span_initialize_dithered_textured()                       \
1420   vec_8x16u dither_offsets;                                                    \
1421   shl_long_8x8b(dither_offsets, dither_offsets_short, 4)                       \
1422
1423 #define setup_blocks_span_initialize_dithered_untextured()                     \
1424   vec_8x8u dither_offsets;                                                     \
1425   add_8x8b(dither_offsets, dither_offsets_short, d64_4)                        \
1426
1427 #define setup_blocks_span_initialize_dithered(texturing)                       \
1428   u32 dither_row = psx_gpu->dither_table[y & 0x3];                             \
1429   u32 dither_shift = (span_edge_data->left_x & 0x3) * 8;                       \
1430   vec_8x8s dither_offsets_short;                                               \
1431                                                                                \
1432   dither_row =                                                                 \
1433    (dither_row >> dither_shift) | (dither_row << (32 - dither_shift));         \
1434   dup_2x32b(vector_cast(vec_2x32u, dither_offsets_short), dither_row);         \
1435   setup_blocks_span_initialize_dithered_##texturing()                          \
1436
1437 #define setup_blocks_span_initialize_undithered(texturing)                     \
1438
1439
1440 #define setup_blocks_span_initialize_shaded_textured()                         \
1441 {                                                                              \
1442   vec_4x32u block_span;                                                        \
1443   u32 offset = span_edge_data->left_x;                                         \
1444                                                                                \
1445   uvrg = *span_uvrg_offset;                                                    \
1446   mla_scalar_4x32b(uvrg, uvrg_dx, offset);                                     \
1447   b = *span_b_offset;                                                          \
1448   b += b_dx * offset;                                                          \
1449                                                                                \
1450   dup_4x32b(u_block, uvrg.e[0]);                                               \
1451   dup_4x32b(v_block, uvrg.e[1]);                                               \
1452   dup_4x32b(r_block, uvrg.e[2]);                                               \
1453   dup_4x32b(g_block, uvrg.e[3]);                                               \
1454   dup_4x32b(b_block, b);                                                       \
1455                                                                                \
1456   block_span = psx_gpu->u_block_span;                                          \
1457   add_4x32b(u_block, u_block, block_span);                                     \
1458   block_span = psx_gpu->v_block_span;                                          \
1459   add_4x32b(v_block, v_block, block_span);                                     \
1460   block_span = psx_gpu->r_block_span;                                          \
1461   add_4x32b(r_block, r_block, block_span);                                     \
1462   block_span = psx_gpu->g_block_span;                                          \
1463   add_4x32b(g_block, g_block, block_span);                                     \
1464   block_span = psx_gpu->b_block_span;                                          \
1465   add_4x32b(b_block, b_block, block_span);                                     \
1466 }
1467
1468 #define setup_blocks_span_initialize_shaded_untextured()                       \
1469 {                                                                              \
1470   vec_4x32u block_span;                                                        \
1471   u32 offset = span_edge_data->left_x;                                         \
1472                                                                                \
1473   rgb.low = span_uvrg_offset->high;                                            \
1474   rgb.high.e[0] = *span_b_offset;                                              \
1475   mla_scalar_4x32b(rgb, rgb_dx, offset);                                       \
1476                                                                                \
1477   dup_4x32b(r_block, rgb.e[0]);                                                \
1478   dup_4x32b(g_block, rgb.e[1]);                                                \
1479   dup_4x32b(b_block, rgb.e[2]);                                                \
1480                                                                                \
1481   block_span = psx_gpu->r_block_span;                                          \
1482   add_4x32b(r_block, r_block, block_span);                                     \
1483   block_span = psx_gpu->g_block_span;                                          \
1484   add_4x32b(g_block, g_block, block_span);                                     \
1485   block_span = psx_gpu->b_block_span;                                          \
1486   add_4x32b(b_block, b_block, block_span);                                     \
1487 }                                                                              \
1488
1489 #define setup_blocks_span_initialize_unshaded_textured()                       \
1490 {                                                                              \
1491   vec_4x32u block_span;                                                        \
1492   u32 offset = span_edge_data->left_x;                                         \
1493                                                                                \
1494   uv = span_uvrg_offset->low;                                                  \
1495   mla_scalar_2x32b(uv, uv_dx, offset);                                         \
1496                                                                                \
1497   dup_4x32b(u_block, uv.e[0]);                                                 \
1498   dup_4x32b(v_block, uv.e[1]);                                                 \
1499                                                                                \
1500   block_span = psx_gpu->u_block_span;                                          \
1501   add_4x32b(u_block, u_block, block_span);                                     \
1502   block_span = psx_gpu->v_block_span;                                          \
1503   add_4x32b(v_block, v_block, block_span);                                     \
1504 }                                                                              \
1505
1506 #define setup_blocks_span_initialize_unshaded_untextured()                     \
1507
1508
1509 #define setup_blocks_texture_swizzled()                                        \
1510 {                                                                              \
1511   vec_8x8u u_saved = u;                                                        \
1512   sli_8x8b(u, v, 4);                                                           \
1513   sri_8x8b(v, u_saved, 4);                                                     \
1514 }                                                                              \
1515
1516 #define setup_blocks_texture_unswizzled()                                      \
1517
1518 #define setup_blocks_store_shaded_textured(swizzling, dithering, target,       \
1519  edge_type)                                                                    \
1520 {                                                                              \
1521   vec_8x16u u_whole;                                                           \
1522   vec_8x16u v_whole;                                                           \
1523   vec_8x16u r_whole;                                                           \
1524   vec_8x16u g_whole;                                                           \
1525   vec_8x16u b_whole;                                                           \
1526                                                                                \
1527   vec_8x8u u;                                                                  \
1528   vec_8x8u v;                                                                  \
1529   vec_8x8u r;                                                                  \
1530   vec_8x8u g;                                                                  \
1531   vec_8x8u b;                                                                  \
1532   vec_8x16u uv;                                                                \
1533                                                                                \
1534   vec_4x32u dx4;                                                               \
1535   vec_4x32u dx8;                                                               \
1536                                                                                \
1537   shr_narrow_4x32b(u_whole.low, u_block, 16);                                  \
1538   shr_narrow_4x32b(v_whole.low, v_block, 16);                                  \
1539   shr_narrow_4x32b(r_whole.low, r_block, 16);                                  \
1540   shr_narrow_4x32b(g_whole.low, g_block, 16);                                  \
1541   shr_narrow_4x32b(b_whole.low, b_block, 16);                                  \
1542                                                                                \
1543   dup_4x32b(dx4, uvrg_dx4.e[0]);                                               \
1544   add_high_narrow_4x32b(u_whole.high, u_block, dx4);                           \
1545   dup_4x32b(dx4, uvrg_dx4.e[1]);                                               \
1546   add_high_narrow_4x32b(v_whole.high, v_block, dx4);                           \
1547   dup_4x32b(dx4, uvrg_dx4.e[2]);                                               \
1548   add_high_narrow_4x32b(r_whole.high, r_block, dx4);                           \
1549   dup_4x32b(dx4, uvrg_dx4.e[3]);                                               \
1550   add_high_narrow_4x32b(g_whole.high, g_block, dx4);                           \
1551   dup_4x32b(dx4, b_dx4);                                                       \
1552   add_high_narrow_4x32b(b_whole.high, b_block, dx4);                           \
1553                                                                                \
1554   mov_narrow_8x16b(u, u_whole);                                                \
1555   mov_narrow_8x16b(v, v_whole);                                                \
1556   mov_narrow_8x16b(r, r_whole);                                                \
1557   mov_narrow_8x16b(g, g_whole);                                                \
1558   mov_narrow_8x16b(b, b_whole);                                                \
1559                                                                                \
1560   dup_4x32b(dx8, uvrg_dx8.e[0]);                                               \
1561   add_4x32b(u_block, u_block, dx8);                                            \
1562   dup_4x32b(dx8, uvrg_dx8.e[1]);                                               \
1563   add_4x32b(v_block, v_block, dx8);                                            \
1564   dup_4x32b(dx8, uvrg_dx8.e[2]);                                               \
1565   add_4x32b(r_block, r_block, dx8);                                            \
1566   dup_4x32b(dx8, uvrg_dx8.e[3]);                                               \
1567   add_4x32b(g_block, g_block, dx8);                                            \
1568   dup_4x32b(dx8, b_dx8);                                                       \
1569   add_4x32b(b_block, b_block, dx8);                                            \
1570                                                                                \
1571   and_8x8b(u, u, texture_mask.low);                                            \
1572   and_8x8b(v, v, texture_mask.high);                                           \
1573   setup_blocks_texture_##swizzling();                                          \
1574                                                                                \
1575   zip_8x16b(uv, u, v);                                                         \
1576   block->uv = uv;                                                              \
1577   block->r = r;                                                                \
1578   block->g = g;                                                                \
1579   block->b = b;                                                                \
1580   block->dither_offsets = vector_cast(vec_8x16u, dither_offsets);              \
1581   block->fb_ptr = fb_ptr;                                                      \
1582 }                                                                              \
1583
1584 #define setup_blocks_store_unshaded_textured(swizzling, dithering, target,     \
1585  edge_type)                                                                    \
1586 {                                                                              \
1587   vec_8x16u u_whole;                                                           \
1588   vec_8x16u v_whole;                                                           \
1589                                                                                \
1590   vec_8x8u u;                                                                  \
1591   vec_8x8u v;                                                                  \
1592   vec_8x16u uv;                                                                \
1593                                                                                \
1594   vec_4x32u dx4;                                                               \
1595   vec_4x32u dx8;                                                               \
1596                                                                                \
1597   shr_narrow_4x32b(u_whole.low, u_block, 16);                                  \
1598   shr_narrow_4x32b(v_whole.low, v_block, 16);                                  \
1599                                                                                \
1600   dup_4x32b(dx4, uv_dx4.e[0]);                                                 \
1601   add_high_narrow_4x32b(u_whole.high, u_block, dx4);                           \
1602   dup_4x32b(dx4, uv_dx4.e[1]);                                                 \
1603   add_high_narrow_4x32b(v_whole.high, v_block, dx4);                           \
1604                                                                                \
1605   mov_narrow_8x16b(u, u_whole);                                                \
1606   mov_narrow_8x16b(v, v_whole);                                                \
1607                                                                                \
1608   dup_4x32b(dx8, uv_dx8.e[0]);                                                 \
1609   add_4x32b(u_block, u_block, dx8);                                            \
1610   dup_4x32b(dx8, uv_dx8.e[1]);                                                 \
1611   add_4x32b(v_block, v_block, dx8);                                            \
1612                                                                                \
1613   and_8x8b(u, u, texture_mask.low);                                            \
1614   and_8x8b(v, v, texture_mask.high);                                           \
1615   setup_blocks_texture_##swizzling();                                          \
1616                                                                                \
1617   zip_8x16b(uv, u, v);                                                         \
1618   block->uv = uv;                                                              \
1619   block->dither_offsets = vector_cast(vec_8x16u, dither_offsets);              \
1620   block->fb_ptr = fb_ptr;                                                      \
1621 }                                                                              \
1622
1623 #define setup_blocks_store_shaded_untextured_dithered()                        \
1624   addq_8x8b(r, r, dither_offsets);                                             \
1625   addq_8x8b(g, g, dither_offsets);                                             \
1626   addq_8x8b(b, b, dither_offsets);                                             \
1627                                                                                \
1628   subq_8x8b(r, r, d64_4);                                                      \
1629   subq_8x8b(g, g, d64_4);                                                      \
1630   subq_8x8b(b, b, d64_4)                                                       \
1631
1632 #define setup_blocks_store_shaded_untextured_undithered()                      \
1633
1634
1635 #define setup_blocks_store_untextured_pixels_indirect_full(_pixels)            \
1636   block->pixels = _pixels;                                                     \
1637   block->fb_ptr = fb_ptr                                                       \
1638
1639 #define setup_blocks_store_untextured_pixels_indirect_edge(_pixels)            \
1640   block->pixels = _pixels;                                                     \
1641   block->fb_ptr = fb_ptr                                                       \
1642
1643 #define setup_blocks_store_shaded_untextured_seed_pixels_indirect()            \
1644   mul_long_8x8b(pixels, r, d64_1)                                              \
1645
1646
1647 #define setup_blocks_store_untextured_pixels_direct_full(_pixels)              \
1648   store_8x16b(_pixels, fb_ptr)                                                 \
1649
1650 #define setup_blocks_store_untextured_pixels_direct_edge(_pixels)              \
1651 {                                                                              \
1652   vec_8x16u fb_pixels;                                                         \
1653   vec_8x16u draw_mask;                                                         \
1654   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
1655                                                                                \
1656   load_8x16b(fb_pixels, fb_ptr);                                               \
1657   dup_8x16b(draw_mask, span_edge_data->right_mask);                            \
1658   tst_8x16b(draw_mask, draw_mask, test_mask);                                  \
1659   bif_8x16b(fb_pixels, _pixels, draw_mask);                                    \
1660   store_8x16b(fb_pixels, fb_ptr);                                              \
1661 }                                                                              \
1662
1663 #define setup_blocks_store_shaded_untextured_seed_pixels_direct()              \
1664   pixels = msb_mask;                                                           \
1665   mla_long_8x8b(pixels, r, d64_1)                                              \
1666
1667
1668 #define setup_blocks_store_shaded_untextured(swizzling, dithering, target,     \
1669  edge_type)                                                                    \
1670 {                                                                              \
1671   vec_8x16u r_whole;                                                           \
1672   vec_8x16u g_whole;                                                           \
1673   vec_8x16u b_whole;                                                           \
1674                                                                                \
1675   vec_8x8u r;                                                                  \
1676   vec_8x8u g;                                                                  \
1677   vec_8x8u b;                                                                  \
1678                                                                                \
1679   vec_4x32u dx4;                                                               \
1680   vec_4x32u dx8;                                                               \
1681                                                                                \
1682   vec_8x16u pixels;                                                            \
1683                                                                                \
1684   shr_narrow_4x32b(r_whole.low, r_block, 16);                                  \
1685   shr_narrow_4x32b(g_whole.low, g_block, 16);                                  \
1686   shr_narrow_4x32b(b_whole.low, b_block, 16);                                  \
1687                                                                                \
1688   dup_4x32b(dx4, rgb_dx4.e[0]);                                                \
1689   add_high_narrow_4x32b(r_whole.high, r_block, dx4);                           \
1690   dup_4x32b(dx4, rgb_dx4.e[1]);                                                \
