notaz.gp2x.de / pcsx_rearmed.git / blame
| Commit | Line | Data |
|---|---|---|
| 75e28f62 E |
1 | /* |
| 2 | * Copyright (C) 2011 Gilead Kutnick "Exophase" <exophase@gmail.com> | |
| 59d15d23 | 3 | * Copyright (C) 2012 GraÅžvydas Ignotas "notaz" <notasas@gmail.com> |
| 75e28f62 E |
4 | * |
| 5 | * This program is free software; you can redistribute it and/or | |
| 6 | * modify it under the terms of the GNU General Public License as | |
| 7 | * published by the Free Software Foundation; either version 2 of | |
| 8 | * the License, or (at your option) any later version. | |
| 9 | * | |
| 10 | * This program is distributed in the hope that it will be useful, | |
| 11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
| 12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
| 13 | * General Public License for more details. | |
| 14 | */ | |
| 15 | ||
| d5c08ed3 | 16 | #define RENDER_INTERLACE_ENABLED 0x1 |
| f0931e56 | 17 | |
| 2d658c89 | 18 | #include "psx_gpu.h" |
| cb88320b | 19 | #include "psx_gpu_offsets.h" |
| 75e28f62 | 20 | |
| cb88320b | 21 | #define psx_gpu_b_dx_offset (psx_gpu_b_block_span_offset + 4) |
| 75e28f62 | 22 | |
| 75e28f62 E |
23 | #define edge_data_left_x_offset 0 |
| 24 | #define edge_data_num_blocks_offset 2 | |
| 25 | #define edge_data_right_mask_offset 4 | |
| 26 | #define edge_data_y_offset 6 | |
| 27 | ||
| ed0fd81d | 28 | .syntax unified |
| 29 | .text | |
| 75e28f62 | 30 | |
| 3f0189c6 | 31 | #if 0 |
| 32 | #define save_abi_regs() \ | |
| 33 | vpush {q4-q7} | |
| 34 | #define restore_abi_regs() \ | |
| 35 | vpop {q4-q7} | |
| 36 | #else | |
| 37 | #define save_abi_regs() | |
| 38 | #define restore_abi_regs() | |
| 39 | #endif | |
| 40 | ||
| 75e28f62 E |
41 | #define psx_gpu r0 |
| 42 | #define v_a r1 | |
| 43 | #define v_b r2 | |
| 44 | #define v_c r3 | |
| 45 | ||
| 46 | #define x0 r4 | |
| 47 | #define x1 r5 | |
| 48 | #define x2 r6 | |
| 49 | #define x0_x1 r5 | |
| 50 | #define x1_x2 r6 | |
| 51 | #define y0 r7 | |
| 52 | #define y1 r8 | |
| 53 | #define y2 r9 | |
| 54 | #define y0_y1 r7 | |
| 55 | #define y1_y2 r8 | |
| 56 | #define b0 r9 | |
| 57 | #define b1 r10 | |
| 58 | #define b2 r11 | |
| 59 | #define b0_b1 r10 | |
| 60 | #define b1_b2 r11 | |
| 61 | ||
| 62 | ||
| 63 | #define area_r_s r5 | |
| 64 | ||
| 65 | #define g_bx0 r2 | |
| 66 | #define g_bx r3 | |
| 67 | #define g_bx2 r4 | |
| 68 | #define g_bx3 r5 | |
| 69 | #define b_base r6 | |
| 70 | #define g_by r8 | |
| 71 | ||
| 72 | #define gs_bx r7 | |
| 73 | #define gs_by r10 | |
| 74 | ||
| 75 | #define ga_bx g_bx | |
| 76 | #define ga_by g_by | |
| 77 | ||
| 78 | #define gw_bx_h g_bx | |
| 79 | #define gw_by_h g_by | |
| 80 | ||
| 81 | #define gw_bx_l r11 | |
| 82 | #define gw_by_l gw_bx_l | |
| 83 | ||
| 84 | #define store_a r0 | |
| 85 | #define store_b r1 | |
| 86 | #define store_inc r5 | |
| 87 | ||
| 88 | ||
| 89 | #define v0 q0 | |
| 90 | #define uvrgb0 d0 | |
| 91 | #define x0_y0 d1 | |
| 92 | ||
| 93 | #define v1 q1 | |
| 94 | #define uvrgb1 d2 | |
| 95 | #define x1_y1 d3 | |
| 96 | ||
| 97 | #define v2 q2 | |
| 98 | #define uvrgb2 d4 | |
| 99 | #define x2_y2 d5 | |
| 100 | ||
| 101 | #define x0_ab q3 | |
| 102 | #define uvrg_xxxx0 q3 | |
| 103 | #define uvrg0 d6 | |
| 104 | #define xxxx0 d7 | |
| 105 | ||
| 106 | #define x1_ab q4 | |
| 107 | #define uvrg_xxxx1 q4 | |
| 108 | #define uvrg1 d8 | |
| 109 | #define xxxx1 d9 | |
| 110 | ||
| 111 | #define x2_ab q5 | |
| 112 | #define uvrg_xxxx2 q5 | |
| 113 | #define uvrg2 d10 | |
| 114 | #define xxxx2 d11 | |
| 115 | ||
| 116 | #define y0_ab q6 | |
| 117 | #define yyyy_uvrg0 q6 | |
| 118 | #define yyyy0 d12 | |
| 119 | #define uvrg0b d13 | |
| 120 | ||
| 121 | #define y1_ab q7 | |
| 122 | #define yyyy_uvrg1 q7 | |
| 123 | #define yyyy1 d14 | |
| 124 | #define uvrg1b d15 | |
| 125 | ||
| 126 | #define y2_ab q8 | |
| 127 | #define yyyy_uvrg2 q8 | |
| 128 | #define yyyy2 d16 | |
| 129 | #define uvrg2b d17 | |
| 130 | ||
| 131 | #define d0_ab q9 | |
| 132 | #define d0_a d18 | |
| 133 | #define d0_b d19 | |
| 134 | ||
| 135 | #define d1_ab q10 | |
| 136 | #define d1_a d20 | |
| 137 | #define d1_b d21 | |
| 138 | ||
| 139 | #define d2_ab q11 | |
| 140 | #define d2_a d22 | |
| 141 | #define d2_b d23 | |
| 142 | ||
| 143 | #define d3_ab q12 | |
| 144 | #define d3_a d24 | |
| 145 | #define d3_b d25 | |
| 146 | ||
| 147 | #define ga_uvrg_x q1 | |
| 148 | #define ga_uvrg_y q4 | |
| 149 | ||
| 150 | #define dx x0_x1 | |
| 151 | #define dy y0_y1 | |
| 152 | #define db b0_b1 | |
| 153 | ||
| 154 | #define uvrg_base q11 | |
| 155 | ||
| 156 | #define gs_uvrg_x q5 | |
| 157 | #define gs_uvrg_y q6 | |
| 158 | ||
| 159 | #define g_uvrg_x q1 | |
| 160 | #define ga_uv_x d2 | |
| 161 | #define g_uv_x d2 | |
| 162 | #define ga_rg_x d3 | |
| 163 | #define g_rg_x d3 | |
| 164 | ||
| 165 | #define g_uvrg_y q4 | |
| 166 | #define ga_uv_y d8 | |
| 167 | #define g_uv_y d8 | |
| 168 | #define ga_rg_y d9 | |
| 169 | #define g_rg_y d9 | |
| 170 | ||
| 171 | #define gw_uv_x q1 | |
| 172 | #define gw_rg_x q2 | |
| 173 | #define gw_uv_y q4 | |
| 174 | #define gw_rg_y q3 | |
| 175 | ||
| 176 | #define w_mask q9 | |
| 177 | #define w_mask_l d18 | |
| 178 | ||
| 179 | #define r_shift q10 | |
| 180 | ||
| 181 | #define uvrg_dx0 q0 | |
| 182 | #define uvrg_dx0l d0 | |
| 183 | #define uvrg_dx0h d1 | |
| 184 | ||
| 185 | #define uvrg_dx1 q1 | |
| 186 | #define uvrg_dx1l d2 | |
| 187 | #define uvrg_dx1h d3 | |
| 188 | ||
| 189 | #define uvrg_dx2 q2 | |
| 190 | #define uvrg_dx2l d4 | |
| 191 | #define uvrg_dx2h d5 | |
| 192 | ||
| 193 | #define uvrg_dx3 q3 | |
| 194 | #define uvrg_dx3l d6 | |
| 195 | #define uvrg_dx3h d7 | |
| 196 | ||
| c6063f89 | 197 | #define uvrgb_phase q13 |
| 75e28f62 E |
198 | |
| 199 | .align 4 | |
| 200 | ||
| 0e4ad319 | 201 | #include "arm_features.h" |
| 8184d7c5 | 202 | |
| 0e4ad319 | 203 | #define function(name) FUNCTION(name): |
| 204 | ||
| 205 | #ifndef TEXRELS_FORBIDDEN | |
| 75e28f62 | 206 | |
| 8184d7c5 | 207 | #define JT_OP_REL(table_label, index_reg, temp) |
| 208 | #define JT_OP(x...) x | |
| 209 | #define JTE(start, target) target | |
| 210 | ||
| 211 | #else | |
| 212 | ||
| 8184d7c5 | 213 | #define JT_OP_REL(table_label, index_reg, temp) \ |
| 214 | adr temp, table_label; \ | |
| e1f6de8f | 215 | ldr temp, [temp, index_reg, lsl #2]; \ |
| 8184d7c5 | 216 | add pc, pc, temp \ |
| 217 | ||
| 218 | #define JT_OP(x...) | |
| 219 | #define JTE(start, target) (target - start) | |
| 220 | ||
| 0e4ad319 | 221 | #endif |
| 4d646738 | 222 | |
| 0e4ad319 | 223 | #ifdef __MACH__ |
| 8184d7c5 | 224 | #define flush_render_block_buffer _flush_render_block_buffer |
| 8184d7c5 | 225 | #define update_texture_8bpp_cache _update_texture_8bpp_cache |
| ee060c58 | 226 | #define setup_blocks_uv_adj_hack _setup_blocks_uv_adj_hack |
| 8184d7c5 | 227 | #endif |
| 228 | ||
| 75e28f62 E |
229 | @ r0: psx_gpu |
| 230 | @ r1: v_a | |
| 231 | @ r2: v_b | |
| 232 | @ r3: v_c | |
| 233 | ||
| 234 | function(compute_all_gradients) | |
| 235 | // First compute the triangle area reciprocal and shift. The division will | |
| 236 | // happen concurrently with much of the work which follows. | |
| 237 | @ r12 = psx_gpu->triangle_area | |
| e1f6de8f | 238 | ldr r12, [psx_gpu, #psx_gpu_triangle_area_offset] |
| 75e28f62 | 239 | stmdb sp!, { r4 - r11, lr } |
| 3f0189c6 | 240 | save_abi_regs() |
| 75e28f62 E |
241 | |
| 242 | @ load exponent of 62 into upper half of double | |
| 243 | movw r4, #0 | |
| 244 | clz r14, r12 @ r14 = shift | |
| 245 | ||
| 246 | movt r4, #((62 + 1023) << 4) | |
| 247 | mov r12, r12, lsl r14 @ r12 = triangle_area_normalized | |
| 248 | ||
| 249 | @ load area normalized into lower half of double | |
| 250 | mov r5, r12, lsr #10 | |
| 251 | vmov.f64 d30, r5, r4 @ d30 = (1 << 62) + ta_n | |
| 252 | ||
| 253 | movt r4, #((1022 + 31) << 4) | |
| 254 | mov r5, r12, lsl #20 | |
| 255 | ||
| 256 | add r4, r4, r12, lsr #11 | |
| 257 | vmov.f64 d31, r5, r4 | |
| 258 | ||
| 259 | vdiv.f64 d30, d30, d31 @ d30 = ((1 << 62) + ta_n) / ta_n | |
| 260 | ||
| 261 | // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) = | |
| 262 | // ( d0 * d1 ) - ( d2 * d3 ) = | |
| 263 | // ( m0 ) - ( m1 ) = gradient | |
| 264 | ||
| 265 | // This is split to do 12 elements at a time over three sets: a, b, and c. | |
| 266 | // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so | |
| 267 | // two of the slots are unused. | |
| 268 | ||
| 269 | // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as | |
| 270 | // is g. | |
| 271 | ||
| 272 | // First type is: uvrg bxxx xxxx | |
| 273 | // Second type is: yyyy ybyy uvrg | |
| 274 | // Since x_a and y_c are the same the same variable is used for both. | |
| 275 | ||
| e1f6de8f | 276 | vld1.u32 { v0 }, [v_a, :128] @ v0 = { uvrg0, b0, x0, y0 } |
| 277 | ldrsh x0, [v_a, #8] @ load x0 | |
| 75e28f62 | 278 | |
| e1f6de8f | 279 | vld1.u32 { v1 }, [v_b, :128] @ v1 = { uvrg1, b1, x1, y1} |
| 280 | ldrh x1, [v_b, #8] @ load x1 | |
| 75e28f62 | 281 | |
| e1f6de8f | 282 | vld1.u32 { v2 }, [v_c, :128] @ v2 = { uvrg2, b2, x2, y2 } |
| 283 | ldrh x2, [v_c, #8] @ load x2 | |
| 75e28f62 E |
284 | |
| 285 | vmovl.u8 uvrg_xxxx0, uvrgb0 @ uvrg_xxxx0 = { uv0, rg0, b0-, -- } | |
| e1f6de8f | 286 | ldrh y0, [v_a, #10] @ load y0 |
| 75e28f62 E |
287 | |
| 288 | vmovl.u8 uvrg_xxxx1, uvrgb1 @ uvrg_xxxx1 = { uv1, rg1, b1-, -- } | |
| e1f6de8f | 289 | ldrh y1, [v_b, #10] @ load y1 |
| 75e28f62 E |
290 | |
| 291 | vmovl.u8 uvrg_xxxx2, uvrgb2 @ uvrg_xxxx2 = { uv2, rg2, b2-, -- } | |
| e1f6de8f | 292 | ldrh y2, [v_c, #10] @ load y2 |
| 75e28f62 E |
293 | |
| 294 | vmov.u8 uvrg0b, uvrg0 @ uvrg0b = { uv0, rg0 } | |
| 295 | vdup.u16 xxxx0, x0_y0[0] @ xxxx0 = { xx0, xx0 } | |
| 296 | ||
| 297 | orr x1_x2, x1, x2, lsl #16 @ x1_x2 = { x1, x2 } | |
| 298 | pkhbt x0_x1, x0, x1, lsl #16 @ x0_x1 = { x0, x1 } | |
| 299 | ||
| 300 | vmov.u8 uvrg1b, uvrg1 @ uvrg1b = { uv1, rg1 } | |
| 301 | vdup.u16 xxxx1, x1_y1[0] @ xxxx1 = { xx1, xx1 } | |
| 302 | ||
| 303 | vmov.u8 uvrg2b, uvrg2 @ uvrg2b = { uv2, rg2 } | |
| 304 | vdup.u16 xxxx2, x2_y2[0] @ xxxx2 = { xx2, xx2 } | |
| 305 | ||
| e1f6de8f | 306 | ldrb b2, [v_c, #4] @ load b2 |
| 75e28f62 E |
307 | orr y0_y1, y0, y1, lsl #16 @ y0_y1 = { y0, y1 } |
| 308 | ||
| e1f6de8f | 309 | ldrb b1, [v_b, #4] @ load b1 |
| 75e28f62 E |
310 | orr y1_y2, y1, y2, lsl #16 @ y1_y2 = { y1, y2 } |
| 311 | ||
| 312 | vdup.u16 yyyy0, x0_y0[1] @ yyyy0 = { yy0, yy0 } | |
| 313 | vsub.s16 d0_ab, x1_ab, x0_ab | |
| 314 | ||
| e1f6de8f | 315 | ldrb b0, [v_a, #4] @ load b0 |
| 75e28f62 E |
316 | orr b1_b2, b1, b2, lsl #16 @ b1_b2 = { b1, b2 } |
| 317 | ||
| 318 | vdup.u16 yyyy1, x1_y1[1] @ yyyy1 = { yy1, yy1 } | |
| 319 | vsub.s16 d2_ab, x2_ab, x1_ab | |
| 320 | ||
| 321 | vdup.u16 yyyy2, x2_y2[1] @ yyyy2 = { yy2, yy2 } | |
| 322 | vsub.s16 d1_ab, y2_ab, y1_ab | |
| 323 | ||
| 324 | orr b0_b1, b0, b1, lsl #16 @ b1_b2 = { b1, b2 } | |
| 325 | ssub16 dx, x1_x2, x0_x1 @ dx = { x1 - x0, x2 - x1 } | |
| 326 | ||
| 327 | ssub16 dy, y1_y2, y0_y1 @ dy = { y1 - y0, y2 - y1 } | |
| 328 | ssub16 db, b1_b2, b0_b1 @ db = { b1 - b0, b2 - b1 } | |
| 329 | ||
| 330 | vsub.s16 d3_ab, y1_ab, y0_ab | |
| 331 | smusdx ga_by, dx, db @ ga_by = ((x1 - x0) * (b2 - b1)) - | |
| 332 | @ ((x2 - X1) * (b1 - b0)) | |
| 333 | vmull.s16 ga_uvrg_x, d0_a, d1_a | |
| 334 | smusdx ga_bx, db, dy @ ga_bx = ((b1 - b0) * (y2 - y1)) - | |
| 335 | @ ((b2 - b1) * (y1 - y0)) | |
| 336 | vmlsl.s16 ga_uvrg_x, d2_a, d3_a | |
| 337 | movs gs_bx, ga_bx, asr #31 | |
| 338 | ||
| 339 | vmull.s16 ga_uvrg_y, d0_b, d1_b | |
| 340 | rsbmi ga_bx, ga_bx, #0 | |
| 341 | ||
| c6063f89 | 342 | @ r12 = psx_gpu->uvrgb_phase |
| e1f6de8f | 343 | ldr r12, [psx_gpu, #psx_gpu_uvrgb_phase_offset] |
| c6063f89 | 344 | |
| 75e28f62 E |
345 | vmlsl.s16 ga_uvrg_y, d2_b, d3_b |
| 346 | movs gs_by, ga_by, asr #31 | |
| 347 | ||
| 348 | vshr.u64 d0, d30, #22 | |
| c6063f89 | 349 | add b_base, r12, b0, lsl #16 |
| 350 | ||
| 351 | vdup.u32 uvrgb_phase, r12 | |
| 75e28f62 E |
352 | |
| 353 | rsbmi ga_by, ga_by, #0 | |
| 354 | vclt.s32 gs_uvrg_x, ga_uvrg_x, #0 @ gs_uvrg_x = ga_uvrg_x < 0 | |
| 355 | ||
| 356 | @ r12 = psx_gpu->triangle_winding_offset | |
| e1f6de8f | 357 | ldrb r12, [psx_gpu, #psx_gpu_triangle_winding_offset] |
| 75e28f62 E |
358 | vclt.s32 gs_uvrg_y, ga_uvrg_y, #0 @ gs_uvrg_y = ga_uvrg_y < 0 |
| 359 | ||
| 75e28f62 E |
360 | rsb r12, r12, #0 @ r12 = -(triangle->winding) |
| 361 | ||
| 362 | vdup.u32 w_mask, r12 @ w_mask = { -w, -w, -w, -w } | |
| 363 | sub r14, r14, #(62 - 12) @ r14 = shift - (62 - FIXED_BITS) | |
| 364 | ||
| 365 | vshll.u16 uvrg_base, uvrg0, #16 @ uvrg_base = uvrg0 << 16 | |
| aafce833 | 366 | vdup.u32 r_shift, r14 @ r_shift = { shift, shift*, shift, shift* } |
| 367 | @ * - vshl.u64: ignored by hw | |
| c6063f89 | 368 | vadd.u32 uvrg_base, uvrgb_phase |
| 75e28f62 E |
369 | vabs.s32 ga_uvrg_x, ga_uvrg_x @ ga_uvrg_x = abs(ga_uvrg_x) |
| 370 | ||
| 371 | vmov area_r_s, s0 @ area_r_s = triangle_reciprocal | |
| 372 | vabs.s32 ga_uvrg_y, ga_uvrg_y @ ga_uvrg_y = abs(ga_uvrg_y) | |
| 373 | ||
| 374 | vmull.u32 gw_rg_x, ga_rg_x, d0[0] | |
| 375 | vmull.u32 gw_uv_x, ga_uv_x, d0[0] | |
| 376 | vmull.u32 gw_rg_y, ga_rg_y, d0[0] | |
| 377 | vmull.u32 gw_uv_y, ga_uv_y, d0[0] | |
| 378 | ||
| 379 | vshl.u64 gw_rg_x, gw_rg_x, r_shift | |
| 380 | vshl.u64 gw_uv_x, gw_uv_x, r_shift | |
| 381 | vshl.u64 gw_rg_y, gw_rg_y, r_shift | |
| 382 | vshl.u64 gw_uv_y, gw_uv_y, r_shift | |
| 383 | ||
| 384 | veor.u32 gs_uvrg_x, gs_uvrg_x, w_mask | |
| 385 | vmovn.u64 g_uv_x, gw_uv_x | |
| 386 | ||
| 387 | veor.u32 gs_uvrg_y, gs_uvrg_y, w_mask | |
| 388 | vmovn.u64 g_rg_x, gw_rg_x | |
| 389 | ||
| 390 | veor.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
| 391 | vmovn.u64 g_uv_y, gw_uv_y | |
| 392 | ||
| 393 | vsub.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
| 394 | vmovn.u64 g_rg_y, gw_rg_y | |
| 395 | ||
| 396 | veor.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
| 397 | mov ga_bx, ga_bx, lsl #13 | |
| 398 | ||
| 399 | vsub.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
| 400 | mov ga_by, ga_by, lsl #13 | |
| 401 | ||
| 402 | vdup.u32 x0_y0, x0 | |
| 403 | umull gw_bx_l, gw_bx_h, ga_bx, area_r_s | |
| 404 | ||
| 405 | vshl.u32 g_uvrg_x, g_uvrg_x, #4 | |
| 406 | vshl.u32 g_uvrg_y, g_uvrg_y, #4 | |
| 407 | ||
| 408 | umull gw_by_l, gw_by_h, ga_by, area_r_s | |
| 409 | vmls.s32 uvrg_base, ga_uvrg_x, x0_y0[0] | |
| 410 | ||
| 411 | eor gs_bx, gs_bx, r12 | |
| 412 | vadd.u32 uvrg_dx2, uvrg_dx1, uvrg_dx1 | |
| 413 | ||
| 414 | veor.u32 uvrg_dx0, uvrg_dx0, uvrg_dx0 | |
| 415 | eor gs_by, gs_by, r12 | |
| 416 | ||
| 417 | rsb r11, r14, #0 @ r11 = negative shift for scalar lsr | |
| 418 | add store_a, psx_gpu, #psx_gpu_uvrg_offset | |
| 419 | ||
| 420 | sub r11, r11, #(32 - 13) | |
| 421 | ||
| 422 | add store_b, store_a, #16 | |
| 423 | mov store_inc, #32 | |
| 424 | ||
| 425 | vadd.u32 uvrg_dx3, uvrg_dx2, uvrg_dx1 | |
| e1f6de8f | 426 | vst1.u32 { uvrg_base }, [store_a, :128], store_inc |
| 75e28f62 | 427 | |
| e1f6de8f | 428 | vst1.u32 { uvrg_dx1 }, [store_b, :128], store_inc |
| 75e28f62 E |
429 | mov g_bx, gw_bx_h, lsr r11 |
| 430 | ||
| e1f6de8f | 431 | vst1.u32 { g_uvrg_y }, [store_a, :128], store_inc |
| 75e28f62 E |
432 | mov g_by, gw_by_h, lsr r11 |
| 433 | ||
| 434 | vst4.u32 { uvrg_dx0l, uvrg_dx1l, uvrg_dx2l, uvrg_dx3l }, \ | |
| e1f6de8f | 435 | [store_b, :128], store_inc |
| 75e28f62 E |
436 | eor g_bx, g_bx, gs_bx |
| 437 | ||
| 438 | vst4.u32 { uvrg_dx0h, uvrg_dx1h, uvrg_dx2h, uvrg_dx3h }, \ | |
| e1f6de8f | 439 | [store_b, :128], store_inc |
| 75e28f62 E |
440 | sub g_bx, g_bx, gs_bx |
| 441 | ||
| 442 | lsl g_bx, g_bx, #4 | |
| 443 | eor g_by, g_by, gs_by | |
| 444 | ||
| 445 | mls b_base, g_bx, x0, b_base | |
| 446 | sub g_by, g_by, gs_by | |
| 447 | ||
| 448 | lsl g_by, g_by, #4 | |
| 449 | mov g_bx0, #0 | |
| 450 | ||
| 451 | add g_bx2, g_bx, g_bx | |
| 452 | add g_bx3, g_bx, g_bx2 | |
| 453 | ||
| 454 | stmia store_b, { g_bx0, g_bx, g_bx2, g_bx3, b_base, g_by } | |
| 455 | ||
| 3f0189c6 | 456 | restore_abi_regs() |
| 75e28f62 E |
457 | ldmia sp!, { r4 - r11, pc } |
| 458 | ||
| 459 | ||
| 460 | #define psx_gpu r0 | |
| 461 | #define v_a r1 | |
| 462 | #define v_b r2 | |
| 463 | #define v_c r3 | |
| 464 | ||
| 465 | #define temp r14 | |
| 466 | ||
| 467 | #define x_a r4 | |
| 468 | #define x_b r5 | |
| 469 | #define x_c r6 | |
| 470 | #define y_a r1 | |
| 471 | #define y_b r2 | |
| 472 | #define y_c r3 | |
| 473 | ||
| 474 | #define height_minor_a r7 | |
| 475 | #define height_minor_b r8 | |
| 476 | #define height_major r9 | |
| 477 | #define height r9 | |
| 478 | ||
| 479 | #define reciprocal_table_ptr r10 | |
| 480 | ||
| 481 | #define edge_alt_low r4 | |
| 482 | #define edge_alt_high r5 | |
| 483 | #define edge_dx_dy_alt r6 | |
| 484 | #define edge_shift_alt r10 | |
| 485 | ||
| 486 | #define edge_dx_dy_alt_low r4 | |
| 487 | #define edge_dx_dy_alt_high r5 | |
| 488 | ||
| 489 | #define span_edge_data r4 | |
| 490 | #define span_uvrg_offset r5 | |
| 491 | #define span_b_offset r6 | |
| 492 | ||
| 493 | #define clip r14 | |
| 494 | ||
| 495 | #define b r11 | |
| 496 | #define b_dy r12 | |
| 497 | ||
| 498 | ||
| 499 | #define alternate_x q0 | |
| 500 | #define alternate_dx_dy q1 | |
| 501 | #define alternate_x_32 q2 | |
| 502 | ||
| 503 | #define alternate_x_low d0 | |
| 504 | #define alternate_x_high d1 | |
| 505 | #define alternate_dx_dy_low d2 | |
| 506 | #define alternate_dx_dy_high d3 | |
| 507 | #define alternate_x_32_low d4 | |
| 508 | #define alternate_x_32_high d5 | |
| 509 | ||
| 510 | #define left_x q3 | |
| 511 | #define right_x q4 | |
| 512 | #define left_dx_dy q5 | |
| 513 | #define right_dx_dy q6 | |
| 514 | #define left_edge q7 | |
| 515 | #define right_edge q8 | |
| 516 | ||
| 517 | #define left_x_low d6 | |
| 518 | #define left_x_high d7 | |
| 519 | #define right_x_low d8 | |
| 520 | #define right_x_high d9 | |
| 521 | #define left_dx_dy_low d10 | |
| 522 | #define left_dx_dy_high d11 | |
| 523 | #define right_dx_dy_low d12 | |
| 524 | #define right_dx_dy_high d13 | |
| 525 | #define left_edge_low d14 | |
| 526 | #define left_edge_high d15 | |
| 527 | #define right_edge_low d16 | |
| 528 | #define right_edge_high d17 | |
| 529 | ||
| 530 | #define y_mid_point d18 | |
| 531 | #define c_0x0004 d19 | |
| 532 | ||
| 533 | #define left_right_x_16 q11 | |
| 534 | #define span_shifts_y q12 | |
| 535 | #define c_0x0001 q13 | |
| 536 | ||
| 537 | #define span_shifts d24 | |
| 538 | #define y_x4 d25 | |
| 539 | #define c_0xFFFE d26 | |
| 540 | #define c_0x0007 d27 | |
| 541 | ||
| 542 | #define left_right_x_16_low d22 | |
| 543 | #define left_right_x_16_high d23 | |
| 544 | ||
| 545 | #define uvrg q14 | |
| 546 | #define uvrg_dy q15 | |
| ee060c58 | 547 | #define uv d28 |
| 75e28f62 E |
548 | |
| 549 | #define alternate_x_16 d4 | |
| 550 | ||
| 551 | #define v_clip q3 | |
| 552 | #define v_clip_low d6 | |
| 553 | ||
| 554 | #define right_x_32 q10 | |
| 555 | #define left_x_32 q11 | |
| 556 | #define alternate_select d24 | |
| 557 | ||
| 558 | #define right_x_32_low d20 | |
| 559 | #define right_x_32_high d21 | |
| 560 | #define left_x_32_low d22 | |
| 561 | #define left_x_32_high d23 | |
| 562 | ||
| 2d658c89 | 563 | #define tmp_max_blocks d20 |
| 564 | ||
| 75e28f62 E |
565 | #define edges_xy q0 |
| 566 | #define edges_dx_dy d2 | |
| 567 | #define edge_shifts d3 | |
| 568 | #define edge_shifts_64 q2 | |
| 569 | ||
| 570 | #define edges_xy_left d0 | |
| 571 | #define edges_xy_right d1 | |
| 572 | ||
| 573 | #define height_reciprocals d6 | |
| 574 | #define heights d7 | |
| 575 | ||
| 576 | #define widths d8 | |
| 577 | #define c_0x01 d9 | |
| 578 | #define x_starts d10 | |
| 579 | #define x_ends d11 | |
| 580 | ||
| 581 | #define heights_b d12 | |
| 582 | #define edges_dx_dy_64 q10 | |
| 583 | ||
| 584 | #define edges_dx_dy_64_left d20 | |
| 585 | #define edges_dx_dy_64_right d21 | |
| 586 | ||
| 587 | ||
| 588 | #define setup_spans_prologue() \ | |
| 589 | stmdb sp!, { r4 - r11, lr }; \ | |
| 3f0189c6 | 590 | save_abi_regs(); \ |
| 75e28f62 | 591 | \ |
| e1f6de8f | 592 | ldrsh x_a, [v_a, #8]; \ |
| 593 | ldrsh x_b, [v_b, #8]; \ | |
| 594 | ldrsh x_c, [v_c, #8]; \ | |
| 595 | ldrsh y_a, [v_a, #10]; \ | |
| 596 | ldrsh y_b, [v_b, #10]; \ | |
| 597 | ldrsh y_c, [v_c, #10]; \ | |
| 75e28f62 E |
598 | \ |
| 599 | add temp, psx_gpu, #psx_gpu_uvrg_offset; \ | |
| e1f6de8f | 600 | vld1.32 { uvrg }, [temp]; \ |
| 75e28f62 | 601 | add temp, psx_gpu, #psx_gpu_uvrg_dy_offset; \ |
| e1f6de8f | 602 | vld1.32 { uvrg_dy }, [temp]; \ |
| 603 | ldr reciprocal_table_ptr, [psx_gpu, #psx_gpu_reciprocal_table_ptr_offset]; \ | |
| 75e28f62 E |
604 | \ |
| 605 | vmov.u32 c_0x01, #0x01 \ | |
| 606 | ||
| 607 | #define setup_spans_load_b() \ | |
| e1f6de8f | 608 | ldr b, [psx_gpu, #psx_gpu_b_offset]; \ |
| 609 | ldr b_dy, [psx_gpu, #psx_gpu_b_dy_offset] \ | |
| 75e28f62 E |
610 | |
| 611 | #define setup_spans_prologue_b() \ | |
| 612 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 613 | add temp, psx_gpu, #psx_gpu_viewport_start_x_offset; \ | |
| 614 | \ | |
| 615 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 616 | vmov.u16 c_0x0004, #0x0004; \ | |
| 617 | \ | |
| 618 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 619 | vmov.u16 c_0x0001, #0x0001; \ | |
| 620 | \ | |
| e1f6de8f | 621 | vld1.u16 { left_edge_low[], left_edge_high[] }, [temp]; \ |
| 75e28f62 E |
622 | add temp, psx_gpu, #psx_gpu_viewport_end_x_offset; \ |
| 623 | \ | |
| e1f6de8f | 624 | vld1.u16 { right_edge_low[], right_edge_high[] }, [temp]; \ |
| 75e28f62 E |
625 | vadd.u16 right_edge, right_edge, c_0x0001; \ |
| 626 | \ | |
| 627 | vmov.u16 c_0x0007, #0x0007; \ | |
| 628 | vmvn.u16 c_0xFFFE, #0x0001 \ | |
| 629 | ||
| 630 | ||
| 631 | #define compute_edge_delta_x2() \ | |
| e1f6de8f | 632 | ldr temp, [reciprocal_table_ptr, height, lsl #2]; \ |
| 75e28f62 E |
633 | \ |
| 634 | vdup.u32 heights, height; \ | |
| 635 | vsub.u32 widths, x_ends, x_starts; \ | |
| 636 | \ | |
| 637 | vdup.u32 edge_shifts, temp; \ | |
| 638 | vsub.u32 heights_b, heights, c_0x01; \ | |
| 7d5140f5 | 639 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
| 75e28f62 E |
640 | \ |
| 641 | vmla.s32 heights_b, x_starts, heights; \ | |
| 642 | vbic.u16 edge_shifts, #0xE0; \ | |
| 643 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
| 644 | vmull.s32 edges_xy, heights_b, height_reciprocals \ | |
| 645 | ||
| 646 | #define width_alt r6 | |
| 647 | #define height_reciprocal_alt r11 | |
| 648 | #define height_b_alt r12 | |
| 649 | ||
| 650 | #define compute_edge_delta_x3(start_c, height_a, height_b) \ | |
| ed0fd81d | 651 | vmov heights, height_a, height_b; \ |
| e1f6de8f | 652 | ldr temp, [reciprocal_table_ptr, height_a, lsl #2]; \ |
| 75e28f62 | 653 | vmov.u32 edge_shifts[0], temp; \ |
| e1f6de8f | 654 | ldr temp, [reciprocal_table_ptr, height_b, lsl #2]; \ |
| 75e28f62 | 655 | vmov.u32 edge_shifts[1], temp; \ |
| e1f6de8f | 656 | ldr edge_shift_alt, [reciprocal_table_ptr, height_minor_b, lsl #2]; \ |
| 75e28f62 E |
657 | \ |
| 658 | vsub.u32 widths, x_ends, x_starts; \ | |
| 659 | sub width_alt, x_c, start_c; \ | |
| 660 | \ | |
| 661 | vsub.u32 heights_b, heights, c_0x01; \ | |
| 662 | sub height_b_alt, height_minor_b, #1; \ | |
| 663 | \ | |
| 7d5140f5 E |
664 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
| 665 | lsr height_reciprocal_alt, edge_shift_alt, #10; \ | |
| 75e28f62 E |
666 | \ |
| 667 | vmla.s32 heights_b, x_starts, heights; \ | |
| 668 | mla height_b_alt, height_minor_b, start_c, height_b_alt; \ | |
| 669 | \ | |
| 670 | vbic.u16 edge_shifts, #0xE0; \ | |
| 671 | and edge_shift_alt, edge_shift_alt, #0x1F; \ | |
| 672 | \ | |
| 673 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
| 674 | mul edge_dx_dy_alt, width_alt, height_reciprocal_alt; \ | |
| 675 | \ | |
| 676 | vmull.s32 edges_xy, heights_b, height_reciprocals; \ | |
| 677 | smull edge_alt_low, edge_alt_high, height_b_alt, height_reciprocal_alt \ | |
| 678 | ||
| 679 | ||
| 680 | #define setup_spans_adjust_y_up() \ | |
| 681 | vsub.u32 y_x4, y_x4, c_0x0004 \ | |
| 682 | ||
| 683 | #define setup_spans_adjust_y_down() \ | |
| 684 | vadd.u32 y_x4, y_x4, c_0x0004 \ | |
| 685 | ||
| 686 | #define setup_spans_adjust_interpolants_up() \ | |
| 687 | vsub.u32 uvrg, uvrg, uvrg_dy; \ | |
| 688 | sub b, b, b_dy \ | |
| 689 | ||
| 690 | #define setup_spans_adjust_interpolants_down() \ | |
| 691 | vadd.u32 uvrg, uvrg, uvrg_dy; \ | |
| 692 | add b, b, b_dy \ | |
| 693 | ||
| 694 | ||
| 695 | #define setup_spans_clip_interpolants_increment() \ | |
| 696 | mla b, b_dy, clip, b; \ | |
| 697 | vmla.s32 uvrg, uvrg_dy, v_clip \ | |
| 698 | ||
| 699 | #define setup_spans_clip_interpolants_decrement() \ | |
| 700 | mls b, b_dy, clip, b; \ | |
| 701 | vmls.s32 uvrg, uvrg_dy, v_clip \ | |
| 702 | ||
| 703 | #define setup_spans_clip_alternate_yes() \ | |
| 704 | smlal edge_alt_low, edge_alt_high, edge_dx_dy_alt, clip \ | |
| 705 | ||
| 706 | #define setup_spans_clip_alternate_no() \ | |
| 707 | ||
| 708 | #define setup_spans_clip(direction, alternate_active) \ | |
| 709 | vdup.u32 v_clip, clip; \ | |
| 710 | setup_spans_clip_alternate_##alternate_active(); \ | |
| 711 | setup_spans_clip_interpolants_##direction(); \ | |
| 712 | vmlal.s32 edges_xy, edges_dx_dy, v_clip_low \ | |
| 713 | ||
| 714 | ||
| 715 | #define setup_spans_adjust_edges_alternate_no(left_index, right_index) \ | |
| 716 | vmovl.s32 edge_shifts_64, edge_shifts; \ | |
| 717 | vmovl.s32 edges_dx_dy_64, edges_dx_dy; \ | |
| 718 | \ | |
| 719 | vshl.s64 edges_xy, edges_xy, edge_shifts_64; \ | |
| 720 | vshl.s64 edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64; \ | |
| 721 | \ | |
| 722 | vmov left_x_low, edges_xy_##left_index; \ | |
| 723 | vmov right_x_low, edges_xy_##right_index; \ | |
| 724 | \ | |
| 725 | vmov left_dx_dy_low, edges_dx_dy_64_##left_index; \ | |
| 726 | vmov left_dx_dy_high, edges_dx_dy_64_##left_index; \ | |
| 727 | vmov right_dx_dy_low, edges_dx_dy_64_##right_index; \ | |
| 728 | vmov right_dx_dy_high, edges_dx_dy_64_##right_index; \ | |
| 729 | \ | |
| 730 | vadd.u64 left_x_high, left_x_low, left_dx_dy_low; \ | |
| 731 | vadd.u64 right_x_high, right_x_low, right_dx_dy_low; \ | |
| 732 | \ | |
| 733 | vadd.u64 left_dx_dy, left_dx_dy, left_dx_dy; \ | |
| 734 | vadd.u64 right_dx_dy, right_dx_dy, right_dx_dy \ | |
| 735 | ||
| 736 | ||
| 737 | #define setup_spans_adjust_edges_alternate_yes(left_index, right_index) \ | |
| 738 | setup_spans_adjust_edges_alternate_no(left_index, right_index); \ | |
| 739 | \ | |
| 740 | vdup.u16 y_mid_point, y_b; \ | |
| 741 | rsb temp, edge_shift_alt, #32; \ | |
| 742 | \ | |
| 743 | lsl edge_alt_high, edge_alt_high, edge_shift_alt; \ | |
| 744 | orr edge_alt_high, edge_alt_high, edge_alt_low, lsr temp; \ | |
| 745 | lsl edge_alt_low, edge_alt_low, edge_shift_alt; \ | |
| 746 | vmov alternate_x_low, edge_alt_low, edge_alt_high; \ | |
| 747 | \ | |
| 748 | asr edge_dx_dy_alt_high, edge_dx_dy_alt, temp; \ | |
| 749 | lsl edge_dx_dy_alt_low, edge_dx_dy_alt, edge_shift_alt; \ | |
| 750 | vmov alternate_dx_dy_low, edge_dx_dy_alt_low, edge_dx_dy_alt_high; \ | |
| 751 | vmov alternate_dx_dy_high, alternate_dx_dy_low; \ | |
| 752 | \ | |
| 753 | vadd.u64 alternate_x_high, alternate_x_low, alternate_dx_dy_low; \ | |
| 754 | vadd.u64 alternate_dx_dy, alternate_dx_dy, alternate_dx_dy \ | |
| 755 | ||
| 756 | ||
| 757 | #define setup_spans_y_select_up() \ | |
| 758 | vclt.s16 alternate_select, y_x4, y_mid_point \ | |
| 759 | ||
| 760 | #define setup_spans_y_select_down() \ | |
| 761 | vcgt.s16 alternate_select, y_x4, y_mid_point \ | |
| 762 | ||
| 763 | ||
| 764 | #define setup_spans_alternate_select_left() \ | |
| 765 | vbit.u16 left_right_x_16_low, alternate_x_16, alternate_select \ | |
| 766 | ||
| 767 | #define setup_spans_alternate_select_right() \ | |
| 768 | vbit.u16 left_right_x_16_high, alternate_x_16, alternate_select \ | |
| 769 | ||
| 770 | ||
| 771 | #define setup_spans_set_x4_alternate_yes(alternate, direction) \ | |
| 772 | vshrn.s64 alternate_x_32_low, alternate_x, #32; \ | |
| 773 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
| 774 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
| 775 | \ | |
| 776 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
| 777 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 778 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 779 | \ | |
| 780 | vshrn.s64 alternate_x_32_high, alternate_x, #32; \ | |
| 781 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
| 782 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
| 783 | \ | |
| 784 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
| 785 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 786 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 787 | \ | |
| 788 | vmovn.u32 alternate_x_16, alternate_x_32; \ | |
| 789 | setup_spans_y_select_##direction(); \ | |
| 790 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
| 791 | \ | |
| 792 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
| 793 | setup_spans_alternate_select_##alternate(); \ | |
| 794 | \ | |
| e1f6de8f | 795 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 796 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
797 | setup_spans_adjust_interpolants_##direction(); \ |
| 798 | \ | |
| 799 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
| 800 | \ | |
| e1f6de8f | 801 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 802 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
803 | setup_spans_adjust_interpolants_##direction(); \ |
| 804 | \ | |
| 805 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
| 806 | \ | |
| e1f6de8f | 807 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 808 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
809 | setup_spans_adjust_interpolants_##direction(); \ |
| 810 | \ | |
| 811 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
| 812 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
| 813 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
| 814 | \ | |
| e1f6de8f | 815 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 816 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
817 | setup_spans_adjust_interpolants_##direction(); \ |
| 818 | \ | |
| 2d658c89 | 819 | vmov.u16 tmp_max_blocks, #MAX_BLOCKS_PER_ROW; \ |
| 75e28f62 E |
820 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ |
| 821 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
| 2d658c89 | 822 | vmin.u16 left_right_x_16_high, left_right_x_16_high, tmp_max_blocks; \ |
