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 |
| 8184d7c5 | 226 | #endif |
| 227 | ||
| 75e28f62 E |
228 | @ r0: psx_gpu |
| 229 | @ r1: v_a | |
| 230 | @ r2: v_b | |
| 231 | @ r3: v_c | |
| 232 | ||
| 233 | function(compute_all_gradients) | |
| 234 | // First compute the triangle area reciprocal and shift. The division will | |
| 235 | // happen concurrently with much of the work which follows. | |
| 236 | @ r12 = psx_gpu->triangle_area | |
| e1f6de8f | 237 | ldr r12, [psx_gpu, #psx_gpu_triangle_area_offset] |
| 75e28f62 | 238 | stmdb sp!, { r4 - r11, lr } |
| 3f0189c6 | 239 | save_abi_regs() |
| 75e28f62 E |
240 | |
| 241 | @ load exponent of 62 into upper half of double | |
| 242 | movw r4, #0 | |
| 243 | clz r14, r12 @ r14 = shift | |
| 244 | ||
| 245 | movt r4, #((62 + 1023) << 4) | |
| 246 | mov r12, r12, lsl r14 @ r12 = triangle_area_normalized | |
| 247 | ||
| 248 | @ load area normalized into lower half of double | |
| 249 | mov r5, r12, lsr #10 | |
| 250 | vmov.f64 d30, r5, r4 @ d30 = (1 << 62) + ta_n | |
| 251 | ||
| 252 | movt r4, #((1022 + 31) << 4) | |
| 253 | mov r5, r12, lsl #20 | |
| 254 | ||
| 255 | add r4, r4, r12, lsr #11 | |
| 256 | vmov.f64 d31, r5, r4 | |
| 257 | ||
| 258 | vdiv.f64 d30, d30, d31 @ d30 = ((1 << 62) + ta_n) / ta_n | |
| 259 | ||
| 260 | // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) = | |
| 261 | // ( d0 * d1 ) - ( d2 * d3 ) = | |
| 262 | // ( m0 ) - ( m1 ) = gradient | |
| 263 | ||
| 264 | // This is split to do 12 elements at a time over three sets: a, b, and c. | |
| 265 | // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so | |
| 266 | // two of the slots are unused. | |
| 267 | ||
| 268 | // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as | |
| 269 | // is g. | |
| 270 | ||
| 271 | // First type is: uvrg bxxx xxxx | |
| 272 | // Second type is: yyyy ybyy uvrg | |
| 273 | // Since x_a and y_c are the same the same variable is used for both. | |
| 274 | ||
| e1f6de8f | 275 | vld1.u32 { v0 }, [v_a, :128] @ v0 = { uvrg0, b0, x0, y0 } |
| 276 | ldrsh x0, [v_a, #8] @ load x0 | |
| 75e28f62 | 277 | |
| e1f6de8f | 278 | vld1.u32 { v1 }, [v_b, :128] @ v1 = { uvrg1, b1, x1, y1} |
| 279 | ldrh x1, [v_b, #8] @ load x1 | |
| 75e28f62 | 280 | |
| e1f6de8f | 281 | vld1.u32 { v2 }, [v_c, :128] @ v2 = { uvrg2, b2, x2, y2 } |
| 282 | ldrh x2, [v_c, #8] @ load x2 | |
| 75e28f62 E |
283 | |
| 284 | vmovl.u8 uvrg_xxxx0, uvrgb0 @ uvrg_xxxx0 = { uv0, rg0, b0-, -- } | |
| e1f6de8f | 285 | ldrh y0, [v_a, #10] @ load y0 |
| 75e28f62 E |
286 | |
| 287 | vmovl.u8 uvrg_xxxx1, uvrgb1 @ uvrg_xxxx1 = { uv1, rg1, b1-, -- } | |
| e1f6de8f | 288 | ldrh y1, [v_b, #10] @ load y1 |
| 75e28f62 E |
289 | |
| 290 | vmovl.u8 uvrg_xxxx2, uvrgb2 @ uvrg_xxxx2 = { uv2, rg2, b2-, -- } | |
| e1f6de8f | 291 | ldrh y2, [v_c, #10] @ load y2 |
| 75e28f62 E |
292 | |
| 293 | vmov.u8 uvrg0b, uvrg0 @ uvrg0b = { uv0, rg0 } | |
| 294 | vdup.u16 xxxx0, x0_y0[0] @ xxxx0 = { xx0, xx0 } | |
| 295 | ||
| 296 | orr x1_x2, x1, x2, lsl #16 @ x1_x2 = { x1, x2 } | |
| 297 | pkhbt x0_x1, x0, x1, lsl #16 @ x0_x1 = { x0, x1 } | |
| 298 | ||
| 299 | vmov.u8 uvrg1b, uvrg1 @ uvrg1b = { uv1, rg1 } | |