1691   add_high_narrow_4x32b(g_whole.high, g_block, dx4);                           \
1692   dup_4x32b(dx4, rgb_dx4.e[2]);                                                \
1693   add_high_narrow_4x32b(b_whole.high, b_block, dx4);                           \
1694                                                                                \
1695   mov_narrow_8x16b(r, r_whole);                                                \
1696   mov_narrow_8x16b(g, g_whole);                                                \
1697   mov_narrow_8x16b(b, b_whole);                                                \
1698                                                                                \
1699   dup_4x32b(dx8, rgb_dx8.e[0]);                                                \
1700   add_4x32b(r_block, r_block, dx8);                                            \
1701   dup_4x32b(dx8, rgb_dx8.e[1]);                                                \
1702   add_4x32b(g_block, g_block, dx8);                                            \
1703   dup_4x32b(dx8, rgb_dx8.e[2]);                                                \
1704   add_4x32b(b_block, b_block, dx8);                                            \
1705                                                                                \
1706   setup_blocks_store_shaded_untextured_##dithering();                          \
1707                                                                                \
1708   shr_8x8b(r, r, 3);                                                           \
1709   bic_8x8b(g, g, d64_0x07);                                                    \
1710   bic_8x8b(b, b, d64_0x07);                                                    \
1711                                                                                \
1712   setup_blocks_store_shaded_untextured_seed_pixels_##target();                 \
1713   mla_long_8x8b(pixels, g, d64_4);                                             \
1714   mla_long_8x8b(pixels, b, d64_128)                                            \
1715                                                                                \
1716   setup_blocks_store_untextured_pixels_##target##_##edge_type(pixels);         \
1717 }                                                                              \
1718
1719 #define setup_blocks_store_unshaded_untextured(swizzling, dithering, target,   \
1720  edge_type)                                                                    \
1721   setup_blocks_store_untextured_pixels_##target##_##edge_type(colors)          \
1722
1723
1724 #define setup_blocks_store_draw_mask_textured_indirect(_block, bits)           \
1725   (_block)->draw_mask_bits = bits                                              \
1726
1727 #define setup_blocks_store_draw_mask_untextured_indirect(_block, bits)         \
1728 {                                                                              \
1729   vec_8x16u bits_mask;                                                         \
1730   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
1731   dup_8x16b(bits_mask, bits);                                                  \
1732   tst_8x16b(bits_mask, bits_mask, test_mask);                                  \
1733   (_block)->draw_mask = bits_mask;                                             \
1734 }                                                                              \
1735
1736 #define setup_blocks_store_draw_mask_untextured_direct(_block, bits)           \
1737
1738
1739 #define setup_blocks_add_blocks_indirect()                                     \
1740   num_blocks += span_num_blocks;                                               \
1741                                                                                \
1742   if(num_blocks > MAX_BLOCKS)                                                  \
1743   {                                                                            \
1744     psx_gpu->num_blocks = num_blocks - span_num_blocks;                        \
1745     flush_render_block_buffer(psx_gpu);                                        \
1746     num_blocks = span_num_blocks;                                              \
1747     block = psx_gpu->blocks;                                                   \
1748   }                                                                            \
1749
1750 #define setup_blocks_add_blocks_direct()                                       \
1751
1752
1753 #define setup_blocks_builder(shading, texturing, dithering, sw, target)        \
1754 void setup_blocks_##shading##_##texturing##_##dithering##_##sw##_##target(     \
1755  psx_gpu_struct *psx_gpu)                                                      \
1756 {                                                                              \
1757   setup_blocks_load_msb_mask_##target();                                       \
1758   setup_blocks_variables_##shading##_##texturing(target);                      \
1759                                                                                \
1760   edge_data_struct *span_edge_data = psx_gpu->span_edge_data;                  \
1761   vec_4x32u *span_uvrg_offset = psx_gpu->span_uvrg_offset;                     \
1762   u32 *span_b_offset = psx_gpu->span_b_offset;                                 \
1763                                                                                \
1764   block_struct *block = psx_gpu->blocks + psx_gpu->num_blocks;                 \
1765                                                                                \
1766   u32 num_spans = psx_gpu->num_spans;                                          \
1767                                                                                \
1768   u16 *fb_ptr;                                                                 \
1769   u32 y;                                                                       \
1770                                                                                \
1771   u32 num_blocks = psx_gpu->num_blocks;                                        \
1772   u32 span_num_blocks;                                                         \
1773                                                                                \
1774   while(num_spans)                                                             \
1775   {                                                                            \
1776     span_num_blocks = span_edge_data->num_blocks;                              \
1777     if(span_num_blocks)                                                        \
1778     {                                                                          \
1779       y = span_edge_data->y;                                                   \
1780       fb_ptr = psx_gpu->vram_ptr + span_edge_data->left_x + (y * 1024);        \
1781                                                                                \
1782       setup_blocks_span_initialize_##shading##_##texturing();                  \
1783       setup_blocks_span_initialize_##dithering(texturing);                     \
1784                                                                                \
1785       setup_blocks_add_blocks_##target();                                      \
1786                                                                                \
1787       s32 pixel_span = span_num_blocks * 8;                                    \
1788       pixel_span -= __builtin_popcount(span_edge_data->right_mask & 0xFF);     \
1789       span_pixels += pixel_span;                                               \
1790       span_pixel_blocks_unaligned += (pixel_span + 7) / 8;                     \
1791                                                                                \
1792       span_num_blocks--;                                                       \
1793       while(span_num_blocks)                                                   \
1794       {                                                                        \
1795         setup_blocks_store_##shading##_##texturing(sw, dithering, target,      \
1796          full);                                                                \
1797         setup_blocks_store_draw_mask_##texturing##_##target(block, 0x00);      \
1798                                                                                \
1799         fb_ptr += 8;                                                           \
1800         block++;                                                               \
1801         span_num_blocks--;                                                     \
1802       }                                                                        \
1803                                                                                \
1804       setup_blocks_store_##shading##_##texturing(sw, dithering, target, edge); \
1805       setup_blocks_store_draw_mask_##texturing##_##target(block,               \
1806        span_edge_data->right_mask);                                            \
1807                                                                                \
1808       block++;                                                                 \
1809     }                                                                          \
1810     else                                                                       \
1811     {                                                                          \
1812       zero_block_spans++;                                                      \
1813     }                                                                          \
1814                                                                                \
1815     num_spans--;                                                               \
1816     span_edge_data++;                                                          \
1817     span_uvrg_offset++;                                                        \
1818     span_b_offset++;                                                           \
1819   }                                                                            \
1820                                                                                \
1821   psx_gpu->num_blocks = num_blocks;                                            \
1822 }                                                                              \
1823
1824 void setup_blocks_shaded_textured_dithered_unswizzled_indirect(psx_gpu_struct
1825  *psx_gpu);
1826
1827 void setup_blocks_shaded_untextured_dithered_unswizzled_indirect(psx_gpu_struct
1828  *psx_gpu);
1829 void setup_blocks_shaded_untextured_undithered_unswizzled_indirect(
1830  psx_gpu_struct *psx_gpu);
1831 void setup_blocks_shaded_untextured_dithered_unswizzled_direct(psx_gpu_struct
1832  *psx_gpu);
1833 void setup_blocks_shaded_untextured_undithered_unswizzled_direct(
1834  psx_gpu_struct *psx_gpu);
1835
1836 void setup_blocks_unshaded_textured_dithered_unswizzled_indirect(psx_gpu_struct
1837  *psx_gpu);
1838 void setup_blocks_unshaded_untextured_undithered_unswizzled_indirect(
1839  psx_gpu_struct *psx_gpu);
1840 void setup_blocks_unshaded_untextured_undithered_unswizzled_direct(
1841  psx_gpu_struct *psx_gpu);
1842
1843 void setup_blocks_shaded_textured_dithered_swizzled_indirect(psx_gpu_struct
1844  *psx_gpu);
1845 void setup_blocks_unshaded_textured_dithered_swizzled_indirect(psx_gpu_struct
1846  *psx_gpu);
1847
1848
1849 //setup_blocks_builder(unshaded, untextured, undithered, unswizzled, direct);
1850
1851 #ifndef NEON_BUILD
1852
1853 setup_blocks_builder(shaded, textured, dithered, swizzled, indirect);
1854 setup_blocks_builder(shaded, textured, dithered, unswizzled, indirect);
1855
1856 setup_blocks_builder(unshaded, textured, dithered, unswizzled, indirect);
1857 setup_blocks_builder(unshaded, textured, dithered, swizzled, indirect);
1858
1859 setup_blocks_builder(shaded, untextured, undithered, unswizzled, indirect);
1860 setup_blocks_builder(shaded, untextured, dithered, unswizzled, indirect);
1861 setup_blocks_builder(shaded, untextured, undithered, unswizzled, direct);
1862 setup_blocks_builder(shaded, untextured, dithered, unswizzled, direct);
1863
1864 setup_blocks_builder(unshaded, untextured, undithered, unswizzled, indirect);
1865 setup_blocks_builder(unshaded, untextured, undithered, unswizzled, direct);
1866
1867 #endif
1868
1869 void texture_blocks_untextured(psx_gpu_struct *psx_gpu);
1870 void texture_blocks_4bpp(psx_gpu_struct *psx_gpu);
1871 void texture_blocks_8bpp(psx_gpu_struct *psx_gpu);
1872 void texture_blocks_16bpp(psx_gpu_struct *psx_gpu);
1873
1874 #ifndef NEON_BUILD
1875
1876 void texture_blocks_untextured(psx_gpu_struct *psx_gpu)
1877 {
1878   if(psx_gpu->primitive_type != PRIMITIVE_TYPE_SPRITE)
1879     texel_blocks_untextured += psx_gpu->num_blocks;
1880 }
1881
1882 void texture_blocks_4bpp(psx_gpu_struct *psx_gpu)
1883 {
1884   block_struct *block = psx_gpu->blocks;
1885   u32 num_blocks = psx_gpu->num_blocks;
1886   texel_blocks_4bpp += num_blocks;
1887
1888   vec_8x8u texels_low;
1889   vec_8x8u texels_high;
1890   vec_8x8u texels;
1891   vec_8x16u pixels;
1892
1893   vec_8x16u clut_a;
1894   vec_8x16u clut_b;
1895   vec_16x8u clut_low;
1896   vec_16x8u clut_high;
1897
1898   u8 *texture_ptr_8bpp = psx_gpu->texture_page_ptr;
1899   u16 *clut_ptr = psx_gpu->clut_ptr;
1900
1901   // Can be done with one deinterleaving load on NEON
1902   load_8x16b(clut_a, clut_ptr);
1903   load_8x16b(clut_b, clut_ptr + 8);
1904   unzip_16x8b(clut_low, clut_high, clut_a, clut_b);
1905
1906   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_4bpp_mask)
1907     update_texture_4bpp_cache(psx_gpu);
1908
1909   while(num_blocks)
1910   {
1911     texels.e[0] = texture_ptr_8bpp[block->uv.e[0]];
1912     texels.e[1] = texture_ptr_8bpp[block->uv.e[1]];
1913     texels.e[2] = texture_ptr_8bpp[block->uv.e[2]];
1914     texels.e[3] = texture_ptr_8bpp[block->uv.e[3]];
1915     texels.e[4] = texture_ptr_8bpp[block->uv.e[4]];
1916     texels.e[5] = texture_ptr_8bpp[block->uv.e[5]];
1917     texels.e[6] = texture_ptr_8bpp[block->uv.e[6]];
1918     texels.e[7] = texture_ptr_8bpp[block->uv.e[7]];
1919
1920     tbl_16(texels_low, texels, clut_low);
1921     tbl_16(texels_high, texels, clut_high);
1922
1923     // Can be done with an interleaving store on NEON
1924     zip_8x16b(pixels, texels_low, texels_high);
1925
1926     block->texels = pixels;
1927
1928     num_blocks--;
1929     block++;
1930   }
1931 }
1932
1933 void texture_blocks_8bpp(psx_gpu_struct *psx_gpu)
1934 {
1935   block_struct *block = psx_gpu->blocks;
1936   u32 num_blocks = psx_gpu->num_blocks;
1937
1938   texel_blocks_8bpp += num_blocks;
1939
1940   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_8bpp_mask)
1941     update_texture_8bpp_cache(psx_gpu);
1942
1943   vec_8x16u texels;
1944   u8 *texture_ptr_8bpp = psx_gpu->texture_page_ptr;
1945
1946   u32 texel;
1947   u32 offset;
1948   u32 i;
1949
1950   while(num_blocks)
1951   {
1952     for(i = 0; i < 8; i++)
1953     {
1954       offset = block->uv.e[i];
1955
1956       texel = texture_ptr_8bpp[offset];
1957       texels.e[i] = psx_gpu->clut_ptr[texel];
1958     }
1959
1960     block->texels = texels;
1961
1962     num_blocks--;
1963     block++;
1964   }
1965 }
1966
1967 void texture_blocks_16bpp(psx_gpu_struct *psx_gpu)
1968 {
1969   block_struct *block = psx_gpu->blocks;
1970   u32 num_blocks = psx_gpu->num_blocks;
1971
1972   texel_blocks_16bpp += num_blocks;
1973
1974   vec_8x16u texels;
1975
1976   u16 *texture_ptr_16bpp = psx_gpu->texture_page_ptr;
1977   u32 offset;
1978   u32 i;
1979
1980   while(num_blocks)
1981   {
1982     for(i = 0; i < 8; i++)
1983     {
1984       offset = block->uv.e[i];
1985       offset += ((offset & 0xFF00) * 3);
1986
1987       texels.e[i] = texture_ptr_16bpp[offset];
1988     }
1989
1990     block->texels = texels;
1991
1992     num_blocks--;
1993     block++;
1994   }
1995 }
1996
1997 #endif
1998
1999
2000 #define shade_blocks_load_msb_mask_indirect()                                  \
2001
2002 #define shade_blocks_load_msb_mask_direct()                                    \
2003   vec_8x16u msb_mask;                                                          \
2004   dup_8x16b(msb_mask, psx_gpu->mask_msb);                                      \
2005
2006 #define shade_blocks_store_indirect(_draw_mask, _pixels)                       \
2007   block->draw_mask = _draw_mask;                                               \
2008   block->pixels = _pixels                                                      \
2009
2010 #define shade_blocks_store_direct(_draw_mask, _pixels)                         \
2011 {                                                                              \
2012   vec_8x16u fb_pixels;                                                         \
2013   or_8x16b(_pixels, _pixels, msb_mask);                                        \
2014   load_8x16b(fb_pixels, block->fb_ptr);                                        \
2015   bif_8x16b(fb_pixels, _pixels, _draw_mask);                                   \
2016   store_8x16b(fb_pixels, block->fb_ptr);                                       \
2017 }                                                                              \
2018
2019
2020 #define shade_blocks_textured_modulated_shaded_primitive_load()                \
2021
2022 #define shade_blocks_textured_modulated_unshaded_primitive_load()              \
2023 {                                                                              \
2024   u32 color = psx_gpu->triangle_color;                                         \
2025   dup_8x8b(colors_r, color);                                                   \
2026   dup_8x8b(colors_g, color >> 8);                                              \
2027   dup_8x8b(colors_b, color >> 16);                                             \
2028   if(psx_gpu->triangle_color == 0x808080)                                      \
2029     false_modulated_triangles++;                                               \
2030 }                                                                              \
2031
2032 #define shade_blocks_textured_modulated_shaded_block_load()                    \
2033   colors_r = block->r;                                                         \
2034   colors_g = block->g;                                                         \
2035   colors_b = block->b                                                          \
2036
2037 #define shade_blocks_textured_modulated_unshaded_block_load()                  \
2038
2039 #define shade_blocks_textured_modulate_dithered(component)                     \
2040   pixels_##component = block->dither_offsets;                                  \
2041   mla_long_8x8b(pixels_##component, texels_##component, colors_##component)    \
2042
2043 #define shade_blocks_textured_modulate_undithered(component)                   \
2044   mul_long_8x8b(pixels_##component, texels_##component, colors_##component)    \
2045
2046 #define shade_blocks_textured_modulated_builder(shading, dithering, target)    \
2047 void shade_blocks_##shading##_textured_modulated_##dithering##_##target(       \
2048  psx_gpu_struct *psx_gpu)                                                      \
2049 {                                                                              \
2050   block_struct *block = psx_gpu->blocks;                                       \
2051   u32 num_blocks = psx_gpu->num_blocks;                                        \
2052   vec_8x16u texels;                                                            \
2053                                                                                \
2054   vec_8x8u texels_r;                                                           \
2055   vec_8x8u texels_g;                                                           \
2056   vec_8x8u texels_b;                                                           \
2057                                                                                \
2058   vec_8x8u colors_r;                                                           \
2059   vec_8x8u colors_g;                                                           \
2060   vec_8x8u colors_b;                                                           \
2061                                                                                \
2062   vec_8x8u pixels_r_low;                                                       \
2063   vec_8x8u pixels_g_low;                                                       \
2064   vec_8x8u pixels_b_low;                                                       \
2065   vec_8x16u pixels;                                                            \
2066                                                                                \
2067   vec_8x16u pixels_r;                                                          \
2068   vec_8x16u pixels_g;                                                          \