| 75e28f62 | 823 | \ |
| e1f6de8f | 824 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
| 75e28f62 E |
825 | \ |
| 826 | setup_spans_adjust_y_##direction() \ | |
| 827 | ||
| 828 | ||
| 829 | #define setup_spans_set_x4_alternate_no(alternate, direction) \ | |
| 830 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
| 831 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
| 832 | \ | |
| 833 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 834 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 835 | \ | |
| 836 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
| 837 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
| 838 | \ | |
| 839 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 840 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 841 | \ | |
| 842 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
| 843 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
| 844 | \ | |
| e1f6de8f | 845 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 846 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
847 | setup_spans_adjust_interpolants_##direction(); \ |
| 848 | \ | |
| 849 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
| 850 | \ | |
| e1f6de8f | 851 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 852 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
853 | setup_spans_adjust_interpolants_##direction(); \ |
| 854 | \ | |
| 855 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
| 856 | \ | |
| e1f6de8f | 857 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 858 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
859 | setup_spans_adjust_interpolants_##direction(); \ |
| 860 | \ | |
| 861 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
| 862 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
| 863 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
| 864 | \ | |
| e1f6de8f | 865 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 866 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
867 | setup_spans_adjust_interpolants_##direction(); \ |
| 868 | \ | |
| 2d658c89 | 869 | vmov.u16 tmp_max_blocks, #MAX_BLOCKS_PER_ROW; \ |
| 75e28f62 | 870 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ |
| 2d658c89 | 871 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ |
| 872 | vmin.u16 left_right_x_16_high, left_right_x_16_high, tmp_max_blocks; \ | |
| 75e28f62 | 873 | \ |
| e1f6de8f | 874 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
| 75e28f62 E |
875 | \ |
| 876 | setup_spans_adjust_y_##direction() \ | |
| 877 | ||
| 878 | ||
| 879 | #define edge_adjust_low r11 | |
| 880 | #define edge_adjust_high r12 | |
| 881 | ||
| 882 | #define setup_spans_alternate_adjust_yes() \ | |
| 883 | smull edge_adjust_low, edge_adjust_high, edge_dx_dy_alt, height_minor_a; \ | |
| 884 | subs edge_alt_low, edge_alt_low, edge_adjust_low; \ | |
| 885 | sbc edge_alt_high, edge_alt_high, edge_adjust_high \ | |
| 886 | ||
| 887 | #define setup_spans_alternate_adjust_no() \ | |
| 888 | ||
| 889 | ||
| 890 | #define setup_spans_down(left_index, right_index, alternate, alternate_active) \ | |
| 891 | setup_spans_alternate_adjust_##alternate_active(); \ | |
| 892 | setup_spans_load_b(); \ | |
| 893 | \ | |
| e1f6de8f | 894 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
| 75e28f62 E |
895 | subs y_c, y_c, temp; \ |
| 896 | subgt height, height, y_c; \ | |
| 897 | addgt height, height, #1; \ | |
| 898 | \ | |
| e1f6de8f | 899 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
| 75e28f62 E |
900 | subs clip, temp, y_a; \ |
| 901 | ble 0f; \ | |
| 902 | \ | |
| 903 | sub height, height, clip; \ | |
| 904 | add y_a, y_a, clip; \ | |
| 905 | setup_spans_clip(increment, alternate_active); \ | |
| 906 | \ | |
| 907 | 0: \ | |
| 908 | cmp height, #0; \ | |
| 909 | ble 1f; \ | |
| 910 | \ | |
| 911 | orr temp, y_a, y_a, lsl #16; \ | |
| 2d658c89 | 912 | cmp height, #512; \ |
| 75e28f62 | 913 | add temp, temp, #(1 << 16); \ |
| 2d658c89 | 914 | movgt height, #512; \ |
| 75e28f62 E |
915 | add y_a, temp, #2; \ |
| 916 | add y_a, y_a, #(2 << 16); \ | |
| ed0fd81d | 917 | vmov y_x4, temp, y_a; \ |
| 75e28f62 E |
918 | \ |
| 919 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
| 920 | right_index); \ | |
| 921 | setup_spans_prologue_b(); \ | |
| 922 | \ | |
| e1f6de8f | 923 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
924 | \ |
| 925 | 2: \ | |
| 926 | setup_spans_set_x4_alternate_##alternate_active(alternate, down); \ | |
| 927 | subs height, height, #4; \ | |
| 928 | bhi 2b; \ | |
| 929 | \ | |
| ee060c58 | 930 | nop; \ |
| 931 | ldr temp, [psx_gpu, #psx_gpu_hacks_active_offset]; \ | |
| 932 | tst temp, #(AHACK_TEXTURE_ADJ_U | AHACK_TEXTURE_ADJ_V); \ | |
| 933 | beq 1f; \ | |
| 934 | add temp, span_uvrg_offset, height, lsl #4; \ | |
| 935 | vldr uv, [temp, #(-16*2)]; \ | |
| 936 | vstr uv, [temp, #(-16)]; \ | |
| 937 | \ | |
| 75e28f62 E |
938 | 1: \ |
| 939 | ||
| 940 | ||
| 941 | #define setup_spans_alternate_pre_increment_yes() \ | |
| 942 | adds edge_alt_low, edge_alt_low, edge_dx_dy_alt; \ | |
| 943 | adc edge_alt_high, edge_alt_high, edge_dx_dy_alt, asr #31 \ | |
| 944 | ||
| 945 | #define setup_spans_alternate_pre_increment_no() \ | |
| 946 | ||
| 947 | ||
| 948 | #define setup_spans_up_decrement_yes() \ | |
| 949 | suble height, height, #1 \ | |
| 950 | ||
| 951 | #define setup_spans_up_decrement_no() \ | |
| 952 | ||
| 953 | ||
| 954 | #define setup_spans_up(left_index, right_index, alternate, alternate_active) \ | |
| 955 | setup_spans_alternate_adjust_##alternate_active(); \ | |
| 956 | setup_spans_load_b(); \ | |
| 957 | sub y_a, y_a, #1; \ | |
| 958 | \ | |
| e1f6de8f | 959 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
| 75e28f62 E |
960 | subs temp, temp, y_c; \ |
| 961 | subgt height, height, temp; \ | |
| 962 | setup_spans_up_decrement_##alternate_active(); \ | |
| 963 | \ | |
| e1f6de8f | 964 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
| 75e28f62 E |
965 | subs clip, y_a, temp; \ |
| 966 | ble 0f; \ | |
| 967 | \ | |
| 968 | sub height, height, clip; \ | |
| 969 | sub y_a, y_a, clip; \ | |
| 970 | setup_spans_clip(decrement, alternate_active); \ | |
| 971 | \ | |
| 972 | 0: \ | |
| 973 | cmp height, #0; \ | |
| 974 | ble 1f; \ | |
| 975 | \ | |
| 976 | orr temp, y_a, y_a, lsl #16; \ | |
| 2d658c89 | 977 | cmp height, #512; \ |
| 75e28f62 | 978 | sub temp, temp, #(1 << 16); \ |
| 2d658c89 | 979 | movgt height, #512; \ |
| 75e28f62 E |
980 | sub y_a, temp, #2; \ |
| 981 | sub y_a, y_a, #(2 << 16); \ | |
| ed0fd81d | 982 | vmov y_x4, temp, y_a; \ |
| 75e28f62 E |
983 | \ |
| 984 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy; \ | |
| 985 | \ | |
| 986 | setup_spans_alternate_pre_increment_##alternate_active(); \ | |
| 987 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
| 988 | right_index); \ | |
| 989 | setup_spans_adjust_interpolants_up(); \ | |
| 990 | setup_spans_prologue_b(); \ | |
| 991 | \ | |
| e1f6de8f | 992 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
993 | \ |
| 994 | 2: \ | |
| 995 | setup_spans_set_x4_alternate_##alternate_active(alternate, up); \ | |
| 996 | subs height, height, #4; \ | |
| 997 | bhi 2b; \ | |
| 998 | \ | |
| ee060c58 | 999 | nop; \ |
| 1000 | ldr temp, [psx_gpu, #psx_gpu_hacks_active_offset]; \ | |
| 1001 | tst temp, #AHACK_TEXTURE_ADJ_V; \ | |
| 1002 | beq 1f; \ | |
| 1003 | add temp, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 1004 | vldr uv, [temp, #16]; \ | |
| 1005 | vstr uv, [temp, #0]; \ | |
| 1006 | \ | |
| 75e28f62 E |
1007 | 1: \ |
| 1008 | ||
| 1009 | ||
| 1010 | #define setup_spans_epilogue() \ | |
| 3f0189c6 | 1011 | restore_abi_regs(); \ |
| 75e28f62 E |
1012 | ldmia sp!, { r4 - r11, pc } \ |
| 1013 | ||
| 1014 | ||
| 1015 | #define setup_spans_up_up(minor, major) \ | |
| 1016 | setup_spans_prologue(); \ | |
| 1017 | sub height_minor_a, y_a, y_b; \ | |
| 1018 | sub height_minor_b, y_b, y_c; \ | |
| 1019 | sub height, y_a, y_c; \ | |
| 1020 | \ | |
| 1021 | vdup.u32 x_starts, x_a; \ | |
| ed0fd81d | 1022 | vmov x_ends, x_c, x_b; \ |
| 75e28f62 E |
1023 | \ |
| 1024 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
| 1025 | setup_spans_up(major, minor, minor, yes); \ | |
| 1026 | setup_spans_epilogue() \ | |
| 1027 | ||
| 1028 | function(setup_spans_up_left) | |
| 1029 | setup_spans_up_up(left, right) | |
| 1030 | ||
| 1031 | function(setup_spans_up_right) | |
| 1032 | setup_spans_up_up(right, left) | |
| 1033 | ||
| 75e28f62 E |
1034 | #define setup_spans_down_down(minor, major) \ |
| 1035 | setup_spans_prologue(); \ | |
| 1036 | sub height_minor_a, y_b, y_a; \ | |
| 1037 | sub height_minor_b, y_c, y_b; \ | |
| 1038 | sub height, y_c, y_a; \ | |
| 1039 | \ | |
| 1040 | vdup.u32 x_starts, x_a; \ | |
| ed0fd81d | 1041 | vmov x_ends, x_c, x_b; \ |
| 75e28f62 E |
1042 | \ |
| 1043 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
| 1044 | setup_spans_down(major, minor, minor, yes); \ | |
| 1045 | setup_spans_epilogue() \ | |
| 1046 | ||
| 1047 | function(setup_spans_down_left) | |
| 1048 | setup_spans_down_down(left, right) | |
| 1049 | ||
| 1050 | function(setup_spans_down_right) | |
| 1051 | setup_spans_down_down(right, left) | |
| 1052 | ||
| 1053 | ||
| 1054 | #define setup_spans_up_flat() \ | |
| 1055 | sub height, y_a, y_c; \ | |
| 1056 | \ | |
| 1057 | compute_edge_delta_x2(); \ | |
| 1058 | setup_spans_up(left, right, none, no); \ | |
| 1059 | setup_spans_epilogue() \ | |
| 1060 | ||
| 1061 | function(setup_spans_up_a) | |
| 1062 | setup_spans_prologue() | |
| 1063 | ||
| ed0fd81d | 1064 | vmov x_starts, x_a, x_b |
| 75e28f62 E |
1065 | vdup.u32 x_ends, x_c |
| 1066 | ||
| 1067 | setup_spans_up_flat() | |
| 1068 | ||
| 1069 | function(setup_spans_up_b) | |
| 1070 | setup_spans_prologue() | |
| 1071 | ||
| 1072 | vdup.u32 x_starts, x_a | |
| ed0fd81d | 1073 | vmov x_ends, x_b, x_c |
| 75e28f62 E |
1074 | |
| 1075 | setup_spans_up_flat() | |
| 1076 | ||
| 1077 | #define setup_spans_down_flat() \ | |
| 1078 | sub height, y_c, y_a; \ | |
| 1079 | \ | |
| 1080 | compute_edge_delta_x2(); \ | |
| 1081 | setup_spans_down(left, right, none, no); \ | |
| 1082 | setup_spans_epilogue() \ | |
| 1083 | ||
| 1084 | function(setup_spans_down_a) | |
| 1085 | setup_spans_prologue() | |
| 1086 | ||
| ed0fd81d | 1087 | vmov x_starts, x_a, x_b |
| 75e28f62 E |
1088 | vdup.u32 x_ends, x_c |
| 1089 | ||
| 1090 | setup_spans_down_flat() | |
| 1091 | ||
| 1092 | function(setup_spans_down_b) | |
| 1093 | setup_spans_prologue() | |
| 1094 | ||
| 1095 | vdup.u32 x_starts, x_a | |
| ed0fd81d | 1096 | vmov x_ends, x_b, x_c |
| 75e28f62 E |
1097 | |
| 1098 | setup_spans_down_flat() | |
| 1099 | ||
| 1100 | ||
| 1101 | #define middle_y r9 | |
| 1102 | ||
| 1103 | #define edges_xy_b q11 | |
| 1104 | #define edges_dx_dy_b d26 | |
| 1105 | #define edge_shifts_b d27 | |
| 1106 | #define edges_dx_dy_and_shifts_b q13 | |
| 1107 | #define height_increment d20 | |
| 1108 | ||
| 1109 | #define edges_dx_dy_and_shifts q1 | |
| 1110 | ||
| 1111 | #define edges_xy_b_left d22 | |
| 1112 | #define edges_xy_b_right d23 | |
| 1113 | ||
| 1114 | #define setup_spans_up_down_load_edge_set_b() \ | |
| 1115 | vmov edges_xy, edges_xy_b; \ | |
| 1116 | vmov edges_dx_dy_and_shifts, edges_dx_dy_and_shifts_b \ | |
| 1117 | ||
| 1118 | ||
| 1119 | function(setup_spans_up_down) | |
| 1120 | setup_spans_prologue() | |
| 1121 | ||
| 1122 | // s32 middle_y = y_a; | |
| 1123 | sub height_minor_a, y_a, y_b | |
| 1124 | sub height_minor_b, y_c, y_a | |
| 1125 | sub height_major, y_c, y_b | |
| 1126 | ||
| ed0fd81d | 1127 | vmov x_starts, x_a, x_c |
| 75e28f62 E |
1128 | vdup.u32 x_ends, x_b |
| 1129 | ||
| 1130 | compute_edge_delta_x3(x_a, height_minor_a, height_major) | |
| 1131 | ||
| 1132 | mov temp, #0 | |
| ed0fd81d | 1133 | vmov height_increment, temp, height_minor_b |
| 75e28f62 E |
1134 | vmlal.s32 edges_xy, edges_dx_dy, height_increment |
| 1135 | ||
| 1136 | vmov edges_xy_b_left, edge_alt_low, edge_alt_high | |
| 1137 | vmov edges_xy_b_right, edges_xy_right | |
| 1138 | ||
| 1139 | vmov edge_shifts_b, edge_shifts | |
| 1140 | vmov.u32 edge_shifts_b[0], edge_shift_alt | |
| 1141 | ||
| 1142 | vneg.s32 edges_dx_dy_b, edges_dx_dy | |
| 1143 | vmov.u32 edges_dx_dy_b[0], edge_dx_dy_alt | |
| 1144 | ||
| 1145 | mov middle_y, y_a | |
| 1146 | ||
| 1147 | setup_spans_load_b() | |
| 1148 | sub y_a, y_a, #1 | |
| 1149 | ||
| e1f6de8f | 1150 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
| 75e28f62 E |
1151 | subs temp, temp, y_b |
| 1152 | subgt height_minor_a, height_minor_a, temp | |
| 1153 | ||
| e1f6de8f | 1154 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
| 75e28f62 E |
1155 | subs clip, y_a, temp |
| 1156 | ble 0f | |
| 1157 | ||
| 1158 | sub height_minor_a, height_minor_a, clip | |
| 1159 | sub y_a, y_a, clip | |
| 1160 | setup_spans_clip(decrement, no) | |
| 1161 | ||
| 1162 | 0: | |
| 1163 | cmp height_minor_a, #0 | |
| 1164 | ble 3f | |
| 1165 | ||
| 1166 | orr temp, y_a, y_a, lsl #16 | |
| 1167 | sub temp, temp, #(1 << 16) | |
| 1168 | sub y_a, temp, #2 | |
| 1169 | sub y_a, y_a, #(2 << 16) | |
| ed0fd81d | 1170 | vmov y_x4, temp, y_a |
| 75e28f62 E |
1171 | |
| 1172 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy | |
| 1173 | ||
| e1f6de8f | 1174 | strh height_minor_a, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
1175 | |
| 1176 | setup_spans_adjust_edges_alternate_no(left, right); | |
| 1177 | setup_spans_adjust_interpolants_up() | |
| 1178 | setup_spans_up_down_load_edge_set_b() | |
| 1179 | ||
| 1180 | setup_spans_prologue_b() | |
| 1181 | ||
| 1182 | ||
| 1183 | 2: | |
| 1184 | setup_spans_set_x4_alternate_no(none, up) | |
| 1185 | subs height_minor_a, height_minor_a, #4 | |
| 1186 | bhi 2b | |
| 1187 | ||
| 1188 | add span_edge_data, span_edge_data, height_minor_a, lsl #3 | |
| 1189 | add span_uvrg_offset, span_uvrg_offset, height_minor_a, lsl #4 | |
| 1190 | add span_b_offset, span_b_offset, height_minor_a, lsl #2 | |
| 1191 | ||
| 1192 | 4: | |
| 1193 | add temp, psx_gpu, #psx_gpu_uvrg_offset | |
| e1f6de8f | 1194 | vld1.32 { uvrg }, [temp] |
| 75e28f62 E |
1195 | mov y_a, middle_y |
| 1196 | ||
| 1197 | setup_spans_load_b() | |
| 1198 | ||
| e1f6de8f | 1199 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
| 75e28f62 E |
1200 | subs y_c, y_c, temp |
| 1201 | subgt height_minor_b, height_minor_b, y_c | |
| 1202 | addgt height_minor_b, height_minor_b, #1 | |
| 1203 | ||
| e1f6de8f | 1204 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
| 75e28f62 E |
1205 | subs clip, temp, y_a |
| 1206 | ble 0f | |
| 1207 | ||
| 1208 | sub height_minor_b, height_minor_b, clip | |
| 1209 | add y_a, y_a, clip | |
| 1210 | setup_spans_clip(increment, no) | |
| 1211 | ||
| 1212 | 0: | |
| 1213 | cmp height_minor_b, #0 | |
| 1214 | ble 1f | |
| 1215 | ||
| 1216 | orr temp, y_a, y_a, lsl #16 | |
| 1217 | add temp, temp, #(1 << 16) | |
| 1218 | add y_a, temp, #2 | |
| 1219 | add y_a, y_a, #(2 << 16) | |
| ed0fd81d | 1220 | vmov y_x4, temp, y_a |
| 75e28f62 E |
1221 | |
| 1222 | setup_spans_adjust_edges_alternate_no(left, right) | |
| 1223 | ||
| e1f6de8f | 1224 | ldrh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 | 1225 | add temp, temp, height_minor_b |
| b7569147 | 1226 | |
| 1227 | cmp temp, #MAX_SPANS | |
| 1228 | beq 5f | |
| 1229 | ||
| e1f6de8f | 1230 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
1231 | |
| 1232 | 2: | |
| 1233 | setup_spans_set_x4_alternate_no(none, down) | |
| 1234 | subs height_minor_b, height_minor_b, #4 | |
| 1235 | bhi 2b | |
| 1236 | ||
| ee060c58 | 1237 | nop |
| 1238 | ldr temp, [psx_gpu, #psx_gpu_hacks_active_offset] | |
| 1239 | tst temp, #(AHACK_TEXTURE_ADJ_U | AHACK_TEXTURE_ADJ_V) | |
| 1240 | beq 1f | |
| 1241 | add temp, span_uvrg_offset, height, lsl #4 | |
| 1242 | vldr uv, [temp, #(-16*2)] | |
| 1243 | vstr uv, [temp, #(-16)] | |
| 1244 | ||
| 75e28f62 E |
1245 | 1: |
| 1246 | setup_spans_epilogue() | |
| 1247 | ||
| 1248 | 3: | |
| 1249 | setup_spans_up_down_load_edge_set_b() | |
| 1250 | setup_spans_prologue_b() | |
| 1251 | bal 4b | |
| 1252 | ||
| b7569147 | 1253 | 5: |
| 1254 | // FIXME: overflow corner case | |
| 1255 | sub temp, temp, height_minor_b | |
| 1256 | bics height_minor_b, #3 | |
| 1257 | add temp, temp, height_minor_b | |
| e1f6de8f | 1258 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
| b7569147 | 1259 | bne 2b |
| 1260 | bal 1b | |
| 1261 | ||
| 75e28f62 E |
1262 | #undef span_uvrg_offset |
| 1263 | #undef span_edge_data | |
| 1264 | #undef span_b_offset | |
| 1265 | #undef left_x | |
| 1266 | #undef b | |
| 1267 | ||
| 1268 | #define psx_gpu r0 | |
| 1269 | #define num_spans r1 | |
| 1270 | #define span_uvrg_offset r2 | |
| 1271 | #define span_edge_data r3 | |
| 1272 | #define span_b_offset r4 | |
| 1273 | #define b_dx r5 | |
| 1274 | #define span_num_blocks r6 | |
| 1275 | #define y r7 | |
| 1276 | #define left_x r8 | |
| 1277 | #define b r9 | |
| 1278 | #define dither_offset_ptr r10 | |
| 1279 | #define block_ptr_a r11 | |
| 1280 | #define fb_ptr r12 | |
| 1281 | #define num_blocks r14 | |
| 1282 | ||
| 1283 | #define uvrg_dx_ptr r2 | |
| 1284 | #define texture_mask_ptr r3 | |
| ee060c58 | 1285 | #define hacks_active r6 |
| 75e28f62 E |
1286 | #define dither_shift r8 |
| 1287 | #define dither_row r10 | |
| 1288 | ||
| 1289 | #define c_32 r7 | |
| 1290 | #define b_dx4 r8 | |
| 1291 | #define b_dx8 r9 | |
| 1292 | #define block_ptr_b r10 | |
| 1293 | ||
| 1294 | #define block_span_ptr r10 | |
| 1295 | #define right_mask r8 | |
| 1296 | ||
| 1297 | #define color r2 | |
| 1298 | #define color_r r3 | |
| 1299 | #define color_g r4 | |
| 1300 | #define color_b r5 | |
| 1301 | ||
| 1302 | #undef uvrg | |
| ee060c58 | 1303 | #undef uv |
| 75e28f62 E |
1304 | |
| 1305 | #define u_block q0 | |
| 1306 | #define v_block q1 | |
| 1307 | #define r_block q2 | |
| 1308 | #define g_block q3 | |
| 1309 | #define b_block q4 | |
| 1310 | ||
| 1311 | #define uv_dx4 d10 | |
| 1312 | #define rg_dx4 d11 | |
| 1313 | #define uv_dx8 d12 | |
| 1314 | #define rg_dx8 d13 | |
| 1315 | #define b_whole_8 d14 | |
| 1316 | #define fb_mask_ptrs d15 | |
| 1317 | ||
| 1318 | #define uvrg_dx4 q5 | |
| 1319 | #define uvrg_dx8 q6 | |
| 1320 | #define uv_dx8 d12 | |
| 1321 | #define rg_dx8 d13 | |
| 1322 | ||
| 1323 | #define u_whole q8 | |
| 1324 | #define v_whole q9 | |
| 1325 | #define r_whole q10 | |
| 1326 | #define g_whole q11 | |
| 1327 | #define b_whole q12 | |
| 1328 | ||
| 1329 | #define u_whole_low d16 | |
| 1330 | #define u_whole_high d17 | |
| 1331 | #define v_whole_low d18 | |
| 1332 | #define v_whole_high d19 | |
| 1333 | #define r_whole_low d20 | |
| 1334 | #define r_whole_high d21 | |
| 1335 | #define g_whole_low d22 | |
| 1336 | #define g_whole_high d23 | |
| 1337 | #define b_whole_low d24 | |
| 1338 | #define b_whole_high d25 | |
| 1339 | ||
| 1340 | #define dx4 q13 | |
| 1341 | #define dx8 q13 | |
| 1342 | ||
| 1343 | #define u_whole_8 d26 | |
| 1344 | #define v_whole_8 d27 | |
| 1345 | #define u_whole_8b d24 | |
| 1346 | #define r_whole_8 d24 | |
| 1347 | #define g_whole_8 d25 | |
| 1348 | ||
| 1349 | #define uv_whole_8 q13 | |
| 1350 | #define uv_whole_8b q14 | |
| 1351 | ||
| 1352 | #define dither_offsets q14 | |
| 1353 | #define texture_mask q15 | |
| 1354 | #define texture_mask_u d30 | |
| 1355 | #define texture_mask_v d31 | |
| 1356 | ||
| 1357 | #define dither_offsets_short d28 | |
| 1358 | ||
| 1359 | #define v_left_x q8 | |
| 1360 | #define uvrg q9 | |
| 1361 | #define block_span q10 | |
| 1362 | ||
| 1363 | #define uv d18 | |
| 1364 | #define rg d19 | |
| 1365 | ||
| 1366 | #define draw_mask q1 | |
| 1367 | #define draw_mask_edge q13 | |
| 1368 | #define test_mask q0 | |
| 1369 | ||
| 1370 | #define uvrg_dx q3 | |
| 1371 | ||
| 1372 | #define colors q2 | |
| 1373 | ||
| 1374 | #define setup_blocks_texture_swizzled() \ | |
| 1375 | vand.u8 u_whole_8b, u_whole_8, texture_mask_u; \ | |
| 1376 | vsli.u8 u_whole_8, v_whole_8, #4; \ | |
| 1377 | vsri.u8 v_whole_8, u_whole_8b, #4 \ | |
| 1378 | ||
| 1379 | #define setup_blocks_texture_unswizzled() \ | |
| 1380 | ||
| ee060c58 | 1381 | #define setup_blocks_uv_adj_hack_textured(hacks_active) \ |
| 1382 | tst hacks_active, #(AHACK_TEXTURE_ADJ_U | AHACK_TEXTURE_ADJ_V); \ | |
| 1383 | beq 91f; \ | |
| ee060c58 | 1384 | \ |
| db2804fb | 1385 | /* pushing odd num of regs here realigns our unaligned stack */ \ |
| 1386 | vstr texture_mask_u, [r0, #psx_gpu_saved_tmp_offset]; \ | |
| 1387 | vstr texture_mask_v, [r0, #psx_gpu_saved_tmp_offset + 8]; \ | |
| 1388 | push { r0 - r4, EXTRA_UNSAVED_REGS r12, r14 }; \ | |
| ee060c58 | 1389 | mov r12, span_uvrg_offset; \ |
| 1390 | sub r1, block_ptr_a, #64; \ | |
| 1391 | mov r2, span_edge_data; \ | |
| 1392 | mov r3, r12; \ | |
| 1393 | bl setup_blocks_uv_adj_hack; /* psx_gpu=r0 */ \ | |
| db2804fb | 1394 | pop { r0 - r4, EXTRA_UNSAVED_REGS r12, r14 }; \ |
| 1395 | vldr texture_mask_u, [r0, #psx_gpu_saved_tmp_offset]; \ | |
| 1396 | vldr texture_mask_v, [r0, #psx_gpu_saved_tmp_offset + 8]; \ | |
| ee060c58 | 1397 | \ |
| ee060c58 | 1398 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ |
| 1399 | 91: \ | |
| 1400 | ||
| 75e28f62 E |
1401 | |
| 1402 | #define setup_blocks_shaded_textured_builder(swizzling) \ | |
| 1403 | .align 3; \ | |
| 1404 | \ | |
| 1405 | function(setup_blocks_shaded_textured_dithered_##swizzling##_indirect) \ | |
| e1f6de8f | 1406 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
1407 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
| 1408 | \ | |
| e1f6de8f | 1409 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
| 75e28f62 E |
1410 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
| 1411 | \ | |
| 1412 | cmp num_spans, #0; \ | |
| 1413 | bxeq lr; \ | |
| 1414 | \ | |
| 1415 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 1416 | save_abi_regs(); \ |
| 75e28f62 E |
1417 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
| 1418 | \ | |
| e1f6de8f | 1419 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
| 75e28f62 E |
1420 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ |
| 1421 | \ | |
| e1f6de8f | 1422 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
| 75e28f62 E |
1423 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
| 1424 | \ | |
| e1f6de8f | 1425 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1426 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
| 1427 | \ | |
| 1428 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 1429 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1430 | \ | |
| 1431 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 1432 | \ | |
| 1433 | 0: \ | |
| 1434 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1435 | \ | |
| e1f6de8f | 1436 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
1437 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 1438 | \ | |
| e1f6de8f | 1439 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 1440 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
1441 | \ |
| 1442 | cmp span_num_blocks, #0; \ | |
| 1443 | beq 1f; \ | |
| 1444 | \ | |
| e1f6de8f | 1445 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
1446 | add num_blocks, span_num_blocks, num_blocks; \ |
| 1447 | \ | |
| 1448 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 1449 | bgt 2f; \ | |
| 1450 | \ | |
| 1451 | 3: \ | |
| e1f6de8f | 1452 | ldr b, [span_b_offset]; \ |
| 75e28f62 E |
1453 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 1454 | \ | |
| 1455 | vdup.u32 v_left_x, left_x; \ | |
| 1456 | and y, y, #0x3; \ | |
| 1457 | \ | |
| e1f6de8f | 1458 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
1459 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 1460 | \ | |
| 1461 | mla b, b_dx, left_x, b; \ | |
| 1462 | and dither_shift, left_x, #0x03; \ | |
| 1463 | \ | |
| e1f6de8f | 1464 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
1465 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
| 1466 | \ | |
| 1467 | mov dither_shift, dither_shift, lsl #3; \ | |
| 1468 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
| 1469 | \ | |
| 1470 | mov c_32, #32; \ | |
| 1471 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1472 | \ | |
| 1473 | mov dither_row, dither_row, ror dither_shift; \ | |
| 1474 | mov b_dx4, b_dx, lsl #2; \ | |
| 1475 | \ | |
| 1476 | vdup.u32 dither_offsets_short, dither_row; \ | |
| 1477 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
| 1478 | \ | |
| 1479 | vdup.u32 b_block, b; \ | |
| 1480 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
| 1481 | \ | |
| 1482 | vdup.u32 u_block, uv[0]; \ | |
| 1483 | mov b_dx8, b_dx, lsl #3; \ | |
| 1484 | \ | |
| 1485 | vdup.u32 v_block, uv[1]; \ | |
| 1486 | vdup.u32 r_block, rg[0]; \ | |
| 1487 | vdup.u32 g_block, rg[1]; \ | |
| 1488 | \ | |
| e1f6de8f | 1489 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1490 | \ |
| 1491 | vadd.u32 u_block, u_block, block_span; \ | |
| e1f6de8f | 1492 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1493 | \ |
| 1494 | vadd.u32 v_block, v_block, block_span; \ | |
| e1f6de8f | 1495 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1496 | \ |
| 1497 | vadd.u32 r_block, r_block, block_span; \ | |
| e1f6de8f | 1498 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1499 | \ |
| 1500 | vadd.u32 g_block, g_block, block_span; \ | |
| e1f6de8f | 1501 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
| 75e28f62 E |
1502 | \ |
| 1503 | vadd.u32 b_block, b_block, block_span; \ | |
| 1504 | add block_ptr_b, block_ptr_a, #16; \ | |
| 1505 | \ | |
| 1506 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1507 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1508 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 1509 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 1510 | \ | |
| 1511 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1512 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 1513 | \ | |
| 1514 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1515 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1516 | \ | |
| 1517 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1518 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 1519 | \ | |
| 1520 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 1521 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 1522 | \ | |
| 1523 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 1524 | vdup.u32 dx4, b_dx4; \ | |
| 1525 | \ | |
| 1526 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 1527 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1528 | \ | |
| 1529 | vadd.u32 u_block, u_block, dx8; \ | |
| 1530 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1531 | \ | |
| 1532 | vadd.u32 v_block, v_block, dx8; \ | |
| 1533 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 1534 | \ | |
| 1535 | vadd.u32 r_block, r_block, dx8; \ | |
| 1536 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 1537 | \ | |
| 1538 | vadd.u32 g_block, g_block, dx8; \ | |
| 1539 | vdup.u32 dx8, b_dx8; \ | |
| 1540 | \ | |
| 1541 | vadd.u32 b_block, b_block, dx8; \ | |
| 1542 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1543 | \ | |
| 1544 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1545 | \ | |
| 1546 | vmovn.u16 b_whole_8, b_whole; \ | |
| e1f6de8f | 1547 | pld [fb_ptr]; \ |
| 75e28f62 E |
1548 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
| 1549 | \ | |
| 1550 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1551 | setup_blocks_texture_##swizzling(); \ | |
| 1552 | \ | |
| 1553 | vmovn.u16 r_whole_8, r_whole; \ | |
| 1554 | beq 5f; \ | |
| 1555 | \ | |
| 1556 | 4: \ | |
| 1557 | vmovn.u16 g_whole_8, g_whole; \ | |
| 1558 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1559 | \ | |
| e1f6de8f | 1560 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1561 | vshrn.u32 v_whole_low, v_block, #16; \ |
| 1562 | \ | |
| e1f6de8f | 1563 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
| 75e28f62 E |
1564 | vshrn.u32 r_whole_low, r_block, #16; \ |
| 1565 | \ | |
| e1f6de8f | 1566 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1567 | vshrn.u32 g_whole_low, g_block, #16; \ |
| 1568 | \ | |
| 1569 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1570 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 1571 | \ | |
| 1572 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1573 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1574 | \ | |
| 1575 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1576 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 1577 | \ | |
| 1578 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 1579 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 1580 | \ | |
| 1581 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 1582 | vdup.u32 dx4, b_dx4; \ | |
| 1583 | \ | |
| 1584 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 1585 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1586 | \ | |
| 1587 | vadd.u32 u_block, u_block, dx8; \ | |
| 1588 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1589 | \ | |
| 1590 | vadd.u32 v_block, v_block, dx8; \ | |
| 1591 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 1592 | \ | |
| 1593 | vadd.u32 r_block, r_block, dx8; \ | |
| 1594 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 1595 | \ | |
| 1596 | vadd.u32 g_block, g_block, dx8; \ | |
| 1597 | vdup.u32 dx8, b_dx8; \ | |
| 1598 | \ | |
| 1599 | vadd.u32 b_block, b_block, dx8; \ | |
| 1600 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1601 | \ | |
| 1602 | add fb_ptr, fb_ptr, #16; \ | |
| 1603 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1604 | \ | |
| e1f6de8f | 1605 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
| 75e28f62 E |
1606 | vmovn.u16 b_whole_8, b_whole; \ |
| 1607 | \ | |
| e1f6de8f | 1608 | pld [fb_ptr]; \ |
| 75e28f62 E |
1609 | \ |
| 1610 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1611 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1612 | \ | |
| 1613 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1614 | setup_blocks_texture_##swizzling(); \ | |
| 1615 | \ | |
| 1616 | vmovn.u16 r_whole_8, r_whole; \ | |
| 1617 | bne 4b; \ | |
| 1618 | \ | |
| 1619 | 5: \ | |
| 1620 | vmovn.u16 g_whole_8, g_whole; \ | |
| e1f6de8f | 1621 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 | 1622 | \ |
| e1f6de8f | 1623 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
1624 | vdup.u8 draw_mask, right_mask; \ |
| 1625 | \ | |
| ee060c58 | 1626 | ldr hacks_active, [psx_gpu, #psx_gpu_hacks_active_offset]; \ |
| 75e28f62 E |