2069   vec_8x16u pixels_b;                                                          \
2070                                                                                \
2071   vec_8x16u draw_mask;                                                         \
2072   vec_8x16u zero_mask;                                                         \
2073                                                                                \
2074   vec_8x8u d64_0x07;                                                           \
2075   vec_8x8u d64_0x1F;                                                           \
2076   vec_8x8u d64_1;                                                              \
2077   vec_8x8u d64_4;                                                              \
2078   vec_8x8u d64_128;                                                            \
2079                                                                                \
2080   vec_8x16u d128_0x8000;                                                       \
2081                                                                                \
2082   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
2083   u32 draw_mask_bits;                                                          \
2084   shade_blocks_load_msb_mask_##target();                                       \
2085                                                                                \
2086   dup_8x8b(d64_0x07, 0x07);                                                    \
2087   dup_8x8b(d64_0x1F, 0x1F);                                                    \
2088   dup_8x8b(d64_1, 1);                                                          \
2089   dup_8x8b(d64_4, 4);                                                          \
2090   dup_8x8b(d64_128, 128);                                                      \
2091                                                                                \
2092   dup_8x16b(d128_0x8000, 0x8000);                                              \
2093                                                                                \
2094   shade_blocks_textured_modulated_##shading##_primitive_load();                \
2095                                                                                \
2096   while(num_blocks)                                                            \
2097   {                                                                            \
2098     draw_mask_bits = block->draw_mask_bits;                                    \
2099     dup_8x16b(draw_mask, draw_mask_bits);                                      \
2100     tst_8x16b(draw_mask, draw_mask, test_mask);                                \
2101                                                                                \
2102     shade_blocks_textured_modulated_##shading##_block_load();                  \
2103                                                                                \
2104     texels = block->texels;                                                    \
2105                                                                                \
2106     mov_narrow_8x16b(texels_r, texels);                                        \
2107     shr_narrow_8x16b(texels_g, texels, 5);                                     \
2108     shr_narrow_8x16b(texels_b, texels, 7);                                     \
2109                                                                                \
2110     and_8x8b(texels_r, texels_r, d64_0x1F);                                    \
2111     and_8x8b(texels_g, texels_g, d64_0x1F);                                    \
2112     shr_8x8b(texels_b, texels_b, 3);                                           \
2113                                                                                \
2114     shade_blocks_textured_modulate_##dithering(r);                             \
2115     shade_blocks_textured_modulate_##dithering(g);                             \
2116     shade_blocks_textured_modulate_##dithering(b);                             \
2117                                                                                \
2118     cmpeqz_8x16b(zero_mask, texels);                                           \
2119     and_8x16b(pixels, texels, d128_0x8000);                                    \
2120                                                                                \
2121     shrq_narrow_signed_8x16b(pixels_r_low, pixels_r, 4);                       \
2122     shrq_narrow_signed_8x16b(pixels_g_low, pixels_g, 4);                       \
2123     shrq_narrow_signed_8x16b(pixels_b_low, pixels_b, 4);                       \
2124                                                                                \
2125     or_8x16b(zero_mask, draw_mask, zero_mask);                                 \
2126                                                                                \
2127     shr_8x8b(pixels_r_low, pixels_r_low, 3);                                   \
2128     bic_8x8b(pixels_g_low, pixels_g_low, d64_0x07);                            \
2129     bic_8x8b(pixels_b_low, pixels_b_low, d64_0x07);                            \
2130                                                                                \
2131     mla_long_8x8b(pixels, pixels_r_low, d64_1);                                \
2132     mla_long_8x8b(pixels, pixels_g_low, d64_4);                                \
2133     mla_long_8x8b(pixels, pixels_b_low, d64_128);                              \
2134                                                                                \
2135     shade_blocks_store_##target(zero_mask, pixels);                            \
2136                                                                                \
2137     num_blocks--;                                                              \
2138     block++;                                                                   \
2139   }                                                                            \
2140 }                                                                              \
2141
2142 void shade_blocks_shaded_textured_modulated_dithered_direct(psx_gpu_struct
2143  *psx_gpu);
2144 void shade_blocks_shaded_textured_modulated_undithered_direct(psx_gpu_struct
2145  *psx_gpu);
2146 void shade_blocks_unshaded_textured_modulated_dithered_direct(psx_gpu_struct
2147  *psx_gpu);
2148 void shade_blocks_unshaded_textured_modulated_undithered_direct(psx_gpu_struct
2149  *psx_gpu);
2150
2151 void shade_blocks_shaded_textured_modulated_dithered_indirect(psx_gpu_struct
2152  *psx_gpu);
2153 void shade_blocks_shaded_textured_modulated_undithered_indirect(psx_gpu_struct
2154  *psx_gpu);
2155 void shade_blocks_unshaded_textured_modulated_dithered_indirect(psx_gpu_struct
2156  *psx_gpu);
2157 void shade_blocks_unshaded_textured_modulated_undithered_indirect(psx_gpu_struct
2158  *psx_gpu);
2159
2160 #ifndef NEON_BUILD
2161
2162 shade_blocks_textured_modulated_builder(shaded, dithered, direct);
2163 shade_blocks_textured_modulated_builder(shaded, undithered, direct);
2164 shade_blocks_textured_modulated_builder(unshaded, dithered, direct);
2165 shade_blocks_textured_modulated_builder(unshaded, undithered, direct);
2166
2167 shade_blocks_textured_modulated_builder(shaded, dithered, indirect);
2168 shade_blocks_textured_modulated_builder(shaded, undithered, indirect);
2169 shade_blocks_textured_modulated_builder(unshaded, dithered, indirect);
2170 shade_blocks_textured_modulated_builder(unshaded, undithered, indirect);
2171
2172 #endif
2173
2174
2175 #define shade_blocks_textured_unmodulated_builder(target)                      \
2176 void shade_blocks_textured_unmodulated_##target(psx_gpu_struct *psx_gpu)       \
2177 {                                                                              \
2178   block_struct *block = psx_gpu->blocks;                                       \
2179   u32 num_blocks = psx_gpu->num_blocks;                                        \
2180   vec_8x16u draw_mask;                                                         \
2181   vec_8x16u test_mask = psx_gpu->test_mask;                                    \
2182   u32 draw_mask_bits;                                                          \
2183                                                                                \
2184   vec_8x16u pixels;                                                            \
2185   shade_blocks_load_msb_mask_##target();                                       \
2186                                                                                \
2187   while(num_blocks)                                                            \
2188   {                                                                            \
2189     vec_8x16u zero_mask;                                                       \
2190                                                                                \
2191     draw_mask_bits = block->draw_mask_bits;                                    \
2192     dup_8x16b(draw_mask, draw_mask_bits);                                      \
2193     tst_8x16b(draw_mask, draw_mask, test_mask);                                \
2194                                                                                \
2195     pixels = block->texels;                                                    \
2196                                                                                \
2197     cmpeqz_8x16b(zero_mask, pixels);                                           \
2198     or_8x16b(zero_mask, draw_mask, zero_mask);                                 \
2199                                                                                \
2200     shade_blocks_store_##target(zero_mask, pixels);                            \
2201                                                                                \
2202     num_blocks--;                                                              \
2203     block++;                                                                   \
2204   }                                                                            \
2205 }                                                                              \
2206
2207 void shade_blocks_textured_unmodulated_indirect(psx_gpu_struct *psx_gpu);
2208 void shade_blocks_textured_unmodulated_direct(psx_gpu_struct *psx_gpu);
2209
2210 #ifndef NEON_BUILD
2211
2212 shade_blocks_textured_unmodulated_builder(indirect)
2213 shade_blocks_textured_unmodulated_builder(direct)
2214
2215 #endif
2216
2217
2218 void shade_blocks_unshaded_untextured_indirect(psx_gpu_struct *psx_gpu);
2219 void shade_blocks_unshaded_untextured_direct(psx_gpu_struct *psx_gpu);
2220
2221 #ifndef NEON_BUILD
2222
2223 void shade_blocks_unshaded_untextured_indirect(psx_gpu_struct *psx_gpu)
2224 {
2225 }
2226
2227 void shade_blocks_unshaded_untextured_direct(psx_gpu_struct *psx_gpu)
2228 {
2229   block_struct *block = psx_gpu->blocks;
2230   u32 num_blocks = psx_gpu->num_blocks;
2231
2232   vec_8x16u pixels = block->pixels;
2233   shade_blocks_load_msb_mask_direct();
2234
2235   while(num_blocks)
2236   {
2237     shade_blocks_store_direct(block->draw_mask, pixels);
2238
2239     num_blocks--;
2240     block++;
2241   }
2242 }
2243
2244 #endif
2245
2246 void shade_blocks_shaded_untextured(psx_gpu_struct *psx_gpu)
2247 {
2248 }
2249
2250
2251 #define blend_blocks_mask_evaluate_on()                                        \
2252   vec_8x16u mask_pixels;                                                       \
2253   cmpltz_8x16b(mask_pixels, framebuffer_pixels);                               \
2254   or_8x16b(draw_mask, draw_mask, mask_pixels)                                  \
2255
2256 #define blend_blocks_mask_evaluate_off()                                       \
2257
2258 #define blend_blocks_average()                                                 \
2259 {                                                                              \
2260   vec_8x16u pixels_no_msb;                                                     \
2261   vec_8x16u fb_pixels_no_msb;                                                  \
2262                                                                                \
2263   vec_8x16u d128_0x0421;                                                       \
2264   vec_8x16u d128_0x8000;                                                       \
2265                                                                                \
2266   dup_8x16b(d128_0x0421, 0x0421);                                              \
2267   dup_8x16b(d128_0x8000, 0x8000);                                              \
2268                                                                                \
2269   eor_8x16b(blend_pixels, pixels, framebuffer_pixels);                         \
2270   bic_8x16b(pixels_no_msb, pixels, d128_0x8000);                               \
2271   and_8x16b(blend_pixels, blend_pixels, d128_0x0421);                          \
2272   sub_8x16b(blend_pixels, pixels_no_msb, blend_pixels);                        \
2273   bic_8x16b(fb_pixels_no_msb, framebuffer_pixels, d128_0x8000);                \
2274   average_8x16b(blend_pixels, fb_pixels_no_msb, blend_pixels);                 \
2275 }                                                                              \
2276
2277 #define blend_blocks_add()                                                     \
2278 {                                                                              \
2279   vec_8x16u pixels_rb, pixels_g;                                               \
2280   vec_8x16u fb_rb, fb_g;                                                       \
2281                                                                                \
2282   vec_8x16u d128_0x7C1F;                                                       \
2283   vec_8x16u d128_0x03E0;                                                       \
2284                                                                                \
2285   dup_8x16b(d128_0x7C1F, 0x7C1F);                                              \
2286   dup_8x16b(d128_0x03E0, 0x03E0);                                              \
2287                                                                                \
2288   and_8x16b(pixels_rb, pixels, d128_0x7C1F);                                   \
2289   and_8x16b(pixels_g, pixels, d128_0x03E0);                                    \
2290                                                                                \
2291   and_8x16b(fb_rb, framebuffer_pixels, d128_0x7C1F);                           \
2292   and_8x16b(fb_g, framebuffer_pixels, d128_0x03E0);                            \
2293                                                                                \
2294   add_8x16b(fb_rb, fb_rb, pixels_rb);                                          \
2295   add_8x16b(fb_g, fb_g, pixels_g);                                             \
2296                                                                                \
2297   min_16x8b(vector_cast(vec_16x8u, fb_rb), vector_cast(vec_16x8u, fb_rb),      \
2298    vector_cast(vec_16x8u, d128_0x7C1F));                                       \
2299   min_8x16b(fb_g, fb_g, d128_0x03E0);                                          \
2300                                                                                \
2301   or_8x16b(blend_pixels, fb_rb, fb_g);                                         \
2302 }                                                                              \
2303
2304 #define blend_blocks_subtract()                                                \
2305 {                                                                              \
2306   vec_8x16u pixels_rb, pixels_g;                                               \
2307   vec_8x16u fb_rb, fb_g;                                                       \
2308                                                                                \
2309   vec_8x16u d128_0x7C1F;                                                       \
2310   vec_8x16u d128_0x03E0;                                                       \
2311                                                                                \
2312   dup_8x16b(d128_0x7C1F, 0x7C1F);                                              \
2313   dup_8x16b(d128_0x03E0, 0x03E0);                                              \
2314                                                                                \
2315   and_8x16b(pixels_rb, pixels, d128_0x7C1F);                                   \
2316   and_8x16b(pixels_g, pixels, d128_0x03E0);                                    \
2317                                                                                \
2318   and_8x16b(fb_rb, framebuffer_pixels, d128_0x7C1F);                           \
2319   and_8x16b(fb_g, framebuffer_pixels, d128_0x03E0);                            \
2320                                                                                \
2321   subs_16x8b(vector_cast(vec_16x8u, fb_rb),                                    \
2322    vector_cast(vec_16x8u, fb_rb), vector_cast(vec_16x8u, pixels_rb));          \
2323   subs_8x16b(fb_g, fb_g, pixels_g);                                            \
2324                                                                                \
2325   or_8x16b(blend_pixels, fb_rb, fb_g);                                         \
2326 }                                                                              \
2327
2328 #define blend_blocks_add_fourth()                                              \
2329 {                                                                              \
2330   vec_8x16u pixels_rb, pixels_g;                                               \
2331   vec_8x16u pixels_fourth;                                                     \
2332   vec_8x16u fb_rb, fb_g;                                                       \
2333                                                                                \
2334   vec_8x16u d128_0x7C1F;                                                       \
2335   vec_8x16u d128_0x1C07;                                                       \
2336   vec_8x16u d128_0x03E0;                                                       \
2337   vec_8x16u d128_0x00E0;                                                       \
2338                                                                                \
2339   dup_8x16b(d128_0x7C1F, 0x7C1F);                                              \
2340   dup_8x16b(d128_0x1C07, 0x1C07);                                              \
2341   dup_8x16b(d128_0x03E0, 0x03E0);                                              \
2342   dup_8x16b(d128_0x00E0, 0x00E0);                                              \
2343                                                                                \
2344   shr_8x16b(pixels_fourth, vector_cast(vec_8x16s, pixels), 2);                 \
2345                                                                                \
2346   and_8x16b(fb_rb, framebuffer_pixels, d128_0x7C1F);                           \
2347   and_8x16b(fb_g, framebuffer_pixels, d128_0x03E0);                            \
2348                                                                                \
2349   and_8x16b(pixels_rb, pixels_fourth, d128_0x1C07);                            \
2350   and_8x16b(pixels_g, pixels_fourth, d128_0x00E0);                             \
2351                                                                                \
2352   add_8x16b(fb_rb, fb_rb, pixels_rb);                                          \
2353   add_8x16b(fb_g, fb_g, pixels_g);                                             \
2354                                                                                \
2355   min_16x8b(vector_cast(vec_16x8u, fb_rb), vector_cast(vec_16x8u, fb_rb),      \
2356    vector_cast(vec_16x8u, d128_0x7C1F));                                       \
2357   min_8x16b(fb_g, fb_g, d128_0x03E0);                                          \
2358                                                                                \
2359   or_8x16b(blend_pixels, fb_rb, fb_g);                                         \
2360 }                                                                              \
2361
2362 #define blend_blocks_blended_combine_textured()                                \
2363 {                                                                              \
2364   vec_8x16u blend_mask;                                                        \
2365   cmpltz_8x16b(blend_mask, pixels);                                            \
2366                                                                                \
2367   or_immediate_8x16b(blend_pixels, blend_pixels, 0x8000);                      \
2368   bif_8x16b(blend_pixels, pixels, blend_mask);                                 \
2369 }                                                                              \
2370
2371 #define blend_blocks_blended_combine_untextured()                              \
2372
2373
2374 #define blend_blocks_body_blend(blend_mode, texturing)                         \
2375 {                                                                              \
2376   blend_blocks_##blend_mode();                                                 \
2377   blend_blocks_blended_combine_##texturing();                                  \
2378 }                                                                              \
2379
2380 #define blend_blocks_body_average(texturing)                                   \
2381   blend_blocks_body_blend(average, texturing)                                  \
2382
2383 #define blend_blocks_body_add(texturing)                                       \
2384   blend_blocks_body_blend(add, texturing)                                      \
2385
2386 #define blend_blocks_body_subtract(texturing)                                  \
2387   blend_blocks_body_blend(subtract, texturing)                                 \
2388
2389 #define blend_blocks_body_add_fourth(texturing)                                \
2390   blend_blocks_body_blend(add_fourth, texturing)                               \
2391
2392 #define blend_blocks_body_unblended(texturing)                                 \
2393   blend_pixels = pixels                                                        \
2394
2395
2396 #define blend_blocks_builder(texturing, blend_mode, mask_evaluate)             \
2397 void                                                                           \
2398  blend_blocks_##texturing##_##blend_mode##_##mask_evaluate(psx_gpu_struct      \
2399  *psx_gpu)                                                                     \
2400 {                                                                              \
2401   block_struct *block = psx_gpu->blocks;                                       \
2402   u32 num_blocks = psx_gpu->num_blocks;                                        \
2403   vec_8x16u draw_mask;                                                         \
2404   vec_8x16u pixels;                                                            \
2405   vec_8x16u blend_pixels;                                                      \
2406   vec_8x16u framebuffer_pixels;                                                \
2407   vec_8x16u msb_mask;                                                          \
2408                                                                                \
2409   u16 *fb_ptr;                                                                 \
2410                                                                                \
2411   dup_8x16b(msb_mask, psx_gpu->mask_msb);                                      \
2412                                                                                \
2413   while(num_blocks)                                                            \
2414   {                                                                            \
2415     pixels = block->pixels;                                                    \
2416     draw_mask = block->draw_mask;                                              \
2417     fb_ptr = block->fb_ptr;                                                    \
2418                                                                                \
2419     load_8x16b(framebuffer_pixels, fb_ptr);                                    \
2420                                                                                \
2421     blend_blocks_mask_evaluate_##mask_evaluate();                              \
2422     blend_blocks_body_##blend_mode(texturing);                                 \
2423                                                                                \
2424     or_8x16b(blend_pixels, blend_pixels, msb_mask);                            \
2425     bif_8x16b(framebuffer_pixels, blend_pixels, draw_mask);                    \
2426     store_8x16b(framebuffer_pixels, fb_ptr);                                   \
2427                                                                                \
2428     blend_blocks++;                                                            \
2429     num_blocks--;                                                              \
2430     block++;                                                                   \
2431   }                                                                            \
2432 }                                                                              \
2433
2434 void blend_blocks_textured_average_off(psx_gpu_struct *psx_gpu);
2435 void blend_blocks_textured_average_on(psx_gpu_struct *psx_gpu);
2436 void blend_blocks_textured_add_off(psx_gpu_struct *psx_gpu);
2437 void blend_blocks_textured_add_on(psx_gpu_struct *psx_gpu);
2438 void blend_blocks_textured_subtract_off(psx_gpu_struct *psx_gpu);
2439 void blend_blocks_textured_subtract_on(psx_gpu_struct *psx_gpu);
2440 void blend_blocks_textured_add_fourth_off(psx_gpu_struct *psx_gpu);
2441 void blend_blocks_textured_add_fourth_on(psx_gpu_struct *psx_gpu);
2442
2443 void blend_blocks_untextured_average_off(psx_gpu_struct *psx_gpu);
2444 void blend_blocks_untextured_average_on(psx_gpu_struct *psx_gpu);
2445 void blend_blocks_untextured_add_off(psx_gpu_struct *psx_gpu);
2446 void blend_blocks_untextured_add_on(psx_gpu_struct *psx_gpu);
2447 void blend_blocks_untextured_subtract_off(psx_gpu_struct *psx_gpu);
2448 void blend_blocks_untextured_subtract_on(psx_gpu_struct *psx_gpu);
2449 void blend_blocks_untextured_add_fourth_off(psx_gpu_struct *psx_gpu);
2450 void blend_blocks_untextured_add_fourth_on(psx_gpu_struct *psx_gpu);
2451
2452 void blend_blocks_textured_unblended_off(psx_gpu_struct *psx_gpu);
2453 void blend_blocks_textured_unblended_on(psx_gpu_struct *psx_gpu);
2454
2455 #ifndef NEON_BUILD
2456
2457 void blend_blocks_textured_unblended_off(psx_gpu_struct *psx_gpu)
2458 {
2459 }
2460
2461 blend_blocks_builder(textured, average, off);
2462 blend_blocks_builder(textured, average, on);
2463 blend_blocks_builder(textured, add, off);
2464 blend_blocks_builder(textured, add, on);
2465 blend_blocks_builder(textured, subtract, off);
2466 blend_blocks_builder(textured, subtract, on);
2467 blend_blocks_builder(textured, add_fourth, off);
2468 blend_blocks_builder(textured, add_fourth, on);
2469
2470 blend_blocks_builder(untextured, average, off);
2471 blend_blocks_builder(untextured, average, on);
2472 blend_blocks_builder(untextured, add, off);
2473 blend_blocks_builder(untextured, add, on);
2474 blend_blocks_builder(untextured, subtract, off);
2475 blend_blocks_builder(untextured, subtract, on);
2476 blend_blocks_builder(untextured, add_fourth, off);
2477 blend_blocks_builder(untextured, add_fourth, on);
2478
2479 blend_blocks_builder(textured, unblended, on);
2480
2481 #endif
2482
2483
2484 #define vertex_swap(_a, _b)                                                    \
2485 {                                                                              \
2486   vertex_struct *temp_vertex = _a;                                             \
2487   _a = _b;                                                                     \
2488   _b = temp_vertex;                                                            \
2489   triangle_winding ^= 1;                                                       \
2490 }                                                                              \
2491
2492
2493 // Setup blocks parametric-variables:
2494 // SHADE  TEXTURE_MAP SWIZZLING
2495 // 0      0           x
2496 // 0      1           0
2497 // 0      1           1
2498 // 1      0           x
2499 // 1      1           0
2500 // 1      1           1
2501 // 8 inputs, 6 combinations
2502
2503 #define setup_blocks_switch_untextured_unshaded(dithering, target)             \
2504   setup_blocks_unshaded_untextured_undithered_unswizzled_##target              \
2505
2506 #define setup_blocks_switch_untextured_shaded(dithering, target)               \
2507   setup_blocks_shaded_untextured_##dithering##_unswizzled_##target             \
2508
2509 #define setup_blocks_switch_untextured(shading, texture_mode, dithering,       \
2510  target)                                                                       \
2511   setup_blocks_switch_untextured_##shading(dithering, target)                  \
2512
2513 #define setup_blocks_switch_texture_mode_4bpp(shading)                         \
2514   setup_blocks_##shading##_textured_dithered_swizzled_indirect                 \
2515
2516 #define setup_blocks_switch_texture_mode_8bpp(shading)                         \
2517   setup_blocks_##shading##_textured_dithered_swizzled_indirect                 \
2518
2519 #define setup_blocks_switch_texture_mode_16bpp(shading)                        \
2520   setup_blocks_##shading##_textured_dithered_unswizzled_indirect               \
2521
2522 #define setup_blocks_switch_textured(shading, texture_mode, dithering, target) \
2523   setup_blocks_switch_texture_mode_##texture_mode(shading)                     \
2524
2525 #define setup_blocks_switch_blended(shading, texturing, texture_mode,          \
2526  dithering, mask_evaluate)                                                     \
2527   setup_blocks_switch_##texturing(shading, texture_mode, dithering, indirect)  \
2528
2529 #define setup_blocks_switch_unblended_on(shading, texturing, texture_mode,     \
2530  dithering)                                                                    \
2531   setup_blocks_switch_##texturing(shading, texture_mode, dithering, indirect)  \
2532
2533 #define setup_blocks_switch_unblended_off(shading, texturing, texture_mode,    \
2534  dithering)                                                                    \
2535   setup_blocks_switch_##texturing(shading, texture_mode, dithering, direct)    \
2536
2537 #define setup_blocks_switch_unblended(shading, texturing, texture_mode,        \
2538  dithering, mask_evaluate)                                                     \
2539   setup_blocks_switch_unblended_##mask_evaluate(shading, texturing,            \
2540    texture_mode, dithering)                                                    \
2541
2542 #define setup_blocks_switch(shading, texturing, texture_mode, dithering,       \
2543  blending, mask_evaluate)                                                      \
2544   setup_blocks_switch_##blending(shading, texturing, texture_mode,             \
2545    dithering, mask_evaluate)                                                   \
2546
2547
2548 // Texture blocks:
2549
2550 #define texture_blocks_switch_untextured(texture_mode)                         \
2551   texture_blocks_untextured                                                    \
2552
2553 #define texture_blocks_switch_textured(texture_mode)                           \
2554   texture_blocks_##texture_mode                                                \
2555
2556 #define texture_blocks_switch(texturing, texture_mode)                         \
2557   texture_blocks_switch_##texturing(texture_mode)                              \
2558
2559
2560 // Shade blocks parametric-variables:
2561 // SHADE  TEXTURE_MAP  MODULATE_TEXELS  dither_mode
2562 // 0      0            x                x
2563 // 0      1            0                0
2564 // 0      1            0                1
2565 // x      1            1                x
2566 // 1      0            x                0
2567 // 1      0            x                1
2568 // 1      1            0                0
2569 // 1      1            0                1
2570 // 16 inputs, 8 combinations
2571
2572 #define shade_blocks_switch_unshaded_untextured(modulation, dithering, target) \
2573   shade_blocks_unshaded_untextured_##target                                    \
2574
2575 #define shade_blocks_switch_unshaded_textured_unmodulated(dithering, target)   \
2576   shade_blocks_textured_unmodulated_##target                                   \
2577
2578 #define shade_blocks_switch_unshaded_textured_modulated(dithering, target)     \
2579   shade_blocks_unshaded_textured_modulated_##dithering##_##target              \
2580
2581 #define shade_blocks_switch_unshaded_textured(modulation, dithering, target)   \
2582   shade_blocks_switch_unshaded_textured_##modulation(dithering, target)        \
2583
2584 #define shade_blocks_switch_unshaded(texturing, modulation, dithering, target) \
2585   shade_blocks_switch_unshaded_##texturing(modulation, dithering, target)      \
2586
2587 #define shade_blocks_switch_shaded_untextured(modulation, dithering, target)   \
2588   shade_blocks_shaded_untextured                                               \
2589
2590 #define shade_blocks_switch_shaded_textured_unmodulated(dithering, target)     \
2591   shade_blocks_textured_unmodulated_##target                                   \
2592
2593 #define shade_blocks_switch_shaded_textured_modulated(dithering, target)       \
2594   shade_blocks_shaded_textured_modulated_##dithering##_##target                \
2595
2596 #define shade_blocks_switch_shaded_textured(modulation, dithering, target)     \
2597   shade_blocks_switch_shaded_textured_##modulation(dithering, target)          \
2598
2599 #define shade_blocks_switch_shaded(texturing, modulation, dithering, target)   \
2600   shade_blocks_switch_shaded_##texturing(modulation, dithering, target)        \
2601
2602 #define shade_blocks_switch_mask_off(shading, texturing, modulation,           \
2603  dithering)                                                                    \
2604   shade_blocks_switch_##shading(texturing, modulation, dithering, direct)      \
2605
2606 #define shade_blocks_switch_mask_on(shading, texturing, modulation,            \
2607  dithering)                                                                    \
2608   shade_blocks_switch_##shading(texturing, modulation, dithering, indirect)    \
2609
2610 #define shade_blocks_switch_blended(shading, texturing, modulation, dithering, \
2611  mask_evaluate)                                                                \
2612   shade_blocks_switch_##shading(texturing, modulation, dithering, indirect)    \
2613
2614 #define shade_blocks_switch_unblended(shading, texturing, modulation,          \
2615  dithering, mask_evaluate)                                                     \
2616   shade_blocks_switch_mask_##mask_evaluate(shading, texturing, modulation,     \
2617    dithering)                                                                  \
2618
2619 #define shade_blocks_switch(shading, texturing, modulation, dithering,         \
2620  blending, mask_evaluate)                                                      \
2621   shade_blocks_switch_##blending(shading, texturing, modulation, dithering,    \
2622    mask_evaluate)                                                              \
2623
2624
2625 // Blend blocks parametric-variables:
2626 // TEXTURE_MAP BLEND  BM_A BM_B mask_evaluate
2627 // x           0      x    x    0
2628 // x           0      x    x    1
2629 // 0           1      0    0    0
2630 // 0           1      0    0    1
2631 // 0           1      0    1    0
2632 // 0           1      0    1    1
2633 // 0           1      1    0    0
2634 // 0           1      1    0    1
2635 // 0           1      1    1    0
2636 // 0           1      1    1    1
2637 // 1           1      0    0    0
2638 // 1           1      0    0    1
2639 // 1           1      0    1    0
2640 // 1           1      0    1    1
2641 // 1           1      1    0    0
2642 // 1           1      1    0    1
2643 // 1           1      1    1    0
2644 // 1           1      1    1    1
2645 // 32 inputs, 18 combinations
2646
2647 #define blend_blocks_switch_unblended(texturing, blend_mode, mask_evaluate)    \
2648   blend_blocks_textured_unblended_##mask_evaluate                              \
2649
2650 #define blend_blocks_switch_blended(texturing, blend_mode, mask_evaluate)      \
2651   blend_blocks_##texturing##_##blend_mode##_##mask_evaluate                    \
2652
2653 #define blend_blocks_switch(texturing, blending, blend_mode, mask_evaluate)    \
2654   blend_blocks_switch_##blending(texturing, blend_mode, mask_evaluate)         \
2655
2656
2657 #define render_blocks_switch_block_modulation(texture_mode, blend_mode,        \
2658  mask_evaluate, shading, dithering, texturing, blending, modulation)           \
2659 {                                                                              \
2660   setup_blocks_switch(shading, texturing, texture_mode, dithering, blending,   \
2661    mask_evaluate),                                                             \
2662   texture_blocks_switch(texturing, texture_mode),                              \
2663   shade_blocks_switch(shading, texturing, modulation, dithering, blending,     \
2664    mask_evaluate),                                                             \
2665   blend_blocks_switch(texturing, blending, blend_mode, mask_evaluate)          \
2666 }                                                                              \
2667
2668 #define render_blocks_switch_block_blending(texture_mode, blend_mode,          \
2669  mask_evaluate, shading, dithering, texturing, blending)                       \
2670   render_blocks_switch_block_modulation(texture_mode, blend_mode,              \
2671    mask_evaluate, shading, dithering, texturing, blending, modulated),         \
2672   render_blocks_switch_block_modulation(texture_mode, blend_mode,              \
2673    mask_evaluate, shading, dithering, texturing, blending, unmodulated)        \
2674
2675 #define render_blocks_switch_block_texturing(texture_mode, blend_mode,         \
2676  mask_evaluate, shading, dithering, texturing)                                 \
2677   render_blocks_switch_block_blending(texture_mode, blend_mode,                \
2678    mask_evaluate, shading, dithering, texturing, unblended),                   \
2679   render_blocks_switch_block_blending(texture_mode, blend_mode,                \
2680    mask_evaluate, shading, dithering, texturing, blended)                      \
2681
2682 #define render_blocks_switch_block_dithering(texture_mode, blend_mode,         \
2683  mask_evaluate, shading, dithering)                                            \
2684   render_blocks_switch_block_texturing(texture_mode, blend_mode,               \
2685    mask_evaluate, shading, dithering, untextured),                             \
2686   render_blocks_switch_block_texturing(texture_mode, blend_mode,               \
2687    mask_evaluate, shading, dithering, textured)                                \
2688
2689 #define render_blocks_switch_block_shading(texture_mode, blend_mode,           \
2690  mask_evaluate, shading)                                                       \
2691   render_blocks_switch_block_dithering(texture_mode, blend_mode,               \
2692    mask_evaluate, shading, undithered),                                        \
2693   render_blocks_switch_block_dithering(texture_mode, blend_mode,               \
2694    mask_evaluate, shading, dithered)                                           \
2695
2696 #define render_blocks_switch_block_mask_evaluate(texture_mode, blend_mode,     \
2697  mask_evaluate)                                                                \
2698   render_blocks_switch_block_shading(texture_mode, blend_mode, mask_evaluate,  \
2699    unshaded),                                                                  \
2700   render_blocks_switch_block_shading(texture_mode, blend_mode, mask_evaluate,  \
2701    shaded)                                                                     \
2702
2703 #define render_blocks_switch_block_blend_mode(texture_mode, blend_mode)        \
2704   render_blocks_switch_block_mask_evaluate(texture_mode, blend_mode, off),     \
2705   render_blocks_switch_block_mask_evaluate(texture_mode, blend_mode, on)       \
2706
2707 #define render_blocks_switch_block_texture_mode(texture_mode)                  \
2708   render_blocks_switch_block_blend_mode(texture_mode, average),                \
2709   render_blocks_switch_block_blend_mode(texture_mode, add),                    \
2710   render_blocks_switch_block_blend_mode(texture_mode, subtract),               \
2711   render_blocks_switch_block_blend_mode(texture_mode, add_fourth)              \
2712
2713 #define render_blocks_switch_block()                                           \
2714   render_blocks_switch_block_texture_mode(4bpp),                               \
2715   render_blocks_switch_block_texture_mode(8bpp),                               \
2716   render_blocks_switch_block_texture_mode(16bpp),                              \
2717   render_blocks_switch_block_texture_mode(4bpp)                                \
2718
2719
2720 render_block_handler_struct render_triangle_block_handlers[] =
2721 {
2722   render_blocks_switch_block()
2723 };
2724
2725 #undef render_blocks_switch_block_modulation
2726
2727 #define render_blocks_switch_block_modulation(texture_mode, blend_mode,        \