1627 | vmov.u32 fb_mask_ptrs[0], right_mask; \ |
| 1628 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 1629 | vzip.u8 u_whole_8, v_whole_8; \ | |
| 1630 | \ | |
| 1631 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
| e1f6de8f | 1632 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
| 1633 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ | |
| 1634 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
| 1635 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
| 75e28f62 | 1636 | \ |
| ee060c58 | 1637 | setup_blocks_uv_adj_hack_textured(hacks_active); \ |
| 1638 | \ | |
| 75e28f62 E |
1639 | 1: \ |
| 1640 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 1641 | add span_b_offset, span_b_offset, #4; \ | |
| 1642 | \ | |
| 1643 | add span_edge_data, span_edge_data, #8; \ | |
| 1644 | subs num_spans, num_spans, #1; \ | |
| 1645 | \ | |
| e1f6de8f | 1646 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1647 | bne 0b; \ |
| 1648 | \ | |
| 3f0189c6 | 1649 | restore_abi_regs(); \ |
| 75e28f62 E |
1650 | ldmia sp!, { r4 - r11, pc }; \ |
| 1651 | \ | |
| 1652 | 2: \ | |
| db2804fb | 1653 | vstr texture_mask_u, [r0, #psx_gpu_saved_tmp_offset]; \ |
| 1654 | vstr texture_mask_v, [r0, #psx_gpu_saved_tmp_offset + 8]; \ | |
| 1655 | /* pushing odd num of regs here realigns our unaligned stack */ \ | |
| 1656 | push { r0 - r3, EXTRA_UNSAVED_REGS r12 }; /* r14=num_blocks */ \ | |
| 75e28f62 | 1657 | bl flush_render_block_buffer; \ |
| db2804fb | 1658 | pop { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ |
| 1659 | vldr texture_mask_u, [r0, #psx_gpu_saved_tmp_offset]; \ | |
| 1660 | vldr texture_mask_v, [r0, #psx_gpu_saved_tmp_offset + 8]; \ | |
| 75e28f62 E |
1661 | \ |
| 1662 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
| 1663 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1664 | \ | |
| 1665 | mov num_blocks, span_num_blocks; \ | |
| 1666 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1667 | bal 3b \ | |
| 1668 | ||
| 1669 | ||
| 1670 | setup_blocks_shaded_textured_builder(swizzled) | |
| 1671 | setup_blocks_shaded_textured_builder(unswizzled) | |
| 1672 | ||
| 1673 | ||
| 1674 | #define setup_blocks_unshaded_textured_builder(swizzling) \ | |
| 1675 | .align 3; \ | |
| 1676 | \ | |
| 1677 | function(setup_blocks_unshaded_textured_dithered_##swizzling##_indirect) \ | |
| e1f6de8f | 1678 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
1679 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
| 1680 | \ | |
| e1f6de8f | 1681 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
| 75e28f62 E |
1682 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
| 1683 | \ | |
| 1684 | cmp num_spans, #0; \ | |
| 1685 | bxeq lr; \ | |
| 1686 | \ | |
| 1687 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 1688 | save_abi_regs(); \ |
| 75e28f62 E |
1689 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
| 1690 | \ | |
| 1691 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
| 1692 | \ | |
| e1f6de8f | 1693 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
| 75e28f62 E |
1694 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
| 1695 | \ | |
| e1f6de8f | 1696 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1697 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
| 1698 | \ | |
| 1699 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1700 | \ | |
| 1701 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 1702 | \ | |
| 1703 | 0: \ | |
| 1704 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1705 | \ | |
| e1f6de8f | 1706 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
1707 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 1708 | \ | |
| e1f6de8f | 1709 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 1710 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
1711 | \ |
| 1712 | cmp span_num_blocks, #0; \ | |
| 1713 | beq 1f; \ | |
| 1714 | \ | |
| e1f6de8f | 1715 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
1716 | add num_blocks, span_num_blocks, num_blocks; \ |
| 1717 | \ | |
| 1718 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 1719 | bgt 2f; \ | |
| 1720 | \ | |
| 1721 | 3: \ | |
| 1722 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
| 1723 | \ | |
| 1724 | vdup.u32 v_left_x, left_x; \ | |
| 1725 | and y, y, #0x3; \ | |
| 1726 | \ | |
| e1f6de8f | 1727 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
1728 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 1729 | \ | |
| 1730 | and dither_shift, left_x, #0x03; \ | |
| 1731 | \ | |
| e1f6de8f | 1732 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
1733 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
| 1734 | \ | |
| 1735 | mov dither_shift, dither_shift, lsl #3; \ | |
| 1736 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
| 1737 | \ | |
| 1738 | mov c_32, #32; \ | |
| 1739 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1740 | \ | |
| 1741 | mov dither_row, dither_row, ror dither_shift; \ | |
| 1742 | \ | |
| 1743 | vdup.u32 dither_offsets_short, dither_row; \ | |
| 1744 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
| 1745 | \ | |
| 1746 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
| 1747 | \ | |
| 1748 | vdup.u32 u_block, uv[0]; \ | |
| 1749 | \ | |
| 1750 | vdup.u32 v_block, uv[1]; \ | |
| e1f6de8f | 1751 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1752 | \ |
| 1753 | vadd.u32 u_block, u_block, block_span; \ | |
| e1f6de8f | 1754 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1755 | \ |
| 1756 | vadd.u32 v_block, v_block, block_span; \ | |
| 1757 | add block_ptr_b, block_ptr_a, #16; \ | |
| 1758 | \ | |
| 1759 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1760 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1761 | \ | |
| 1762 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1763 | \ | |
| 1764 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1765 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1766 | \ | |
| 1767 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1768 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1769 | \ | |
| 1770 | vadd.u32 u_block, u_block, dx8; \ | |
| 1771 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1772 | \ | |
| 1773 | vadd.u32 v_block, v_block, dx8; \ | |
| 1774 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1775 | \ | |
| 1776 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1777 | \ | |
| e1f6de8f | 1778 | pld [fb_ptr]; \ |
| 75e28f62 E |
1779 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
| 1780 | \ | |
| 1781 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1782 | setup_blocks_texture_##swizzling(); \ | |
| 1783 | \ | |
| 1784 | beq 5f; \ | |
| 1785 | \ | |
| 1786 | 4: \ | |
| 1787 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1788 | \ | |
| e1f6de8f | 1789 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1790 | vshrn.u32 v_whole_low, v_block, #16; \ |
| 1791 | \ | |
| 1792 | add block_ptr_b, block_ptr_b, #32; \ | |
| e1f6de8f | 1793 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1794 | \ |
| 1795 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1796 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1797 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1798 | \ | |
| 1799 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1800 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1801 | \ | |
| 1802 | vadd.u32 u_block, u_block, dx8; \ | |
| 1803 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1804 | \ | |
| 1805 | vadd.u32 v_block, v_block, dx8; \ | |
| 1806 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1807 | \ | |
| 1808 | add fb_ptr, fb_ptr, #16; \ | |
| 1809 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1810 | \ | |
| e1f6de8f | 1811 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
| 1812 | pld [fb_ptr]; \ | |
| 75e28f62 E |
1813 | \ |
| 1814 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1815 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1816 | \ | |
| 1817 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1818 | setup_blocks_texture_##swizzling(); \ | |
| 1819 | \ | |
| 1820 | bne 4b; \ | |
| 1821 | \ | |
| 1822 | 5: \ | |
| e1f6de8f | 1823 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 | 1824 | \ |
| e1f6de8f | 1825 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
1826 | vdup.u8 draw_mask, right_mask; \ |
| 1827 | \ | |
| ee060c58 | 1828 | ldr hacks_active, [psx_gpu, #psx_gpu_hacks_active_offset]; \ |
| 75e28f62 E |
1829 | vmov.u32 fb_mask_ptrs[0], right_mask; \ |
| 1830 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 1831 | vzip.u8 u_whole_8, v_whole_8; \ | |
| 1832 | \ | |
| 1833 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
| 1834 | add block_ptr_b, block_ptr_b, #32; \ | |
| e1f6de8f | 1835 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ |
| 1836 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
| 1837 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
| 75e28f62 | 1838 | \ |
| ee060c58 | 1839 | setup_blocks_uv_adj_hack_textured(hacks_active); \ |
| 1840 | \ | |
| 75e28f62 E |
1841 | 1: \ |
| 1842 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 1843 | add span_edge_data, span_edge_data, #8; \ | |
| 1844 | subs num_spans, num_spans, #1; \ | |
| 1845 | \ | |
| e1f6de8f | 1846 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1847 | bne 0b; \ |
| 1848 | \ | |
| 3f0189c6 | 1849 | restore_abi_regs(); \ |
| 75e28f62 E |
1850 | ldmia sp!, { r4 - r11, pc }; \ |
| 1851 | \ | |
| 1852 | 2: \ | |
| db2804fb | 1853 | vstr texture_mask_u, [r0, #psx_gpu_saved_tmp_offset]; \ |
| 1854 | vstr texture_mask_v, [r0, #psx_gpu_saved_tmp_offset + 8]; \ | |
| 1855 | push { r0 - r3, EXTRA_UNSAVED_REGS r12 }; /* r14=num_blocks */ \ | |
| 75e28f62 | 1856 | bl flush_render_block_buffer; \ |
| db2804fb | 1857 | pop { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ |
| 1858 | vldr texture_mask_u, [r0, #psx_gpu_saved_tmp_offset]; \ | |
| 1859 | vldr texture_mask_v, [r0, #psx_gpu_saved_tmp_offset + 8]; \ | |
| 75e28f62 E |
1860 | \ |
| 1861 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
| 1862 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1863 | \ | |
| 1864 | mov num_blocks, span_num_blocks; \ | |
| 1865 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1866 | bal 3b \ | |
| 1867 | ||
| 1868 | ||
| 1869 | setup_blocks_unshaded_textured_builder(swizzled) | |
| 1870 | setup_blocks_unshaded_textured_builder(unswizzled) | |
| 1871 | ||
| 1872 | ||
| 1873 | .align 3 | |
| 1874 | ||
| 1875 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_indirect) | |
| e1f6de8f | 1876 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
1877 | veor.u32 draw_mask, draw_mask, draw_mask |
| 1878 | ||
| 1879 | cmp num_spans, #0 | |
| 1880 | bxeq lr | |
| 1881 | ||
| 1882 | stmdb sp!, { r4 - r11, r14 } | |
| 3f0189c6 | 1883 | save_abi_regs() |
| e1f6de8f | 1884 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 | 1885 | |
| e1f6de8f | 1886 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
| 75e28f62 E |
1887 | |
| 1888 | ubfx color_r, color, #3, #5 | |
| 1889 | ubfx color_g, color, #11, #5 | |
| 1890 | ubfx color_b, color, #19, #5 | |
| 1891 | ||
| 1892 | orr color, color_r, color_b, lsl #10 | |
| 1893 | orr color, color, color_g, lsl #5 | |
| 1894 | ||
| 1895 | vdup.u16 colors, color | |
| 1896 | ||
| e1f6de8f | 1897 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
1898 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset |
| 1899 | ||
| 1900 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
| 1901 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6 | |
| 1902 | ||
| 1903 | 0: | |
| e1f6de8f | 1904 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
| 1905 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
| 75e28f62 | 1906 | |
| e1f6de8f | 1907 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 75e28f62 E |
1908 | |
| 1909 | cmp span_num_blocks, #0 | |
| 1910 | beq 1f | |
| 1911 | ||
| e1f6de8f | 1912 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
| 75e28f62 E |
1913 | add num_blocks, span_num_blocks, num_blocks |
| 1914 | ||
| 1915 | cmp num_blocks, #MAX_BLOCKS | |
| 1916 | bgt 2f | |
| 1917 | ||
| 1918 | 3: | |
| 1919 | add fb_ptr, fb_ptr, y, lsl #11 | |
| 1920 | and y, y, #0x3 | |
| 1921 | ||
| 1922 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
| 1923 | mov c_32, #32 | |
| 1924 | ||
| 1925 | subs span_num_blocks, span_num_blocks, #1 | |
| 1926 | ||
| 1927 | add block_ptr_b, block_ptr_a, #16 | |
| e1f6de8f | 1928 | pld [fb_ptr] |
| 75e28f62 E |
1929 | |
| 1930 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
| 1931 | beq 5f | |
| 1932 | ||
| 1933 | 4: | |
| e1f6de8f | 1934 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_32 |
| 1935 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 | |
| 1936 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 | |
| 75e28f62 E |
1937 | |
| 1938 | add fb_ptr, fb_ptr, #16 | |
| 1939 | add block_ptr_b, block_ptr_b, #32 | |
| 1940 | ||
| e1f6de8f | 1941 | pld [fb_ptr] |
| 75e28f62 E |
1942 | |
| 1943 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
| 1944 | subs span_num_blocks, span_num_blocks, #1 | |
| 1945 | ||
| 1946 | bne 4b | |
| 1947 | ||
| 1948 | 5: | |
| e1f6de8f | 1949 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset] |
| 75e28f62 E |
1950 | |
| 1951 | vdup.u8 draw_mask_edge, right_mask | |
| 1952 | vtst.u16 draw_mask_edge, draw_mask_edge, test_mask | |
| 1953 | ||
| e1f6de8f | 1954 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 |
| 1955 | vst1.u32 { draw_mask_edge }, [block_ptr_a, :128], c_32 | |
| 75e28f62 | 1956 | add block_ptr_b, block_ptr_b, #32 |
| e1f6de8f | 1957 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 |
| 75e28f62 E |
1958 | |
| 1959 | 1: | |
| 1960 | add span_edge_data, span_edge_data, #8 | |
| 1961 | subs num_spans, num_spans, #1 | |
| 1962 | ||
| e1f6de8f | 1963 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
1964 | bne 0b |
| 1965 | ||
| 3f0189c6 | 1966 | restore_abi_regs() |
| 75e28f62 E |
1967 | ldmia sp!, { r4 - r11, pc } |
| 1968 | ||
| 1969 | 2: | |
| db2804fb | 1970 | vstr d4, [r0, #psx_gpu_saved_tmp_offset] /* colors */ |
| 1971 | vstr d5, [r0, #psx_gpu_saved_tmp_offset + 8] | |
| 1972 | push { r0 - r3, EXTRA_UNSAVED_REGS r12 } | |
| 75e28f62 | 1973 | bl flush_render_block_buffer |
| db2804fb | 1974 | pop { r0 - r3, EXTRA_UNSAVED_REGS r12 } |
| 1975 | vldr d4, [r0, #psx_gpu_saved_tmp_offset] | |
| 1976 | vldr d5, [r0, #psx_gpu_saved_tmp_offset + 8] | |
| 75e28f62 | 1977 | |
| e1f6de8f | 1978 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 E |
1979 | veor.u32 draw_mask, draw_mask, draw_mask |
| 1980 | ||
| 1981 | mov num_blocks, span_num_blocks | |
| 1982 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
| 1983 | bal 3b | |
| 1984 | ||
| 1985 | ||
| 1986 | #define mask_msb_scalar r14 | |
| 1987 | ||
| 1988 | #define msb_mask q15 | |
| 1989 | ||
| 1990 | #define pixels_low d16 | |
| 1991 | ||
| 1992 | #define msb_mask_low d30 | |
| 1993 | #define msb_mask_high d31 | |
| 1994 | ||
| 1995 | ||
| 1996 | .align 3 | |
| 1997 | ||
| 1998 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_direct) | |
| e1f6de8f | 1999 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
2000 | |
| 2001 | cmp num_spans, #0 | |
| 2002 | bxeq lr | |
| 2003 | ||
| 2004 | stmdb sp!, { r4 - r11, r14 } | |
| 2005 | ||
| e1f6de8f | 2006 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
| 75e28f62 E |
2007 | |
| 2008 | ubfx color_r, color, #3, #5 | |
| 2009 | ubfx color_g, color, #11, #5 | |
| 2010 | ||
| e1f6de8f | 2011 | ldrh mask_msb_scalar, [psx_gpu, #psx_gpu_mask_msb_offset] |
| 75e28f62 E |
2012 | ubfx color_b, color, #19, #5 |
| 2013 | ||
| 2014 | orr color, color_r, color_b, lsl #10 | |
| 2015 | orr color, color, color_g, lsl #5 | |
| 2016 | orr color, color, mask_msb_scalar | |
| 2017 | ||
| 2018 | vdup.u16 colors, color | |
| 2019 | ||
| 2020 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
| ed0fd81d | 2021 | orr color, color, color, lsl #16 |
| 3867c6ef | 2022 | |
| 75e28f62 E |
2023 | |
| 2024 | 0: | |
| e1f6de8f | 2025 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
| 2026 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
| 75e28f62 | 2027 | |
| e1f6de8f | 2028 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 75e28f62 E |
2029 | |
| 2030 | cmp span_num_blocks, #0 | |
| 2031 | beq 1f | |
| 2032 | ||
| e1f6de8f | 2033 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
| 75e28f62 E |
2034 | |
| 2035 | add fb_ptr, fb_ptr, y, lsl #11 | |
| 2036 | subs span_num_blocks, span_num_blocks, #1 | |
| 2037 | ||
| 2038 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
| 2039 | beq 3f | |
| 2040 | ||
| 2041 | 2: | |
| e1f6de8f | 2042 | vst1.u32 { colors }, [fb_ptr]! |
| 75e28f62 E |
2043 | subs span_num_blocks, span_num_blocks, #1 |
| 2044 | ||
| 2045 | bne 2b | |
| 2046 | ||
| 2047 | 3: | |
| e1f6de8f | 2048 | ldrb right_mask, [span_edge_data, #edge_data_right_mask_offset] |
| 75e28f62 | 2049 | |
| 3867c6ef E |
2050 | cmp right_mask, #0x0 |
| 2051 | beq 5f | |
| 2052 | ||
| 2053 | tst right_mask, #0xF | |
| e1f6de8f | 2054 | streq color, [fb_ptr], #4 |
| 3867c6ef | 2055 | moveq right_mask, right_mask, lsr #4 |
| e1f6de8f | 2056 | streq color, [fb_ptr], #4 |
| 3867c6ef E |
2057 | |
| 2058 | tst right_mask, #0x3 | |
| e1f6de8f | 2059 | streq color, [fb_ptr], #4 |
| 3867c6ef E |
2060 | moveq right_mask, right_mask, lsr #2 |
| 2061 | ||
| 2062 | tst right_mask, #0x1 | |
| e1f6de8f | 2063 | strheq color, [fb_ptr] |
| 75e28f62 E |
2064 | |
| 2065 | 1: | |
| 2066 | add span_edge_data, span_edge_data, #8 | |
| 2067 | subs num_spans, num_spans, #1 | |
| 75e28f62 E |
2068 | bne 0b |
| 2069 | ||
| 2070 | ldmia sp!, { r4 - r11, pc } | |
| 2071 | ||
| 3867c6ef | 2072 | 5: |
| e1f6de8f | 2073 | vst1.u32 { colors }, [fb_ptr] |
| 3867c6ef | 2074 | bal 1b |
| 75e28f62 E |
2075 | |
| 2076 | ||
| 2077 | #undef c_64 | |
| 2078 | ||
| 2079 | #define c_64 r7 | |
| 2080 | #define rg_dx_ptr r2 | |
| 2081 | ||
| 2082 | ||
| 2083 | #undef r_block | |
| 2084 | #undef g_block | |
| 2085 | #undef b_block | |
| 2086 | #undef r_whole | |
| 2087 | #undef g_whole | |
| 2088 | #undef b_whole | |
| 2089 | #undef r_whole_low | |
| 2090 | #undef r_whole_high | |
| 2091 | #undef g_whole_low | |
| 2092 | #undef g_whole_high | |
| 2093 | #undef b_whole_low | |
| 2094 | #undef b_whole_high | |
| 2095 | #undef r_whole_8 | |
| 2096 | #undef g_whole_8 | |
| 2097 | #undef b_whole_8 | |
| 2098 | #undef dither_offsets | |
| 2099 | #undef rg_dx4 | |
| 2100 | #undef rg_dx8 | |
| 2101 | #undef dx4 | |
| 2102 | #undef dx8 | |
| 2103 | #undef v_left_x | |
| 2104 | #undef uvrg | |
| 2105 | #undef block_span | |
| 2106 | #undef rg | |
| 2107 | #undef draw_mask | |
| 2108 | #undef test_mask | |
| 2109 | ||
| 2110 | #define r_block q0 | |
| 2111 | #define g_block q1 | |
| 2112 | #define b_block q2 | |
| 2113 | ||
| 2114 | #define r_whole q3 | |
| 2115 | #define g_whole q4 | |
| 2116 | #define b_whole q5 | |
| 2117 | ||
| 2118 | #define r_whole_low d6 | |
| 2119 | #define r_whole_high d7 | |
| 2120 | #define g_whole_low d8 | |
| 2121 | #define g_whole_high d9 | |
| 2122 | #define b_whole_low d10 | |
| 2123 | #define b_whole_high d11 | |
| 2124 | ||
| 2125 | #define gb_whole_8 q6 | |
| 2126 | ||
| 2127 | #define g_whole_8 d12 | |
| 2128 | #define b_whole_8 d13 | |
| 2129 | ||
| 2130 | #define r_whole_8 d14 | |
| 2131 | ||
| 2132 | #define pixels q8 | |
| 2133 | ||
| 2134 | #define rg_dx4 d18 | |
| 2135 | #define rg_dx8 d19 | |
| 2136 | ||
| 2137 | #define dx4 q10 | |
| 2138 | #define dx8 q10 | |
| 2139 | ||
| 2140 | #define v_left_x d6 | |
| 2141 | #define uvrg q4 | |
| 2142 | #define block_span q5 | |
| 2143 | ||
| 2144 | #define rg d9 | |
| 2145 | ||
| 2146 | #define d64_1 d22 | |
| 2147 | #define d64_128 d23 | |
| 2148 | ||
| 2149 | #define d128_4 q12 | |
| 2150 | #define d128_0x7 q13 | |
| 2151 | ||
| 2152 | #define d64_4 d24 | |
| 2153 | ||
| 2154 | #define dither_offsets q14 | |
| 2155 | #define draw_mask q15 | |
| 2156 | ||
| 2157 | #define dither_offsets_low d28 | |
| 2158 | ||
| 2159 | #define rg_dx d0 | |
| 2160 | #define test_mask q10 | |
| 2161 | ||
| 2162 | ||
| 2163 | #define setup_blocks_shaded_untextured_dither_a_dithered() \ | |
| 2164 | vqadd.u8 r_whole_8, r_whole_8, dither_offsets_low; \ | |
| 2165 | vqadd.u8 gb_whole_8, gb_whole_8, dither_offsets; \ | |
| 2166 | ||
| 2167 | #define setup_blocks_shaded_untextured_dither_b_dithered() \ | |
| 2168 | vqsub.u8 r_whole_8, r_whole_8, d64_4; \ | |
| 2169 | vqsub.u8 gb_whole_8, gb_whole_8, d128_4 \ | |
| 2170 | ||
| 2171 | #define setup_blocks_shaded_untextured_dither_a_undithered() \ | |
| 2172 | ||
| 2173 | #define setup_blocks_shaded_untextured_dither_b_undithered() \ | |
| 2174 | ||
| 2175 | ||
| 2176 | #define setup_blocks_shaded_untextured_indirect_builder(dithering) \ | |
| 2177 | .align 3; \ | |
| 2178 | \ | |
| 2179 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_indirect) \ | |
| e1f6de8f | 2180 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
2181 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
| 2182 | \ | |
| e1f6de8f | 2183 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
| 75e28f62 E |
2184 | \ |
| 2185 | cmp num_spans, #0; \ | |
| 2186 | bxeq lr; \ | |
| 2187 | \ | |
| 2188 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 2189 | save_abi_regs(); \ |
| 75e28f62 E |
2190 | vshl.u32 rg_dx4, rg_dx, #2; \ |
| 2191 | \ | |
| e1f6de8f | 2192 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
| 75e28f62 E |
2193 | vshl.u32 rg_dx8, rg_dx, #3; \ |
| 2194 | \ | |
| 2195 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 2196 | \ | |
| e1f6de8f | 2197 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
2198 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
| 2199 | \ | |
| 2200 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 2201 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 2202 | \ | |
| 2203 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 2204 | vmov.u8 d64_1, #1; \ | |
| 2205 | \ | |
| 2206 | vmov.u8 d128_4, #4; \ | |
| 2207 | vmov.u8 d64_128, #128; \ | |
| 2208 | \ | |
| 2209 | vmov.u8 d128_0x7, #0x7; \ | |
| 2210 | \ | |
| 2211 | 0: \ | |
| e1f6de8f | 2212 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
2213 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 2214 | \ | |
| e1f6de8f | 2215 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 2216 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
2217 | \ |
| 2218 | cmp span_num_blocks, #0; \ | |
| 2219 | beq 1f; \ | |
| 2220 | \ | |
| e1f6de8f | 2221 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
2222 | add num_blocks, span_num_blocks, num_blocks; \ |
| 2223 | \ | |
| 2224 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 2225 | bgt 2f; \ | |
| 2226 | \ | |
| 2227 | 3: \ | |
| e1f6de8f | 2228 | ldr b, [span_b_offset]; \ |
| 75e28f62 E |
2229 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 2230 | \ | |
| 2231 | vdup.u32 v_left_x, left_x; \ | |
| 2232 | and y, y, #0x3; \ | |
| 2233 | \ | |
| e1f6de8f | 2234 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
2235 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 2236 | \ | |
| 2237 | mla b, b_dx, left_x, b; \ | |
| 2238 | and dither_shift, left_x, #0x03; \ | |
| 2239 | \ | |
| e1f6de8f | 2240 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
2241 | vshr.u32 rg_dx, rg_dx4, #2; \ |
| 2242 | \ | |
| 2243 | mov dither_shift, dither_shift, lsl #3; \ | |
| 2244 | vmla.u32 rg, rg_dx, v_left_x; \ | |
| 2245 | \ | |
| 2246 | mov c_64, #64; \ | |
| 2247 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2248 | \ | |
| 2249 | mov dither_row, dither_row, ror dither_shift; \ | |
| 2250 | mov b_dx4, b_dx, lsl #2; \ | |
| 2251 | \ | |
| 2252 | vdup.u32 dither_offsets, dither_row; \ | |
| 2253 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
| 2254 | \ | |
| 2255 | vdup.u32 b_block, b; \ | |
| 2256 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
| 2257 | \ | |
| 2258 | mov b_dx8, b_dx, lsl #3; \ | |
| 2259 | vdup.u32 r_block, rg[0]; \ | |
| 2260 | vdup.u32 g_block, rg[1]; \ | |
| 2261 | \ | |
| e1f6de8f | 2262 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2263 | \ |
| 2264 | vadd.u32 r_block, r_block, block_span; \ | |
| e1f6de8f | 2265 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2266 | \ |
| 2267 | vadd.u32 g_block, g_block, block_span; \ | |
| e1f6de8f | 2268 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
| 75e28f62 E |
2269 | \ |
| 2270 | vadd.u32 b_block, b_block, block_span; \ | |
| 2271 | add block_ptr_b, block_ptr_a, #16; \ | |
| 2272 | \ | |
| 2273 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2274 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2275 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2276 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2277 | \ | |
| 2278 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2279 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2280 | \ | |
| 2281 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2282 | vdup.u32 dx4, b_dx4; \ | |
| 2283 | \ | |
| 2284 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2285 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2286 | \ | |
| 2287 | vadd.u32 r_block, r_block, dx8; \ | |
| 2288 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2289 | \ | |
| 2290 | vadd.u32 g_block, g_block, dx8; \ | |
| 2291 | vdup.u32 dx8, b_dx8; \ | |
| 2292 | \ | |
| 2293 | vadd.u32 b_block, b_block, dx8; \ | |
| 2294 | \ | |
| 2295 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2296 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2297 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2298 | \ | |
| 2299 | beq 5f; \ | |
| 2300 | veor.u32 draw_mask, draw_mask, draw_mask; \ | |
| 2301 | \ | |
| 2302 | 4: \ | |
| 2303 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2304 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2305 | \ | |
| 2306 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2307 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2308 | \ | |
| 2309 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| e1f6de8f | 2310 | str fb_ptr, [block_ptr_a, #44]; \ |
| 75e28f62 E |
2311 | \ |
| 2312 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2313 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2314 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2315 | \ | |
| 2316 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2317 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2318 | \ | |
| 2319 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2320 | vdup.u32 dx4, b_dx4; \ | |
| 2321 | \ | |
| 2322 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2323 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2324 | \ | |
| 2325 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
| 2326 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2327 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2328 | \ | |
| 2329 | vadd.u32 r_block, r_block, dx8; \ | |
| 2330 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2331 | \ | |
| 2332 | vadd.u32 g_block, g_block, dx8; \ | |
| 2333 | vdup.u32 dx8, b_dx8; \ | |
| 2334 | \ | |
| 2335 | vadd.u32 b_block, b_block, dx8; \ | |
| 2336 | add fb_ptr, fb_ptr, #16; \ | |
| 2337 | \ | |
| 2338 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2339 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2340 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2341 | \ | |
| e1f6de8f | 2342 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
| 2343 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
| 75e28f62 | 2344 | \ |
| e1f6de8f | 2345 | pld [fb_ptr]; \ |
| 75e28f62 E |
2346 | \ |
| 2347 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2348 | bne 4b; \ | |
| 2349 | \ | |
| 2350 | 5: \ | |
| e1f6de8f | 2351 | str fb_ptr, [block_ptr_a, #44]; \ |
| 75e28f62 E |
2352 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ |
| 2353 | \ | |
| e1f6de8f | 2354 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 E |
2355 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
| 2356 | \ | |
| 2357 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2358 | vdup.u8 draw_mask, right_mask; \ | |
| 2359 | \ | |
| 2360 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| e1f6de8f | 2361 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
2362 | \ |
| 2363 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 2364 | \ | |
| 2365 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
| 2366 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2367 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2368 | \ | |
| e1f6de8f | 2369 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
| 2370 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
| 75e28f62 E |
2371 | \ |
| 2372 | 1: \ | |
| 2373 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 2374 | add span_b_offset, span_b_offset, #4; \ | |
| 2375 | \ | |
| 2376 | add span_edge_data, span_edge_data, #8; \ | |
| 2377 | subs num_spans, num_spans, #1; \ | |
| 2378 | \ | |
| e1f6de8f | 2379 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
2380 | bne 0b; \ |
| 2381 | \ | |
| 3f0189c6 | 2382 | restore_abi_regs(); \ |
| db2804fb | 2383 | pop { r4 - r11, pc }; \ |
| 75e28f62 E |
2384 | \ |
| 2385 | 2: \ | |
| db2804fb | 2386 | vstr rg_dx4, [r0, #psx_gpu_saved_tmp_offset]; \ |
| 2387 | push { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ | |
| 75e28f62 | 2388 | bl flush_render_block_buffer; \ |
| db2804fb | 2389 | pop { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ |
| 2390 | vldr rg_dx4, [r0, #psx_gpu_saved_tmp_offset]; \ | |
| 75e28f62 E |
2391 | \ |
| 2392 | vmov.u8 d64_1, #1; \ | |
| 2393 | vmov.u8 d128_4, #4; \ | |
| 2394 | vmov.u8 d64_128, #128; \ | |
| 2395 | vmov.u8 d128_0x7, #0x7; \ | |
| 2396 | \ | |
| 2397 | vadd.u32 rg_dx8, rg_dx4, rg_dx4; \ | |
| 2398 | \ | |
| 2399 | mov num_blocks, span_num_blocks; \ | |
| 2400 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 2401 | bal 3b \ | |
| 2402 | ||
| 2403 | ||
| 2404 | setup_blocks_shaded_untextured_indirect_builder(undithered) | |
| 2405 | setup_blocks_shaded_untextured_indirect_builder(dithered) | |
| 2406 | ||
| 2407 | ||
| 2408 | #undef draw_mask | |
| 2409 | ||
| 2410 | #define mask_msb_ptr r14 | |
| 2411 | ||
| 2412 | #define draw_mask q0 | |
| 2413 | #define pixels_low d16 | |
| 3867c6ef | 2414 | #define pixels_high d17 |
| 75e28f62 E |
2415 | |
| 2416 | ||
| 2417 | ||
| 2418 | #define setup_blocks_shaded_untextured_direct_builder(dithering) \ | |
| 2419 | .align 3; \ | |
| 2420 | \ | |
| 2421 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_direct) \ | |
| e1f6de8f | 2422 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
2423 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
| 2424 | \ | |
| e1f6de8f | 2425 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
| 75e28f62 E |
2426 | \ |
| 2427 | cmp num_spans, #0; \ | |
| 2428 | bxeq lr; \ | |
| 2429 | \ | |
| 2430 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 2431 | save_abi_regs(); \ |
| 75e28f62 E |
2432 | vshl.u32 rg_dx4, rg_dx, #2; \ |
| 2433 | \ | |
| e1f6de8f | 2434 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
| 75e28f62 E |
2435 | vshl.u32 rg_dx8, rg_dx, #3; \ |
| 2436 | \ | |
| 2437 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 2438 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 2439 | \ | |
| 2440 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 2441 | vmov.u8 d64_1, #1; \ | |
| 2442 | \ | |
| 2443 | vmov.u8 d128_4, #4; \ | |
| 2444 | vmov.u8 d64_128, #128; \ | |
| 2445 | \ | |
| 2446 | vmov.u8 d128_0x7, #0x7; \ | |
| 2447 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| e1f6de8f | 2448 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
2449 | \ |
| 2450 | 0: \ | |
| e1f6de8f | 2451 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
2452 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 2453 | \ | |
| e1f6de8f | 2454 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 2455 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
2456 | \ |
| 2457 | cmp span_num_blocks, #0; \ | |
| 2458 | beq 1f; \ | |
| 2459 | \ | |
| e1f6de8f | 2460 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
2461 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 2462 | \ | |
| e1f6de8f | 2463 | ldr b, [span_b_offset]; \ |
| 75e28f62 E |
2464 | vdup.u32 v_left_x, left_x; \ |
| 2465 | and y, y, #0x3; \ | |
| 2466 | \ | |
| e1f6de8f | 2467 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
2468 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 2469 | \ | |
| 2470 | mla b, b_dx, left_x, b; \ | |
| 2471 | and dither_shift, left_x, #0x03; \ | |
| 2472 | \ | |
| e1f6de8f | 2473 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
2474 | vshr.u32 rg_dx, rg_dx4, #2; \ |
| 2475 | \ | |
| 2476 | mov dither_shift, dither_shift, lsl #3; \ | |
| 2477 | vmla.u32 rg, rg_dx, v_left_x; \ | |
| 2478 | \ | |
| 2479 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2480 | \ | |
| 2481 | mov dither_row, dither_row, ror dither_shift; \ | |
| 2482 | mov b_dx4, b_dx, lsl #2; \ | |
| 2483 | \ | |
| 2484 | vdup.u32 dither_offsets, dither_row; \ | |
| 2485 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
| 2486 | \ | |
| 2487 | vdup.u32 b_block, b; \ | |
| 2488 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
| 2489 | \ | |
| 2490 | mov b_dx8, b_dx, lsl #3; \ | |
| 2491 | vdup.u32 r_block, rg[0]; \ | |
| 2492 | vdup.u32 g_block, rg[1]; \ | |
| 2493 | \ | |
| e1f6de8f | 2494 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2495 | \ |
| 2496 | vadd.u32 r_block, r_block, block_span; \ | |
| e1f6de8f | 2497 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2498 | \ |
| 2499 | vadd.u32 g_block, g_block, block_span; \ | |
| e1f6de8f | 2500 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
| 75e28f62 E |
2501 | \ |
| 2502 | vadd.u32 b_block, b_block, block_span; \ | |
| 2503 | add block_ptr_b, block_ptr_a, #16; \ | |
| 2504 | \ | |
| 2505 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2506 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2507 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2508 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2509 | \ | |
| 2510 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2511 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2512 | \ | |
| 2513 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2514 | vdup.u32 dx4, b_dx4; \ | |
| 2515 | \ | |
| 2516 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2517 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2518 | \ | |
| 2519 | vadd.u32 r_block, r_block, dx8; \ | |
| 2520 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2521 | \ | |
| 2522 | vadd.u32 g_block, g_block, dx8; \ | |
| 2523 | vdup.u32 dx8, b_dx8; \ | |
| 2524 | \ | |
| 2525 | vadd.u32 b_block, b_block, dx8; \ | |
| 2526 | \ | |
| 2527 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2528 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2529 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2530 | \ | |
| 2531 | beq 3f; \ | |
| 2532 | \ | |
| 2533 | 2: \ | |
| 2534 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2535 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2536 | \ | |
| 2537 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2538 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2539 | \ | |
| 2540 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2541 | \ | |
| 2542 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2543 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2544 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2545 | \ | |
| 2546 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2547 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2548 | \ | |
| 2549 | vmov pixels, msb_mask; \ | |
| 2550 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2551 | vdup.u32 dx4, b_dx4; \ | |
| 2552 | \ | |
| 2553 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2554 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2555 | \ | |
| 2556 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
| 2557 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2558 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2559 | \ | |
| 2560 | vadd.u32 r_block, r_block, dx8; \ | |
| 2561 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2562 | \ | |
| 2563 | vadd.u32 g_block, g_block, dx8; \ | |
| 2564 | vdup.u32 dx8, b_dx8; \ | |
| 2565 | \ | |
| 2566 | vadd.u32 b_block, b_block, dx8; \ | |
| 2567 | \ | |
| 2568 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2569 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2570 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2571 | \ | |
| e1f6de8f | 2572 | vst1.u32 { pixels }, [fb_ptr]!; \ |
| 75e28f62 E |
2573 | subs span_num_blocks, span_num_blocks, #1; \ |
| 2574 | bne 2b; \ | |
| 2575 | \ | |
| 2576 | 3: \ | |
| 2577 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2578 | \ | |
| e1f6de8f | 2579 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 E |
2580 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
| 2581 | \ | |
| 2582 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 3867c6ef | 2583 | rbit right_mask, right_mask; \ |
| 75e28f62 E |
2584 | vmov pixels, msb_mask; \ |
| 2585 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 3867c6ef | 2586 | clz right_mask, right_mask; \ |
| 75e28f62 E |
2587 | \ |
| 2588 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
| 2589 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2590 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2591 | \ | |
| 8184d7c5 | 2592 | JT_OP_REL(100f, right_mask, temp); \ |
| e1f6de8f | 2593 | JT_OP(ldr pc, [pc, right_mask, lsl #2]); \ |
| 3867c6ef | 2594 | nop; \ |
| 8184d7c5 | 2595 | 100: \ |
| 3867c6ef | 2596 | nop; \ |
| 8184d7c5 | 2597 | .word JTE(100b, 4f); \ |
| 2598 | .word JTE(100b, 5f); \ | |
| 2599 | .word JTE(100b, 6f); \ | |
| 2600 | .word JTE(100b, 7f); \ | |
| 2601 | .word JTE(100b, 8f); \ | |
| 2602 | .word JTE(100b, 9f); \ | |
| 2603 | .word JTE(100b, 10f); \ | |
| 2604 | .word JTE(100b, 11f); \ | |
| 3867c6ef | 2605 | \ |
| 75e28f62 | 2606 | 4: \ |
| e1f6de8f | 2607 | vst1.u16 { pixels_low[0] }, [fb_ptr]; \ |
| 3867c6ef E |
2608 | bal 1f; \ |
| 2609 | \ | |
| 2610 | 5: \ | |
| e1f6de8f | 2611 | vst1.u32 { pixels_low[0] }, [fb_ptr]; \ |
| 3867c6ef E |
2612 | bal 1f; \ |
| 2613 | \ | |
| 2614 | 6: \ | |
| e1f6de8f | 2615 | vst1.u32 { pixels_low[0] }, [fb_ptr]!; \ |
| 2616 | vst1.u16 { pixels_low[2] }, [fb_ptr]; \ | |
| 3867c6ef E |
2617 | bal 1f; \ |
| 2618 | \ | |
| 2619 | 7: \ | |
| e1f6de8f | 2620 | vst1.u32 { pixels_low }, [fb_ptr]; \ |
| 3867c6ef E |
2621 | bal 1f; \ |
| 2622 | \ | |
| 2623 | 8: \ | |
| e1f6de8f | 2624 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
| 2625 | vst1.u16 { pixels_high[0] }, [fb_ptr]; \ | |
| 3867c6ef E |
2626 | bal 1f; \ |
| 2627 | \ | |
| 2628 | 9: \ | |
| e1f6de8f | 2629 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
| 2630 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
| 3867c6ef E |
2631 | bal 1f; \ |
| 2632 | \ | |
| 2633 | 10: \ | |
| e1f6de8f | 2634 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
| 2635 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
| 2636 | vst1.u16 { pixels_high[2] }, [fb_ptr]; \ | |
| 3867c6ef E |
2637 | bal 1f; \ |
| 2638 | \ | |
| 2639 | 11: \ | |
| e1f6de8f | 2640 | vst1.u32 { pixels }, [fb_ptr]; \ |
| 3867c6ef | 2641 | bal 1f; \ |
| 75e28f62 E |
2642 | \ |
| 2643 | 1: \ | |
| 2644 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 2645 | add span_b_offset, span_b_offset, #4; \ | |
| 2646 | \ | |
| 2647 | add span_edge_data, span_edge_data, #8; \ | |
| 2648 | subs num_spans, num_spans, #1; \ | |
| 2649 | \ | |
| 2650 | bne 0b; \ | |
| 2651 | \ | |
| 3f0189c6 | 2652 | restore_abi_regs(); \ |
| 75e28f62 E |
2653 | ldmia sp!, { r4 - r11, pc } \ |
| 2654 | ||
| 2655 | setup_blocks_shaded_untextured_direct_builder(undithered) | |
| 2656 | setup_blocks_shaded_untextured_direct_builder(dithered) | |
| 2657 | ||
| 2658 | ||
| 2659 | #undef psx_gpu | |
| 2660 | #undef num_blocks | |
| 2661 | #undef triangle | |
| 2662 | #undef c_64 | |
| 2663 | ||
| 2664 | #define psx_gpu r0 | |
| 2665 | #define block_ptr r1 | |
| 2666 | #define num_blocks r2 | |
| 2667 | #define uv_01 r3 | |
| 2668 | #define uv_23 r4 | |
| 2669 | #define uv_45 r5 | |
| 2670 | #define uv_67 r6 | |
| 2671 | #define uv_0 r7 | |
| 2672 | #define uv_1 r3 | |
| 2673 | #define uv_2 r8 | |
| 2674 | #define uv_3 r4 | |
| 2675 | #define uv_4 r9 | |
| 2676 | #define uv_5 r5 | |
| 2677 | #define uv_6 r10 | |
| 2678 | #define uv_7 r6 | |
| 2679 | #define texture_ptr r11 | |
| 2680 | ||
| 2681 | #define pixel_0 r7 | |
| 2682 | #define pixel_1 r3 | |
| 2683 | #define pixel_2 r8 | |
| 2684 | #define pixel_3 r4 | |
| 2685 | #define pixel_4 r9 | |
| 2686 | #define pixel_5 r5 | |
| 2687 | #define pixel_6 r10 | |
| 2688 | #define pixel_7 r6 | |
| 2689 | ||
| 2690 | #define pixels_a r7 | |
| 2691 | #define pixels_b r9 | |
| 2692 | #define pixels_c r8 | |
| 2693 | #define pixels_d r10 | |
| 2694 | ||
| 2695 | #define c_64 r0 | |
| 2696 | ||
| 2697 | #define clut_ptr r12 | |
| 2698 | #define current_texture_mask r5 | |
| 2699 | #define dirty_textures_mask r6 | |
| 2700 | ||
| 2701 | #define texels d0 | |
| 2702 | ||
| 2703 | #define clut_low_a d2 | |
| 2704 | #define clut_low_b d3 | |
| 2705 | #define clut_high_a d4 | |
| 2706 | #define clut_high_b d5 | |
| 2707 | ||
| 2708 | #define clut_a q1 | |
| 2709 | #define clut_b q2 | |
| 2710 | ||
| 2711 | #define texels_low d6 | |
| 2712 | #define texels_high d7 | |
| 2713 | ||
| 2714 | .align 3 | |
| 2715 | ||
| 2716 | function(texture_blocks_untextured) | |
| 2717 | bx lr | |
| 2718 | ||
| 2719 | ||
| 2720 | .align 3 | |
| 2721 | ||
| 2722 | function(texture_blocks_4bpp) | |
| 2723 | stmdb sp!, { r3 - r11, r14 } | |
| 2724 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2725 | ||
| e1f6de8f | 2726 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 2727 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
| 75e28f62 | 2728 | |
| e1f6de8f | 2729 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
| 2730 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128] | |
| 75e28f62 | 2731 | |
| e1f6de8f | 2732 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
| 75e28f62 E |
2733 | vuzp.u8 clut_a, clut_b |
| 2734 | ||
| e1f6de8f | 2735 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
| 75e28f62 E |
2736 | tst dirty_textures_mask, current_texture_mask |
| 2737 | ||
| 2738 | bne 1f | |
| 2739 | mov c_64, #64 | |
| 2740 | ||
| 2741 | 0: | |
| 2742 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
| 2743 | ||
| 2744 | uxtah uv_0, texture_ptr, uv_01 | |
| 2745 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
| 2746 | ||
| 2747 | uxtah uv_2, texture_ptr, uv_23 | |
| 2748 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
| 2749 | ||
| 2750 | uxtah uv_4, texture_ptr, uv_45 | |
| e1f6de8f | 2751 | ldrb pixel_0, [uv_0] |
| 75e28f62 E |
2752 | |
| 2753 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
| e1f6de8f | 2754 | ldrb pixel_1, [uv_1] |
| 75e28f62 E |
2755 | |
| 2756 | uxtah uv_6, texture_ptr, uv_67 | |
| e1f6de8f | 2757 | ldrb pixel_2, [uv_2] |
| 75e28f62 E |
2758 | |
| 2759 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
| e1f6de8f | 2760 | ldrb pixel_3, [uv_3] |
| 75e28f62 | 2761 | |
| e1f6de8f | 2762 | ldrb pixel_4, [uv_4] |
| 75e28f62 E |
2763 | subs num_blocks, num_blocks, #1 |
| 2764 | ||
| e1f6de8f | 2765 | ldrb pixel_5, [uv_5] |
| 75e28f62 E |
2766 | orr pixels_a, pixel_0, pixel_1, lsl #8 |
| 2767 | ||
| e1f6de8f | 2768 | ldrb pixel_6, [uv_6] |
| 75e28f62 E |
2769 | orr pixels_b, pixel_4, pixel_5, lsl #8 |
| 2770 | ||
| e1f6de8f | 2771 | ldrb pixel_7, [uv_7] |
| 75e28f62 E |
2772 | orr pixels_a, pixels_a, pixel_2, lsl #16 |
| 2773 | ||
| 2774 | orr pixels_b, pixels_b, pixel_6, lsl #16 | |
| 2775 | orr pixels_a, pixels_a, pixel_3, lsl #24 | |
| 2776 | ||
| 2777 | orr pixels_b, pixels_b, pixel_7, lsl #24 | |
| ed0fd81d | 2778 | vmov texels, pixels_a, pixels_b |
| 75e28f62 E |
2779 | |
| 2780 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels | |
| 2781 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels | |
| 2782 | ||
| e1f6de8f | 2783 | vst2.u8 { texels_low, texels_high }, [block_ptr, :128], c_64 |
| 75e28f62 E |
2784 | bne 0b |
| 2785 | ||
| 2786 | ldmia sp!, { r3 - r11, pc } | |
| 2787 | ||
| 2788 | 1: | |
| 2789 | stmdb sp!, { r1 - r2 } | |
| 2790 | bl update_texture_4bpp_cache | |
| 2791 | ||
| 2792 | mov c_64, #64 | |
| 2793 | ldmia sp!, { r1 - r2 } | |
| 2794 | bal 0b | |
| 2795 | ||
| 2796 | ||
| 2797 | .align 3 | |
| 2798 | ||
| 2799 | function(texture_blocks_8bpp) | |
| db2804fb | 2800 | push { r4 - r11, lr } |
| 75e28f62 E |
2801 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset |
| 2802 | ||
| e1f6de8f | 2803 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 2804 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
| 75e28f62 | 2805 | |
| e1f6de8f | 2806 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
| 2807 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] | |
| 75e28f62 | 2808 | |
| e1f6de8f | 2809 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset] |
| 75e28f62 E |
2810 | tst dirty_textures_mask, current_texture_mask |
| 2811 | ||
| 2812 | bne 1f | |
| 2813 | nop | |
| 2814 | ||
| 2815 | 0: | |
| 2816 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
| 2817 | ||
| 2818 | uxtah uv_0, texture_ptr, uv_01 | |
| 2819 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
| 2820 | ||
| 2821 | uxtah uv_2, texture_ptr, uv_23 | |
| 2822 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
| 2823 | ||
| 2824 | uxtah uv_4, texture_ptr, uv_45 | |
| e1f6de8f | 2825 | ldrb pixel_0, [uv_0] |
| 75e28f62 E |
2826 | |
| 2827 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
| e1f6de8f | 2828 | ldrb pixel_1, [uv_1] |
| 75e28f62 E |
2829 | |
| 2830 | uxtah uv_6, texture_ptr, uv_67 | |
| e1f6de8f | 2831 | ldrb pixel_2, [uv_2] |
| 75e28f62 E |
2832 | |
| 2833 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
| e1f6de8f | 2834 | ldrb pixel_3, [uv_3] |
| 75e28f62 | 2835 | |
| e1f6de8f | 2836 | ldrb pixel_4, [uv_4] |
| 75e28f62 E |
2837 | add pixel_0, pixel_0, pixel_0 |
| 2838 | ||
| e1f6de8f | 2839 | ldrb pixel_5, [uv_5] |
| 75e28f62 E |
2840 | add pixel_1, pixel_1, pixel_1 |
| 2841 | ||
| e1f6de8f | 2842 | ldrb pixel_6, [uv_6] |
| 75e28f62 E |
2843 | add pixel_2, pixel_2, pixel_2 |
| 2844 | ||
| e1f6de8f | 2845 | ldrb pixel_7, [uv_7] |
| 75e28f62 E |
2846 | add pixel_3, pixel_3, pixel_3 |
| 2847 | ||
| e1f6de8f | 2848 | ldrh pixel_0, [clut_ptr, pixel_0] |
| 75e28f62 E |
2849 | add pixel_4, pixel_4, pixel_4 |
| 2850 | ||
| e1f6de8f | 2851 | ldrh pixel_1, [clut_ptr, pixel_1] |
| 75e28f62 E |
2852 | add pixel_5, pixel_5, pixel_5 |
| 2853 | ||
| e1f6de8f | 2854 | ldrh pixel_2, [clut_ptr, pixel_2] |
| 75e28f62 E |
2855 | add pixel_6, pixel_6, pixel_6 |
| 2856 | ||
| e1f6de8f | 2857 | ldrh pixel_3, [clut_ptr, pixel_3] |
| 75e28f62 E |
2858 | add pixel_7, pixel_7, pixel_7 |
| 2859 | ||
| e1f6de8f | 2860 | ldrh pixel_4, [clut_ptr, pixel_4] |
| 75e28f62 E |
2861 | orr pixels_a, pixel_0, pixel_1, lsl #16 |
| 2862 | ||
| e1f6de8f | 2863 | ldrh pixel_5, [clut_ptr, pixel_5] |
| 75e28f62 E |
2864 | orr pixels_c, pixel_2, pixel_3, lsl #16 |
| 2865 | ||
| e1f6de8f | 2866 | ldrh pixel_6, [clut_ptr, pixel_6] |
| 75e28f62 E |
2867 | subs num_blocks, num_blocks, #1 |
| 2868 | ||
| e1f6de8f | 2869 | ldrh pixel_7, [clut_ptr, pixel_7] |
| 75e28f62 E |
2870 | orr pixels_b, pixel_4, pixel_5, lsl #16 |
| 2871 | ||
| 2872 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
| 2873 | stm block_ptr, { pixels_a, pixels_c, pixels_b, pixels_d } | |
| 2874 | ||
| 2875 | add block_ptr, block_ptr, #64 | |
| 2876 | bne 0b | |
| 2877 | ||
| db2804fb | 2878 | pop { r4 - r11, pc } |
| 75e28f62 E |
2879 | |
| 2880 | 1: | |
| db2804fb | 2881 | /* pushing odd num of regs here realigns our unaligned stack */ |
| 2882 | push { r1 - r2, EXTRA_UNSAVED_REGS r12 } | |
| 2883 | bl update_texture_8bpp_cache | |
| 2884 | pop { r1 - r2, EXTRA_UNSAVED_REGS r12 } | |
| 2885 | bal 0b | |
| 75e28f62 E |
2886 | |
| 2887 | ||
| 2888 | #undef uv_0 | |
| 2889 | #undef uv_1 | |
| 2890 | #undef uv_2 | |
| 2891 | #undef uv_3 | |
| 2892 | #undef uv_4 | |
| 2893 | #undef uv_5 | |
| 2894 | #undef uv_6 | |
| 2895 | #undef uv_7 | |
| 2896 | ||
| 2897 | #undef pixel_0 | |
| 2898 | #undef pixel_1 | |
| 2899 | #undef pixel_2 | |
| 2900 | #undef pixel_3 | |
| 2901 | #undef pixel_4 | |
| 2902 | #undef pixel_5 | |
| 2903 | #undef pixel_6 | |
| 2904 | #undef pixel_7 | |
| 2905 | ||
| 2906 | #undef texture_ptr | |
| 2907 | ||
| 2908 | #undef pixels_a | |
| 2909 | #undef pixels_b | |
| 2910 | #undef pixels_c | |
| 2911 | #undef pixels_d | |
| 2912 | ||
| 2913 | #define psx_gpu r0 | |
| 2914 | #define block_ptr r1 | |
| 2915 | #define num_blocks r2 | |
| 2916 | ||
| 2917 | #define uv_0 r3 | |
| 2918 | #define uv_1 r4 | |
| 2919 | #define u_0 r3 | |
| 2920 | #define u_1 r4 | |
| 2921 | #define v_0 r5 | |
| 2922 | #define v_1 r6 | |
| 2923 | ||
| 2924 | #define uv_2 r5 | |
| 2925 | #define uv_3 r6 | |
| 2926 | #define u_2 r5 | |
| 2927 | #define u_3 r6 | |
| 2928 | #define v_2 r7 | |
| 2929 | #define v_3 r8 | |
| 2930 | ||
| 2931 | #define uv_4 r7 | |
| 2932 | #define uv_5 r8 | |
| 2933 | #define u_4 r7 | |
| 2934 | #define u_5 r8 | |
| 2935 | #define v_4 r9 | |
| 2936 | #define v_5 r10 | |
| 2937 | ||
| 2938 | #define uv_6 r9 | |
| 2939 | #define uv_7 r10 | |
| 2940 | #define u_6 r9 | |
| 2941 | #define u_7 r10 | |
| 2942 | #define v_6 r11 | |
| 2943 | #define v_7 r0 | |
| 2944 | ||
| 2945 | #define pixel_0 r3 | |
| 2946 | #define pixel_1 r4 | |
| 2947 | #define pixel_2 r5 | |
| 2948 | #define pixel_3 r6 | |
| 2949 | #define pixel_4 r7 | |
| 2950 | #define pixel_5 r8 | |
| 2951 | #define pixel_6 r9 | |
| 2952 | #define pixel_7 r10 | |
| 2953 | ||
| 2954 | #define pixels_a r3 | |
| 2955 | #define pixels_b r5 | |
| 2956 | #define pixels_c r7 | |
| 2957 | #define pixels_d r9 | |
| 2958 | ||
| 2959 | #define texture_ptr r12 | |
| 2960 | ||
| 2961 | ||
| 2962 | .align 3 | |
| 2963 | ||
| 2964 | function(texture_blocks_16bpp) | |
| 2965 | stmdb sp!, { r3 - r11, r14 } | |
| 2966 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2967 | ||
| e1f6de8f | 2968 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 2969 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] | |
| 75e28f62 E |
2970 | |
| 2971 | 0: | |
| e1f6de8f | 2972 | ldrh uv_0, [block_ptr] |
| 75e28f62 E |
2973 | subs num_blocks, num_blocks, #1 |
| 2974 | ||
| e1f6de8f | 2975 | ldrh uv_1, [block_ptr, #2] |
| 75e28f62 E |
2976 | |
| 2977 | and v_0, uv_0, #0xFF00 | |
| 2978 | and v_1, uv_1, #0xFF00 | |
| 2979 | ||
| 2980 | and u_0, uv_0, #0xFF | |
| 2981 | and u_1, uv_1, #0xFF | |
| 2982 | ||
| 2983 | add uv_0, u_0, v_0, lsl #2 | |
| e1f6de8f | 2984 | ldrh uv_2, [block_ptr, #4] |
| 75e28f62 E |
2985 | |
| 2986 | add uv_1, u_1, v_1, lsl #2 | |
| e1f6de8f | 2987 | ldrh uv_3, [block_ptr, #6] |
| 75e28f62 E |
2988 | |
| 2989 | add uv_0, uv_0, uv_0 | |
| 2990 | add uv_1, uv_1, uv_1 | |
| 2991 | ||
| 2992 | and v_2, uv_2, #0xFF00 | |
| 2993 | and v_3, uv_3, #0xFF00 | |
| 2994 | ||
| 2995 | and u_2, uv_2, #0xFF | |
| 2996 | and u_3, uv_3, #0xFF | |
| 2997 | ||
| 2998 | add uv_2, u_2, v_2, lsl #2 | |
| e1f6de8f | 2999 | ldrh uv_4, [block_ptr, #8] |
| 75e28f62 E |
3000 | |
| 3001 | add uv_3, u_3, v_3, lsl #2 | |
| e1f6de8f | 3002 | ldrh uv_5, [block_ptr, #10] |
| 75e28f62 E |
3003 | |
| 3004 | add uv_2, uv_2, uv_2 | |
| 3005 | add uv_3, uv_3, uv_3 | |
| 3006 | ||
| 3007 | and v_4, uv_4, #0xFF00 | |
| 3008 | and v_5, uv_5, #0xFF00 | |
| 3009 | ||
| 3010 | and u_4, uv_4, #0xFF | |
| 3011 | and u_5, uv_5, #0xFF | |
| 3012 | ||
| 3013 | add uv_4, u_4, v_4, lsl #2 | |
| e1f6de8f | 3014 | ldrh uv_6, [block_ptr, #12] |
| 75e28f62 E |
3015 | |
| 3016 | add uv_5, u_5, v_5, lsl #2 | |
| e1f6de8f | 3017 | ldrh uv_7, [block_ptr, #14] |
| 75e28f62 E |
3018 | |
| 3019 | add uv_4, uv_4, uv_4 | |
| e1f6de8f | 3020 | ldrh pixel_0, [texture_ptr, uv_0] |
| 75e28f62 E |
3021 | |
| 3022 | add uv_5, uv_5, uv_5 | |
| e1f6de8f | 3023 | ldrh pixel_1, [texture_ptr, uv_1] |
| 75e28f62 E |
3024 | |
| 3025 | and v_6, uv_6, #0xFF00 | |
| e1f6de8f | 3026 | ldrh pixel_2, [texture_ptr, uv_2] |
| 75e28f62 E |
3027 | |
| 3028 | and v_7, uv_7, #0xFF00 | |
| e1f6de8f | 3029 | ldrh pixel_3, [texture_ptr, uv_3] |
| 75e28f62 E |
3030 | |
| 3031 | and u_6, uv_6, #0xFF | |
| e1f6de8f | 3032 | ldrh pixel_4, [texture_ptr, uv_4] |
| 75e28f62 E |
3033 | |
| 3034 | and u_7, uv_7, #0xFF | |
| e1f6de8f | 3035 | ldrh pixel_5, [texture_ptr, uv_5] |
| 75e28f62 E |
3036 | |
| 3037 | add uv_6, u_6, v_6, lsl #2 | |
| 3038 | add uv_7, u_7, v_7, lsl #2 | |
| 3039 | ||
| 3040 | add uv_6, uv_6, uv_6 | |
| 3041 | add uv_7, uv_7, uv_7 | |
| 3042 | ||
| 3043 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
| 3044 | orr pixels_b, pixel_2, pixel_3, lsl #16 | |
| 3045 | ||
| e1f6de8f | 3046 | ldrh pixel_6, [texture_ptr, uv_6] |
| 75e28f62 E |
3047 | orr pixels_c, pixel_4, pixel_5, lsl #16 |
| 3048 | ||
| e1f6de8f | 3049 | ldrh pixel_7, [texture_ptr, uv_7] |
| 75e28f62 E |
3050 | orr pixels_d, pixel_6, pixel_7, lsl #16 |
| 3051 | ||
| 3052 | stm block_ptr, { pixels_a, pixels_b, pixels_c, pixels_d } | |
| 3053 | add block_ptr, block_ptr, #64 | |
| 3054 | ||
| 3055 | bne 0b | |
| 3056 | ||
| 3057 | ldmia sp!, { r3 - r11, pc } | |
| 3058 | ||
| 3059 | ||
| 3060 | #undef num_blocks | |
| 3061 | ||
| 3062 | #undef test_mask | |
| 3063 | #undef texels | |
| 3064 | #undef pixels_b | |
| 3065 | #undef pixels | |
| 3066 | #undef d64_1 | |
| 3067 | #undef d64_4 | |
| 3068 | #undef d64_128 | |
| 3069 | #undef draw_mask | |
| 3070 | #undef msb_mask | |
| 3071 | #undef msb_mask_low | |
| 3072 | #undef msb_mask_high | |
| 3073 | #undef fb_pixels | |
| 3074 | ||
| 3075 | #undef c_32 | |
| 3076 | #undef fb_ptr | |
| 3077 | #undef mask_msb_ptr | |
| 3078 | ||
| 3079 | #define psx_gpu r0 | |
| 3080 | #define num_blocks r1 | |
| 3081 | #define color_ptr r2 | |
| 3867c6ef E |
3082 | #define colors_scalar r2 |
| 3083 | #define colors_scalar_compare r3 | |
| 75e28f62 E |
3084 | #define mask_msb_ptr r2 |
| 3085 | ||
| 3086 | #define block_ptr_load_a r0 | |
| 3087 | #define block_ptr_store r3 | |
| 3088 | #define block_ptr_load_b r12 | |
| 3089 | #define c_32 r2 | |
| 3090 | ||
| 3091 | #define c_48 r4 | |
| 3092 | #define fb_ptr r14 | |
| 3093 | #define draw_mask_bits_scalar r5 | |
| 3094 | ||
| 3095 | #define d128_0x07 q0 | |
| 3096 | #define d128_0x1F q1 | |
| 3097 | #define d128_0x8000 q2 | |
| 3098 | #define test_mask q3 | |
| 3099 | #define texels q4 | |
| 3100 | #define colors_rg q5 | |
| 3101 | #define colors_b_dm_bits q6 | |
| 3102 | #define texels_rg q7 | |
| 3103 | #define pixels_r q8 | |
| 3104 | #define pixels_g q9 | |
| 3105 | #define pixels_b q10 | |
| 3106 | #define pixels q11 | |
| 3107 | #define zero_mask q4 | |
| 3108 | #define draw_mask q12 | |
| 3109 | #define msb_mask q13 | |
| 3110 | ||
| 3111 | #define fb_pixels q8 | |
| 3112 | ||
| 3113 | #define pixels_gb_low q9 | |
| 3114 | ||
| 3115 | #define colors_r d10 | |
| 3116 | #define colors_g d11 | |
| 3117 | #define colors_b d12 | |
| 3118 | #define draw_mask_bits d13 | |
| 3119 | #define texels_r d14 | |
| 3120 | #define texels_g d15 | |
| 3121 | #define pixels_r_low d16 | |
| 3122 | #define pixels_g_low d18 | |
| 3123 | #define pixels_b_low d19 | |
| 3124 | #define msb_mask_low d26 | |
| 3125 | #define msb_mask_high d27 | |
| 3126 | ||
| 3127 | #define d64_1 d28 | |
| 3128 | #define d64_4 d29 | |
| 3129 | #define d64_128 d30 | |
| 3130 | #define texels_b d31 | |
| 3131 | ||
| 3132 | #define shade_blocks_textured_modulated_prologue_indirect() \ | |
| 3133 | mov c_48, #48; \ | |
| 3134 | add block_ptr_store, psx_gpu, #psx_gpu_blocks_offset \ | |
| 3135 | ||
| 3136 | #define shade_blocks_textured_modulated_prologue_direct() \ | |
| 3137 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| e1f6de8f | 3138 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] \ |
| 75e28f62 | 3139 | |
| 75e28f62 | 3140 | |
| 3867c6ef E |
3141 | #define shade_blocks_textured_modulated_prologue_shaded(dithering, target) \ |
| 3142 | ||
| 3143 | #define shade_blocks_textured_false_modulation_check_undithered(target) \ | |
| e1f6de8f | 3144 | ldr colors_scalar, [psx_gpu, #psx_gpu_triangle_color_offset]; \ |
| 3867c6ef E |
3145 | movw colors_scalar_compare, #0x8080; \ |
| 3146 | \ | |
| 3147 | movt colors_scalar_compare, #0x80; \ | |
| 3148 | cmp colors_scalar, colors_scalar_compare; \ | |
| 3149 | beq shade_blocks_textured_unmodulated_##target \ | |
| 3150 | ||
| 3151 | #define shade_blocks_textured_false_modulation_check_dithered(target) \ | |
| 3152 | ||
| 3153 | #define shade_blocks_textured_modulated_prologue_unshaded(dithering, target) \ | |
| 3154 | shade_blocks_textured_false_modulation_check_##dithering(target); \ | |
| 75e28f62 | 3155 | add color_ptr, psx_gpu, #psx_gpu_triangle_color_offset; \ |
| e1f6de8f | 3156 | vld1.u32 { colors_r[] }, [color_ptr, :32]; \ |
| 75e28f62 E |
3157 | vdup.u8 colors_g, colors_r[1]; \ |
| 3158 | vdup.u8 colors_b, colors_r[2]; \ | |
| 3159 | vdup.u8 colors_r, colors_r[0] \ | |
| 3160 | ||
| 3161 | ||
| 3162 | #define shade_blocks_textured_modulated_load_dithered(target) \ | |
| e1f6de8f | 3163 | vld1.u32 { target }, [block_ptr_load_b, :128] \ |
| 75e28f62 E |
3164 | |
| 3165 | #define shade_blocks_textured_modulated_load_last_dithered(target) \ | |
| e1f6de8f | 3166 | vld1.u32 { target }, [block_ptr_load_b, :128], c_32 \ |
| 75e28f62 E |
3167 | |
| 3168 | #define shade_blocks_textured_modulated_load_undithered(target) \ | |
| 3169 | ||
| 3170 | #define shade_blocks_textured_modulated_load_last_undithered(target) \ | |
| 3171 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
| 3172 | ||
| 3173 | #define shade_blocks_textured_modulate_dithered(channel) \ | |
| 3174 | vmlal.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
| 3175 | ||
| 3176 | #define shade_blocks_textured_modulate_undithered(channel) \ | |
| 3177 | vmull.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
| 3178 | ||
| 3179 | ||
| 3180 | #define shade_blocks_textured_modulated_store_draw_mask_indirect(offset) \ | |
| e1f6de8f | 3181 | vst1.u32 { draw_mask }, [block_ptr_store, :128]! \ |
| 75e28f62 E |
3182 | |
| 3183 | #define shade_blocks_textured_modulated_store_draw_mask_direct(offset) \ | |
| e1f6de8f | 3184 | ldr fb_ptr, [block_ptr_load_b, #(offset - 64)]; \ |
| 3185 | vld1.u32 { fb_pixels }, [fb_ptr]; \ | |
| 75e28f62 E |
3186 | vbit.u16 pixels, fb_pixels, draw_mask \ |
| 3187 | ||
| 3188 | #define shade_blocks_textured_modulated_store_pixels_indirect() \ | |
| e1f6de8f | 3189 | vst1.u32 { pixels }, [block_ptr_store, :128], c_48 \ |
| 75e28f62 E |
3190 | |
| 3191 | #define shade_blocks_textured_modulated_store_pixels_direct() \ | |
| e1f6de8f | 3192 | vst1.u32 { pixels }, [fb_ptr] \ |
| 75e28f62 E |
3193 | |
| 3194 | ||
| 3195 | #define shade_blocks_textured_modulated_load_rg_shaded() \ | |
| e1f6de8f | 3196 | vld1.u32 { colors_r, colors_g }, [block_ptr_load_b, :128], c_32 \ |
| 75e28f62 E |
3197 | |
| 3198 | #define shade_blocks_textured_modulated_load_rg_unshaded() \ | |
| 3199 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
| 3200 | ||
| 3201 | #define shade_blocks_textured_modulated_load_bdm_shaded() \ | |
| e1f6de8f | 3202 | vld1.u32 { colors_b, draw_mask_bits }, [block_ptr_load_a, :128], c_32 \ |
| 75e28f62 E |
3203 | |
| 3204 | #define shade_blocks_textured_modulated_load_bdm_unshaded() \ | |
| e1f6de8f | 3205 | ldr draw_mask_bits_scalar, [block_ptr_load_a, #8]; \ |
| 75e28f62 E |
3206 | add block_ptr_load_a, block_ptr_load_a, #32 \ |
| 3207 | ||
| 3208 | #define shade_blocks_textured_modulated_expand_draw_mask_shaded() \ | |
| 3209 | vdup.u16 draw_mask, draw_mask_bits[0] \ | |
| 3210 | ||
| 3211 | #define shade_blocks_textured_modulated_expand_draw_mask_unshaded() \ | |
| 3212 | vdup.u16 draw_mask, draw_mask_bits_scalar \ | |
| 3213 | ||
| 3214 | ||
| 3215 | #define shade_blocks_textured_modulated_apply_msb_mask_indirect() \ | |
| 3216 | ||
| 3217 | #define shade_blocks_textured_modulated_apply_msb_mask_direct() \ | |
| 3218 | vorr.u16 pixels, pixels, msb_mask \ | |
| 3219 | ||
| 3220 | ||
| 3221 | #define shade_blocks_textured_modulated_builder(shading, dithering, target) \ | |
| 3222 | .align 3; \ | |
| 3223 | \ | |
| 3224 | function(shade_blocks_##shading##_textured_modulated_##dithering##_##target) \ | |
| fab27ba2 | 3225 | save_abi_regs(); \ |
| 3867c6ef | 3226 | shade_blocks_textured_modulated_prologue_##shading(dithering, target); \ |
| 75e28f62 | 3227 | stmdb sp!, { r4 - r5, lr }; \ |
| e1f6de8f | 3228 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 | 3229 | \ |
| e1f6de8f | 3230 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
3231 | \ |