2728  mask_evaluate, shading, dithering, texturing, blending, modulation)           \
2729   "render flags:\n"                                                            \
2730   "texture mode:     " #texture_mode "\n"                                      \
2731   "blend mode:       " #blend_mode "\n"                                        \
2732   "mask evaluation:  " #mask_evaluate "\n"                                     \
2733   #shading "\n"                                                                \
2734   #dithering "\n"                                                              \
2735   #texturing "\n"                                                              \
2736   #blending "\n"                                                               \
2737   #modulation "\n"                                                             \
2738
2739 char *render_block_flag_strings[] =
2740 {
2741   render_blocks_switch_block()
2742 };
2743
2744
2745 #define triangle_y_direction_up   1
2746 #define triangle_y_direction_flat 2
2747 #define triangle_y_direction_down 0
2748
2749 #define triangle_winding_positive 0
2750 #define triangle_winding_negative 1
2751
2752 #define triangle_set_direction(direction_variable, value)                      \
2753   u32 direction_variable = (u32)(value) >> 31;                                 \
2754   if(value == 0)                                                               \
2755     direction_variable = 2                                                     \
2756
2757 #define triangle_case(direction_a, direction_b, direction_c, winding)          \
2758   case (triangle_y_direction_##direction_a |                                   \
2759    (triangle_y_direction_##direction_b << 2) |                                 \
2760    (triangle_y_direction_##direction_c << 4) |                                 \
2761    (triangle_winding_##winding << 6))                                          \
2762
2763 void render_triangle(psx_gpu_struct *psx_gpu, vertex_struct *vertexes,
2764  u32 flags)
2765 {
2766   s32 y_top, y_bottom;
2767   s32 triangle_area;
2768   u32 triangle_winding = 0;
2769
2770   vertex_struct *a = &(vertexes[0]);
2771   vertex_struct *b = &(vertexes[1]);
2772   vertex_struct *c = &(vertexes[2]);
2773
2774   triangle_area = triangle_signed_area_x2(a->x, a->y, b->x, b->y, c->x, c->y);
2775
2776   triangles++;
2777
2778   if(triangle_area == 0)
2779   {
2780     trivial_rejects++;
2781     return;
2782   }
2783
2784   if(b->y < a->y)
2785     vertex_swap(a, b);
2786
2787   if(c->y < b->y)
2788   {
2789     vertex_swap(b, c);
2790
2791     if(b->y < a->y)
2792       vertex_swap(a, b);
2793   }
2794
2795   y_bottom = c->y;
2796   y_top = a->y;
2797
2798   if((y_bottom - y_top) >= 512)
2799   {
2800     trivial_rejects++;
2801     return;
2802   }
2803
2804   if(triangle_area < 0)
2805   {
2806     triangle_area = -triangle_area;
2807     triangle_winding ^= 1;
2808     vertex_swap(a, c);
2809   }
2810
2811   if(b->x < a->x)
2812     vertex_swap(a, b);
2813
2814   if(c->x < b->x)
2815   {
2816     vertex_swap(b, c);
2817
2818     if(b->x < a->x)
2819       vertex_swap(a, b);
2820   }
2821
2822   if((c->x - a->x) >= 1024)
2823   {
2824     trivial_rejects++;
2825     return;
2826   }
2827
2828   if(invalidate_texture_cache_region_viewport(psx_gpu, a->x, y_top, c->x,
2829    y_bottom) == 0)
2830   {
2831     trivial_rejects++;
2832     return;
2833   }
2834
2835   psx_gpu->num_spans = 0;
2836   psx_gpu->triangle_area = triangle_area;
2837   psx_gpu->triangle_winding = triangle_winding;
2838
2839   s32 y_delta_a = b->y - a->y;
2840   s32 y_delta_b = c->y - b->y;
2841   s32 y_delta_c = c->y - a->y;
2842
2843   triangle_set_direction(y_direction_a, y_delta_a);
2844   triangle_set_direction(y_direction_b, y_delta_b);
2845   triangle_set_direction(y_direction_c, y_delta_c);
2846
2847   compute_all_gradients(psx_gpu, a, b, c);
2848
2849   switch(y_direction_a | (y_direction_b << 2) | (y_direction_c << 4) |
2850    (triangle_winding << 6))
2851   {
2852     triangle_case(up, up, up, negative):
2853     triangle_case(up, up, flat, negative):
2854     triangle_case(up, up, down, negative):
2855       setup_spans_up_right(psx_gpu, a, b, c);
2856       break;
2857
2858     triangle_case(flat, up, up, negative):
2859     triangle_case(flat, up, flat, negative):
2860     triangle_case(flat, up, down, negative):
2861       setup_spans_up_a(psx_gpu, a, b, c);
2862       break;
2863
2864     triangle_case(down, up, up, negative):
2865       setup_spans_up_down(psx_gpu, a, c, b);
2866       break;
2867
2868     triangle_case(down, up, flat, negative):
2869       setup_spans_down_a(psx_gpu, a, c, b);
2870       break;
2871
2872     triangle_case(down, up, down, negative):
2873       setup_spans_down_right(psx_gpu, a, c, b);
2874       break;
2875
2876     triangle_case(down, flat, up, negative):
2877     triangle_case(down, flat, flat, negative):
2878     triangle_case(down, flat, down, negative):
2879       setup_spans_down_b(psx_gpu, a, b, c);
2880       break;
2881
2882     triangle_case(down, down, up, negative):
2883     triangle_case(down, down, flat, negative):
2884     triangle_case(down, down, down, negative):
2885       setup_spans_down_left(psx_gpu, a, b, c);
2886       break;
2887
2888     triangle_case(up, up, up, positive):
2889     triangle_case(up, up, flat, positive):
2890     triangle_case(up, up, down, positive):
2891       setup_spans_up_left(psx_gpu, a, b, c);
2892       break;
2893
2894     triangle_case(up, flat, up, positive):
2895     triangle_case(up, flat, flat, positive):
2896     triangle_case(up, flat, down, positive):
2897       setup_spans_up_b(psx_gpu, a, b, c);
2898       break;
2899
2900     triangle_case(up, down, up, positive):
2901       setup_spans_up_right(psx_gpu, a, c, b);
2902       break;
2903
2904     triangle_case(up, down, flat, positive):
2905       setup_spans_up_a(psx_gpu, a, c, b);
2906       break;
2907
2908     triangle_case(up, down, down, positive):
2909       setup_spans_up_down(psx_gpu, a, b, c);
2910       break;
2911
2912     triangle_case(flat, down, up, positive):
2913     triangle_case(flat, down, flat, positive):
2914     triangle_case(flat, down, down, positive):
2915       setup_spans_down_a(psx_gpu, a, b, c);
2916       break;
2917
2918     triangle_case(down, down, up, positive):
2919     triangle_case(down, down, flat, positive):
2920     triangle_case(down, down, down, positive):
2921       setup_spans_down_right(psx_gpu, a, b, c);
2922       break;
2923   }
2924
2925   spans += psx_gpu->num_spans;
2926
2927   u32 render_state = flags &
2928    (RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND |
2929    RENDER_FLAGS_TEXTURE_MAP | RENDER_FLAGS_SHADE);
2930   render_state |= psx_gpu->render_state_base;
2931
2932   if((psx_gpu->render_state != render_state) ||
2933    (psx_gpu->primitive_type != PRIMITIVE_TYPE_TRIANGLE))
2934   {
2935     psx_gpu->render_state = render_state;
2936     flush_render_block_buffer(psx_gpu);
2937     state_changes++;
2938   }
2939
2940   psx_gpu->primitive_type = PRIMITIVE_TYPE_TRIANGLE;
2941
2942   psx_gpu->render_block_handler =
2943    &(render_triangle_block_handlers[render_state]);
2944   ((setup_blocks_function_type *)psx_gpu->render_block_handler->setup_blocks)
2945    (psx_gpu);
2946 }
2947
2948
2949 void texture_sprite_blocks_8bpp(psx_gpu_struct *psx_gpu);
2950
2951 #ifndef NEON_BUILD
2952
2953 void texture_sprite_blocks_8bpp(psx_gpu_struct *psx_gpu)
2954 {
2955   block_struct *block = psx_gpu->blocks;
2956   u32 num_blocks = psx_gpu->num_blocks;
2957
2958   vec_8x16u texels;
2959   vec_8x8u texel_indexes;
2960
2961   u16 *clut_ptr = psx_gpu->clut_ptr;
2962   u32 i;
2963
2964   while(num_blocks)
2965   {
2966     texel_indexes = block->r;
2967
2968     for(i = 0; i < 8; i++)
2969     {
2970       texels.e[i] = clut_ptr[texel_indexes.e[i]];
2971     }
2972
2973     block->texels = texels;
2974
2975     num_blocks--;
2976     block++;
2977   }
2978 }
2979
2980 #endif
2981
2982
2983 #define setup_sprite_tiled_initialize_4bpp()                                   \
2984   u16 *clut_ptr = psx_gpu->clut_ptr;                                           \
2985   vec_8x16u clut_a, clut_b;                                                    \
2986   vec_16x8u clut_low, clut_high;                                               \
2987                                                                                \
2988   load_8x16b(clut_a, clut_ptr);                                                \
2989   load_8x16b(clut_b, clut_ptr + 8);                                            \
2990   unzip_16x8b(clut_low, clut_high, clut_a, clut_b);                            \
2991                                                                                \
2992   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_4bpp_mask)        \
2993     update_texture_4bpp_cache(psx_gpu)                                         \
2994
2995 #define setup_sprite_tiled_initialize_8bpp()                                   \
2996   if(psx_gpu->current_texture_mask & psx_gpu->dirty_textures_8bpp_mask)        \
2997     update_texture_8bpp_cache(psx_gpu)                                         \
2998
2999
3000 #define setup_sprite_tile_fetch_texel_block_8bpp(offset)                       \
3001   texture_block_ptr = psx_gpu->texture_page_ptr +                              \
3002    ((texture_offset + offset) & texture_mask);                                 \
3003                                                                                \
3004   load_64b(texels, texture_block_ptr)                                          \
3005
3006
3007 #define setup_sprite_tile_setup_block_yes(side, offset, texture_mode)          \
3008
3009 #define setup_sprite_tile_setup_block_no(side, offset, texture_mode)           \
3010
3011 #define setup_sprite_tile_add_blocks(tile_num_blocks)                          \
3012   num_blocks += tile_num_blocks;                                               \
3013   sprite_blocks += tile_num_blocks;                                            \
3014                                                                                \
3015   if(num_blocks > MAX_BLOCKS)                                                  \
3016   {                                                                            \
3017     flush_render_block_buffer(psx_gpu);                                        \
3018     num_blocks = tile_num_blocks;                                              \
3019     block = psx_gpu->blocks;                                                   \
3020   }                                                                            \
3021
3022 #define setup_sprite_tile_full_4bpp(edge)                                      \
3023 {                                                                              \
3024   vec_8x8u texels_low, texels_high;                                            \
3025   vec_8x16u pixels;                                                            \
3026   setup_sprite_tile_add_blocks(sub_tile_height * 2);                           \
3027                                                                                \
3028   while(sub_tile_height)                                                       \
3029   {                                                                            \
3030     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3031     tbl_16(texels_low, texels, clut_low);                                      \
3032     tbl_16(texels_high, texels, clut_high);                                    \
3033     zip_8x16b(pixels, texels_low, texels_high);                                \
3034                                                                                \
3035     block->texels = pixels;                                                    \
3036     block->draw_mask_bits = left_mask_bits;                                    \
3037     block->fb_ptr = fb_ptr;                                                    \
3038     block++;                                                                   \
3039                                                                                \
3040     setup_sprite_tile_fetch_texel_block_8bpp(8);                               \
3041     tbl_16(texels_low, texels, clut_low);                                      \
3042     tbl_16(texels_high, texels, clut_high);                                    \
3043     zip_8x16b(pixels, texels_low, texels_high);                                \
3044                                                                                \
3045     block->texels = pixels;                                                    \
3046     block->draw_mask_bits = right_mask_bits;                                   \
3047     block->fb_ptr = fb_ptr + 8;                                                \
3048     block++;                                                                   \
3049                                                                                \
3050     fb_ptr += 1024;                                                            \
3051     texture_offset += 0x10;                                                    \
3052     sub_tile_height--;                                                         \
3053   }                                                                            \
3054   texture_offset += 0xF00;                                                     \
3055   psx_gpu->num_blocks = num_blocks;                                            \
3056 }                                                                              \
3057
3058 #define setup_sprite_tile_half_4bpp(edge)                                      \
3059 {                                                                              \
3060   vec_8x8u texels_low, texels_high;                                            \
3061   vec_8x16u pixels;                                                            \
3062   setup_sprite_tile_add_blocks(sub_tile_height);                               \
3063                                                                                \
3064   while(sub_tile_height)                                                       \
3065   {                                                                            \
3066     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3067     tbl_16(texels_low, texels, clut_low);                                      \
3068     tbl_16(texels_high, texels, clut_high);                                    \
3069     zip_8x16b(pixels, texels_low, texels_high);                                \
3070                                                                                \
3071     block->texels = pixels;                                                    \
3072     block->draw_mask_bits = edge##_mask_bits;                                  \
3073     block->fb_ptr = fb_ptr;                                                    \
3074     block++;                                                                   \
3075                                                                                \
3076     fb_ptr += 1024;                                                            \
3077     texture_offset += 0x10;                                                    \
3078     sub_tile_height--;                                                         \
3079   }                                                                            \
3080   texture_offset += 0xF00;                                                     \
3081   psx_gpu->num_blocks = num_blocks;                                            \
3082 }                                                                              \
3083
3084
3085 #define setup_sprite_tile_full_8bpp(edge)                                      \
3086 {                                                                              \
3087   setup_sprite_tile_add_blocks(sub_tile_height * 2);                           \
3088                                                                                \
3089   while(sub_tile_height)                                                       \
3090   {                                                                            \
3091     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3092     block->r = texels;                                                         \
3093     block->draw_mask_bits = left_mask_bits;                                    \
3094     block->fb_ptr = fb_ptr;                                                    \
3095     block++;                                                                   \
3096                                                                                \
3097     setup_sprite_tile_fetch_texel_block_8bpp(8);                               \
3098     block->r = texels;                                                         \
3099     block->draw_mask_bits = right_mask_bits;                                   \
3100     block->fb_ptr = fb_ptr + 8;                                                \
3101     block++;                                                                   \
3102                                                                                \
3103     fb_ptr += 1024;                                                            \
3104     texture_offset += 0x10;                                                    \
3105     sub_tile_height--;                                                         \
3106   }                                                                            \
3107   texture_offset += 0xF00;                                                     \
3108   psx_gpu->num_blocks = num_blocks;                                            \
3109 }                                                                              \
3110
3111 #define setup_sprite_tile_half_8bpp(edge)                                      \
3112 {                                                                              \
3113   setup_sprite_tile_add_blocks(sub_tile_height * 2);                           \
3114                                                                                \
3115   while(sub_tile_height)                                                       \
3116   {                                                                            \
3117     setup_sprite_tile_fetch_texel_block_8bpp(0);                               \
3118     block->r = texels;                                                         \
3119     block->draw_mask_bits = edge##_mask_bits;                                  \
3120     block->fb_ptr = fb_ptr;                                                    \
3121     block++;                                                                   \
3122                                                                                \
3123     fb_ptr += 1024;                                                            \
3124     texture_offset += 0x10;                                                    \
3125     sub_tile_height--;                                                         \
3126   }                                                                            \
3127   texture_offset += 0xF00;                                                     \
3128   psx_gpu->num_blocks = num_blocks;                                            \
3129 }                                                                              \
3130
3131
3132 #define setup_sprite_tile_column_edge_pre_adjust_half_right()                  \
3133   texture_offset = texture_offset_base + 8;                                    \
3134   fb_ptr += 8                                                                  \
3135
3136 #define setup_sprite_tile_column_edge_pre_adjust_half_left()                   \
3137   texture_offset = texture_offset_base                                         \
3138
3139 #define setup_sprite_tile_column_edge_pre_adjust_half(edge)                    \
3140   setup_sprite_tile_column_edge_pre_adjust_half_##edge()                       \
3141
3142 #define setup_sprite_tile_column_edge_pre_adjust_full(edge)                    \
3143   texture_offset = texture_offset_base                                         \
3144
3145 #define setup_sprite_tile_column_edge_post_adjust_half_right()                 \
3146   fb_ptr -= 8                                                                  \
3147
3148 #define setup_sprite_tile_column_edge_post_adjust_half_left()                  \
3149
3150 #define setup_sprite_tile_column_edge_post_adjust_half(edge)                   \
3151   setup_sprite_tile_column_edge_post_adjust_half_##edge()                      \
3152
3153 #define setup_sprite_tile_column_edge_post_adjust_full(edge)                   \
3154
3155
3156 #define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode)  \
3157 do                                                                             \