| 3232 | shade_blocks_textured_modulated_prologue_##target(); \ | |
| 75e28f62 E |
3233 | \ |
| 3234 | add block_ptr_load_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3235 | mov c_32, #32; \ | |
| 3236 | \ | |
| 3237 | add block_ptr_load_b, block_ptr_load_a, #16; \ | |
| 3238 | vmov.u8 d64_1, #1; \ | |
| 3239 | vmov.u8 d64_4, #4; \ | |
| 3240 | vmov.u8 d64_128, #128; \ | |
| 3241 | \ | |
| e1f6de8f | 3242 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
| 75e28f62 E |
3243 | vmov.u8 d128_0x07, #0x07; \ |
| 3244 | \ | |
| 3245 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
| 3246 | vmov.u8 d128_0x1F, #0x1F; \ | |
| 3247 | \ | |
| 3248 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
| 3249 | vmov.u16 d128_0x8000, #0x8000; \ | |
| 3250 | \ | |
| 3251 | vmovn.u16 texels_r, texels; \ | |
| 3252 | vshrn.u16 texels_g, texels, #5; \ | |
| 3253 | \ | |
| 3254 | vshrn.u16 texels_b, texels, #7; \ | |
| 3255 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
| 3256 | \ | |
| 3257 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
| 3258 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 3259 | \ | |
| 3260 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
| 3261 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
| 3262 | \ | |
| 3263 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
| 3264 | vshr.u8 texels_b, texels_b, #3; \ | |
| 3265 | \ | |
| 3266 | shade_blocks_textured_modulate_##dithering(r); \ | |
| 3267 | shade_blocks_textured_modulate_##dithering(g); \ | |
| 3268 | shade_blocks_textured_modulate_##dithering(b); \ | |
| 3269 | \ | |
| 3270 | vand.u16 pixels, texels, d128_0x8000; \ | |
| 3271 | vceq.u16 zero_mask, texels, #0; \ | |
| 3272 | \ | |
| 3273 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
| 3274 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
| 3275 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
| 3276 | \ | |
| 3277 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
| 3278 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
| 3279 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
| 3280 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
| 3281 | \ | |
| 3282 | subs num_blocks, num_blocks, #1; \ | |
| 3283 | beq 1f; \ | |
| 3284 | \ | |
| 3285 | .align 3; \ | |
| 3286 | \ | |
| 3287 | 0: \ | |
| e1f6de8f | 3288 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
| 75e28f62 E |
3289 | shade_blocks_textured_modulated_load_rg_##shading(); \ |
| 3290 | vshrn.u16 texels_g, texels, #5; \ | |
| 3291 | \ | |
| 3292 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
| 3293 | vshrn.u16 texels_b, texels, #7; \ | |
| 3294 | \ | |
| e1f6de8f | 3295 | pld [block_ptr_load_a]; \ |
| 75e28f62 E |
3296 | vmovn.u16 texels_r, texels; \ |
| 3297 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
| 3298 | \ | |
| 3299 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
| 3300 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
| 3301 | shade_blocks_textured_modulated_store_draw_mask_##target(-4); \ | |
| 3302 | \ | |
| 3303 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
| 3304 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
| 3305 | \ | |
| 3306 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
| 3307 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
| 3308 | \ | |
| 3309 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
| 3310 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 3311 | \ | |
| 3312 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
| 3313 | vshr.u8 texels_b, texels_b, #3; \ | |
| 3314 | \ | |
| 3315 | shade_blocks_textured_modulate_##dithering(r); \ | |
| 3316 | shade_blocks_textured_modulate_##dithering(g); \ | |
| 3317 | shade_blocks_textured_modulate_##dithering(b); \ | |
| 3318 | \ | |
| 3319 | vand.u16 pixels, texels, d128_0x8000; \ | |
| 3320 | vceq.u16 zero_mask, texels, #0; \ | |
| 3321 | \ | |
| 3322 | subs num_blocks, num_blocks, #1; \ | |
| 3323 | \ | |
| 3324 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
| 3325 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
| 3326 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
| 3327 | \ | |
| 3328 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
| 3329 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
| 3330 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
| 3331 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
| 3332 | \ | |
| 3333 | bne 0b; \ | |
| 3334 | \ | |
| 3335 | 1: \ | |
| 3336 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
| 3337 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
| 3338 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
| 3339 | \ | |
| 3340 | shade_blocks_textured_modulated_store_draw_mask_##target(28); \ | |
| 3341 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
| 3342 | \ | |
| fab27ba2 | 3343 | ldmia sp!, { r4 - r5, lr }; \ |
| 3f0189c6 | 3344 | restore_abi_regs(); \ |
| fab27ba2 | 3345 | bx lr \ |
| 75e28f62 E |
3346 | |
| 3347 | ||
| 3348 | shade_blocks_textured_modulated_builder(shaded, dithered, direct); | |
| 3349 | shade_blocks_textured_modulated_builder(shaded, undithered, direct); | |
| 3350 | shade_blocks_textured_modulated_builder(unshaded, dithered, direct); | |
| 3351 | shade_blocks_textured_modulated_builder(unshaded, undithered, direct); | |
| 3352 | ||
| 3353 | shade_blocks_textured_modulated_builder(shaded, dithered, indirect); | |
| 3354 | shade_blocks_textured_modulated_builder(shaded, undithered, indirect); | |
| 3355 | shade_blocks_textured_modulated_builder(unshaded, dithered, indirect); | |
| 3356 | shade_blocks_textured_modulated_builder(unshaded, undithered, indirect); | |
| 3357 | ||
| 3358 | ||
| 3359 | #undef c_64 | |
| 3360 | #undef fb_ptr | |
| 3361 | #undef color_ptr | |
| 3362 | ||
| 3363 | #undef color_r | |
| 3364 | #undef color_g | |
| 3365 | #undef color_b | |
| 3366 | ||
| 3367 | #undef test_mask | |
| 3368 | #undef pixels | |
| 3369 | #undef draw_mask | |
| 3370 | #undef zero_mask | |
| 3371 | #undef fb_pixels | |
| 3372 | #undef msb_mask | |
| 3373 | #undef msb_mask_low | |
| 3374 | #undef msb_mask_high | |
| 3375 | ||
| 3376 | #define psx_gpu r0 | |
| 3377 | #define num_blocks r1 | |
| 3378 | #define mask_msb_ptr r2 | |
| 3379 | #define color_ptr r3 | |
| 3380 | ||
| 3381 | #define block_ptr_load r0 | |
| 3382 | #define draw_mask_store_ptr r3 | |
| 3383 | #define draw_mask_bits_ptr r12 | |
| 3384 | #define draw_mask_ptr r12 | |
| 3385 | #define pixel_store_ptr r14 | |
| 3386 | ||
| 3387 | #define fb_ptr_cmp r4 | |
| 3388 | ||
| 3389 | #define fb_ptr r3 | |
| 3390 | #define fb_ptr_next r14 | |
| 3391 | ||
| 3392 | #define c_64 r2 | |
| 3393 | ||
| 3394 | #define test_mask q0 | |
| 3395 | #define pixels q1 | |
| 3396 | #define draw_mask q2 | |
| 3397 | #define zero_mask q3 | |
| 3398 | #define draw_mask_combined q4 | |
| 3399 | #define fb_pixels q5 | |
| 3400 | #define fb_pixels_next q6 | |
| 3401 | #define msb_mask q7 | |
| 3402 | ||
| 3403 | #define draw_mask_low d4 | |
| 3404 | #define draw_mask_high d5 | |
| 3405 | #define msb_mask_low d14 | |
| 3406 | #define msb_mask_high d15 | |
| 3407 | ||
| 3408 | .align 3 | |
| 3409 | function(shade_blocks_textured_unmodulated_indirect) | |
| 3f0189c6 | 3410 | stmdb sp!, { r4, r14 } |
| 3411 | save_abi_regs() | |
| 75e28f62 E |
3412 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
| 3413 | ||
| e1f6de8f | 3414 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
3415 | add pixel_store_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
| 3416 | ||
| e1f6de8f | 3417 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 E |
3418 | add draw_mask_store_ptr, psx_gpu, #psx_gpu_blocks_offset |
| 3419 | ||
| 3420 | mov c_64, #64 | |
| 3421 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
| 3422 | ||
| e1f6de8f | 3423 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 75e28f62 | 3424 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ |
| e1f6de8f | 3425 | [draw_mask_bits_ptr, :16], c_64 |
| 75e28f62 E |
3426 | vceq.u16 zero_mask, pixels, #0 |
| 3427 | ||
| 3428 | vtst.u16 draw_mask, draw_mask, test_mask | |
| e1f6de8f | 3429 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
| 75e28f62 E |
3430 | |
| 3431 | subs num_blocks, num_blocks, #1 | |
| 3432 | beq 1f | |
| 3433 | ||
| 3434 | 0: | |
| e1f6de8f | 3435 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 75e28f62 E |
3436 | vorr.u16 draw_mask_combined, draw_mask, zero_mask |
| 3437 | ||
| 3438 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| e1f6de8f | 3439 | [draw_mask_bits_ptr, :16], c_64 |
| 75e28f62 E |
3440 | vceq.u16 zero_mask, pixels, #0 |
| 3441 | ||
| 3442 | vtst.u16 draw_mask, draw_mask, test_mask | |
| e1f6de8f | 3443 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
| 75e28f62 | 3444 | |
| e1f6de8f | 3445 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
| 75e28f62 E |
3446 | subs num_blocks, num_blocks, #1 |
| 3447 | ||
| 3448 | bne 0b | |
| 3449 | ||
| 3450 | 1: | |
| 3451 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| e1f6de8f | 3452 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
| 75e28f62 | 3453 | |
| 3f0189c6 | 3454 | restore_abi_regs() |
| 3455 | ldmia sp!, { r4, pc } | |
| 75e28f62 E |
3456 | |
| 3457 | ||
| 3458 | .align 3 | |
| 3459 | ||
| 3460 | function(shade_blocks_textured_unmodulated_direct) | |
| 3461 | stmdb sp!, { r4, r14 } | |
| 3f0189c6 | 3462 | save_abi_regs() |
| 75e28f62 E |
3463 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
| 3464 | ||
| e1f6de8f | 3465 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
3466 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
| 3467 | ||
| e1f6de8f | 3468 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
| 75e28f62 E |
3469 | mov c_64, #64 |
| 3470 | ||
| e1f6de8f | 3471 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 E |
3472 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset |
| 3473 | ||
| 3474 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| e1f6de8f | 3475 | [draw_mask_bits_ptr, :16], c_64 |
| 3476 | ldr fb_ptr_next, [block_ptr_load, #44] | |
| 75e28f62 | 3477 | |
| e1f6de8f | 3478 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 3479 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
| 75e28f62 E |
3480 | vceq.u16 zero_mask, pixels, #0 |
| 3481 | vtst.u16 draw_mask, draw_mask, test_mask | |
| 3482 | ||
| 3483 | subs num_blocks, num_blocks, #1 | |
| 3484 | beq 1f | |
| 3485 | ||
| 3486 | 0: | |
| 3487 | mov fb_ptr, fb_ptr_next | |
| e1f6de8f | 3488 | ldr fb_ptr_next, [block_ptr_load, #44] |
| 75e28f62 E |
3489 | |
| 3490 | vorr.u16 pixels, pixels, msb_mask | |
| 3491 | ||
| 3492 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3493 | vmov fb_pixels, fb_pixels_next | |
| 3494 | ||
| 3495 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| e1f6de8f | 3496 | [draw_mask_bits_ptr, :16], c_64 |
| 75e28f62 E |
3497 | vbif.u16 fb_pixels, pixels, draw_mask_combined |
| 3498 | ||
| 75e28f62 | 3499 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr |
| e1f6de8f | 3500 | pld [fb_ptr_next, #64] |
| 8438c3c7 | 3501 | |
| 75e28f62 | 3502 | add fb_ptr_cmp, fb_ptr_cmp, #14 |
| e1f6de8f | 3503 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 8438c3c7 | 3504 | |
| 75e28f62 E |
3505 | cmp fb_ptr_cmp, #28 |
| 3506 | bls 4f | |
| 3507 | ||
| e1f6de8f | 3508 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 E |
3509 | vceq.u16 zero_mask, pixels, #0 |
| 3510 | ||
| e1f6de8f | 3511 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 E |
3512 | vtst.u16 draw_mask, draw_mask, test_mask |
| 3513 | ||
| 3514 | 3: | |
| 3515 | subs num_blocks, num_blocks, #1 | |
| 3516 | bne 0b | |
| 3517 | ||
| 3518 | 1: | |
| 3519 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3520 | vbif.u16 fb_pixels_next, pixels, draw_mask_combined | |
| 3521 | ||
| e1f6de8f | 3522 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 | 3523 | |
| 3f0189c6 | 3524 | restore_abi_regs() |
| 75e28f62 E |
3525 | ldmia sp!, { r4, pc } |
| 3526 | ||
| 3527 | 4: | |
| e1f6de8f | 3528 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 E |
3529 | vceq.u16 zero_mask, pixels, #0 |
| 3530 | ||
| e1f6de8f | 3531 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 E |
3532 | vtst.u16 draw_mask, draw_mask, test_mask |
| 3533 | ||
| 3534 | bal 3b | |
| 3535 | ||
| 3536 | ||
| 3537 | function(shade_blocks_unshaded_untextured_indirect) | |
| 3538 | bx lr | |
| 3539 | ||
| 3540 | .align 3 | |
| 3541 | ||
| 3542 | function(shade_blocks_unshaded_untextured_direct) | |
| 3543 | stmdb sp!, { r4, r14 } | |
| 3f0189c6 | 3544 | save_abi_regs() |
| 75e28f62 E |
3545 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset |
| 3546 | ||
| e1f6de8f | 3547 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
3548 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
| 3549 | ||
| e1f6de8f | 3550 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
| 75e28f62 E |
3551 | add color_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
| 3552 | ||
| 3553 | add block_ptr_load, psx_gpu, #(psx_gpu_blocks_offset + 44) | |
| e1f6de8f | 3554 | vld1.u16 { pixels }, [color_ptr, :128] |
| 75e28f62 E |
3555 | |
| 3556 | mov c_64, #64 | |
| e1f6de8f | 3557 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
| 75e28f62 E |
3558 | |
| 3559 | vorr.u16 pixels, pixels, msb_mask | |
| 3560 | subs num_blocks, num_blocks, #1 | |
| 3561 | ||
| e1f6de8f | 3562 | ldr fb_ptr_next, [block_ptr_load], #64 |
| 75e28f62 | 3563 | |
| e1f6de8f | 3564 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 E |
3565 | beq 1f |
| 3566 | ||
| 3567 | 0: | |
| 3568 | vmov fb_pixels, fb_pixels_next | |
| 3569 | mov fb_ptr, fb_ptr_next | |
| e1f6de8f | 3570 | ldr fb_ptr_next, [block_ptr_load], #64 |
| 75e28f62 E |
3571 | |
| 3572 | vbif.u16 fb_pixels, pixels, draw_mask | |
| e1f6de8f | 3573 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
| 75e28f62 E |
3574 | |
| 3575 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr | |
| 3576 | add fb_ptr_cmp, fb_ptr_cmp, #14 | |
| 3577 | cmp fb_ptr_cmp, #28 | |
| 3578 | bls 4f | |
| 3579 | ||
| e1f6de8f | 3580 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 3581 | vst1.u16 { fb_pixels }, [fb_ptr] | |
| 75e28f62 E |
3582 | |
| 3583 | 3: | |
| 3584 | subs num_blocks, num_blocks, #1 | |
| 3585 | bne 0b | |
| 3586 | ||
| 3587 | 1: | |
| 3588 | vbif.u16 fb_pixels_next, pixels, draw_mask | |
| e1f6de8f | 3589 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 | 3590 | |
| 3f0189c6 | 3591 | restore_abi_regs() |
| 75e28f62 E |
3592 | ldmia sp!, { r4, pc } |
| 3593 | ||
| 3594 | 4: | |
| e1f6de8f | 3595 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 3596 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
| 75e28f62 E |
3597 | bal 3b |
| 3598 | ||
| 3599 | ||
| 3600 | #undef draw_mask_ptr | |
| 3601 | #undef c_64 | |
| 3602 | #undef fb_ptr | |
| 3603 | #undef fb_ptr_next | |
| 3604 | #undef fb_ptr_cmp | |
| 3605 | ||
| 3606 | #define psx_gpu r0 | |
| 3607 | #define num_blocks r1 | |
| 3608 | #define msb_mask_ptr r2 | |
| 3609 | #define pixel_ptr r3 | |
| 3610 | #define draw_mask_ptr r0 | |
| 3611 | #define c_64 r2 | |
| 3612 | #define fb_ptr r12 | |
| 3613 | #define fb_ptr_next r14 | |
| 3614 | #define fb_ptr_cmp r4 | |
| 3615 | ||
| 3616 | #undef msb_mask | |
| 3617 | #undef draw_mask | |
| 3618 | #undef pixels | |
| 3619 | #undef fb_pixels | |
| 3620 | #undef d128_0x8000 | |
| 3621 | #undef msb_mask_low | |
| 3622 | #undef msb_mask_high | |
| 3623 | #undef draw_mask_next | |
| 3624 | #undef pixels_g | |
| 3625 | #undef blend_pixels | |
| 3626 | #undef fb_pixels_next | |
| 3627 | ||
| 3628 | #define msb_mask q0 | |
| 3629 | #define draw_mask q1 | |
| 3630 | #define pixels q2 | |
| 3631 | #define fb_pixels q3 | |
| 3632 | #define blend_pixels q4 | |
| 3633 | #define pixels_no_msb q5 | |
| 3634 | #define blend_mask q6 | |
| 3635 | #define fb_pixels_no_msb q7 | |
| 3636 | #define d128_0x8000 q8 | |
| 3637 | #define d128_0x0421 q9 | |
| 3638 | #define fb_pixels_next q10 | |
| 3639 | #define blend_pixels_next q11 | |
| 3640 | #define pixels_next q12 | |
| 3641 | #define draw_mask_next q13 | |
| 3642 | #define write_mask q14 | |
| 3643 | ||
| 3644 | #define pixels_rb q5 | |
| 3645 | #define pixels_mg q7 | |
| 3646 | #define pixels_g q7 | |
| 3647 | #define d128_0x7C1F q8 | |
| 3648 | #define d128_0x03E0 q9 | |
| 3649 | #define fb_pixels_rb q10 | |
| 3650 | #define fb_pixels_g q11 | |
| 3651 | #define fb_pixels_masked q11 | |
| 3652 | #define d128_0x83E0 q15 | |
| 3653 | #define pixels_fourth q7 | |
| 3654 | #define d128_0x1C07 q12 | |
| 3655 | #define d128_0x00E0 q13 | |
| 3656 | #define d128_0x80E0 q13 | |
| 3657 | ||
| 3658 | #define msb_mask_low d0 | |
| 3659 | #define msb_mask_high d1 | |
| 3660 | ||
| 3661 | #define blend_blocks_average_set_blend_mask_textured(source) \ | |
| 3662 | vclt.s16 blend_mask, source, #0 \ | |
| 3663 | ||
| 3664 | #define blend_blocks_average_set_stp_bit_textured() \ | |
| 3665 | vorr.u16 blend_pixels, #0x8000 \ | |
| 3666 | ||
| 3667 | #define blend_blocks_average_combine_textured(source) \ | |
| 3668 | vbif.u16 blend_pixels, source, blend_mask \ | |
| 3669 | ||
| 3670 | #define blend_blocks_average_set_blend_mask_untextured(source) \ | |
| 3671 | ||
| 3672 | #define blend_blocks_average_set_stp_bit_untextured() \ | |
| 3673 | ||
| 3674 | #define blend_blocks_average_combine_untextured(source) \ | |
| 3675 | ||
| 3676 | #define blend_blocks_average_mask_set_on() \ | |
| 3677 | vclt.s16 write_mask, fb_pixels_next, #0 \ | |
| 3678 | ||
| 3679 | #define blend_blocks_average_mask_copy_on() \ | |
| 3680 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
| 3681 | ||
| 3682 | #define blend_blocks_average_mask_copy_b_on() \ | |
| 3683 | vorr.u16 draw_mask_next, draw_mask_next, write_mask \ | |
| 3684 | ||
| 3685 | #define blend_blocks_average_mask_set_off() \ | |
| 3686 | ||
| 3687 | #define blend_blocks_average_mask_copy_off() \ | |
| 3688 | vmov draw_mask, draw_mask_next \ | |
| 3689 | ||
| 3690 | #define blend_blocks_average_mask_copy_b_off() \ | |
| 3691 | ||
| 3692 | #define blend_blocks_average_builder(texturing, mask_evaluate) \ | |
| 3693 | .align 3; \ | |
| 3694 | \ | |
| 3695 | function(blend_blocks_##texturing##_average_##mask_evaluate) \ | |
| 3696 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 3697 | save_abi_regs(); \ |
| 75e28f62 | 3698 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 3699 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
3700 | \ |
| 3701 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 3702 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
3703 | \ |
| 3704 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3705 | mov c_64, #64; \ | |
| 3706 | \ | |
| 3707 | vmov.u16 d128_0x8000, #0x8000; \ | |
| e1f6de8f | 3708 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 3709 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 75e28f62 E |
3710 | \ |
| 3711 | vmov.u16 d128_0x0421, #0x0400; \ | |
| e1f6de8f | 3712 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3713 | \ |
| 3714 | vorr.u16 d128_0x0421, #0x0021; \ | |
| e1f6de8f | 3715 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 E |
3716 | \ |
| 3717 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
| 3718 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3719 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3720 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3721 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
| 3722 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3723 | \ | |
| 3724 | subs num_blocks, num_blocks, #1; \ | |
| 3725 | beq 1f; \ | |
| 3726 | \ | |
| 3727 | 0: \ | |
| 3728 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 3729 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 E |
3730 | \ |
| 3731 | vmov pixels, pixels_next; \ | |
| e1f6de8f | 3732 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3733 | \ |
| 3734 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
| 3735 | \ | |
| 3736 | blend_blocks_average_mask_copy_##mask_evaluate(); \ | |
| e1f6de8f | 3737 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
3738 | \ |
| 3739 | blend_blocks_average_set_blend_mask_##texturing(pixels); \ | |
| 3740 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
| 3741 | vmov fb_pixels, fb_pixels_next; \ | |
| 3742 | blend_blocks_average_combine_##texturing(pixels); \ | |
| 3743 | \ | |
| 3744 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3745 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3746 | cmp fb_ptr_cmp, #28; \ | |
| 3747 | bls 2f; \ | |
| 3748 | \ | |
| e1f6de8f | 3749 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 E |
3750 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
| 3751 | \ | |
| 3752 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3753 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3754 | \ | |
| 3755 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3756 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
| 3757 | \ | |
| 3758 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3759 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3760 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
| e1f6de8f | 3761 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3762 | \ |
| 3763 | 3: \ | |
| 3764 | subs num_blocks, num_blocks, #1; \ | |
| 3765 | bne 0b; \ | |
| 3766 | \ | |
| 3767 | 1: \ | |
| 3768 | blend_blocks_average_mask_copy_b_##mask_evaluate(); \ | |
| 3769 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
| 3770 | \ | |
| 3771 | blend_blocks_average_set_blend_mask_##texturing(pixels_next); \ | |
| 3772 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
| 3773 | blend_blocks_average_combine_##texturing(pixels_next); \ | |
| 3774 | \ | |
| 3775 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3776 | vbif.u16 fb_pixels_next, blend_pixels, draw_mask_next; \ | |
| e1f6de8f | 3777 | vst1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 | 3778 | \ |
| 3f0189c6 | 3779 | restore_abi_regs(); \ |
| 75e28f62 E |
3780 | ldmia sp!, { r4, pc }; \ |
| 3781 | \ | |
| 3782 | 2: \ | |
| 3783 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3784 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
| e1f6de8f | 3785 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
| 75e28f62 | 3786 | \ |
| e1f6de8f | 3787 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 E |
3788 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
| 3789 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3790 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3791 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3792 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3793 | \ | |
| 3794 | bal 3b \ | |
| 3795 | ||
| 3796 | blend_blocks_average_builder(textured, off) | |
| 3797 | blend_blocks_average_builder(untextured, off) | |
| 3798 | blend_blocks_average_builder(textured, on) | |
| 3799 | blend_blocks_average_builder(untextured, on) | |
| 3800 | ||
| 3801 | ||
| 3802 | #define blend_blocks_add_mask_set_on() \ | |
| 3803 | vclt.s16 write_mask, fb_pixels, #0 \ | |
| 3804 | ||
| 3805 | #define blend_blocks_add_mask_copy_on() \ | |
| 3806 | vorr.u16 draw_mask, draw_mask, write_mask \ | |
| 3807 | ||
| 3808 | #define blend_blocks_add_mask_set_off() \ | |
| 3809 | ||
| 3810 | #define blend_blocks_add_mask_copy_off() \ | |
| 3811 | ||
| 3812 | ||
| 3813 | #define blend_blocks_add_textured_builder(mask_evaluate) \ | |
| 3814 | .align 3; \ | |
| 3815 | \ | |
| 3816 | function(blend_blocks_textured_add_##mask_evaluate) \ | |
| 3817 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 3818 | save_abi_regs(); \ |
| 75e28f62 | 3819 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 3820 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
3821 | \ |
| 3822 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 3823 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
3824 | \ |
| 3825 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3826 | mov c_64, #64; \ | |
| 3827 | \ | |
| 3828 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 3829 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 3830 | vmov.u16 d128_0x83E0, #0x8000; \ | |
| 3831 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 3832 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 3833 | vorr.u16 d128_0x83E0, d128_0x83E0, d128_0x03E0; \ | |
| 3834 | \ | |
| e1f6de8f | 3835 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 3836 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 3837 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 75e28f62 | 3838 | vclt.s16 blend_mask, pixels, #0; \ |
| e1f6de8f | 3839 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3840 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3841 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3842 | \ | |
| 3843 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3844 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 3845 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3846 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
| 3847 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 3848 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 3849 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3850 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 3851 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3852 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 3853 | \ | |
| 3854 | subs num_blocks, num_blocks, #1; \ | |
| 3855 | beq 1f; \ | |
| 3856 | \ | |
| 3857 | 0: \ | |
| 3858 | mov fb_ptr, fb_ptr_next; \ | |
| 3859 | \ | |
| e1f6de8f | 3860 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 3861 | \ |
| e1f6de8f | 3862 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3863 | vclt.s16 blend_mask, pixels, #0; \ |
| 3864 | \ | |
| 3865 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 3866 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3867 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
| 3868 | \ | |
| 8438c3c7 | 3869 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ |
| e1f6de8f | 3870 | pld [fb_ptr_next, #64]; \ |
| 75e28f62 E |
3871 | \ |
| 3872 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 8438c3c7 | 3873 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
| 3874 | \ | |
| 75e28f62 | 3875 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ |
| e1f6de8f | 3876 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 8438c3c7 | 3877 | \ |
| 75e28f62 E |
3878 | cmp fb_ptr_cmp, #28; \ |
| 3879 | bls 2f; \ | |
| 3880 | \ | |
| e1f6de8f | 3881 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3882 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3883 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3884 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3885 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3886 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| e1f6de8f | 3887 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3888 | \ |
| 3889 | 3: \ | |
| 3890 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 3891 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3892 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 3893 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3894 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 3895 | \ | |
| 3896 | subs num_blocks, num_blocks, #1; \ | |
| 3897 | bne 0b; \ | |
| 3898 | \ | |
| 3899 | 1: \ | |
| 3900 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3901 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 3902 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 3903 | \ |
| 3f0189c6 | 3904 | restore_abi_regs(); \ |
| 75e28f62 E |
3905 | ldmia sp!, { r4, pc }; \ |
| 3906 | \ | |
| 3907 | 2: \ | |
| e1f6de8f | 3908 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3909 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
| 3910 | \ | |
| e1f6de8f | 3911 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3912 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3913 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3914 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3915 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 3916 | bal 3b \ | |
| 3917 | ||
| 3918 | ||
| 3919 | #define blend_blocks_add_untextured_builder(mask_evaluate) \ | |
| 3920 | .align 3; \ | |
| 3921 | \ | |
| 3922 | function(blend_blocks_untextured_add_##mask_evaluate) \ | |
| 3923 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 3924 | save_abi_regs(); \ |
| 75e28f62 | 3925 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 3926 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
3927 | \ |
| 3928 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 3929 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
3930 | \ |
| 3931 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3932 | mov c_64, #64; \ | |
| 3933 | \ | |
| 3934 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 3935 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 3936 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 3937 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 3938 | \ | |
| e1f6de8f | 3939 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 3940 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 3941 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 3942 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
| 75e28f62 E |
3943 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3944 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3945 | \ | |
| 3946 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3947 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
| 3948 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3949 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3950 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3951 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3952 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3953 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 3954 | \ | |
| 3955 | subs num_blocks, num_blocks, #1; \ | |
| 3956 | beq 1f; \ | |
| 3957 | \ | |
| 3958 | 0: \ | |
| 3959 | mov fb_ptr, fb_ptr_next; \ | |
| 3960 | \ | |
| e1f6de8f | 3961 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 3962 | \ |
| e1f6de8f | 3963 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3964 | \ |
| 3965 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3966 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3967 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
| 3968 | \ | |
| 3969 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 3970 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
3971 | \ |
| 3972 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3973 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3974 | cmp fb_ptr_cmp, #28; \ | |
| 3975 | bls 2f; \ | |
| 3976 | \ | |
| e1f6de8f | 3977 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3978 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3979 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3980 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3981 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| e1f6de8f | 3982 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3983 | \ |
| 3984 | 3: \ | |
| 3985 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3986 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3987 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3988 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3989 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 3990 | \ | |
| 3991 | subs num_blocks, num_blocks, #1; \ | |
| 3992 | bne 0b; \ | |
| 3993 | \ | |
| 3994 | 1: \ | |
| 3995 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3996 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3997 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 3998 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 3999 | \ |
| 3f0189c6 | 4000 | restore_abi_regs(); \ |
| 75e28f62 E |
4001 | ldmia sp!, { r4, pc }; \ |
| 4002 | \ | |
| 4003 | 2: \ | |
| e1f6de8f | 4004 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4005 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
| 4006 | \ | |
| e1f6de8f | 4007 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4008 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4009 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4010 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4011 | bal 3b \ | |
| 4012 | ||
| 4013 | ||
| 4014 | blend_blocks_add_textured_builder(off) | |
| 4015 | blend_blocks_add_textured_builder(on) | |
| 4016 | blend_blocks_add_untextured_builder(off) | |
| 4017 | blend_blocks_add_untextured_builder(on) | |
| 4018 | ||
| 4019 | #define blend_blocks_subtract_set_blend_mask_textured() \ | |
| 4020 | vclt.s16 blend_mask, pixels_next, #0 \ | |
| 4021 | ||
| 4022 | #define blend_blocks_subtract_combine_textured() \ | |
| 4023 | vbif.u16 blend_pixels, pixels, blend_mask \ | |
| 4024 | ||
| 718a9e58 | 4025 | #define blend_blocks_subtract_set_stp_textured() \ |
| 75e28f62 E |
4026 | vorr.u16 blend_pixels, #0x8000 \ |
| 4027 | ||
| 4028 | #define blend_blocks_subtract_msb_mask_textured() \ | |
| 4029 | vorr.u16 pixels, pixels_next, msb_mask \ | |
| 4030 | ||
| 4031 | #define blend_blocks_subtract_set_blend_mask_untextured() \ | |
| 4032 | ||
| 4033 | #define blend_blocks_subtract_combine_untextured() \ | |
| 4034 | ||
| 718a9e58 | 4035 | #define blend_blocks_subtract_set_stp_untextured() \ |
| 75e28f62 E |
4036 | vorr.u16 blend_pixels, blend_pixels, msb_mask \ |
| 4037 | ||
| 4038 | #define blend_blocks_subtract_msb_mask_untextured() \ | |
| 4039 | ||
| 4040 | ||
| 4041 | #define blend_blocks_subtract_mask_set_on() \ | |
| 4042 | vclt.s16 write_mask, fb_pixels, #0 \ | |
| 4043 | ||
| 4044 | #define blend_blocks_subtract_mask_copy_on() \ | |
| 4045 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
| 4046 | ||
| 4047 | #define blend_blocks_subtract_mask_set_off() \ | |
| 4048 | ||
| 4049 | #define blend_blocks_subtract_mask_copy_off() \ | |
| 4050 | vmov draw_mask, draw_mask_next \ | |
| 4051 | ||
| 4052 | ||
| 4053 | #define blend_blocks_subtract_builder(texturing, mask_evaluate) \ | |
| 4054 | .align 3; \ | |
| 4055 | \ | |
| 4056 | function(blend_blocks_##texturing##_subtract_##mask_evaluate) \ | |
| 4057 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 4058 | save_abi_regs(); \ |
| 75e28f62 | 4059 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 4060 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
4061 | \ |
| 4062 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 4063 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
4064 | \ |
| 4065 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4066 | mov c_64, #64; \ | |
| 4067 | \ | |
| 4068 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4069 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 4070 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 4071 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 4072 | \ | |
| e1f6de8f | 4073 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 4074 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 4075 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ | |
| 75e28f62 | 4076 | blend_blocks_subtract_set_blend_mask_##texturing(); \ |
| e1f6de8f | 4077 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4078 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
| 4079 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
| 4080 | \ | |
| 4081 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
| 4082 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4083 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4084 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4085 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4086 | \ | |
| 4087 | subs num_blocks, num_blocks, #1; \ | |
| 4088 | beq 1f; \ | |
| 4089 | \ | |
| 4090 | 0: \ | |
| 4091 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
| 4092 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 4093 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 4094 | \ |
| e1f6de8f | 4095 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
4096 | blend_blocks_subtract_msb_mask_##texturing(); \ |
| 4097 | \ | |
| e1f6de8f | 4098 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
4099 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
| 4100 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
| 718a9e58 | 4101 | blend_blocks_subtract_set_stp_##texturing(); \ |
| 75e28f62 E |
4102 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ |
| 4103 | blend_blocks_subtract_combine_##texturing(); \ | |
| 4104 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
| 4105 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 4106 | \ | |
| 4107 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4108 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4109 | cmp fb_ptr_cmp, #28; \ | |
| 4110 | bls 2f; \ | |
| 4111 | \ | |
| e1f6de8f | 4112 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4113 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
| 4114 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4115 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4116 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| e1f6de8f | 4117 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4118 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
| 4119 | \ | |
| 4120 | 3: \ | |
| 4121 | subs num_blocks, num_blocks, #1; \ | |
| 4122 | bne 0b; \ | |
| 4123 | \ | |
| 4124 | 1: \ | |
| 4125 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
| 4126 | \ | |
| 4127 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
| 4128 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 718a9e58 | 4129 | blend_blocks_subtract_set_stp_##texturing(); \ |
| 75e28f62 E |
4130 | blend_blocks_subtract_combine_##texturing(); \ |
| 4131 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4132 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4133 | \ |
| 3f0189c6 | 4134 | restore_abi_regs(); \ |
| 75e28f62 E |
4135 | ldmia sp!, { r4, pc }; \ |
| 4136 | \ | |
| 4137 | 2: \ | |
| e1f6de8f | 4138 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 4139 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
| 75e28f62 E |
4140 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
| 4141 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4142 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4143 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4144 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4145 | bal 3b \ | |
| 4146 | ||
| 4147 | ||
| 4148 | blend_blocks_subtract_builder(textured, off) | |
| 4149 | blend_blocks_subtract_builder(textured, on) | |
| 4150 | blend_blocks_subtract_builder(untextured, off) | |
| 4151 | blend_blocks_subtract_builder(untextured, on) | |
| 4152 | ||
| 4153 | ||
| 4154 | #define blend_blocks_add_fourth_textured_builder(mask_evaluate) \ | |
| 4155 | .align 3; \ | |
| 4156 | \ | |
| 4157 | function(blend_blocks_textured_add_fourth_##mask_evaluate) \ | |
| 4158 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 4159 | save_abi_regs(); \ |
| 75e28f62 | 4160 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 4161 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
4162 | \ |
| 4163 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 4164 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
4165 | \ |
| 4166 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4167 | mov c_64, #64; \ | |
| 4168 | \ | |
| 4169 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4170 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 75e28f62 | 4171 | vmov.u16 d128_0x1C07, #0x1C00; \ |
| d1c75d1e | 4172 | vmov.u16 d128_0x00E0, #0x00E0; \ |
| 75e28f62 E |
4173 | vorr.u16 d128_0x7C1F, #0x001F; \ |
| 4174 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 75e28f62 | 4175 | vorr.u16 d128_0x1C07, #0x0007; \ |
| 75e28f62 | 4176 | \ |
| e1f6de8f | 4177 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 4178 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 4179 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 75e28f62 | 4180 | vclt.s16 blend_mask, pixels, #0; \ |
| e1f6de8f | 4181 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4182 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4183 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| d1c75d1e | 4184 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
| 75e28f62 E |
4185 | \ |
| 4186 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| d1c75d1e E |
4187 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 4188 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4189 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 75e28f62 | 4190 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
| d1c75d1e | 4191 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
| 75e28f62 | 4192 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
| d1c75d1e | 4193 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
| 75e28f62 E |
4194 | \ |
| 4195 | subs num_blocks, num_blocks, #1; \ | |
| 4196 | beq 1f; \ | |
| 4197 | \ | |
| 4198 | 0: \ | |
| 4199 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 4200 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 4201 | \ |
| d1c75d1e | 4202 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
| 718a9e58 | 4203 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
| d1c75d1e E |
4204 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
| 4205 | \ | |
| e1f6de8f | 4206 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 | 4207 | vclt.s16 blend_mask, pixels, #0; \ |
| 75e28f62 | 4208 | vshr.s16 pixels_fourth, pixels, #2; \ |
| d1c75d1e | 4209 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
| 75e28f62 E |
4210 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
| 4211 | \ | |
| 4212 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4213 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
4214 | \ |
| 4215 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4216 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4217 | cmp fb_ptr_cmp, #28; \ | |
| 4218 | bls 2f; \ | |
| 4219 | \ | |
| e1f6de8f | 4220 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4221 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 75e28f62 | 4222 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
| d1c75d1e E |
4223 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 4224 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| e1f6de8f | 4225 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4226 | \ |
| 4227 | 3: \ | |
| d1c75d1e | 4228 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ |
| 75e28f62 | 4229 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
| d1c75d1e | 4230 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
| 75e28f62 | 4231 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
| d1c75d1e | 4232 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
| 75e28f62 E |
4233 | \ |
| 4234 | subs num_blocks, num_blocks, #1; \ | |
| 4235 | bne 0b; \ | |
| 4236 | \ | |
| 4237 | 1: \ | |
| 4238 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 718a9e58 | 4239 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
| d1c75d1e | 4240 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
| 718a9e58 | 4241 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
| 75e28f62 | 4242 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
| e1f6de8f | 4243 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4244 | \ |
| 3f0189c6 | 4245 | restore_abi_regs(); \ |
| 75e28f62 E |
4246 | ldmia sp!, { r4, pc }; \ |
| 4247 | \ | |
| 4248 | 2: \ | |
| e1f6de8f | 4249 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| d1c75d1e | 4250 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 75e28f62 | 4251 | \ |
| e1f6de8f | 4252 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4253 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 75e28f62 | 4254 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
| d1c75d1e | 4255 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ |
| 75e28f62 E |
4256 | bal 3b \ |
| 4257 | ||
| 4258 | ||
| d1c75d1e | 4259 | |
| 75e28f62 E |
4260 | #define blend_blocks_add_fourth_untextured_builder(mask_evaluate) \ |
| 4261 | .align 3; \ | |
| 4262 | \ | |
| 4263 | function(blend_blocks_untextured_add_fourth_##mask_evaluate) \ | |
| 4264 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 4265 | save_abi_regs(); \ |
| 75e28f62 | 4266 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 4267 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
4268 | \ |
| 4269 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 4270 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
4271 | \ |
| 4272 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4273 | mov c_64, #64; \ | |
| 4274 | \ | |
| 4275 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4276 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 75e28f62 E |
4277 | vmov.u16 d128_0x1C07, #0x1C00; \ |
| 4278 | vmov.u16 d128_0x00E0, #0x00E0; \ | |
| 4279 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 4280 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 75e28f62 E |
4281 | vorr.u16 d128_0x1C07, #0x0007; \ |
| 4282 | \ | |
| e1f6de8f | 4283 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 4284 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 4285 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 4286 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
| 75e28f62 E |
4287 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4288 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4289 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4290 | \ | |
| 4291 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4292 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
| 4293 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4294 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4295 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4296 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4297 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4298 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 4299 | \ | |
| 4300 | subs num_blocks, num_blocks, #1; \ | |
| 4301 | beq 1f; \ | |
| 4302 | \ | |
| 4303 | 0: \ | |
| 4304 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 4305 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 4306 | \ |
| e1f6de8f | 4307 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
4308 | \ |
| 4309 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4310 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4311 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 4312 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4313 | \ | |
| 4314 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4315 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
4316 | \ |
| 4317 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4318 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4319 | cmp fb_ptr_cmp, #28; \ | |
| 4320 | bls 2f; \ | |
| 4321 | \ | |
| e1f6de8f | 4322 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4323 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4324 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4325 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
| 4326 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| e1f6de8f | 4327 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4328 | \ |
| 4329 | 3: \ | |
| 4330 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4331 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4332 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4333 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4334 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 4335 | \ | |
| 4336 | subs num_blocks, num_blocks, #1; \ | |
| 4337 | bne 0b; \ | |
| 4338 | \ | |
| 4339 | 1: \ | |
| 4340 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4341 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 4342 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4343 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4344 | \ |
| 3f0189c6 | 4345 | restore_abi_regs(); \ |
| 75e28f62 E |
4346 | ldmia sp!, { r4, pc }; \ |
| 4347 | \ | |
| 4348 | 2: \ | |
| e1f6de8f | 4349 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4350 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 4351 | \ | |
| e1f6de8f | 4352 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4353 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4354 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4355 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4356 | bal 3b \ | |
| 4357 | ||
| 4358 | ||
| 4359 | blend_blocks_add_fourth_textured_builder(off) | |
| 4360 | blend_blocks_add_fourth_textured_builder(on) | |
| 4361 | blend_blocks_add_fourth_untextured_builder(off) | |
| 4362 | blend_blocks_add_fourth_untextured_builder(on) | |
| 4363 | ||
| 4364 | // TODO: Optimize this more. Need a scene that actually uses it for | |
| 4365 | // confirmation.. | |
| 4366 | ||
| 4367 | .align 3 | |
| 4368 | ||
| 4369 | function(blend_blocks_textured_unblended_on) | |
| 4370 | stmdb sp!, { r4, r14 } | |
| 3f0189c6 | 4371 | save_abi_regs() |
| 75e28f62 | 4372 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
| e1f6de8f | 4373 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
4374 | |
| 4375 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
| e1f6de8f | 4376 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
| 75e28f62 E |
4377 | |
| 4378 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 4379 | mov c_64, #64 | |
| 4380 | ||
| e1f6de8f | 4381 | ldr fb_ptr, [pixel_ptr, #28] |
| 4382 | vld1.u16 { fb_pixels }, [fb_ptr] | |
| 4383 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
| 75e28f62 | 4384 | vclt.s16 write_mask, fb_pixels, #0 |
| e1f6de8f | 4385 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
| 75e28f62 E |
4386 | |
| 4387 | subs num_blocks, num_blocks, #1 | |
| 4388 | beq 1f | |
| 4389 | ||
| 4390 | 0: | |
| 134f81ec | 4391 | vorr.u16 pixels, pixels, msb_mask |
| 75e28f62 E |
4392 | vorr.u16 draw_mask, draw_mask, write_mask |
| 4393 | vbif.u16 fb_pixels, pixels, draw_mask | |
| e1f6de8f | 4394 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 | 4395 | |
| e1f6de8f | 4396 | ldr fb_ptr, [pixel_ptr, #28] |
| 4397 | vld1.u16 { fb_pixels }, [fb_ptr] | |
| 4398 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
| 75e28f62 | 4399 | vclt.s16 write_mask, fb_pixels, #0 |
| e1f6de8f | 4400 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
| 75e28f62 E |
4401 | |
| 4402 | subs num_blocks, num_blocks, #1 | |
| 4403 | bne 0b | |
| 4404 | ||
| 4405 | 1: | |
| 134f81ec | 4406 | vorr.u16 pixels, pixels, msb_mask |
| 75e28f62 E |
4407 | vorr.u16 draw_mask, draw_mask, write_mask |
| 4408 | vbif.u16 fb_pixels, pixels, draw_mask | |
| e1f6de8f | 4409 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 | 4410 | |
| 3f0189c6 | 4411 | restore_abi_regs() |
| 75e28f62 E |
4412 | ldmia sp!, { r4, pc } |
| 4413 | ||
| 4414 | ||
| 4415 | function(blend_blocks_textured_unblended_off) | |
| 4416 | bx lr | |
| 4417 | ||
| 4418 | ||
| 4419 | function(warmup) | |
| 4420 | mov r3, #64 | |
| 4421 | cmp r0, #0 | |
| 4422 | bxeq lr | |
| 4423 | ||
| 4424 | 0: | |
| e1f6de8f | 4425 | vld1.u32 { u_whole_8, v_whole_8 }, [r1, :128], r3 |
| 75e28f62 E |
4426 | |
| 4427 | subs r0, r0, #1 | |
| 4428 | bne 0b | |
| 4429 | ||
| 4430 | bx lr | |
| 4431 | ||
| 6c4a10c4 | 4432 | #undef vram_ptr |
| 75e28f62 | 4433 | #undef color |
| 75e28f62 E |
4434 | #undef x |
| 4435 | #undef y | |
| 4436 | #undef width | |
| 4437 | #undef height | |
| 4438 | #undef fb_ptr | |
| 4439 | #undef texture_mask | |
| 4440 | #undef num_blocks | |
| 4441 | #undef temp | |
| 4442 | #undef dirty_textures_mask | |
| 4443 | #undef clut_ptr | |
| 4444 | #undef current_texture_mask | |
| 4445 | ||
| 4446 | #define psx_gpu r0 | |
| 4447 | #define x r1 | |
| 4448 | #define y r2 | |
| 4449 | #define u r3 | |
| 4450 | #define v r4 | |
| 4451 | #define width r5 | |
| 4452 | #define height r6 | |
| 4453 | #define offset_u r8 | |
| 4454 | #define offset_v r9 | |
| 4455 | #define offset_u_right r10 | |
| 4456 | #define width_rounded r11 | |
| 4457 | #define height_rounded r12 | |
| 4458 | ||
| 4459 | #define texture_offset_base r1 | |
| 4460 | #define tile_width r2 | |
| 4461 | #define tile_height r3 | |
| 4462 | #define num_blocks r4 | |
| 4463 | #define block r5 | |
| 4464 | #define sub_tile_height r6 | |
| 4465 | #define fb_ptr r7 | |
| 4466 | #define texture_mask r8 | |
| 4467 | #define column_data r9 | |
| 4468 | #define texture_offset r10 | |
| 4469 | #define tiles_remaining r11 | |
| 4470 | #define fb_ptr_advance_column r12 | |
| 4471 | #define texture_block_ptr r14 | |
| 4472 | ||
| 8184d7c5 | 4473 | #define temp r14 |
| 4474 | ||
| 75e28f62 E |
4475 | #define texture_page_ptr r3 |
| 4476 | #define left_block_mask r4 | |
| 4477 | #define right_block_mask r5 | |
| 4478 | #define texture_mask_rev r10 | |
| 4479 | #define control_mask r11 | |
| 4480 | ||
| 4481 | #define dirty_textures_mask r4 | |
| 4482 | #define clut_ptr r5 | |
| 4483 | #define current_texture_mask r6 | |
| 4484 | ||
| 4485 | ||
| 4486 | #undef texels | |
| 4487 | #undef clut_low_a | |
| 4488 | #undef clut_low_b | |
| 4489 | #undef clut_high_a | |
| 4490 | #undef clut_high_b | |
| 4491 | #undef clut_a | |
| 4492 | #undef clut_b | |
| 4493 | #undef texels_low | |
| 4494 | #undef texels_high | |
| 4495 | ||
| 4496 | #define texels d0 | |
| 4497 | #define draw_masks_fb_ptrs q1 | |
| 4498 | ||
| 4499 | #define draw_mask_fb_ptr_left d2 | |
| 4500 | #define draw_mask_fb_ptr_right d3 | |
| 4501 | ||
| 59d15d23 | 4502 | #define draw_mask_fb_ptr_left_a d2 |
| 4503 | #define draw_mask_fb_ptr_left_b d3 | |
| 4504 | #define draw_mask_fb_ptr_right_a d10 | |
| 4505 | #define draw_mask_fb_ptr_right_b d11 | |
| 4506 | #define draw_masks_fb_ptrs2 q5 | |
| 4507 | ||
| 75e28f62 E |
4508 | #define clut_low_a d4 |
| 4509 | #define clut_low_b d5 | |
| 4510 | #define clut_high_a d6 | |
| 4511 | #define clut_high_b d7 | |
| 4512 | ||
| 4513 | #define block_masks d8 | |
| 4514 | #define block_masks_shifted d9 | |
| 4515 | ||
| 4516 | #define clut_a q2 | |
| 4517 | #define clut_b q3 | |
| 4518 | ||
| 59d15d23 | 4519 | #define texels_low d12 |
| 4520 | #define texels_high d13 | |
| 75e28f62 | 4521 | |
| 59d15d23 | 4522 | #define texels_wide_low d14 |
| 4523 | #define texels_wide_high d15 | |
| 4524 | #define texels_wide q7 | |
| 75e28f62 | 4525 | |
| 03eb3b69 | 4526 | .align 3 |
| 75e28f62 | 4527 | |
| 59d15d23 | 4528 | setup_sprite_flush_blocks: |
| db2804fb | 4529 | push { r0 - r3, EXTRA_UNSAVED_REGS r12, lr } |
| 4530 | add block, r0, #psx_gpu_saved_tmp_offset /* r5 */ | |
| 4531 | vstmia block, { q1 - q3 } | |
| 4532 | bl flush_render_block_buffer | |
| 4533 | vldmia block, { q1 - q3 } | |
| 4534 | pop { r0 - r3, EXTRA_UNSAVED_REGS r12, lr } | |
| 75e28f62 | 4535 | |
| db2804fb | 4536 | add block, psx_gpu, #psx_gpu_blocks_offset |
| 4537 | bx lr | |
| 75e28f62 E |
4538 | |
| 4539 | ||
| 4540 | setup_sprite_update_texture_4bpp_cache: | |
| db2804fb | 4541 | push { r0 - r4, lr } |
| 75e28f62 | 4542 | bl update_texture_4bpp_cache |
| db2804fb | 4543 | pop { r0 - r4, pc } |
| 75e28f62 E |
4544 | |
| 4545 | ||
| 4546 | setup_sprite_update_texture_8bpp_cache: | |
| db2804fb | 4547 | push { r0 - r4, EXTRA_UNSAVED_REGS lr } |
| 75e28f62 | 4548 | bl update_texture_8bpp_cache |
| db2804fb | 4549 | pop { r0 - r4, EXTRA_UNSAVED_REGS pc } |
| 75e28f62 E |
4550 | |
| 4551 | ||
| 4552 | #define setup_sprite_tiled_initialize_4bpp() \ | |
| 4553 | ldr dirty_textures_mask, \ | |
| e1f6de8f | 4554 | [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset]; \ |
| 4555 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset]; \ | |
| 75e28f62 | 4556 | \ |
| e1f6de8f | 4557 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset]; \ |
| 4558 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128]; \ | |
| 75e28f62 E |
4559 | \ |
| 4560 | tst current_texture_mask, dirty_textures_mask; \ | |
| 4561 | vuzp.u8 clut_a, clut_b; \ | |
| 4562 | \ | |
| 4563 | blne setup_sprite_update_texture_4bpp_cache \ | |
| 4564 | ||
| 4565 | #define setup_sprite_tiled_initialize_8bpp() \ | |
| 4566 | ldr dirty_textures_mask, \ | |
| e1f6de8f | 4567 | [psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset]; \ |
| 4568 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset]; \ | |
| 75e28f62 E |
4569 | \ |
| 4570 | tst current_texture_mask, dirty_textures_mask; \ | |
| 4571 | blne setup_sprite_update_texture_8bpp_cache \ | |
| 4572 | ||
| 4573 | ||
| 75e28f62 E |
4574 | #define setup_sprite_block_count_single() \ |
| 4575 | sub_tile_height \ | |
| 4576 | ||
| 4577 | #define setup_sprite_block_count_double() \ | |
| 4578 | sub_tile_height, lsl #1 \ | |
| 4579 | ||
| 4580 | #define setup_sprite_tile_add_blocks(type) \ | |
| 4581 | add num_blocks, num_blocks, setup_sprite_block_count_##type(); \ | |
| 4582 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 4583 | \ | |
| 59d15d23 | 4584 | movgt num_blocks, setup_sprite_block_count_##type(); \ |
| 4585 | blgt setup_sprite_flush_blocks \ | |
| 75e28f62 E |
4586 | |
| 4587 | ||
| 4588 | #define setup_sprite_tile_full_4bpp(edge) \ | |
| 4589 | setup_sprite_tile_add_blocks(double); \ | |
| 4590 | \ | |
| 4591 | 4: \ | |
| 4592 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4593 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
| 4594 | \ | |
| e1f6de8f | 4595 | pld [fb_ptr]; \ |
| 75e28f62 | 4596 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
| e1f6de8f | 4597 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 75e28f62 E |
4598 | \ |
| 4599 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
| 4600 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 4601 | \ | |
| e1f6de8f | 4602 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
| 75e28f62 E |
4603 | add texture_block_ptr, texture_offset, #8; \ |
| 4604 | \ | |
| 4605 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
| 4606 | add block, block, #40; \ | |
| 4607 | \ | |
| 4608 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4609 | add fb_ptr, fb_ptr, #16; \ | |
| 4610 | \ | |
| e1f6de8f | 4611 | vst1.u32 { draw_mask_fb_ptr_left }, [block, :64]; \ |
| 75e28f62 E |
4612 | add block, block, #24; \ |
| 4613 | \ | |
| e1f6de8f | 4614 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 75e28f62 E |
4615 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ |
| 4616 | \ | |
| e1f6de8f | 4617 | pld [fb_ptr]; \ |
| 75e28f62 E |
4618 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ |
| 4619 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 4620 | \ | |
| e1f6de8f | 4621 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
| 75e28f62 E |
4622 | add block, block, #40; \ |
| 4623 | \ | |
| 4624 | add texture_offset, texture_offset, #0x10; \ | |
| 4625 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
| 4626 | \ | |
| e1f6de8f | 4627 | vst1.u32 { draw_mask_fb_ptr_right }, [block, :64]; \ |
| 75e28f62 E |
4628 | add block, block, #24; \ |
| 4629 | \ | |
| 4630 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4631 | bne 4b; \ | |
| 4632 | \ | |
| 4633 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 4634 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 75e28f62 E |
4635 | |
| 4636 | ||
| 4637 | #define setup_sprite_tile_half_4bpp(edge) \ | |
| 4638 | setup_sprite_tile_add_blocks(single); \ | |
| 4639 | \ | |
| 4640 | 4: \ | |
| 4641 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4642 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
| 4643 | \ | |
| e1f6de8f | 4644 | pld [fb_ptr]; \ |
| 75e28f62 | 4645 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
| e1f6de8f | 4646 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 75e28f62 E |
4647 | \ |
| 4648 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
| 4649 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 4650 | \ | |
| e1f6de8f | 4651 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
| 75e28f62 E |
4652 | add block, block, #40; \ |
| 4653 | \ | |
| 4654 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| e1f6de8f | 4655 | vst1.u32 { draw_mask_fb_ptr_##edge }, [block, :64]; \ |
| 75e28f62 E |
4656 | \ |
| 4657 | add block, block, #24; \ | |
| 4658 | add texture_offset, texture_offset, #0x10; \ | |
| 4659 | \ | |
| 4660 | add fb_ptr, fb_ptr, #2048; \ | |
| 4661 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4662 | \ | |
| 4663 | bne 4b; \ | |
| 4664 | \ | |
| 4665 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 4666 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 75e28f62 E |
4667 | |
| 4668 | ||
| 4669 | #define setup_sprite_tile_full_8bpp(edge) \ | |
| 4670 | setup_sprite_tile_add_blocks(double); \ | |
| 4671 | add block, block, #16; \ | |
| 4672 | \ | |
| 4673 | 4: \ | |
| 4674 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4675 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
| 4676 | \ | |
| e1f6de8f | 4677 | pld [fb_ptr]; \ |