3158 {                                                                              \
3159   sub_tile_height = column_data;                                               \
3160   setup_sprite_tile_column_edge_pre_adjust_##edge_mode(edge);                  \
3161   setup_sprite_tile_##edge_mode##_##texture_mode(edge);                        \
3162   setup_sprite_tile_column_edge_post_adjust_##edge_mode(edge);                 \
3163 } while(0)                                                                     \
3164
3165 #define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode)   \
3166 do                                                                             \
3167 {                                                                              \
3168   u32 tiles_remaining = column_data >> 16;                                     \
3169   sub_tile_height = column_data & 0xFF;                                        \
3170   setup_sprite_tile_column_edge_pre_adjust_##edge_mode(edge);                  \
3171   setup_sprite_tile_##edge_mode##_##texture_mode(edge);                        \
3172   tiles_remaining -= 1;                                                        \
3173                                                                                \
3174   while(tiles_remaining)                                                       \
3175   {                                                                            \
3176     sub_tile_height = 16;                                                      \
3177     setup_sprite_tile_##edge_mode##_##texture_mode(edge);                      \
3178     tiles_remaining--;                                                         \
3179   }                                                                            \
3180                                                                                \
3181   sub_tile_height = (column_data >> 8) & 0xFF;                                 \
3182   setup_sprite_tile_##edge_mode##_##texture_mode(edge);                        \
3183   setup_sprite_tile_column_edge_post_adjust_##edge_mode(edge);                 \
3184 } while(0)                                                                     \
3185
3186
3187 #define setup_sprite_column_data_single()                                      \
3188   column_data = height                                                         \
3189
3190 #define setup_sprite_column_data_multi()                                       \
3191   column_data = 16 - offset_v;                                                 \
3192   column_data |= ((height_rounded & 0xF) + 1) << 8;                            \
3193   column_data |= (tile_height - 1) << 16                                       \
3194
3195
3196 #define setup_sprite_tile_column_width_single(texture_mode, multi_height,      \
3197  edge_mode, edge)                                                              \
3198 {                                                                              \
3199   setup_sprite_column_data_##multi_height();                                   \
3200   left_mask_bits = left_block_mask | right_block_mask;                         \
3201   right_mask_bits = left_mask_bits >> 8;                                       \
3202                                                                                \
3203   setup_sprite_tile_column_height_##multi_height(edge_mode, edge,              \
3204    texture_mode);                                                              \
3205 }                                                                              \
3206
3207 #define setup_sprite_tiled_advance_column()                                    \
3208   texture_offset_base += 0x100;                                                \
3209   if((texture_offset_base & 0xF00) == 0)                                       \
3210     texture_offset_base -= (0x100 + 0xF00)                                     \
3211
3212 #define setup_sprite_tile_column_width_multi(texture_mode, multi_height,       \
3213  left_mode, right_mode)                                                        \
3214 {                                                                              \
3215   setup_sprite_column_data_##multi_height();                                   \
3216   s32 fb_ptr_advance_column = 16 - (1024 * height);                            \
3217                                                                                \
3218   tile_width -= 2;                                                             \
3219   left_mask_bits = left_block_mask;                                            \
3220   right_mask_bits = left_mask_bits >> 8;                                       \
3221                                                                                \
3222   setup_sprite_tile_column_height_##multi_height(left_mode, right,             \
3223    texture_mode);                                                              \
3224   fb_ptr += fb_ptr_advance_column;                                             \
3225                                                                                \
3226   left_mask_bits = 0x00;                                                       \
3227   right_mask_bits = 0x00;                                                      \
3228                                                                                \
3229   while(tile_width)                                                            \
3230   {                                                                            \
3231     setup_sprite_tiled_advance_column();                                       \
3232     setup_sprite_tile_column_height_##multi_height(full, none, texture_mode);  \
3233     fb_ptr += fb_ptr_advance_column;                                           \
3234     tile_width--;                                                              \
3235   }                                                                            \
3236                                                                                \
3237   left_mask_bits = right_block_mask;                                           \
3238   right_mask_bits = left_mask_bits >> 8;                                       \
3239                                                                                \
3240   setup_sprite_tiled_advance_column();                                         \
3241   setup_sprite_tile_column_height_##multi_height(right_mode, left,             \
3242    texture_mode);                                                              \
3243 }                                                                              \
3244
3245
3246 #define setup_sprite_tiled_builder(texture_mode)                               \
3247 void setup_sprite_##texture_mode(psx_gpu_struct *psx_gpu, s32 x, s32 y,        \
3248  s32 u, s32 v, s32 width, s32 height, u32 color)                               \
3249 {                                                                              \
3250   s32 offset_u = u & 0xF;                                                      \
3251   s32 offset_v = v & 0xF;                                                      \
3252                                                                                \
3253   s32 width_rounded = offset_u + width + 15;                                   \
3254   s32 height_rounded = offset_v + height + 15;                                 \
3255   s32 tile_height = height_rounded / 16;                                       \
3256   s32 tile_width = width_rounded / 16;                                         \
3257   u32 offset_u_right = width_rounded & 0xF;                                    \
3258                                                                                \
3259   u32 left_block_mask = ~(0xFFFF << offset_u);                                 \
3260   u32 right_block_mask = 0xFFFE << offset_u_right;                             \
3261                                                                                \
3262   u32 left_mask_bits;                                                          \
3263   u32 right_mask_bits;                                                         \
3264                                                                                \
3265   u32 sub_tile_height;                                                         \
3266   u32 column_data;                                                             \
3267                                                                                \
3268   u32 texture_mask = (psx_gpu->texture_mask_width & 0xF) |                     \
3269    ((psx_gpu->texture_mask_height & 0xF) << 4) |                               \
3270    ((psx_gpu->texture_mask_width >> 4) << 8) |                                 \
3271    ((psx_gpu->texture_mask_height >> 4) << 12);                                \
3272   u32 texture_offset = ((v & 0xF) << 4) | ((u & 0xF0) << 4) |                  \
3273    ((v & 0xF0) << 8);                                                          \
3274   u32 texture_offset_base = texture_offset;                                    \
3275   u32 control_mask;                                                            \
3276                                                                                \
3277   u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + (x - offset_u);               \
3278   u32 num_blocks = psx_gpu->num_blocks;                                        \
3279   block_struct *block = psx_gpu->blocks + num_blocks;                          \
3280                                                                                \
3281   u16 *texture_block_ptr;                                                      \
3282   vec_8x8u texels;                                                             \
3283                                                                                \
3284   setup_sprite_tiled_initialize_##texture_mode();                              \
3285                                                                                \
3286   control_mask = tile_width == 1;                                              \
3287   control_mask |= (tile_height == 1) << 1;                                     \
3288   control_mask |= ((left_block_mask & 0xFF) == 0xFF) << 2;                     \
3289   control_mask |= (((right_block_mask >> 8) & 0xFF) == 0xFF) << 3;             \
3290                                                                                \
3291   sprites_##texture_mode++;                                                    \
3292                                                                                \
3293   switch(control_mask)                                                         \
3294   {                                                                            \
3295     default:                                                                   \
3296     case 0x0:                                                                  \
3297       setup_sprite_tile_column_width_multi(texture_mode, multi, full, full);   \
3298       break;                                                                   \
3299                                                                                \
3300     case 0x1:                                                                  \
3301       setup_sprite_tile_column_width_single(texture_mode, multi, full, none);  \
3302       break;                                                                   \
3303                                                                                \
3304     case 0x2:                                                                  \
3305       setup_sprite_tile_column_width_multi(texture_mode, single, full, full);  \
3306       break;                                                                   \
3307                                                                                \
3308     case 0x3:                                                                  \
3309       setup_sprite_tile_column_width_single(texture_mode, single, full, none); \
3310       break;                                                                   \
3311                                                                                \
3312     case 0x4:                                                                  \
3313       setup_sprite_tile_column_width_multi(texture_mode, multi, half, full);   \
3314       break;                                                                   \
3315                                                                                \
3316     case 0x5:                                                                  \
3317       setup_sprite_tile_column_width_single(texture_mode, multi, half, right); \
3318       break;                                                                   \
3319                                                                                \
3320     case 0x6:                                                                  \
3321       setup_sprite_tile_column_width_multi(texture_mode, single, half, full);  \
3322       break;                                                                   \
3323                                                                                \
3324     case 0x7:                                                                  \
3325       setup_sprite_tile_column_width_single(texture_mode, single, half, right);\
3326       break;                                                                   \
3327                                                                                \
3328     case 0x8:                                                                  \
3329       setup_sprite_tile_column_width_multi(texture_mode, multi, full, half);   \
3330       break;                                                                   \
3331                                                                                \
3332     case 0x9:                                                                  \
3333       setup_sprite_tile_column_width_single(texture_mode, multi, half, left);  \
3334       break;                                                                   \
3335                                                                                \
3336     case 0xA:                                                                  \
3337       setup_sprite_tile_column_width_multi(texture_mode, single, full, half);  \
3338       break;                                                                   \
3339                                                                                \
3340     case 0xB:                                                                  \
3341       setup_sprite_tile_column_width_single(texture_mode, single, half, left); \
3342       break;                                                                   \
3343                                                                                \
3344     case 0xC:                                                                  \
3345       setup_sprite_tile_column_width_multi(texture_mode, multi, half, half);   \
3346       break;                                                                   \
3347                                                                                \
3348     case 0xE:                                                                  \
3349       setup_sprite_tile_column_width_multi(texture_mode, single, half, half);  \
3350       break;                                                                   \
3351   }                                                                            \
3352 }                                                                              \
3353
3354
3355 void setup_sprite_4bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3356  s32 width, s32 height, u32 color);
3357 void setup_sprite_8bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3358  s32 width, s32 height, u32 color);
3359 void setup_sprite_16bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u, s32 v,
3360  s32 width, s32 height, u32 color);
3361
3362 #ifndef NEON_BUILD
3363 setup_sprite_tiled_builder(4bpp);
3364 setup_sprite_tiled_builder(8bpp);
3365
3366 void setup_sprite_16bpp(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
3367  s32 v, s32 width, s32 height, u32 color)
3368 {
3369   u32 left_offset = u & 0x7;
3370   u32 width_rounded = width + left_offset + 7;
3371
3372   u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + (x - left_offset);
3373   u32 right_width = width_rounded & 0x7;
3374   u32 block_width = width_rounded / 8;
3375   u32 fb_ptr_pitch = (1024 + 8) - (block_width * 8);
3376
3377   u32 left_mask_bits = ~(0xFF << left_offset);
3378   u32 right_mask_bits = 0xFE << right_width;
3379
3380   u32 texture_offset_base = u + (v * 1024);
3381   u32 texture_mask =
3382    psx_gpu->texture_mask_width | (psx_gpu->texture_mask_height * 1024);
3383
3384   u32 blocks_remaining;
3385   u32 num_blocks = psx_gpu->num_blocks;
3386   block_struct *block = psx_gpu->blocks + num_blocks;
3387
3388   u16 *texture_page_ptr = psx_gpu->texture_page_ptr;
3389   u16 *texture_block_ptr;
3390
3391   texture_offset_base &= ~0x7;
3392
3393   sprites_16bpp++;
3394
3395   if(block_width == 1)
3396   {
3397     u32 mask_bits = left_mask_bits | right_mask_bits;
3398
3399     while(height)
3400     {
3401       num_blocks++;
3402       sprite_blocks++;
3403
3404       if(num_blocks > MAX_BLOCKS)
3405       {
3406         flush_render_block_buffer(psx_gpu);
3407         num_blocks = 1;
3408         block = psx_gpu->blocks;
3409       }
3410
3411       texture_block_ptr =
3412        texture_page_ptr + (texture_offset_base & texture_mask);
3413
3414       load_128b(block->texels, texture_block_ptr);
3415       block->draw_mask_bits = mask_bits;
3416       block->fb_ptr = fb_ptr;
3417
3418       block++;
3419
3420       texture_offset_base += 1024;
3421       fb_ptr += 1024;
3422
3423       height--;
3424       psx_gpu->num_blocks = num_blocks;
3425     }
3426   }
3427   else
3428   {
3429     u32 texture_offset;
3430
3431     while(height)
3432     {
3433       blocks_remaining = block_width - 2;
3434       num_blocks += block_width;
3435       sprite_blocks += block_width;
3436
3437       if(num_blocks > MAX_BLOCKS)
3438       {
3439         flush_render_block_buffer(psx_gpu);
3440         num_blocks = block_width;
3441         block = psx_gpu->blocks;
3442       }
3443
3444       texture_offset = texture_offset_base;
3445       texture_offset_base += 1024;
3446
3447       texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);
3448       load_128b(block->texels, texture_block_ptr);
3449
3450       block->draw_mask_bits = left_mask_bits;
3451       block->fb_ptr = fb_ptr;
3452
3453       texture_offset += 8;
3454       fb_ptr += 8;
3455       block++;
3456
3457       while(blocks_remaining)
3458       {
3459         texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);
3460         load_128b(block->texels, texture_block_ptr);
3461
3462         block->draw_mask_bits = 0;
3463         block->fb_ptr = fb_ptr;
3464
3465         texture_offset += 8;
3466         fb_ptr += 8;
3467         block++;
3468
3469         blocks_remaining--;
3470       }
3471
3472       texture_block_ptr = texture_page_ptr + (texture_offset & texture_mask);
3473       load_128b(block->texels, texture_block_ptr);
3474
3475       block->draw_mask_bits = right_mask_bits;
3476       block->fb_ptr = fb_ptr;
3477
3478       fb_ptr += fb_ptr_pitch;
3479       block++;
3480
3481       height--;
3482       psx_gpu->num_blocks = num_blocks;
3483     }
3484   }
3485 }
3486
3487 #endif
3488
3489 void setup_sprite_untextured(psx_gpu_struct *psx_gpu, s32 x, s32 y, s32 u,
3490  s32 v, s32 width, s32 height, u32 color)
3491 {
3492   u32 right_width = ((width - 1) & 0x7) + 1;
3493   u32 right_mask_bits = (0xFF << right_width);
3494   u16 *fb_ptr = psx_gpu->vram_ptr + (y * 1024) + x;
3495   u32 block_width = (width + 7) / 8;
3496   u32 fb_ptr_pitch = 1024 - ((block_width - 1) * 8);
3497   u32 blocks_remaining;
3498   u32 num_blocks = psx_gpu->num_blocks;
3499   block_struct *block = psx_gpu->blocks + num_blocks;
3500
3501   u32 color_r = color & 0xFF;
3502   u32 color_g = (color >> 8) & 0xFF;
3503   u32 color_b = (color >> 16) & 0xFF;
3504   vec_8x16u colors;
3505   vec_8x16u right_mask;
3506   vec_8x16u test_mask = psx_gpu->test_mask;
3507   vec_8x16u zero_mask;
3508
3509   sprites_untextured++;
3510
3511   color = (color_r >> 3) | ((color_g >> 3) << 5) | ((color_b >> 3) << 10);
3512
3513   dup_8x16b(colors, color);
3514   dup_8x16b(zero_mask, 0x00);
3515   dup_8x16b(right_mask, right_mask_bits);
3516   tst_8x16b(right_mask, right_mask, test_mask);
3517
3518   while(height)
3519   {
3520     blocks_remaining = block_width - 1;
3521     num_blocks += block_width;
3522     sprite_blocks += block_width;
3523
3524     if(num_blocks > MAX_BLOCKS)
3525     {
3526       flush_render_block_buffer(psx_gpu);
3527       num_blocks = block_width;
3528       block = psx_gpu->blocks;
3529     }
3530
3531     while(blocks_remaining)
3532     {
3533       block->pixels = colors;
3534       block->draw_mask = zero_mask;
3535       block->fb_ptr = fb_ptr;
3536
3537       fb_ptr += 8;
3538       block++;
3539       blocks_remaining--;
3540     }
3541
3542     block->pixels = colors;
3543     block->draw_mask = right_mask;
3544     block->fb_ptr = fb_ptr;
3545
3546     block++;
3547     fb_ptr += fb_ptr_pitch;
3548
3549     height--;
3550     psx_gpu->num_blocks = num_blocks;
3551   }
3552 }
3553
3554
3555
3556 #define setup_sprite_blocks_switch_textured(texture_mode)                      \
3557   setup_sprite_##texture_mode                                                  \
3558
3559 #define setup_sprite_blocks_switch_untextured(texture_mode)                    \
3560   setup_sprite_untextured                                                      \
3561
3562 #define setup_sprite_blocks_switch(texturing, texture_mode)                    \
3563   setup_sprite_blocks_switch_##texturing(texture_mode)                         \
3564
3565
3566 #define texture_sprite_blocks_switch_4bpp()                                    \
3567   texture_blocks_untextured                                                    \
3568
3569 #define texture_sprite_blocks_switch_8bpp()                                    \
3570   texture_sprite_blocks_8bpp                                                   \
3571
3572 #define texture_sprite_blocks_switch_16bpp()                                   \
3573   texture_blocks_untextured                                                    \
3574
3575 #define texture_sprite_blocks_switch_untextured(texture_mode)                  \
3576   texture_blocks_untextured                                                    \
3577
3578 #define texture_sprite_blocks_switch_textured(texture_mode)                    \
3579   texture_sprite_blocks_switch_##texture_mode()                                \
3580
3581 #define render_sprite_blocks_switch_block_modulation(texture_mode, blend_mode, \
3582  mask_evaluate, shading, dithering, texturing, blending, modulation)           \
3583 {                                                                              \
3584   setup_sprite_blocks_switch(texturing, texture_mode),                         \
3585   texture_sprite_blocks_switch_##texturing(texture_mode),                      \
3586   shade_blocks_switch(unshaded, texturing, modulation, undithered, blending,   \
3587    mask_evaluate),                                                             \
3588   blend_blocks_switch(texturing, blending, blend_mode, mask_evaluate)          \
3589 }                                                                              \
3590
3591 #define render_sprite_blocks_switch_block_blending(texture_mode, blend_mode,   \
3592  mask_evaluate, shading, dithering, texturing, blending)                       \
3593   render_sprite_blocks_switch_block_modulation(texture_mode, blend_mode,       \
3594    mask_evaluate, shading, dithering, texturing, blending, modulated),         \
3595   render_sprite_blocks_switch_block_modulation(texture_mode, blend_mode,       \
3596    mask_evaluate, shading, dithering, texturing, blending, unmodulated)        \
3597
3598 #define render_sprite_blocks_switch_block_texturing(texture_mode, blend_mode,  \
3599  mask_evaluate, shading, dithering, texturing)                                 \
3600   render_sprite_blocks_switch_block_blending(texture_mode, blend_mode,         \
3601    mask_evaluate, shading, dithering, texturing, unblended),                   \
3602   render_sprite_blocks_switch_block_blending(texture_mode, blend_mode,         \
3603    mask_evaluate, shading, dithering, texturing, blended)                      \
3604
3605 #define render_sprite_blocks_switch_block_dithering(texture_mode, blend_mode,  \
3606  mask_evaluate, shading, dithering)                                            \
3607   render_sprite_blocks_switch_block_texturing(texture_mode, blend_mode,        \
3608    mask_evaluate, shading, dithering, untextured),                             \
3609   render_sprite_blocks_switch_block_texturing(texture_mode, blend_mode,        \
3610    mask_evaluate, shading, dithering, textured)                                \
3611
3612 #define render_sprite_blocks_switch_block_shading(texture_mode, blend_mode,    \
3613  mask_evaluate, shading)                                                       \
3614   render_sprite_blocks_switch_block_dithering(texture_mode, blend_mode,        \
3615    mask_evaluate, shading, undithered),                                        \
3616   render_sprite_blocks_switch_block_dithering(texture_mode, blend_mode,        \
3617    mask_evaluate, shading, dithered)                                           \
3618
3619 #define render_sprite_blocks_switch_block_mask_evaluate(texture_mode,          \
3620  blend_mode, mask_evaluate)                                                    \
3621   render_sprite_blocks_switch_block_shading(texture_mode, blend_mode,          \
3622    mask_evaluate, unshaded),                                                   \
3623   render_sprite_blocks_switch_block_shading(texture_mode, blend_mode,          \
3624    mask_evaluate, shaded)                                                      \
3625
3626 #define render_sprite_blocks_switch_block_blend_mode(texture_mode, blend_mode) \
3627   render_sprite_blocks_switch_block_mask_evaluate(texture_mode, blend_mode,    \
3628    off),                                                                       \
3629   render_sprite_blocks_switch_block_mask_evaluate(texture_mode, blend_mode,    \
3630    on)                                                                         \
3631
3632 #define render_sprite_blocks_switch_block_texture_mode(texture_mode)           \
3633   render_sprite_blocks_switch_block_blend_mode(texture_mode, average),         \
3634   render_sprite_blocks_switch_block_blend_mode(texture_mode, add),             \
3635   render_sprite_blocks_switch_block_blend_mode(texture_mode, subtract),        \
3636   render_sprite_blocks_switch_block_blend_mode(texture_mode, add_fourth)       \
3637
3638 #define render_sprite_blocks_switch_block()                                    \
3639   render_sprite_blocks_switch_block_texture_mode(4bpp),                        \
3640   render_sprite_blocks_switch_block_texture_mode(8bpp),                        \
3641   render_sprite_blocks_switch_block_texture_mode(16bpp),                       \
3642   render_sprite_blocks_switch_block_texture_mode(4bpp)                         \
3643
3644
3645 render_block_handler_struct render_sprite_block_handlers[] =
3646 {
3647   render_sprite_blocks_switch_block()
3648 };
3649
3650
3651 void render_sprite(psx_gpu_struct *psx_gpu, s32 x, s32 y, u32 u, u32 v,
3652  s32 width, s32 height, u32 flags, u32 color)
3653 {
3654   s32 x_right = x + width - 1;
3655   s32 y_bottom = y + height - 1;
3656
3657   if(invalidate_texture_cache_region_viewport(psx_gpu, x, y, x_right,
3658    y_bottom) == 0)
3659   {
3660     return;
3661   }
3662
3663   if(x < psx_gpu->viewport_start_x)
3664   {
3665     u32 clip = psx_gpu->viewport_start_x - x;
3666     x += clip;
3667     u += clip;
3668     width -= clip;
3669   }
3670
3671   if(y < psx_gpu->viewport_start_y)
3672   {
3673     s32 clip = psx_gpu->viewport_start_y - y;
3674     y += clip;
3675     v += clip;
3676     height -= clip;
3677   }
3678
3679   if(x_right > psx_gpu->viewport_end_x)
3680     width -= x_right - psx_gpu->viewport_end_x;
3681
3682   if(y_bottom > psx_gpu->viewport_end_y)
3683     height -= y_bottom - psx_gpu->viewport_end_y;
3684
3685   if((width <= 0) || (height <= 0))
3686     return;
3687
3688   sprites++;
3689
3690   span_pixels += width * height;
3691   spans += height;
3692
3693   u32 render_state = flags &
3694    (RENDER_FLAGS_MODULATE_TEXELS | RENDER_FLAGS_BLEND |
3695    RENDER_FLAGS_TEXTURE_MAP);
3696   render_state |=
3697    (psx_gpu->render_state_base & ~RENDER_STATE_DITHER);
3698
3699   if((psx_gpu->render_state != render_state) ||
3700    (psx_gpu->primitive_type != PRIMITIVE_TYPE_SPRITE))
3701   {
3702     psx_gpu->render_state = render_state;
3703     flush_render_block_buffer(psx_gpu);
3704     state_changes++;
3705   }
3706
3707   psx_gpu->primitive_type = PRIMITIVE_TYPE_SPRITE;
3708
3709   color &= 0xFFFFFF;
3710
3711   if(psx_gpu->triangle_color != color)
3712   {
3713     flush_render_block_buffer(psx_gpu);
3714     psx_gpu->triangle_color = color;
3715   }
3716
3717   if(color == 0x808080)
3718     render_state |= RENDER_FLAGS_MODULATE_TEXELS;
3719
3720   render_block_handler_struct *render_block_handler =
3721    &(render_sprite_block_handlers[render_state]);
3722   psx_gpu->render_block_handler = render_block_handler;
3723
3724   ((setup_sprite_function_type *)render_block_handler->setup_blocks)
3725    (psx_gpu, x, y, u, v, width, height, color);
3726 }
3727
3728 #define draw_pixel_line_mask_evaluate_yes()                                    \
3729   if(*vram_ptr & 0x8000)                                                       \
3730
3731 #define draw_pixel_line_mask_evaluate_no()                                     \
3732
3733
3734 #define draw_pixel_line_shaded()                                               \
3735 {                                                                              \
3736   color_r = fixed_to_int(current_r);                                           \
3737   color_g = fixed_to_int(current_g);                                           \
3738   color_b = fixed_to_int(current_b);                                           \
3739                                                                                \
3740   current_r += gradient_r;                                                     \
3741   current_g += gradient_g;                                                     \
3742   current_b += gradient_b;                                                     \
3743 }                                                                              \
3744
3745 #define draw_pixel_line_unshaded()                                             \
3746 {                                                                              \
3747   color_r = color & 0xFF;                                                      \
3748   color_g = (color >> 8) & 0xFF;                                               \
3749   color_b = (color >> 16) & 0xFF;                                              \
3750 }                                                                              \
3751
3752
3753 #define draw_pixel_line_dithered(_x, _y)                                       \
3754 {                                                                              \
3755   u32 dither_xor = _x ^ _y;                                                    \
3756   s32 dither_offset = (dither_xor >> 1) & 0x1;                                 \
3757   dither_offset |= (_y & 0x1) << 1;                                            \
3758   dither_offset |= (dither_xor & 0x1) << 2;                                    \
3759   dither_offset -= 4;                                                          \
3760                                                                                \
3761   color_r += dither_offset;                                                    \
3762   color_g += dither_offset;                                                    \
3763   color_b += dither_offset;                                                    \
3764                                                                                \
3765   if(color_r < 0)                                                              \
3766     color_r = 0;                                                               \
3767                                                                                \
3768   if(color_g < 0)                                                              \
3769     color_g = 0;                                                               \
3770                                                                                \
3771   if(color_b < 0)                                                              \
3772     color_b = 0;                                                               \
3773                                                                                \
3774   if(color_r > 255)                                                            \
3775     color_r = 255;                                                             \
3776                                                                                \
3777   if(color_g > 255)                                                            \
3778     color_g = 255;                                                             \
3779                                                                                \
3780   if(color_b > 255)                                                            \
3781     color_b = 255;                                                             \
3782 }                                                                              \
3783
3784 #define draw_pixel_line_undithered(_x, _y)                                     \
3785
3786
3787 #define draw_pixel_line_average()                                              \
3788   color_r = (color_r + fb_r) / 2;                                              \
3789   color_g = (color_g + fb_g) / 2;                                              \
3790   color_b = (color_b + fb_b) / 2                                               \
3791
3792 #define draw_pixel_line_add()                                                  \
3793   color_r += fb_r;                                                             \
3794   color_g += fb_g;                                                             \
3795   color_b += fb_b;                                                             \
3796                                                                                \
3797   if(color_r > 31)                                                             \
3798     color_r = 31;                                                              \
3799                                                                                \
3800   if(color_g > 31)                                                             \
3801     color_g = 31;                                                              \
3802                                                                                \
3803   if(color_b > 31)                                                             \
3804     color_b = 31                                                               \
3805                                                                                \
3806
3807 #define draw_pixel_line_subtract()                                             \
3808   color_r = fb_r - color_r;                                                    \
3809   color_g = fb_g - color_g;                                                    \
3810   color_b = fb_b - color_b;                                                    \
3811                                                                                \
3812   if(color_r < 0)                                                              \
3813     color_r = 0;                                                               \
3814                                                                                \
3815   if(color_g < 0)                                                              \
3816     color_g = 0;                                                               \
3817                                                                                \
3818   if(color_b < 0)                                                              \
3819     color_b = 0                                                                \
3820
3821 #define draw_pixel_line_add_fourth()                                           \
3822   color_r = fb_r + (color_r / 4);                                              \
3823   color_g = fb_g + (color_g / 4);                                              \
3824   color_b = fb_b + (color_b / 4);                                              \
3825                                                                                \
3826   if(color_r > 31)                                                             \
3827     color_r = 31;                                                              \
3828                                                                                \
3829   if(color_g > 31)                                                             \
3830     color_g = 31;                                                              \
3831                                                                                \
3832   if(color_b > 31)                                                             \
3833     color_b = 31                                                               \
3834
3835
3836 #define draw_pixel_line_blended(blend_mode)                                    \
3837   s32 fb_pixel = *vram_ptr;                                                    \
3838   s32 fb_r = fb_pixel & 0x1F;                                                  \
3839   s32 fb_g = (fb_pixel >> 5) & 0x1F;                                           \
3840   s32 fb_b = (fb_pixel >> 10) & 0x1F;                                          \
3841                                                                                \
3842   draw_pixel_line_##blend_mode()                                               \
3843
3844 #define draw_pixel_line_unblended(blend_mode)                                  \
3845
3846
3847 #define draw_pixel_line(_x, _y, shading, blending, dithering, mask_evaluate,   \
3848  blend_mode)                                                                   \
3849   if((_x >= psx_gpu->viewport_start_x) && (_y >= psx_gpu->viewport_start_y) && \
3850    (_x <= psx_gpu->viewport_end_x) && (_y <= psx_gpu->viewport_end_y))         \
3851   {                                                                            \
3852     draw_pixel_line_mask_evaluate_##mask_evaluate()                            \
3853     {                                                                          \
3854       draw_pixel_line_##shading();                                             \
3855       draw_pixel_line_##dithering(_x, _y);                                     \
3856                                                                                \
3857       color_r >>= 3;                                                           \
3858       color_g >>= 3;                                                           \
3859       color_b >>= 3;                                                           \
3860                                                                                \
3861       draw_pixel_line_##blending(blend_mode);                                  \
3862                                                                                \
3863       *vram_ptr = color_r | (color_g << 5) | (color_b << 10) |                 \
3864        psx_gpu->mask_msb;                                                      \
3865     }                                                                          \
3866   }                                                                            \
3867
3868 #define update_increment(value)                                                \
3869   value++                                                                      \
3870
3871 #define update_decrement(value)                                                \
3872   value--                                                                      \
3873
3874 #define update_vram_row_increment(value)                                       \
3875   vram_ptr += 1024                                                             \
3876
3877 #define update_vram_row_decrement(value)                                       \
3878   vram_ptr -= 1024                                                             \
3879
3880 #define compare_increment(a, b)                                                \
3881   (a <= b)                                                                     \
3882
3883 #define compare_decrement(a, b)                                                \
3884   (a >= b)                                                                     \
3885
3886 #define set_line_gradients(minor)                                              \
3887 {                                                                              \
3888   s32 gradient_divisor = delta_##minor;                                        \
3889   if(gradient_divisor != 0)                                                    \
3890   {                                                                            \
3891     gradient_r = int_to_fixed(vertex_b->r - vertex_a->r) / gradient_divisor;   \