| 75e28f62 | 4678 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
| e1f6de8f | 4679 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 75e28f62 E |
4680 | \ |
| 4681 | add texture_block_ptr, texture_offset, #8; \ | |
| e1f6de8f | 4682 | vst1.u32 { texels }, [block, :64]; \ |
| 75e28f62 E |
4683 | \ |
| 4684 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
| 4685 | add block, block, #24; \ | |
| 4686 | \ | |
| 4687 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 4688 | \ | |
| 4689 | add fb_ptr, fb_ptr, #16; \ | |
| e1f6de8f | 4690 | vst1.u32 { draw_mask_fb_ptr_left }, [block, :64]; \ |
| 75e28f62 E |
4691 | \ |
| 4692 | add block, block, #40; \ | |
| e1f6de8f | 4693 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 4694 | pld [fb_ptr]; \ | |
| 75e28f62 E |
4695 | \ |
| 4696 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ | |
| e1f6de8f | 4697 | vst1.u32 { texels }, [block, :64]; \ |
| 75e28f62 E |
4698 | add block, block, #24; \ |
| 4699 | \ | |
| 4700 | add texture_offset, texture_offset, #0x10; \ | |
| 4701 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
| 4702 | \ | |
| e1f6de8f | 4703 | vst1.u32 { draw_mask_fb_ptr_right }, [block, :64]; \ |
| 75e28f62 E |
4704 | add block, block, #40; \ |
| 4705 | \ | |
| 4706 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4707 | bne 4b; \ | |
| 4708 | \ | |
| 4709 | sub block, block, #16; \ | |
| 4710 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 4711 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 75e28f62 E |
4712 | |
| 4713 | ||
| 4714 | #define setup_sprite_tile_half_8bpp(edge) \ | |
| 4715 | setup_sprite_tile_add_blocks(single); \ | |
| 4716 | add block, block, #16; \ | |
| 4717 | \ | |
| 4718 | 4: \ | |
| 4719 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| 4720 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
| e1f6de8f | 4721 | pld [fb_ptr]; \ |
| 75e28f62 E |
4722 | \ |
| 4723 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| e1f6de8f | 4724 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 75e28f62 | 4725 | \ |
| e1f6de8f | 4726 | vst1.u32 { texels }, [block, :64]; \ |
| 75e28f62 E |
4727 | add block, block, #24; \ |
| 4728 | \ | |
| e1f6de8f | 4729 | vst1.u32 { draw_mask_fb_ptr_##edge }, [block, :64]; \ |
| 75e28f62 E |
4730 | add block, block, #40; \ |
| 4731 | \ | |
| 4732 | add texture_offset, texture_offset, #0x10; \ | |
| 4733 | add fb_ptr, fb_ptr, #2048; \ | |
| 4734 | \ | |
| 4735 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 4736 | bne 4b; \ | |
| 4737 | \ | |
| 4738 | sub block, block, #16; \ | |
| 4739 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 4740 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 75e28f62 E |
4741 | |
| 4742 | ||
| 4743 | #define setup_sprite_tile_column_edge_pre_adjust_half_right() \ | |
| 4744 | add texture_offset, texture_offset_base, #8; \ | |
| 4745 | add fb_ptr, fb_ptr, #16 \ | |
| 4746 | ||
| 4747 | #define setup_sprite_tile_column_edge_pre_adjust_half_left() \ | |
| 4748 | mov texture_offset, texture_offset_base \ | |
| 4749 | ||
| 4750 | #define setup_sprite_tile_column_edge_pre_adjust_half(edge) \ | |
| 4751 | setup_sprite_tile_column_edge_pre_adjust_half_##edge() \ | |
| 4752 | ||
| 4753 | #define setup_sprite_tile_column_edge_pre_adjust_full(edge) \ | |
| 4754 | mov texture_offset, texture_offset_base \ | |
| 4755 | ||
| 4756 | #define setup_sprite_tile_column_edge_post_adjust_half_right() \ | |
| 4757 | sub fb_ptr, fb_ptr, #16 \ | |
| 4758 | ||
| 4759 | #define setup_sprite_tile_column_edge_post_adjust_half_left() \ | |
| 4760 | ||
| 4761 | #define setup_sprite_tile_column_edge_post_adjust_half(edge) \ | |
| 4762 | setup_sprite_tile_column_edge_post_adjust_half_##edge() \ | |
| 4763 | ||
| 4764 | #define setup_sprite_tile_column_edge_post_adjust_full(edge) \ | |
| 4765 | ||
| 4766 | ||
| 59d15d23 | 4767 | #define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode, \ |
| 4768 | x4mode) \ | |
| 75e28f62 | 4769 | mov sub_tile_height, column_data; \ |
| 59d15d23 | 4770 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
| 4771 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
| 4772 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
| 75e28f62 | 4773 | |
| 59d15d23 | 4774 | #define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode, \ |
| 4775 | x4mode) \ | |
| 75e28f62 E |
4776 | and sub_tile_height, column_data, #0xFF; \ |
| 4777 | mov tiles_remaining, column_data, lsr #16; \ | |
| 59d15d23 | 4778 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
| 4779 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
| 75e28f62 E |
4780 | \ |
| 4781 | subs tiles_remaining, tiles_remaining, #1; \ | |
| 4782 | beq 2f; \ | |
| 4783 | \ | |
| 4784 | 3: \ | |
| 4785 | mov sub_tile_height, #16; \ | |
| 59d15d23 | 4786 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
| 75e28f62 E |
4787 | subs tiles_remaining, tiles_remaining, #1; \ |
| 4788 | bne 3b; \ | |
| 4789 | \ | |
| 4790 | 2: \ | |
| 4791 | uxtb sub_tile_height, column_data, ror #8; \ | |
| 59d15d23 | 4792 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
| 4793 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
| 75e28f62 E |
4794 | |
| 4795 | ||
| 4796 | #define setup_sprite_column_data_single() \ | |
| 4797 | mov column_data, height; \ | |
| e1f6de8f | 4798 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] \ |
| 75e28f62 E |
4799 | |
| 4800 | #define setup_sprite_column_data_multi() \ | |
| 4801 | and height_rounded, height_rounded, #0xF; \ | |
| 4802 | rsb column_data, offset_v, #16; \ | |
| 4803 | \ | |
| 4804 | add height_rounded, height_rounded, #1; \ | |
| 4805 | sub tile_height, tile_height, #1; \ | |
| 4806 | \ | |
| 4807 | orr column_data, column_data, tile_height, lsl #16; \ | |
| e1f6de8f | 4808 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset]; \ |
| 75e28f62 E |
4809 | \ |
| 4810 | orr column_data, column_data, height_rounded, lsl #8 \ | |
| 4811 | ||
| 59d15d23 | 4812 | #define setup_sprite_setup_left_draw_mask_fb_ptr() \ |
| 4813 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
| 4814 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ | |
| 4815 | ||
| 4816 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column() \ | |
| 4817 | mov fb_ptr_advance_column, #32; \ | |
| 4818 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
| 4819 | \ | |
| ed0fd81d | 4820 | sub fb_ptr_advance_column, fb_ptr_advance_column, height, lsl #11; \ |
| 59d15d23 | 4821 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ |
| 4822 | ||
| 4823 | #define setup_sprite_setup_right_draw_mask_fb_ptr() \ | |
| 4824 | vdup.u8 draw_mask_fb_ptr_left, block_masks[4]; \ | |
| 4825 | vdup.u8 draw_mask_fb_ptr_right, block_masks[5] \ | |
| 4826 | ||
| 4827 | #define setup_sprite_tile_column_width_single(tm, multi_height, edge_mode, \ | |
| 4828 | edge, x4mode) \ | |
| 4829 | setup_sprite_##tm##_single_##multi_height##_##edge_mode##_##edge##x4mode: \ | |
| 75e28f62 E |
4830 | setup_sprite_column_data_##multi_height(); \ |
| 4831 | vext.32 block_masks_shifted, block_masks, block_masks, #1; \ | |
| 4832 | vorr.u32 block_masks, block_masks, block_masks_shifted; \ | |
| 59d15d23 | 4833 | setup_sprite_setup_left_draw_mask_fb_ptr##x4mode(); \ |
| 75e28f62 | 4834 | \ |
| 59d15d23 | 4835 | setup_sprite_tile_column_height_##multi_height(edge_mode, edge, tm, x4mode); \ |
| db2804fb | 4836 | vpop { q4 - q7 }; \ |
| 4837 | pop { r3 - r11, pc } \ | |
| 75e28f62 E |
4838 | |
| 4839 | #define setup_sprite_tiled_advance_column() \ | |
| 4840 | add texture_offset_base, texture_offset_base, #0x100; \ | |
| 4841 | tst texture_offset_base, #0xF00; \ | |
| 4842 | subeq texture_offset_base, texture_offset_base, #(0x100 + 0xF00) \ | |
| 4843 | ||
| 4844 | #define setup_sprite_tile_column_width_multi(tm, multi_height, left_mode, \ | |
| 59d15d23 | 4845 | right_mode, x4mode) \ |
| 4846 | setup_sprite_##tm##_multi_##multi_height##_##left_mode##_##right_mode##x4mode:\ | |
| 75e28f62 | 4847 | setup_sprite_column_data_##multi_height(); \ |
| 75e28f62 | 4848 | \ |
| 59d15d23 | 4849 | setup_sprite_setup_left_draw_mask_fb_ptr_advance_column##x4mode(); \ |
| 75e28f62 | 4850 | \ |
| 59d15d23 | 4851 | setup_sprite_tile_column_height_##multi_height(left_mode, right, tm, x4mode);\ |
| 75e28f62 E |
4852 | \ |
| 4853 | subs tile_width, tile_width, #2; \ | |
| 4854 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ | |
| 4855 | \ | |
| 75e28f62 E |
4856 | beq 1f; \ |
| 4857 | \ | |
| 59d15d23 | 4858 | vmov.u8 draw_masks_fb_ptrs, #0; \ |
| 4859 | vmov.u8 draw_masks_fb_ptrs2, #0; \ | |
| 4860 | \ | |
| 75e28f62 E |
4861 | 0: \ |
| 4862 | setup_sprite_tiled_advance_column(); \ | |
| 59d15d23 | 4863 | setup_sprite_tile_column_height_##multi_height(full, none, tm, x4mode); \ |
| 75e28f62 E |
4864 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ |
| 4865 | subs tile_width, tile_width, #1; \ | |
| 4866 | bne 0b; \ | |
| 4867 | \ | |
| 4868 | 1: \ | |
| 59d15d23 | 4869 | setup_sprite_setup_right_draw_mask_fb_ptr##x4mode(); \ |
| 75e28f62 E |
4870 | \ |
| 4871 | setup_sprite_tiled_advance_column(); \ | |
| 59d15d23 | 4872 | setup_sprite_tile_column_height_##multi_height(right_mode, left, tm, x4mode);\ |
| db2804fb | 4873 | vpop { q4 - q7 }; \ |
| 4874 | pop { r3 - r11, pc } \ | |
| 75e28f62 E |
4875 | |
| 4876 | ||
| 59d15d23 | 4877 | #define setup_sprite_offset_u_adjust() \ |
| 4878 | ||
| 4879 | #define setup_sprite_get_left_block_mask() \ | |
| 4880 | and left_block_mask, left_block_mask, #0xFF \ | |
| 4881 | ||
| 4882 | #define setup_sprite_compare_left_block_mask() \ | |
| 4883 | cmp left_block_mask, #0xFF \ | |
| 4884 | ||
| 4885 | #define setup_sprite_get_right_block_mask() \ | |
| 4886 | uxtb right_block_mask, right_block_mask, ror #8 \ | |
| 4887 | ||
| 4888 | #define setup_sprite_compare_right_block_mask() \ | |
| 4889 | cmp right_block_mask, #0xFF \ | |
| 4890 | ||
| 4891 | ||
| 4892 | ||
| 4893 | /* 4x stuff */ | |
| 4894 | #define fb_ptr2 column_data | |
| 4895 | ||
| 4896 | #define setup_sprite_offset_u_adjust_4x() \ | |
| 4897 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
| 4898 | lsl offset_u_right, #1; \ | |
| 4899 | lsl offset_u, #1; \ | |
| 4900 | add offset_u_right, #1 \ | |
| 4901 | ||
| 4902 | #define setup_sprite_get_left_block_mask_4x() \ | |
| 4903 | sxth left_block_mask, left_block_mask \ | |
| 4904 | ||
| 4905 | #define setup_sprite_compare_left_block_mask_4x() \ | |
| 4906 | cmp left_block_mask, #0xFFFFFFFF \ | |
| 4907 | ||
| 4908 | #define setup_sprite_get_right_block_mask_4x() \ | |
| 4909 | sxth right_block_mask, right_block_mask, ror #16 \ | |
| 4910 | ||
| 4911 | #define setup_sprite_compare_right_block_mask_4x() \ | |
| 4912 | cmp right_block_mask, #0xFFFFFFFF \ | |
| 4913 | ||
| 4914 | ||
| 4915 | #define widen_texels_16bpp(texels_) \ | |
| 4916 | vmov texels_wide_low, texels_; \ | |
| 4917 | vmov texels_wide_high, texels_; \ | |
| 4918 | vzip.16 texels_wide_low, texels_wide_high \ | |
| 4919 | ||
| 4920 | #define widen_texels_8bpp(texels_) \ | |
| 4921 | vmov texels_wide_low, texels_; \ | |
| 4922 | vmov texels_wide_high, texels_; \ | |
| 4923 | vzip.8 texels_wide_low, texels_wide_high \ | |
| 4924 | ||
| 4925 | #define write_block_16bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
| e1f6de8f | 4926 | vst1.u32 { texels_ }, [block_, :128]; \ |
| 59d15d23 | 4927 | add block_, block_, #40; \ |
| 4928 | \ | |
| 4929 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
| e1f6de8f | 4930 | vst1.u32 { draw_mask_fb_ptr_ }, [block_, :64]; \ |
| 59d15d23 | 4931 | add block_, block_, #24 \ |
| 4932 | ||
| 4933 | /* assumes 16-byte offset already added to block_ */ | |
| 4934 | #define write_block_8bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
| e1f6de8f | 4935 | vst1.u32 { texels_ }, [block_, :64]; \ |
| 59d15d23 | 4936 | add block_, block_, #24; \ |
| 4937 | \ | |
| 4938 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
| e1f6de8f | 4939 | vst1.u32 { draw_mask_fb_ptr_ }, [block_, :64]; \ |
| 59d15d23 | 4940 | add block_, block_, #40 \ |
| 4941 | ||
| 4942 | #define do_texture_block_16bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
| 4943 | draw_mask_fb_ptr_b_) \ | |
| 4944 | widen_texels_16bpp(texels_low); \ | |
| 4945 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
| 4946 | \ | |
| 4947 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
| 4948 | \ | |
| 4949 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
| 4950 | widen_texels_16bpp(texels_high); \ | |
| 4951 | \ | |
| 4952 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
| 4953 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
| 4954 | \ | |
| 4955 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
| 4956 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
| 4957 | ||
| 4958 | #define do_texture_block_8bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
| 4959 | draw_mask_fb_ptr_b_) \ | |
| 4960 | widen_texels_8bpp(texels); \ | |
| 4961 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
| 4962 | \ | |
| 4963 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
| 4964 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
| 4965 | \ | |
| 4966 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
| 4967 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
| 4968 | \ | |
| 4969 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
| 4970 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
| 4971 | ||
| 4972 | ||
| 4973 | #define setup_sprite_tiled_initialize_4bpp_4x() \ | |
| e1f6de8f | 4974 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset]; \ |
| 4975 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128]; \ | |
| 59d15d23 | 4976 | \ |
| 4977 | vuzp.u8 clut_a, clut_b \ | |
| 4978 | ||
| 4979 | #define setup_sprite_tiled_initialize_8bpp_4x() \ | |
| 4980 | ||
| 4981 | ||
| 4982 | #define setup_sprite_block_count_single_4x() \ | |
| 4983 | sub_tile_height, lsl #2 \ | |
| 4984 | ||
| 4985 | #define setup_sprite_block_count_double_4x() \ | |
| 4986 | sub_tile_height, lsl #(1+2) \ | |
| 4987 | ||
| 4988 | #define setup_sprite_tile_full_4bpp_4x(edge) \ | |
| 4989 | setup_sprite_tile_add_blocks(double_4x); \ | |
| 4990 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
| 4991 | \ | |
| 4992 | 4: \ | |
| 4993 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| e1f6de8f | 4994 | pld [fb_ptr]; \ |
| 59d15d23 | 4995 | \ |
| 4996 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| e1f6de8f | 4997 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 59d15d23 | 4998 | \ |
| 4999 | add texture_block_ptr, texture_offset, #8; \ | |
| 5000 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
| 5001 | \ | |
| 5002 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
| 5003 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 5004 | \ | |
| 5005 | vzip.8 texels_low, texels_high; \ | |
| 5006 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
| 5007 | draw_mask_fb_ptr_left_b); \ | |
| 5008 | \ | |
| 5009 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| e1f6de8f | 5010 | pld [fb_ptr, #2048]; \ |
| 59d15d23 | 5011 | \ |
| e1f6de8f | 5012 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 8438c3c7 | 5013 | add fb_ptr, fb_ptr, #16*2; \ |
| 59d15d23 | 5014 | \ |
| 8438c3c7 | 5015 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ |
| 59d15d23 | 5016 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ |
| 5017 | \ | |
| 5018 | vzip.8 texels_low, texels_high; \ | |
| 5019 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
| 5020 | draw_mask_fb_ptr_right_b); \ | |
| 5021 | \ | |
| 5022 | add texture_offset, texture_offset, #0x10; \ | |
| 5023 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
| 5024 | \ | |
| 5025 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 5026 | bne 4b; \ | |
| 5027 | \ | |
| 5028 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
| 5029 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 5030 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 59d15d23 | 5031 | |
| 5032 | ||
| 5033 | #define setup_sprite_tile_half_4bpp_4x(edge) \ | |
| 5034 | setup_sprite_tile_add_blocks(single_4x); \ | |
| 5035 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
| 5036 | \ | |
| 5037 | 4: \ | |
| 5038 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| e1f6de8f | 5039 | pld [fb_ptr]; \ |
| 59d15d23 | 5040 | \ |
| 5041 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| e1f6de8f | 5042 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 59d15d23 | 5043 | \ |
| 5044 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 5045 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
| 5046 | \ | |
| 5047 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
| 5048 | add texture_offset, texture_offset, #0x10; \ | |
| 5049 | \ | |
| 5050 | vzip.8 texels_low, texels_high; \ | |
| 5051 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ | |
| 5052 | draw_mask_fb_ptr_##edge##_b); \ | |
| 5053 | \ | |
| e1f6de8f | 5054 | pld [fb_ptr, #2048]; \ |
| 59d15d23 | 5055 | add fb_ptr, fb_ptr, #2048 * 2; \ |
| 59d15d23 | 5056 | \ |
| 8438c3c7 | 5057 | subs sub_tile_height, sub_tile_height, #1; \ |
| 59d15d23 | 5058 | bne 4b; \ |
| 5059 | \ | |
| 5060 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
| 5061 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 5062 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 59d15d23 | 5063 | |
| 5064 | ||
| 5065 | #define setup_sprite_tile_full_8bpp_4x(edge) \ | |
| 5066 | setup_sprite_tile_add_blocks(double_4x); \ | |
| 5067 | add block, block, #16; \ | |
| 5068 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
| 5069 | \ | |
| 5070 | 4: \ | |
| 5071 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| e1f6de8f | 5072 | pld [fb_ptr]; \ |
| 59d15d23 | 5073 | \ |
| 5074 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| e1f6de8f | 5075 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 59d15d23 | 5076 | \ |
| 5077 | add texture_block_ptr, texture_offset, #8; \ | |
| 5078 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
| 5079 | draw_mask_fb_ptr_left_b); \ | |
| 5080 | \ | |
| e1f6de8f | 5081 | pld [fb_ptr, #2048]; \ |
| 59d15d23 | 5082 | and texture_block_ptr, texture_block_ptr, texture_mask; \ |
| 5083 | \ | |
| 5084 | add fb_ptr, fb_ptr, #16*2; \ | |
| 5085 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| 5086 | \ | |
| e1f6de8f | 5087 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 59d15d23 | 5088 | \ |
| 5089 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
| 5090 | draw_mask_fb_ptr_right_b); \ | |
| 5091 | \ | |
| 5092 | add texture_offset, texture_offset, #0x10; \ | |
| 5093 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
| 5094 | \ | |
| 5095 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 5096 | bne 4b; \ | |
| 5097 | \ | |
| 5098 | sub block, block, #16; \ | |
| 5099 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
| 5100 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 5101 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 59d15d23 | 5102 | |
| 5103 | ||
| 5104 | #define setup_sprite_tile_half_8bpp_4x(edge) \ | |
| 5105 | setup_sprite_tile_add_blocks(single_4x); \ | |
| 5106 | add block, block, #16; \ | |
| 5107 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
| 5108 | \ | |
| 5109 | 4: \ | |
| 5110 | and texture_block_ptr, texture_offset, texture_mask; \ | |
| e1f6de8f | 5111 | pld [fb_ptr]; \ |
| 59d15d23 | 5112 | \ |
| 5113 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
| e1f6de8f | 5114 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
| 59d15d23 | 5115 | \ |
| e1f6de8f | 5116 | pld [fb_ptr, #2048]; \ |
| 59d15d23 | 5117 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ |
| 5118 | draw_mask_fb_ptr_##edge##_b); \ | |
| 5119 | \ | |
| 5120 | add texture_offset, texture_offset, #0x10; \ | |
| 5121 | add fb_ptr, fb_ptr, #2048 * 2; \ | |
| 5122 | \ | |
| 5123 | subs sub_tile_height, sub_tile_height, #1; \ | |
| 5124 | bne 4b; \ | |
| 5125 | \ | |
| 5126 | sub block, block, #16; \ | |
| 5127 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
| 5128 | add texture_offset, texture_offset, #0xF00; \ | |
| e1f6de8f | 5129 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
| 59d15d23 | 5130 | |
| 5131 | ||
| 5132 | #define setup_sprite_tile_column_edge_pre_adjust_half_right_4x() \ | |
| 5133 | add texture_offset, texture_offset_base, #8; \ | |
| 5134 | add fb_ptr, fb_ptr, #16 * 2 \ | |
| 5135 | ||
| 5136 | #define setup_sprite_tile_column_edge_pre_adjust_half_left_4x() \ | |
| 5137 | mov texture_offset, texture_offset_base \ | |
| 5138 | ||
| 5139 | #define setup_sprite_tile_column_edge_pre_adjust_half_4x(edge) \ | |
| 5140 | setup_sprite_tile_column_edge_pre_adjust_half_##edge##_4x() \ | |
| 5141 | ||
| 5142 | #define setup_sprite_tile_column_edge_pre_adjust_full_4x(edge) \ | |
| 5143 | mov texture_offset, texture_offset_base \ | |
| 5144 | ||
| 5145 | #define setup_sprite_tile_column_edge_post_adjust_half_right_4x() \ | |
| 5146 | sub fb_ptr, fb_ptr, #16 * 2 \ | |
| 5147 | ||
| 5148 | #define setup_sprite_tile_column_edge_post_adjust_half_left_4x() \ | |
| 5149 | ||
| 5150 | #define setup_sprite_tile_column_edge_post_adjust_half_4x(edge) \ | |
| 5151 | setup_sprite_tile_column_edge_post_adjust_half_##edge##_4x() \ | |
| 5152 | ||
| 5153 | #define setup_sprite_tile_column_edge_post_adjust_full_4x(edge) \ | |
| 5154 | ||
| 5155 | ||
| 5156 | #define setup_sprite_setup_left_draw_mask_fb_ptr_4x() \ | |
| 5157 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
| 5158 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
| 5159 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ | |
| 5160 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
| 5161 | ||
| 5162 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column_4x() \ | |
| 5163 | mov fb_ptr_advance_column, #32 * 2; \ | |
| 5164 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
| 5165 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
| ed0fd81d | 5166 | sub fb_ptr_advance_column, fb_ptr_advance_column, height, lsl #11 + 1; \ |
| 59d15d23 | 5167 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ |
| 5168 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
| 5169 | ||
| 5170 | #define setup_sprite_setup_right_draw_mask_fb_ptr_4x() \ | |
| 5171 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[4]; \ | |
| 5172 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[5]; \ | |
| 5173 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[6]; \ | |
| 5174 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[7] \ | |
| 5175 | ||
| 5176 | ||
| 75e28f62 E |
5177 | // r0: psx_gpu |
| 5178 | // r1: x | |
| 5179 | // r2: y | |
| 5180 | // r3: u | |
| e1f6de8f | 5181 | // [sp]: v |
| 5182 | // [sp + 4]: width | |
| 5183 | // [sp + 8]: height | |
| 5184 | // [sp + 12]: color (unused) | |
| 75e28f62 | 5185 | |
| 59d15d23 | 5186 | #define setup_sprite_tiled_builder(texture_mode, x4mode) \ |
| 5187 | \ | |
| 5188 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, full, \ | |
| 5189 | x4mode); \ | |
| 5190 | setup_sprite_tile_column_width_single(texture_mode, multi, full, none, \ | |
| 5191 | x4mode); \ | |
| 5192 | setup_sprite_tile_column_width_multi(texture_mode, single, full, full, \ | |
| 5193 | x4mode); \ | |
| 5194 | setup_sprite_tile_column_width_single(texture_mode, single, full, none, \ | |
| 5195 | x4mode); \ | |
| 5196 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, full, \ | |
| 5197 | x4mode); \ | |
| 5198 | setup_sprite_tile_column_width_single(texture_mode, multi, half, right, \ | |
| 5199 | x4mode); \ | |
| 5200 | setup_sprite_tile_column_width_multi(texture_mode, single, half, full, \ | |
| 5201 | x4mode); \ | |
| 5202 | setup_sprite_tile_column_width_single(texture_mode, single, half, right, \ | |
| 5203 | x4mode); \ | |
| 5204 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, half, \ | |
| 5205 | x4mode); \ | |
| 5206 | setup_sprite_tile_column_width_single(texture_mode, multi, half, left, \ | |
| 5207 | x4mode); \ | |
| 5208 | setup_sprite_tile_column_width_multi(texture_mode, single, full, half, \ | |
| 5209 | x4mode); \ | |
| 5210 | setup_sprite_tile_column_width_single(texture_mode, single, half, left, \ | |
| 5211 | x4mode); \ | |
| 5212 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, half, \ | |
| 5213 | x4mode); \ | |
| 5214 | setup_sprite_tile_column_width_multi(texture_mode, single, half, half, \ | |
| 5215 | x4mode); \ | |
| 75e28f62 E |
5216 | \ |
| 5217 | .align 4; \ | |
| 5218 | \ | |
| 59d15d23 | 5219 | function(setup_sprite_##texture_mode##x4mode) \ |
| db2804fb | 5220 | push { r3 - r11, lr }; \ |
| 59d15d23 | 5221 | setup_sprite_tiled_initialize_##texture_mode##x4mode(); \ |
| 75e28f62 | 5222 | \ |
| db2804fb | 5223 | ldr v, [sp, #4*(10+0)]; \ |
| 75e28f62 E |
5224 | and offset_u, u, #0xF; \ |
| 5225 | \ | |
| db2804fb | 5226 | ldr width, [sp, #4*(10+1)]; \ |
| e1f6de8f | 5227 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ |
| 75e28f62 | 5228 | \ |
| db2804fb | 5229 | ldr height, [sp, #4*(10+2)]; \ |
| 75e28f62 E |
5230 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 5231 | \ | |
| db2804fb | 5232 | vpush { q4 - q7 }; \ |
| 3f0189c6 | 5233 | \ |
| 75e28f62 E |
5234 | add fb_ptr, fb_ptr, x, lsl #1; \ |
| 5235 | and offset_v, v, #0xF; \ | |
| 5236 | \ | |
| 5237 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
| 5238 | add width_rounded, offset_u, width; \ | |
| 5239 | \ | |
| 5240 | add height_rounded, offset_v, height; \ | |
| 5241 | add width_rounded, width_rounded, #15; \ | |
| 5242 | \ | |
| 5243 | add height_rounded, height_rounded, #15; \ | |
| 5244 | mov tile_width, width_rounded, lsr #4; \ | |
| 5245 | \ | |
| 5246 | /* texture_offset_base = VH-VL-00-00 */\ | |
| 5247 | mov texture_offset_base, v, lsl #8; \ | |
| 5248 | and offset_u_right, width_rounded, #0xF; \ | |
| 5249 | \ | |
| 5250 | /* texture_offset_base = VH-UH-UL-00 */\ | |
| 5251 | bfi texture_offset_base, u, #4, #8; \ | |
| 59d15d23 | 5252 | mov right_block_mask, #0xFFFFFFFE; \ |
| 5253 | \ | |
| 5254 | setup_sprite_offset_u_adjust##x4mode(); \ | |
| 75e28f62 E |
5255 | \ |
| 5256 | /* texture_offset_base = VH-UH-VL-00 */\ | |
| 5257 | bfi texture_offset_base, v, #4, #4; \ | |
| 59d15d23 | 5258 | mov left_block_mask, #0xFFFFFFFF; \ |
| 75e28f62 E |
5259 | \ |
| 5260 | mov tile_height, height_rounded, lsr #4; \ | |
| 5261 | mvn left_block_mask, left_block_mask, lsl offset_u; \ | |
| 5262 | \ | |
| 5263 | /* texture_mask = HH-HL-WH-WL */\ | |
| e1f6de8f | 5264 | ldrh texture_mask, [psx_gpu, #psx_gpu_texture_mask_width_offset]; \ |
| 75e28f62 E |
5265 | mov right_block_mask, right_block_mask, lsl offset_u_right; \ |
| 5266 | \ | |
| 5267 | /* texture_mask_rev = WH-WL-HH-HL */\ | |
| 5268 | rev16 texture_mask_rev, texture_mask; \ | |
| 5269 | vmov block_masks, left_block_mask, right_block_mask; \ | |
| 5270 | \ | |
| 5271 | /* texture_mask = HH-HL-HL-WL */\ | |
| 5272 | bfi texture_mask, texture_mask_rev, #4, #4; \ | |
| 5273 | /* texture_mask_rev = 00-00-00-WH */\ | |
| 5274 | mov texture_mask_rev, texture_mask_rev, lsr #12; \ | |
| 5275 | \ | |
| 5276 | /* texture_mask = HH-WH-HL-WL */\ | |
| 5277 | bfi texture_mask, texture_mask_rev, #8, #4; \ | |
| 59d15d23 | 5278 | setup_sprite_get_left_block_mask##x4mode(); \ |
| 75e28f62 E |
5279 | \ |
| 5280 | mov control_mask, #0; \ | |
| 59d15d23 | 5281 | setup_sprite_compare_left_block_mask##x4mode(); \ |
| 75e28f62 | 5282 | \ |
| 59d15d23 | 5283 | setup_sprite_get_right_block_mask##x4mode(); \ |
| 75e28f62 E |
5284 | orreq control_mask, control_mask, #0x4; \ |
| 5285 | \ | |
| e1f6de8f | 5286 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 59d15d23 | 5287 | setup_sprite_compare_right_block_mask##x4mode(); \ |
| 75e28f62 E |
5288 | \ |
| 5289 | orreq control_mask, control_mask, #0x8; \ | |
| 5290 | cmp tile_width, #1; \ | |
| 5291 | \ | |
| 5292 | add block, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 5293 | orreq control_mask, control_mask, #0x1; \ | |
| 5294 | \ | |
| 5295 | cmp tile_height, #1; \ | |
| 5296 | add block, block, num_blocks, lsl #6; \ | |
| 5297 | \ | |
| 5298 | orreq control_mask, control_mask, #0x2; \ | |
| 8184d7c5 | 5299 | JT_OP_REL(9f, control_mask, temp); \ |
| e1f6de8f | 5300 | JT_OP(ldr pc, [pc, control_mask, lsl #2]); \ |
| 75e28f62 E |
5301 | nop; \ |
| 5302 | \ | |
| 8184d7c5 | 5303 | 9: \ |
| 5304 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_full_full##x4mode); \ | |
| 5305 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_full_none##x4mode); \ | |
| 5306 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_full_full##x4mode); \ | |
| 5307 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_full_none##x4mode); \ | |
| 5308 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_half_full##x4mode); \ | |
| 5309 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_half_right##x4mode); \ | |
| 5310 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_half_full##x4mode); \ | |
| 5311 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_half_right##x4mode);\ | |
| 5312 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_full_half##x4mode); \ | |
| 5313 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_half_left##x4mode); \ | |
| 5314 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_full_half##x4mode); \ | |
| 5315 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_half_left##x4mode); \ | |
| 5316 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_half_half##x4mode); \ | |
| 75e28f62 | 5317 | .word 0x00000000; \ |
| 8184d7c5 | 5318 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_half_half##x4mode); \ |
| 59d15d23 | 5319 | |
| 5320 | ||
| 5321 | setup_sprite_tiled_builder(4bpp,); | |
| 5322 | setup_sprite_tiled_builder(8bpp,); | |
| 75e28f62 | 5323 | |
| 59d15d23 | 5324 | #undef draw_mask_fb_ptr_left |
| 5325 | #undef draw_mask_fb_ptr_right | |
| 75e28f62 | 5326 | |
| 59d15d23 | 5327 | setup_sprite_tiled_builder(4bpp, _4x); |
| 5328 | setup_sprite_tiled_builder(8bpp, _4x); | |
| 75e28f62 E |
5329 | |
| 5330 | ||
| 5331 | #undef block_ptr | |
| 5332 | #undef num_blocks | |
| 5333 | #undef clut_ptr | |
| 5334 | ||
| 5335 | #define psx_gpu r0 | |
| 5336 | #define block_ptr r0 | |
| 5337 | #define num_blocks r1 | |
| 5338 | #define clut_ptr r2 | |
| 5339 | #define texel_shift_mask r3 | |
| 5340 | #define block_pixels_a r4 | |
| 5341 | #define block_pixels_b r5 | |
| 5342 | #define texel_0 r6 | |
| 5343 | #define texel_2 r7 | |
| 5344 | #define texel_4 r8 | |
| 5345 | #define texel_6 r9 | |