3892     gradient_g = int_to_fixed(vertex_b->g - vertex_a->g) / gradient_divisor;   \
3893     gradient_b = int_to_fixed(vertex_b->b - vertex_a->b) / gradient_divisor;   \
3894   }                                                                            \
3895   else                                                                         \
3896   {                                                                            \
3897     gradient_r = 0;                                                            \
3898     gradient_g = 0;                                                            \
3899     gradient_b = 0;                                                            \
3900   }                                                                            \
3901   current_r = fixed_center(vertex_a->r);                                       \
3902   current_g = fixed_center(vertex_a->g);                                       \
3903   current_b = fixed_center(vertex_a->b);                                       \
3904 }
3905
3906 #define draw_line_span_horizontal(direction, shading, blending, dithering,     \
3907  mask_evaluate, blend_mode)                                                    \
3908 do                                                                             \
3909 {                                                                              \
3910   error_step = delta_y * 2;                                                    \
3911   error_wrap = delta_x * 2;                                                    \
3912   error = delta_x;                                                             \
3913                                                                                \
3914   current_y = y_a;                                                             \
3915   set_line_gradients(x);                                                       \
3916                                                                                \
3917   for(current_x = x_a; current_x <= x_b; current_x++)                          \
3918   {                                                                            \
3919     draw_pixel_line(current_x, current_y, shading, blending, dithering,        \
3920      mask_evaluate, blend_mode);                                               \
3921     error += error_step;                                                       \
3922     vram_ptr++;                                                                \
3923                                                                                \
3924     if(error >= error_wrap)                                                    \
3925     {                                                                          \
3926       update_##direction(current_y);                                           \
3927       update_vram_row_##direction();                                           \
3928       error -= error_wrap;                                                     \
3929     }                                                                          \
3930   }                                                                            \
3931 } while(0)                                                                     \
3932
3933 #define draw_line_span_vertical(direction, shading, blending, dithering,       \
3934  mask_evaluate, blend_mode)                                                    \
3935 do                                                                             \
3936 {                                                                              \
3937   error_step = delta_x * 2;                                                    \
3938   error_wrap = delta_y * 2;                                                    \
3939   error = delta_y;                                                             \
3940                                                                                \
3941   current_x = x_a;                                                             \
3942   set_line_gradients(y);                                                       \
3943                                                                                \
3944   for(current_y = y_a; compare_##direction(current_y, y_b);                    \
3945    update_##direction(current_y))                                              \
3946   {                                                                            \
3947     draw_pixel_line(current_x, current_y, shading, blending, dithering,        \
3948      mask_evaluate, blend_mode);                                               \
3949     error += error_step;                                                       \
3950     update_vram_row_##direction();                                             \
3951                                                                                \
3952     if(error > error_wrap)                                                     \
3953     {                                                                          \
3954       vram_ptr++;                                                              \
3955       current_x++;                                                             \
3956       error -= error_wrap;                                                     \
3957     }                                                                          \
3958   }                                                                            \
3959 } while(0)                                                                     \
3960
3961
3962 #define render_line_body(shading, blending, dithering, mask_evaluate,          \
3963  blend_mode)                                                                   \
3964   if(delta_y < 0)                                                              \
3965   {                                                                            \
3966     delta_y *= -1;                                                             \
3967                                                                                \
3968     if(delta_y >= 512)                                                         \
3969       return;                                                                  \
3970                                                                                \
3971     if(delta_x > delta_y)                                                      \
3972     {                                                                          \
3973       draw_line_span_horizontal(decrement, shading, blending, dithering,       \
3974        mask_evaluate, blend_mode);                                             \
3975     }                                                                          \
3976     else                                                                       \
3977     {                                                                          \
3978       draw_line_span_vertical(decrement, shading, blending, dithering,         \
3979        mask_evaluate, blend_mode);                                             \
3980     }                                                                          \
3981   }                                                                            \
3982   else                                                                         \
3983   {                                                                            \
3984     if(delta_y >= 512)                                                         \
3985       return;                                                                  \
3986                                                                                \
3987     if(delta_x > delta_y)                                                      \
3988     {                                                                          \
3989       draw_line_span_horizontal(increment, shading, blending, dithering,       \
3990        mask_evaluate, blend_mode);                                             \
3991     }                                                                          \
3992     else                                                                       \
3993     {                                                                          \
3994       draw_line_span_vertical(increment, shading, blending, dithering,         \
3995        mask_evaluate, blend_mode);                                             \
3996     }                                                                          \
3997   }                                                                            \
3998
3999
4000 void render_line(psx_gpu_struct *psx_gpu, vertex_struct *vertexes, u32 flags,
4001  u32 color)
4002 {
4003   s32 color_r, color_g, color_b;
4004   u32 triangle_winding = 0;
4005
4006   fixed_type gradient_r = 0;
4007   fixed_type gradient_g = 0;
4008   fixed_type gradient_b = 0;
4009   fixed_type current_r = 0;
4010   fixed_type current_g = 0;
4011   fixed_type current_b = 0;
4012
4013   s32 y_a, y_b;
4014   s32 x_a, x_b;
4015
4016   s32 delta_x, delta_y;
4017
4018   s32 current_x;
4019   s32 current_y;
4020
4021   u32 error_step;
4022   u32 error;
4023   u32 error_wrap;
4024
4025   u16 *vram_ptr;
4026
4027   flush_render_block_buffer(psx_gpu);
4028   psx_gpu->primitive_type = PRIMITIVE_TYPE_LINE;
4029
4030   vertex_struct *vertex_a = &(vertexes[0]);
4031   vertex_struct *vertex_b = &(vertexes[1]);
4032
4033   u32 control_mask;
4034
4035   lines++;
4036
4037   if(vertex_a->x >= vertex_b->x)
4038   {
4039     vertex_swap(vertex_a, vertex_b);
4040   }
4041
4042   x_a = vertex_a->x;
4043   x_b = vertex_b->x;
4044
4045   y_a = vertex_a->y;
4046   y_b = vertex_b->y;
4047
4048   delta_x = x_b - x_a;
4049   delta_y = y_b - y_a;
4050
4051   if(delta_x >= 1024)
4052     return;
4053
4054   flags &= ~RENDER_FLAGS_TEXTURE_MAP;
4055
4056   vram_ptr = psx_gpu->vram_ptr + (y_a * 1024) + x_a;
4057
4058   control_mask = 0x0;
4059
4060   if(flags & RENDER_FLAGS_SHADE)
4061     control_mask |= 0x1;
4062
4063   if(flags & RENDER_FLAGS_BLEND)
4064   {
4065     control_mask |= 0x2;
4066     control_mask |= ((psx_gpu->render_state_base >> 6) & 0x3) << 4;
4067   }
4068
4069   if(psx_gpu->render_state_base & RENDER_STATE_DITHER)
4070     control_mask |= 0x4;
4071
4072   if(psx_gpu->render_state_base & RENDER_STATE_MASK_EVALUATE)
4073     control_mask |= 0x8;
4074
4075   switch(control_mask)
4076   {
4077     case 0x0:
4078       render_line_body(unshaded, unblended, undithered, no, none);
4079       break;
4080
4081     case 0x1:
4082       render_line_body(shaded, unblended, undithered, no, none);
4083       break;
4084
4085     case 0x2:
4086       render_line_body(unshaded, blended, undithered, no, average);
4087       break;
4088
4089     case 0x3:
4090       render_line_body(shaded, blended, undithered, no, average);
4091       break;
4092
4093     case 0x4:
4094       render_line_body(unshaded, unblended, dithered, no, none);
4095       break;
4096
4097     case 0x5:
4098       render_line_body(shaded, unblended, dithered, no, none);
4099       break;
4100
4101     case 0x6:
4102       render_line_body(unshaded, blended, dithered, no, average);
4103       break;
4104
4105     case 0x7:
4106       render_line_body(shaded, blended, dithered, no, average);
4107       break;
4108
4109     case 0x8:
4110       render_line_body(unshaded, unblended, undithered, yes, none);
4111       break;
4112
4113     case 0x9:
4114       render_line_body(shaded, unblended, undithered, yes, none);
4115       break;
4116
4117     case 0xA:
4118       render_line_body(unshaded, blended, undithered, yes, average);
4119       break;
4120
4121     case 0xB:
4122       render_line_body(shaded, blended, undithered, yes, average);
4123       break;
4124
4125     case 0xC:
4126       render_line_body(unshaded, unblended, dithered, yes, none);
4127       break;
4128
4129     case 0xD:
4130       render_line_body(shaded, unblended, dithered, yes, none);
4131       break;
4132
4133     case 0xE:
4134       render_line_body(unshaded, blended, dithered, yes, average);
4135       break;
4136
4137     case 0xF:
4138       render_line_body(shaded, blended, dithered, yes, average);
4139       break;
4140
4141     case 0x12:
4142       render_line_body(unshaded, blended, undithered, no, add);
4143       break;
4144
4145     case 0x13:
4146       render_line_body(shaded, blended, undithered, no, add);
4147       break;
4148
4149     case 0x16:
4150       render_line_body(unshaded, blended, dithered, no, add);
4151       break;
4152
4153     case 0x17:
4154       render_line_body(shaded, blended, dithered, no, add);
4155       break;
4156
4157     case 0x1A:
4158       render_line_body(unshaded, blended, undithered, yes, add);
4159       break;
4160
4161     case 0x1B:
4162       render_line_body(shaded, blended, undithered, yes, add);
4163       break;
4164
4165     case 0x1E:
4166       render_line_body(unshaded, blended, dithered, yes, add);
4167       break;
4168
4169     case 0x1F:
4170       render_line_body(shaded, blended, dithered, yes, add);
4171       break;
4172
4173     case 0x22:
4174       render_line_body(unshaded, blended, undithered, no, subtract);
4175       break;
4176
4177     case 0x23:
4178       render_line_body(shaded, blended, undithered, no, subtract);
4179       break;
4180
4181     case 0x26:
4182       render_line_body(unshaded, blended, dithered, no, subtract);
4183       break;
4184
4185     case 0x27:
4186       render_line_body(shaded, blended, dithered, no, subtract);
4187       break;
4188
4189     case 0x2A:
4190       render_line_body(unshaded, blended, undithered, yes, subtract);
4191       break;
4192
4193     case 0x2B:
4194       render_line_body(shaded, blended, undithered, yes, subtract);
4195       break;
4196
4197     case 0x2E:
4198       render_line_body(unshaded, blended, dithered, yes, subtract);
4199       break;
4200
4201     case 0x2F:
4202       render_line_body(shaded, blended, dithered, yes, subtract);
4203       break;
4204
4205     case 0x32:
4206       render_line_body(unshaded, blended, undithered, no, add_fourth);
4207       break;
4208
4209     case 0x33:
4210       render_line_body(shaded, blended, undithered, no, add_fourth);
4211       break;
4212
4213     case 0x36:
4214       render_line_body(unshaded, blended, dithered, no, add_fourth);
4215       break;
4216
4217     case 0x37:
4218       render_line_body(shaded, blended, dithered, no, add_fourth);
4219       break;
4220
4221     case 0x3A:
4222       render_line_body(unshaded, blended, undithered, yes, add_fourth);
4223       break;
4224
4225     case 0x3B:
4226       render_line_body(shaded, blended, undithered, yes, add_fourth);
4227       break;
4228
4229     case 0x3E:
4230       render_line_body(unshaded, blended, dithered, yes, add_fourth);
4231       break;
4232
4233     case 0x3F:
4234       render_line_body(shaded, blended, dithered, yes, add_fourth);
4235       break;
4236   }
4237 }
4238
4239
4240 void render_block_fill(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
4241  u32 width, u32 height)
4242 {
4243   invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
4244
4245 #ifndef NEON_BUILD
4246   u32 r = color & 0xFF;
4247   u32 g = (color >> 8) & 0xFF;
4248   u32 b = (color >> 16) & 0xFF;
4249   u32 color_16bpp = (r >> 3) | ((g >> 3) << 5) | ((b >> 3) << 10);
4250
4251   u16 *vram_ptr = psx_gpu->vram_ptr + x + (y * 1024);
4252   u32 draw_x, draw_y;
4253
4254   for(draw_y = 0; draw_y < height; draw_y++)
4255   {
4256     for(draw_x = 0; draw_x < width; draw_x++)
4257     {
4258       vram_ptr[draw_x] = color_16bpp;
4259     }
4260
4261     vram_ptr += 1024;
4262   }
4263 #else
4264   void render_block_fill_body(psx_gpu_struct *psx_gpu, u32 color, u32 x, u32 y,
4265    u32 width, u32 height);
4266
4267   render_block_fill_body(psx_gpu, color, x, y, width, height);
4268 #endif
4269 }
4270
4271 void render_block_copy(psx_gpu_struct *psx_gpu, u16 *source, u32 x, u32 y,
4272  u32 width, u32 height, u32 pitch)
4273 {
4274   u16 *vram_ptr = psx_gpu->vram_ptr + x + (y * 1024);
4275   u32 draw_x, draw_y;
4276
4277   invalidate_texture_cache_region(psx_gpu, x, y, x + width - 1, y + height - 1);
4278
4279   //printf("copy for %d, %d\n", width, height);
4280
4281   for(draw_y = 0; draw_y < height; draw_y++)
4282   {
4283     for(draw_x = 0; draw_x < width; draw_x++)
4284     {
4285       vram_ptr[draw_x] = source[draw_x];
4286     }
4287
4288     source += pitch;
4289     vram_ptr += 1024;
4290   }
4291 }
4292
4293 void render_block_move(psx_gpu_struct *psx_gpu, u32 source_x, u32 source_y,
4294  u32 dest_x, u32 dest_y, u32 width, u32 height)
4295 {
4296   render_block_copy(psx_gpu, psx_gpu->vram_ptr + source_x + (source_y * 1024),
4297    dest_x, dest_y, width, height, 1024);
4298 }
4299
4300
4301 void initialize_reciprocal_table(void)
4302 {
4303   u32 height;
4304   u32 height_normalized;
4305   u32 height_reciprocal;
4306   s32 shift;
4307
4308   for(height = 1; height < 512; height++)
4309   {
4310     shift = __builtin_clz(height);
4311     height_normalized = height << shift;
4312     height_reciprocal = ((1ULL << 50) + (height_normalized - 1)) /
4313      height_normalized;
4314
4315     shift = 32 - (50 - shift);
4316
4317     reciprocal_table[height] = (height_reciprocal << 12) | shift;
4318   }
4319 }
4320
4321
4322 #define dither_table_row(a, b, c, d)                                           \
4323  ((a & 0xFF) | ((b & 0xFF) << 8) | ((c & 0xFF) << 16) | ((d & 0xFF) << 24))    \
4324
4325 void initialize_psx_gpu(psx_gpu_struct *psx_gpu, u16 *vram)
4326 {
4327   vec_8x16u test_mask =
4328    { { { 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80 } } };
4329
4330   psx_gpu->test_mask = test_mask;
4331
4332   psx_gpu->pixel_count_mode = 0;
4333   psx_gpu->pixel_compare_mode = 0;
4334
4335   psx_gpu->vram_pixel_counts_a = malloc(sizeof(u8) * 1024 * 512);
4336   psx_gpu->vram_pixel_counts_b = malloc(sizeof(u8) * 1024 * 512);
4337   memset(psx_gpu->vram_pixel_counts_a, 0, sizeof(u8) * 1024 * 512);
4338   memset(psx_gpu->vram_pixel_counts_b, 0, sizeof(u8) * 1024 * 512);
4339   psx_gpu->compare_vram = malloc(sizeof(u16) * 1024 * 512);
4340
4341   psx_gpu->dirty_textures_4bpp_mask = 0xFFFFFFFF;
4342   psx_gpu->dirty_textures_8bpp_mask = 0xFFFFFFFF;
4343   psx_gpu->dirty_textures_8bpp_alternate_mask = 0xFFFFFFFF;
4344   psx_gpu->viewport_mask = 0;
4345   psx_gpu->current_texture_page = 0;
4346   psx_gpu->current_texture_mask = 0;
4347   psx_gpu->last_8bpp_texture_page = 0;
4348
4349   psx_gpu->clut_settings = 0;
4350   psx_gpu->texture_settings = 0;
4351   psx_gpu->render_state = 0;
4352   psx_gpu->render_state_base = 0;
4353   psx_gpu->num_blocks = 0;
4354
4355   psx_gpu->vram_ptr = vram;
4356
4357   psx_gpu->texture_page_ptr = psx_gpu->vram_ptr;
4358   psx_gpu->clut_ptr = psx_gpu->vram_ptr;
4359
4360   psx_gpu->mask_msb = 0;
4361
4362   memset(psx_gpu->vram_ptr, 0, sizeof(u16) * 1024 * 512);
4363
4364   initialize_reciprocal_table();
4365
4366   //    00 01 10 11
4367   // 00  0  4  1  5
4368   // 01  6  2  7  3
4369   // 10  1  5  0  4
4370   // 11  7  3  6  2
4371   // (minus ones(4) * 4)
4372
4373   // d0: (1 3 5 7): x1 ^ y1
4374   // d1: (2 3 6 7): y0
4375   // d2: (4 5 6 7): x0 ^ y0
4376
4377
4378   psx_gpu->dither_table[0] = dither_table_row(-4, 0, -3, 1);
4379   psx_gpu->dither_table[1] = dither_table_row(2, -2, 3, -1);
4380   psx_gpu->dither_table[2] = dither_table_row(-3, 1, -4, 0);
4381   psx_gpu->dither_table[3] = dither_table_row(3, -1, 2, -2);
4382
4383   psx_gpu->primitive_type = PRIMITIVE_TYPE_UNKNOWN;
4384 }
4385
4386 u64 get_us(void)
4387 {
4388   struct timeval tv;
4389   gettimeofday(&tv, NULL);
4390
4391   return (tv.tv_sec * 1000000ULL) + tv.tv_usec;
4392 }
4393
4394 #ifdef NEON_BUILD
4395
4396 u32 get_counter()
4397 {
4398   u32 counter;
4399   __asm__ volatile("mrc p15, 0, %0, c9, c13, 0" : "=r"(counter));
4400
4401   return counter;
4402 }
4403
4404 void init_counter(void)
4405 {
4406   u32 value;
4407   asm volatile("mrc p15, 0, %0, c9, c12, 0" : "=r"(value));
4408   value |= 5; // master enable, ccnt reset
4409   value &= ~8; // ccnt divider 0
4410   asm volatile("mcr p15, 0, %0, c9, c12, 0" :: "r"(value));
4411   // enable cycle counter
4412   asm volatile("mcr p15, 0, %0, c9, c12, 1" :: "r"(1 << 31));
4413 }
4414
4415 void triangle_benchmark(psx_gpu_struct *psx_gpu)
4416 {
4417   u32 i;
4418
4419   u32 ticks;
4420   u32 ticks_elapsed;
4421
4422   const u32 iterations = 500000;
4423
4424   psx_gpu->num_blocks = 64;
4425   psx_gpu->clut_ptr = psx_gpu->vram_ptr;
4426
4427   for(i = 0; i < 64; i++)
4428   {
4429     memset(&(psx_gpu->blocks[i].r), 0, 16);
4430   }
4431
4432   init_counter();
4433
4434   ticks = get_counter();
4435
4436   for(i = 0; i < iterations; i++)
4437   {
4438     texture_sprite_blocks_8bpp(psx_gpu);
4439   }
4440
4441   ticks_elapsed = get_counter() - ticks;
4442
4443   printf("benchmark: %lf cycles\n", (double)ticks_elapsed / (iterations * 64));
4444 }
4445
4446 #endif
4447