| 5346 | #define texel_1 r10 | |
| 5347 | #define texel_3 r11 | |
| 5348 | #define texel_5 r12 | |
| 5349 | #define texel_7 r14 | |
| 5350 | #define texels_01 r6 | |
| 5351 | #define texels_23 r7 | |
| 5352 | #define texels_45 r8 | |
| 5353 | #define texels_67 r9 | |
| 5354 | ||
| 5355 | function(texture_sprite_blocks_8bpp) | |
| db2804fb | 5356 | push { r4 - r11, r14 } |
| 75e28f62 E |
5357 | movw texel_shift_mask, #(0xFF << 1) |
| 5358 | ||
| e1f6de8f | 5359 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 5360 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] | |
| 75e28f62 E |
5361 | |
| 5362 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| e1f6de8f | 5363 | ldr block_pixels_a, [block_ptr, #16] |
| 75e28f62 E |
5364 | |
| 5365 | 0: | |
| 5366 | and texel_0, texel_shift_mask, block_pixels_a, lsl #1 | |
| e1f6de8f | 5367 | ldr block_pixels_b, [block_ptr, #20] |
| 75e28f62 E |
5368 | |
| 5369 | and texel_1, texel_shift_mask, block_pixels_a, lsr #7 | |
| e1f6de8f | 5370 | ldrh texel_0, [clut_ptr, texel_0] |
| 75e28f62 E |
5371 | |
| 5372 | and texel_2, texel_shift_mask, block_pixels_a, lsr #15 | |
| e1f6de8f | 5373 | ldrh texel_1, [clut_ptr, texel_1] |
| 75e28f62 E |
5374 | |
| 5375 | and texel_3, texel_shift_mask, block_pixels_a, lsr #23 | |
| e1f6de8f | 5376 | ldr block_pixels_a, [block_ptr, #(64 + 16)] |
| 75e28f62 | 5377 | |
| e1f6de8f | 5378 | ldrh texel_2, [clut_ptr, texel_2] |
| 75e28f62 E |
5379 | and texel_4, texel_shift_mask, block_pixels_b, lsl #1 |
| 5380 | ||
| e1f6de8f | 5381 | ldrh texel_3, [clut_ptr, texel_3] |
| 75e28f62 E |
5382 | and texel_5, texel_shift_mask, block_pixels_b, lsr #7 |
| 5383 | ||
| e1f6de8f | 5384 | ldrh texel_4, [clut_ptr, texel_4] |
| 75e28f62 E |
5385 | and texel_6, texel_shift_mask, block_pixels_b, lsr #15 |
| 5386 | ||
| e1f6de8f | 5387 | ldrh texel_5, [clut_ptr, texel_5] |
| 75e28f62 E |
5388 | and texel_7, texel_shift_mask, block_pixels_b, lsr #23 |
| 5389 | ||
| e1f6de8f | 5390 | ldrh texel_6, [clut_ptr, texel_6] |
| 75e28f62 E |
5391 | orr texels_01, texel_0, texel_1, lsl #16 |
| 5392 | ||
| e1f6de8f | 5393 | ldrh texel_7, [clut_ptr, texel_7] |
| 75e28f62 E |
5394 | orr texels_23, texel_2, texel_3, lsl #16 |
| 5395 | ||
| 5396 | orr texels_45, texel_4, texel_5, lsl #16 | |
| e1f6de8f | 5397 | str texels_01, [block_ptr, #0] |
| 75e28f62 E |
5398 | |
| 5399 | orr texels_67, texel_6, texel_7, lsl #16 | |
| e1f6de8f | 5400 | str texels_23, [block_ptr, #4] |
| 75e28f62 E |
5401 | |
| 5402 | subs num_blocks, num_blocks, #1 | |
| e1f6de8f | 5403 | str texels_45, [block_ptr, #8] |
| 75e28f62 | 5404 | |
| e1f6de8f | 5405 | str texels_67, [block_ptr, #12] |
| 75e28f62 E |
5406 | add block_ptr, block_ptr, #64 |
| 5407 | ||
| 5408 | bne 0b | |
| db2804fb | 5409 | nop |
| 75e28f62 | 5410 | |
| db2804fb | 5411 | pop { r4 - r11, pc } |
| 75e28f62 E |
5412 | |
| 5413 | ||
| 5414 | #undef width_rounded | |
| 5415 | #undef texture_mask | |
| 5416 | #undef num_blocks | |
| 5417 | #undef texture_offset | |
| 59d15d23 | 5418 | #undef texels_low |
| 5419 | #undef texels_high | |
| 5420 | #undef texels_wide_low | |
| 5421 | #undef texels_wide_high | |
| 5422 | #undef texels_wide | |
| 5423 | #undef fb_ptr2 | |
| 8184d7c5 | 5424 | #undef temp |
| 75e28f62 E |
5425 | |
| 5426 | #define psx_gpu r0 | |
| 5427 | #define x r1 | |
| 5428 | #define y r2 | |
| 5429 | #define u r3 | |
| 5430 | #define v r4 | |
| 5431 | #define width r5 | |
| 5432 | #define height r6 | |
| 5433 | #define left_offset r8 | |
| 5434 | #define width_rounded r9 | |
| 5435 | #define right_width r10 | |
| 59d15d23 | 5436 | |
| 75e28f62 E |
5437 | #define block_width r11 |
| 5438 | ||
| 5439 | #define texture_offset_base r1 | |
| 5440 | #define texture_mask r2 | |
| 5441 | #define texture_page_ptr r3 | |
| 5442 | #define num_blocks r4 | |
| 5443 | #define block r5 | |
| 5444 | #define fb_ptr r7 | |
| 5445 | #define texture_offset r8 | |
| 5446 | #define blocks_remaining r9 | |
| 59d15d23 | 5447 | #define fb_ptr2 r10 |
| 75e28f62 E |
5448 | #define fb_ptr_pitch r12 |
| 5449 | #define texture_block_ptr r14 | |
| 5450 | ||
| 5451 | #define texture_mask_width r2 | |
| 5452 | #define texture_mask_height r3 | |
| 5453 | #define left_mask_bits r4 | |
| 5454 | #define right_mask_bits r5 | |
| 5455 | ||
| 5456 | ||
| 5457 | #undef block_masks | |
| 5458 | #undef block_masks_shifted | |
| 5459 | #undef texels | |
| 5460 | ||
| 5461 | #define block_masks d0 | |
| 5462 | #define block_masks_shifted d1 | |
| 5463 | #define draw_mask_fb_ptr d2 | |
| 5464 | #define texels q2 | |
| 5465 | ||
| 59d15d23 | 5466 | #define draw_mask_fb_ptr_a d2 |
| 5467 | #define draw_mask_fb_ptr_b d3 | |
| 5468 | #define texels_low d4 | |
| 5469 | #define texels_high d5 | |
| 5470 | #define texels_wide_low d6 | |
| 5471 | #define texels_wide_high d7 | |
| 5472 | #define texels_wide q3 | |
| 75e28f62 | 5473 | |
| 75e28f62 | 5474 | |
| 59d15d23 | 5475 | setup_sprites_16bpp_flush: |
| db2804fb | 5476 | push { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
| 5477 | add r1, r0, #psx_gpu_saved_tmp_offset | |
| 5478 | vstmia r1, { d0 - d3 } | |
| 5479 | bl flush_render_block_buffer | |
| 5480 | pop { r0 - r3, EXTRA_UNSAVED_REGS r12 } | |
| 5481 | add lr, r0, #psx_gpu_saved_tmp_offset | |
| 5482 | vldmia lr, { d0 - d3 } | |
| 75e28f62 E |
5483 | |
| 5484 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 5485 | mov num_blocks, block_width | |
| 5486 | ||
| db2804fb | 5487 | pop { pc } |
| 75e28f62 E |
5488 | |
| 5489 | function(setup_sprite_16bpp) | |
| db2804fb | 5490 | push { r3 - r11, lr } |
| e1f6de8f | 5491 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 75e28f62 | 5492 | |
| db2804fb | 5493 | ldr v, [sp, #4*(10+0)] |
| 75e28f62 E |
5494 | add fb_ptr, fb_ptr, y, lsl #11 |
| 5495 | ||
| db2804fb | 5496 | ldr width, [sp, #4*(10+1)] |
| 75e28f62 E |
5497 | add fb_ptr, fb_ptr, x, lsl #1 |
| 5498 | ||
| db2804fb | 5499 | ldr height, [sp, #4*(10+2)] |
| 75e28f62 E |
5500 | and left_offset, u, #0x7 |
| 5501 | ||
| 5502 | add texture_offset_base, u, u | |
| 5503 | add width_rounded, width, #7 | |
| 5504 | ||
| ed0fd81d | 5505 | add texture_offset_base, texture_offset_base, v, lsl #11 |
| 75e28f62 E |
5506 | mov left_mask_bits, #0xFF |
| 5507 | ||
| e1f6de8f | 5508 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
| 75e28f62 E |
5509 | add width_rounded, width_rounded, left_offset |
| 5510 | ||
| e1f6de8f | 5511 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
| 75e28f62 E |
5512 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
| 5513 | ||
| 5514 | add texture_mask, texture_mask_width, texture_mask_width | |
| 5515 | mov right_mask_bits, #0xFE | |
| 5516 | ||
| 5517 | and right_width, width_rounded, #0x7 | |
| 5518 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
| 5519 | ||
| ed0fd81d | 5520 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
| 75e28f62 E |
5521 | mov block_width, width_rounded, lsr #3 |
| 5522 | ||
| 5523 | mov right_mask_bits, right_mask_bits, lsl right_width | |
| 5524 | movw fb_ptr_pitch, #(2048 + 16) | |
| 5525 | ||
| 5526 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4 | |
| 5527 | vmov block_masks, left_mask_bits, right_mask_bits | |
| 5528 | ||
| e1f6de8f | 5529 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
5530 | add block, psx_gpu, #psx_gpu_blocks_offset |
| 5531 | ||
| 6ea0f7bf | 5532 | bic texture_offset_base, texture_offset_base, #0xF |
| 75e28f62 E |
5533 | cmp block_width, #1 |
| 5534 | ||
| e1f6de8f | 5535 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 75e28f62 E |
5536 | add block, block, num_blocks, lsl #6 |
| 5537 | ||
| 5538 | bne 0f | |
| 5539 | ||
| 5540 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
| 5541 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
| 5542 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
| 5543 | ||
| 5544 | 1: | |
| 5545 | add num_blocks, num_blocks, #1 | |
| 5546 | cmp num_blocks, #MAX_BLOCKS | |
| 59d15d23 | 5547 | blgt setup_sprites_16bpp_flush |
| 75e28f62 E |
5548 | |
| 5549 | and texture_block_ptr, texture_offset_base, texture_mask | |
| 5550 | subs height, height, #1 | |
| 5551 | ||
| 5552 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5553 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 | 5554 | |
| e1f6de8f | 5555 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5556 | add block, block, #40 |
| 5557 | ||
| 5558 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5559 | pld [fb_ptr] |
| 75e28f62 | 5560 | |
| e1f6de8f | 5561 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5562 | |
| 5563 | add block, block, #24 | |
| 5564 | add texture_offset_base, texture_offset_base, #2048 | |
| 5565 | add fb_ptr, fb_ptr, #2048 | |
| e1f6de8f | 5566 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
5567 | bne 1b |
| 5568 | ||
| db2804fb | 5569 | pop { r3 - r11, pc } |
| 75e28f62 E |
5570 | |
| 5571 | 0: | |
| 5572 | add num_blocks, num_blocks, block_width | |
| 5573 | mov texture_offset, texture_offset_base | |
| 5574 | ||
| 5575 | cmp num_blocks, #MAX_BLOCKS | |
| 59d15d23 | 5576 | blgt setup_sprites_16bpp_flush |
| 75e28f62 E |
5577 | |
| 5578 | add texture_offset_base, texture_offset_base, #2048 | |
| 5579 | and texture_block_ptr, texture_offset, texture_mask | |
| 5580 | ||
| 5581 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5582 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 | 5583 | |
| e1f6de8f | 5584 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5585 | add block, block, #40 |
| 5586 | ||
| 5587 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
| 5588 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5589 | pld [fb_ptr] |
| 75e28f62 | 5590 | |
| e1f6de8f | 5591 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5592 | subs blocks_remaining, block_width, #2 |
| 5593 | ||
| 5594 | add texture_offset, texture_offset, #16 | |
| 5595 | add fb_ptr, fb_ptr, #16 | |
| 5596 | ||
| 5597 | vmov.u8 draw_mask_fb_ptr, #0 | |
| 5598 | ||
| 5599 | add block, block, #24 | |
| 5600 | beq 2f | |
| 5601 | ||
| 5602 | 1: | |
| 5603 | and texture_block_ptr, texture_offset, texture_mask | |
| 5604 | subs blocks_remaining, blocks_remaining, #1 | |
| 5605 | ||
| 5606 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5607 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 | 5608 | |
| e1f6de8f | 5609 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5610 | add block, block, #40 |
| 5611 | ||
| 5612 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5613 | pld [fb_ptr] |
| 75e28f62 | 5614 | |
| e1f6de8f | 5615 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5616 | |
| 5617 | add texture_offset, texture_offset, #16 | |
| 5618 | add fb_ptr, fb_ptr, #16 | |
| 5619 | ||
| 5620 | add block, block, #24 | |
| 5621 | bne 1b | |
| 5622 | ||
| 5623 | 2: | |
| 5624 | and texture_block_ptr, texture_offset, texture_mask | |
| 5625 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5626 | ||
| e1f6de8f | 5627 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 E |
5628 | vdup.u8 draw_mask_fb_ptr, block_masks[4] |
| 5629 | ||
| e1f6de8f | 5630 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5631 | add block, block, #40 |
| 5632 | ||
| 5633 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5634 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5635 | |
| 5636 | add block, block, #24 | |
| 5637 | subs height, height, #1 | |
| 5638 | ||
| 5639 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
| e1f6de8f | 5640 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
5641 | |
| 5642 | bne 0b | |
| db2804fb | 5643 | nop |
| 75e28f62 | 5644 | |
| db2804fb | 5645 | pop { r3 - r11, pc } |
| 75e28f62 E |
5646 | |
| 5647 | ||
| 59d15d23 | 5648 | // 4x version |
| 5649 | // FIXME: duplicate code with normal version :( | |
| 5650 | #undef draw_mask_fb_ptr | |
| 5651 | ||
| 5652 | function(setup_sprite_16bpp_4x) | |
| db2804fb | 5653 | push { r3 - r11, lr } |
| e1f6de8f | 5654 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 59d15d23 | 5655 | |
| db2804fb | 5656 | ldr v, [sp, #4*(10+0)] |
| 59d15d23 | 5657 | add fb_ptr, fb_ptr, y, lsl #11 |
| 5658 | ||
| db2804fb | 5659 | ldr width, [sp, #4*(10+1)] |
| 59d15d23 | 5660 | add fb_ptr, fb_ptr, x, lsl #1 |
| 5661 | ||
| db2804fb | 5662 | ldr height, [sp, #4*(10+2)] |
| 59d15d23 | 5663 | and left_offset, u, #0x7 |
| 5664 | ||
| 5665 | add texture_offset_base, u, u | |
| 5666 | add width_rounded, width, #7 | |
| 5667 | ||
| ed0fd81d | 5668 | add texture_offset_base, texture_offset_base, v, lsl #11 |
| 59d15d23 | 5669 | movw left_mask_bits, #0xFFFF |
| 5670 | ||
| e1f6de8f | 5671 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
| 59d15d23 | 5672 | add width_rounded, width_rounded, left_offset |
| 5673 | ||
| 5674 | lsl left_offset, #1 | |
| 5675 | ||
| e1f6de8f | 5676 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
| 59d15d23 | 5677 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
| 5678 | ||
| 5679 | add texture_mask, texture_mask_width, texture_mask_width | |
| 5680 | movw right_mask_bits, #0xFFFC | |
| 5681 | ||
| 5682 | and right_width, width_rounded, #0x7 | |
| 5683 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
| 5684 | ||
| 5685 | lsl right_width, #1 | |
| 5686 | ||
| ed0fd81d | 5687 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
| 59d15d23 | 5688 | mov block_width, width_rounded, lsr #3 |
| 5689 | ||
| 5690 | mov right_mask_bits, right_mask_bits, lsl right_width | |
| 5691 | movw fb_ptr_pitch, #(2048 + 16) * 2 | |
| 5692 | ||
| 5693 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4+1 | |
| 5694 | vmov block_masks, left_mask_bits, right_mask_bits | |
| 5695 | ||
| e1f6de8f | 5696 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 59d15d23 | 5697 | add block, psx_gpu, #psx_gpu_blocks_offset |
| 5698 | ||
| 5699 | bic texture_offset_base, texture_offset_base, #0xF | |
| 5700 | cmp block_width, #1 | |
| 5701 | ||
| e1f6de8f | 5702 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 59d15d23 | 5703 | add block, block, num_blocks, lsl #6 |
| 5704 | ||
| 5705 | lsl block_width, #2 | |
| 5706 | bne 0f | |
| 5707 | ||
| 5708 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
| 5709 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
| 5710 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] | |
| 5711 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
| 5712 | ||
| 5713 | 1: | |
| 5714 | add num_blocks, num_blocks, block_width | |
| 5715 | cmp num_blocks, #MAX_BLOCKS | |
| 5716 | blgt setup_sprites_16bpp_flush | |
| 5717 | ||
| 5718 | and texture_block_ptr, texture_offset_base, texture_mask | |
| 5719 | subs height, height, #1 | |
| 5720 | ||
| 5721 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5722 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5723 | |
| 5724 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5725 | ||
| 5726 | add texture_offset_base, texture_offset_base, #2048 | |
| 5727 | add fb_ptr, fb_ptr, #2048*2 | |
| e1f6de8f | 5728 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 59d15d23 | 5729 | bne 1b |
| 5730 | ||
| db2804fb | 5731 | pop { r3 - r11, pc } |
| 59d15d23 | 5732 | |
| 5733 | 0: | |
| 5734 | add num_blocks, num_blocks, block_width | |
| 5735 | mov texture_offset, texture_offset_base | |
| 5736 | ||
| 5737 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] // left_mask_bits | |
| 5738 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
| 5739 | ||
| 5740 | cmp num_blocks, #MAX_BLOCKS | |
| 5741 | blgt setup_sprites_16bpp_flush | |
| 5742 | ||
| 5743 | add texture_offset_base, texture_offset_base, #2048 | |
| 5744 | and texture_block_ptr, texture_offset, texture_mask | |
| 5745 | ||
| 5746 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5747 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5748 | |
| 5749 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5750 | ||
| 5751 | subs blocks_remaining, block_width, #2*4 | |
| 5752 | add texture_offset, texture_offset, #16 | |
| 5753 | ||
| 5754 | vmov.u8 draw_mask_fb_ptr_a, #0 | |
| 5755 | vmov.u8 draw_mask_fb_ptr_b, #0 | |
| 5756 | ||
| 5757 | add fb_ptr, fb_ptr, #16*2 | |
| 5758 | beq 2f | |
| 5759 | ||
| 5760 | 1: | |
| 5761 | and texture_block_ptr, texture_offset, texture_mask | |
| 5762 | subs blocks_remaining, blocks_remaining, #4 | |
| 5763 | ||
| 5764 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5765 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5766 | |
| 5767 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5768 | add texture_offset, texture_offset, #16 | |
| 5769 | ||
| 5770 | add fb_ptr, fb_ptr, #16*2 | |
| 5771 | bgt 1b | |
| 5772 | ||
| 5773 | 2: | |
| 5774 | vdup.u8 draw_mask_fb_ptr_a, block_masks[4] // right_mask_bits | |
| 5775 | vdup.u8 draw_mask_fb_ptr_b, block_masks[5] | |
| 5776 | ||
| 5777 | and texture_block_ptr, texture_offset, texture_mask | |
| 5778 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5779 | ||
| e1f6de8f | 5780 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5781 | |
| 5782 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5783 | subs height, height, #1 | |
| 5784 | ||
| 5785 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
| e1f6de8f | 5786 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 59d15d23 | 5787 | |
| 5788 | bne 0b | |
| db2804fb | 5789 | nop |
| 59d15d23 | 5790 | |
| db2804fb | 5791 | pop { r3 - r11, pc } |
| 59d15d23 | 5792 | |
| 5793 | ||
| f0931e56 | 5794 | #undef width |
| 5795 | #undef right_width | |
| 5796 | #undef right_mask_bits | |
| 5797 | #undef color | |
| 5798 | #undef height | |
| 5799 | #undef blocks_remaining | |
| 5800 | #undef colors | |
| 5801 | #undef right_mask | |
| 5802 | #undef test_mask | |
| 5803 | #undef draw_mask | |
| 5804 | ||
| 5805 | #define psx_gpu r0 | |
| 5806 | #define x r1 | |
| 5807 | #define y r2 | |
| 5808 | #define width r3 | |
| 5809 | #define right_width r5 | |
| 5810 | #define right_mask_bits r6 | |
| 5811 | #define fb_ptr r7 | |
| 5812 | #define color r8 | |
| 5813 | #define height r9 | |
| 5814 | #define fb_ptr_pitch r12 | |
| 5815 | ||
| 5816 | // referenced by setup_sprites_16bpp_flush | |
| 5817 | #define num_blocks r4 | |
| 5818 | #define block r5 | |
| 5819 | #define block_width r11 | |
| 5820 | ||
| 5821 | #define color_r r1 | |
| 5822 | #define color_g r2 | |
| 5823 | #define color_b r8 | |
| 5824 | #define blocks_remaining r6 | |
| 5825 | ||
| 5826 | #define colors q0 | |
| 5827 | #define right_mask q1 | |
| 5828 | #define test_mask q2 | |
| 5829 | #define draw_mask q2 | |
| 5830 | #define draw_mask_bits_fb_ptr d6 | |
| 5831 | ||
| 5832 | ||
| 5833 | .align 3 | |
| 5834 | ||
| 2d658c89 | 5835 | function(setup_sprite_untextured_512) |
| db2804fb | 5836 | push { r4 - r11, r14 } |
| f0931e56 | 5837 | |
| db2804fb | 5838 | ldr width, [sp, #4*(9+1)] |
| e1f6de8f | 5839 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| f0931e56 | 5840 | |
| db2804fb | 5841 | ldr height, [sp, #4*(9+2)] |
| f0931e56 | 5842 | add fb_ptr, fb_ptr, y, lsl #11 |
| 5843 | ||
| 5844 | add fb_ptr, fb_ptr, x, lsl #1 | |
| 5845 | sub right_width, width, #1 | |
| 5846 | ||
| db2804fb | 5847 | ldr color, [sp, #4*(9+3)] |
| f0931e56 | 5848 | and right_width, #7 |
| 5849 | ||
| 5850 | add block_width, width, #7 | |
| 5851 | add right_width, #1 | |
| 5852 | ||
| 5853 | lsr block_width, #3 | |
| 5854 | mov right_mask_bits, #0xff | |
| 5855 | ||
| 5856 | sub fb_ptr_pitch, block_width, #1 | |
| 5857 | lsl right_mask_bits, right_width | |
| 5858 | ||
| 5859 | lsl fb_ptr_pitch, #3+1 | |
| 5860 | ubfx color_r, color, #3, #5 | |
| 5861 | ||
| 5862 | rsb fb_ptr_pitch, #1024*2 | |
| 5863 | ubfx color_g, color, #11, #5 | |
| 5864 | ||
| e1f6de8f | 5865 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| f0931e56 | 5866 | ubfx color_b, color, #19, #5 |
| 5867 | ||
| 5868 | vdup.u16 right_mask, right_mask_bits | |
| 5869 | orr color, color_r, color_b, lsl #10 | |
| 5870 | ||
| e1f6de8f | 5871 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| f0931e56 | 5872 | orr color, color, color_g, lsl #5 |
| 5873 | ||
| 5874 | vtst.u16 right_mask, right_mask, test_mask | |
| 5875 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 5876 | ||
| 5877 | vdup.u16 colors, color | |
| 5878 | add block, block, num_blocks, lsl #6 | |
| 5879 | ||
| 5880 | ||
| 5881 | setup_sprite_untextured_height_loop: | |
| 5882 | add num_blocks, block_width | |
| 5883 | sub blocks_remaining, block_width, #1 | |
| 5884 | ||
| 5885 | cmp num_blocks, #MAX_BLOCKS | |
| 5886 | blgt setup_sprites_16bpp_flush | |
| 5887 | ||
| 5888 | cmp blocks_remaining, #0 | |
| 5889 | ble 1f | |
| 5890 | ||
| 5891 | vmov.u8 draw_mask, #0 /* zero_mask */ | |
| 5892 | vmov.u8 draw_mask_bits_fb_ptr, #0 | |
| 5893 | ||
| 5894 | 0: | |
| e1f6de8f | 5895 | vst1.u32 { draw_mask }, [block, :128]! |
| f0931e56 | 5896 | subs blocks_remaining, #1 |
| 5897 | ||
| e1f6de8f | 5898 | vst1.u32 { colors }, [block, :128] |
| f0931e56 | 5899 | add block, block, #24 |
| 5900 | ||
| 5901 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5902 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
| f0931e56 | 5903 | |
| 5904 | add block, block, #24 | |
| 5905 | add fb_ptr, #8*2 | |
| 5906 | bgt 0b | |
| 5907 | ||
| 5908 | 1: | |
| e1f6de8f | 5909 | vst1.u32 { right_mask }, [block, :128]! |
| f0931e56 | 5910 | subs height, #1 |
| 5911 | ||
| e1f6de8f | 5912 | vst1.u32 { colors }, [block, :128] |
| f0931e56 | 5913 | add block, block, #24 |
| 5914 | ||
| 5915 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5916 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
| f0931e56 | 5917 | |
| 5918 | add block, block, #24 | |
| 5919 | add fb_ptr, fb_ptr_pitch | |
| 5920 | ||
| e1f6de8f | 5921 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| f0931e56 | 5922 | bgt setup_sprite_untextured_height_loop |
| 5923 | ||
| db2804fb | 5924 | pop { r4 - r11, pc } |
| f0931e56 | 5925 | |
| 5926 | ||
| 5927 | ||
| 75e28f62 E |
5928 | #undef texture_page_ptr |
| 5929 | #undef vram_ptr | |
| 5930 | #undef dirty_textures_mask | |
| 5931 | #undef current_texture_mask | |
| 5932 | ||
| 5933 | #define psx_gpu r0 | |
| 5934 | #define current_texture_page r1 | |
| 5935 | #define texture_page_ptr r2 | |
| 5936 | #define vram_ptr_a r3 | |
| 5937 | #define current_texture_page_x r12 | |
| 5938 | #define current_texture_page_y r4 | |
| 5939 | #define dirty_textures_mask r5 | |
| 5940 | #define tile_y r6 | |
| 5941 | #define tile_x r7 | |
| 5942 | #define sub_y r8 | |
| 5943 | #define current_texture_mask r9 | |
| 5944 | #define c_4096 r10 | |
| 5945 | #define vram_ptr_b r11 | |
| 5946 | ||
| 5947 | #define texel_block_a d0 | |
| 5948 | #define texel_block_b d1 | |
| 5949 | #define texel_block_expanded_a q1 | |
| 5950 | #define texel_block_expanded_b q2 | |
| 5951 | #define texel_block_expanded_ab q2 | |
| 5952 | #define texel_block_expanded_c q3 | |
| fab27ba2 | 5953 | #define texel_block_expanded_d q0 |
| 75e28f62 E |
5954 | #define texel_block_expanded_cd q3 |
| 5955 | ||
| 5956 | function(update_texture_4bpp_cache) | |
| db2804fb | 5957 | push { r3 - r11, r14 } |
| 75e28f62 E |
5958 | vpush { q0 - q3 } |
| 5959 | ||
| e1f6de8f | 5960 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
| 75e28f62 | 5961 | |
| e1f6de8f | 5962 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
| 5963 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
| 75e28f62 E |
5964 | |
| 5965 | and current_texture_page_x, current_texture_page, #0xF | |
| e1f6de8f | 5966 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
| 75e28f62 E |
5967 | |
| 5968 | mov current_texture_page_y, current_texture_page, lsr #4 | |
| e1f6de8f | 5969 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
| 75e28f62 E |
5970 | |
| 5971 | add vram_ptr_a, vram_ptr_a, current_texture_page_y, lsl #19 | |
| 5972 | mov tile_y, #16 | |
| 5973 | ||
| 5974 | add vram_ptr_a, vram_ptr_a, current_texture_page_x, lsl #7 | |
| 5975 | bic dirty_textures_mask, current_texture_mask | |
| 5976 | ||
| 5977 | mov tile_x, #16 | |
| e1f6de8f | 5978 | str dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
| 75e28f62 E |
5979 | |
| 5980 | mov sub_y, #8 | |
| 5981 | movw c_4096, #4096 | |
| 5982 | ||
| 5983 | add vram_ptr_b, vram_ptr_a, #2048 | |
| 5984 | ||
| 5985 | 0: | |
| e1f6de8f | 5986 | vld1.u32 { texel_block_a }, [vram_ptr_a, :64], c_4096 |
| 5987 | vld1.u32 { texel_block_b }, [vram_ptr_b, :64], c_4096 | |
| 75e28f62 E |
5988 | |
| 5989 | vmovl.u8 texel_block_expanded_a, texel_block_a | |
| 5990 | vshll.u8 texel_block_expanded_b, texel_block_a, #4 | |
| 5991 | vmovl.u8 texel_block_expanded_c, texel_block_b | |
| 5992 | vshll.u8 texel_block_expanded_d, texel_block_b, #4 | |
| 5993 | ||
| 5994 | vbic.u16 texel_block_expanded_a, #0x00F0 | |
| 5995 | vbic.u16 texel_block_expanded_b, #0x00F0 | |
| 5996 | vbic.u16 texel_block_expanded_c, #0x00F0 | |
| 5997 | vbic.u16 texel_block_expanded_d, #0x00F0 | |
| 5998 | ||
| 5999 | vorr.u16 texel_block_expanded_ab, texel_block_expanded_a, \ | |
| 6000 | texel_block_expanded_b | |
| 6001 | vorr.u16 texel_block_expanded_cd, texel_block_expanded_c, \ | |
| 6002 | texel_block_expanded_d | |
| 6003 | ||
| 6004 | vst1.u32 { texel_block_expanded_ab, texel_block_expanded_cd }, \ | |
| e1f6de8f | 6005 | [texture_page_ptr, :256]! |
| 75e28f62 E |
6006 | |
| 6007 | subs sub_y, sub_y, #1 | |
| 6008 | bne 0b | |
| 6009 | ||
| 6010 | mov sub_y, #8 | |
| 6011 | add vram_ptr_a, vram_ptr_a, #8 | |
| 6012 | add vram_ptr_b, vram_ptr_b, #8 | |
| 6013 | ||
| 6014 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6015 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6016 | ||
| 6017 | subs tile_x, tile_x, #1 | |
| 6018 | bne 0b | |
| 6019 | ||
| 6020 | mov tile_x, #16 | |
| 6021 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6022 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6023 | ||
| 6024 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
| 6025 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
| 6026 | ||
| 6027 | subs tile_y, tile_y, #1 | |
| 6028 | bne 0b | |
| 6029 | ||
| 6030 | vpop { q0 - q3 } | |
| db2804fb | 6031 | pop { r3 - r11, pc } |
| 75e28f62 E |
6032 | |
| 6033 | ||
| 6034 | #undef current_texture_page | |
| 6035 | ||
| 6036 | #define psx_gpu r0 | |
| 6037 | #define texture_page r1 | |
| 6038 | #define texture_page_ptr r2 | |
| 6039 | #define vram_ptr_a r3 | |
| 6040 | #define texture_page_x r12 | |
| 6041 | #define texture_page_y r4 | |
| 6042 | #define current_texture_page r5 | |
| 6043 | #define tile_y r6 | |
| 6044 | #define tile_x r7 | |
| 6045 | #define sub_y r8 | |
| 6046 | #define c_4096 r10 | |
| 6047 | #define vram_ptr_b r11 | |
| 6048 | ||
| 6049 | ||
| 6050 | #undef texels_a | |
| 6051 | #undef texels_b | |
| 6052 | ||
| 6053 | #define texels_a q0 | |
| 6054 | #define texels_b q1 | |
| 6055 | #define texels_c q2 | |
| 6056 | #define texels_d q3 | |
| 6057 | ||
| 6058 | ||
| 6059 | function(update_texture_8bpp_cache_slice) | |
| 6060 | stmdb sp!, { r4 - r11, r14 } | |
| 75e28f62 | 6061 | |
| e1f6de8f | 6062 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
| 6063 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
| 75e28f62 | 6064 | |
| e1f6de8f | 6065 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
| 75e28f62 E |
6066 | mov tile_y, #16 |
| 6067 | ||
| 6068 | and texture_page_x, texture_page, #0xF | |
| 6069 | mov texture_page_y, texture_page, lsr #4 | |
| 6070 | ||
| 6071 | add vram_ptr_a, vram_ptr_a, texture_page_x, lsl #7 | |
| 6072 | mov tile_x, #8 | |
| 6073 | ||
| 6074 | add vram_ptr_a, vram_ptr_a, texture_page_y, lsl #19 | |
| 6075 | eor current_texture_page, current_texture_page, texture_page | |
| 6076 | ||
| 6077 | ands current_texture_page, current_texture_page, #0x1 | |
| 6078 | mov sub_y, #4 | |
| 6079 | ||
| 6080 | addne texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
| 6081 | movw c_4096, #4096 | |
| 6082 | ||
| 6083 | add vram_ptr_b, vram_ptr_a, #2048 | |
| 6084 | ||
| 6085 | 0: | |
| e1f6de8f | 6086 | vld1.u32 { texels_a }, [vram_ptr_a, :128], c_4096 |
| 6087 | vld1.u32 { texels_b }, [vram_ptr_b, :128], c_4096 | |
| 6088 | vld1.u32 { texels_c }, [vram_ptr_a, :128], c_4096 | |
| 6089 | vld1.u32 { texels_d }, [vram_ptr_b, :128], c_4096 | |
| 75e28f62 | 6090 | |
| e1f6de8f | 6091 | vst1.u32 { texels_a, texels_b }, [texture_page_ptr, :256]! |
| 6092 | vst1.u32 { texels_c, texels_d }, [texture_page_ptr, :256]! | |
| 75e28f62 E |
6093 | |
| 6094 | subs sub_y, sub_y, #1 | |
| 6095 | bne 0b | |
| 6096 | ||
| 6097 | mov sub_y, #4 | |
| 6098 | ||
| 6099 | add vram_ptr_a, vram_ptr_a, #16 | |
| 6100 | add vram_ptr_b, vram_ptr_b, #16 | |
| 6101 | ||
| 6102 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6103 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6104 | ||
| 6105 | subs tile_x, tile_x, #1 | |
| 6106 | bne 0b | |
| 6107 | ||
| 6108 | mov tile_x, #8 | |
| 6109 | ||
| 6110 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6111 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6112 | ||
| 6113 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
| 6114 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
| 6115 | ||
| 6116 | subs tile_y, tile_y, #1 | |
| 6117 | add texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
| 6118 | ||
| 6119 | bne 0b | |
| 6120 | ||
| 75e28f62 E |
6121 | ldmia sp!, { r4 - r11, pc } |
| 6122 | ||
| 50f9355a | 6123 | |
| 6124 | /* void scale2x_tiles8(void *dst, const void *src, int w8, int h) */ | |
| 6125 | function(scale2x_tiles8) | |
| 6126 | push { r4, r14 } | |
| 6127 | ||
| 6128 | mov r4, r1 | |
| 6129 | add r12, r0, #1024*2 | |
| 6130 | mov r14, r2 | |
| 6131 | ||
| 6132 | 0: | |
| c2a25f67 | 6133 | pld [r1, #1024*2] |
| e1f6de8f | 6134 | vld1.u16 { q0 }, [r1, :128]! |
| 6135 | vld1.u16 { q2 }, [r1, :128]! | |
| 50f9355a | 6136 | vmov q1, q0 |
| 6137 | vmov q3, q2 | |
| 6138 | vzip.16 q0, q1 | |
| 6139 | vzip.16 q2, q3 | |
| 6140 | subs r14, #2 | |
| e1f6de8f | 6141 | vst1.u16 { q0, q1 }, [r0, :128]! |
| 6142 | vst1.u16 { q0, q1 }, [r12, :128]! | |
| 50f9355a | 6143 | blt 1f |
| e1f6de8f | 6144 | vst1.u16 { q2, q3 }, [r0, :128]! |
| 6145 | vst1.u16 { q2, q3 }, [r12, :128]! | |
| 50f9355a | 6146 | bgt 0b |
| 6147 | 1: | |
| 6148 | subs r3, #1 | |
| 6149 | mov r14, r2 | |
| 6150 | add r0, #1024*2*2 | |
| 6151 | add r4, #1024*2 | |
| ed0fd81d | 6152 | sub r0, r0, r2, lsl #4+1 |
| 50f9355a | 6153 | mov r1, r4 |
| 6154 | add r12, r0, #1024*2 | |
| 6155 | bgt 0b | |
| 6156 | nop | |
| 6157 | ||
| 6158 | pop { r4, pc } | |
| 59d15d23 | 6159 | |
| 6160 | // vim:filetype=armasm |