| 300 | vdup.u16 xxxx1, x1_y1[0] @ xxxx1 = { xx1, xx1 } | |
| 301 | ||
| 302 | vmov.u8 uvrg2b, uvrg2 @ uvrg2b = { uv2, rg2 } | |
| 303 | vdup.u16 xxxx2, x2_y2[0] @ xxxx2 = { xx2, xx2 } | |
| 304 | ||
| e1f6de8f | 305 | ldrb b2, [v_c, #4] @ load b2 |
| 75e28f62 E |
306 | orr y0_y1, y0, y1, lsl #16 @ y0_y1 = { y0, y1 } |
| 307 | ||
| e1f6de8f | 308 | ldrb b1, [v_b, #4] @ load b1 |
| 75e28f62 E |
309 | orr y1_y2, y1, y2, lsl #16 @ y1_y2 = { y1, y2 } |
| 310 | ||
| 311 | vdup.u16 yyyy0, x0_y0[1] @ yyyy0 = { yy0, yy0 } | |
| 312 | vsub.s16 d0_ab, x1_ab, x0_ab | |
| 313 | ||
| e1f6de8f | 314 | ldrb b0, [v_a, #4] @ load b0 |
| 75e28f62 E |
315 | orr b1_b2, b1, b2, lsl #16 @ b1_b2 = { b1, b2 } |
| 316 | ||
| 317 | vdup.u16 yyyy1, x1_y1[1] @ yyyy1 = { yy1, yy1 } | |
| 318 | vsub.s16 d2_ab, x2_ab, x1_ab | |
| 319 | ||
| 320 | vdup.u16 yyyy2, x2_y2[1] @ yyyy2 = { yy2, yy2 } | |
| 321 | vsub.s16 d1_ab, y2_ab, y1_ab | |
| 322 | ||
| 323 | orr b0_b1, b0, b1, lsl #16 @ b1_b2 = { b1, b2 } | |
| 324 | ssub16 dx, x1_x2, x0_x1 @ dx = { x1 - x0, x2 - x1 } | |
| 325 | ||
| 326 | ssub16 dy, y1_y2, y0_y1 @ dy = { y1 - y0, y2 - y1 } | |
| 327 | ssub16 db, b1_b2, b0_b1 @ db = { b1 - b0, b2 - b1 } | |
| 328 | ||
| 329 | vsub.s16 d3_ab, y1_ab, y0_ab | |
| 330 | smusdx ga_by, dx, db @ ga_by = ((x1 - x0) * (b2 - b1)) - | |
| 331 | @ ((x2 - X1) * (b1 - b0)) | |
| 332 | vmull.s16 ga_uvrg_x, d0_a, d1_a | |
| 333 | smusdx ga_bx, db, dy @ ga_bx = ((b1 - b0) * (y2 - y1)) - | |
| 334 | @ ((b2 - b1) * (y1 - y0)) | |
| 335 | vmlsl.s16 ga_uvrg_x, d2_a, d3_a | |
| 336 | movs gs_bx, ga_bx, asr #31 | |
| 337 | ||
| 338 | vmull.s16 ga_uvrg_y, d0_b, d1_b | |
| 339 | rsbmi ga_bx, ga_bx, #0 | |
| 340 | ||
| c6063f89 | 341 | @ r12 = psx_gpu->uvrgb_phase |
| e1f6de8f | 342 | ldr r12, [psx_gpu, #psx_gpu_uvrgb_phase_offset] |
| c6063f89 | 343 | |
| 75e28f62 E |
344 | vmlsl.s16 ga_uvrg_y, d2_b, d3_b |
| 345 | movs gs_by, ga_by, asr #31 | |
| 346 | ||
| 347 | vshr.u64 d0, d30, #22 | |
| c6063f89 | 348 | add b_base, r12, b0, lsl #16 |
| 349 | ||
| 350 | vdup.u32 uvrgb_phase, r12 | |
| 75e28f62 E |
351 | |
| 352 | rsbmi ga_by, ga_by, #0 | |
| 353 | vclt.s32 gs_uvrg_x, ga_uvrg_x, #0 @ gs_uvrg_x = ga_uvrg_x < 0 | |
| 354 | ||
| 355 | @ r12 = psx_gpu->triangle_winding_offset | |
| e1f6de8f | 356 | ldrb r12, [psx_gpu, #psx_gpu_triangle_winding_offset] |
| 75e28f62 E |
357 | vclt.s32 gs_uvrg_y, ga_uvrg_y, #0 @ gs_uvrg_y = ga_uvrg_y < 0 |
| 358 | ||
| 75e28f62 E |
359 | rsb r12, r12, #0 @ r12 = -(triangle->winding) |
| 360 | ||
| 361 | vdup.u32 w_mask, r12 @ w_mask = { -w, -w, -w, -w } | |
| 362 | sub r14, r14, #(62 - 12) @ r14 = shift - (62 - FIXED_BITS) | |
| 363 | ||
| 364 | vshll.u16 uvrg_base, uvrg0, #16 @ uvrg_base = uvrg0 << 16 | |
| aafce833 | 365 | vdup.u32 r_shift, r14 @ r_shift = { shift, shift*, shift, shift* } |
| 366 | @ * - vshl.u64: ignored by hw | |
| c6063f89 | 367 | vadd.u32 uvrg_base, uvrgb_phase |
| 75e28f62 E |
368 | vabs.s32 ga_uvrg_x, ga_uvrg_x @ ga_uvrg_x = abs(ga_uvrg_x) |
| 369 | ||
| 370 | vmov area_r_s, s0 @ area_r_s = triangle_reciprocal | |
| 371 | vabs.s32 ga_uvrg_y, ga_uvrg_y @ ga_uvrg_y = abs(ga_uvrg_y) | |
| 372 | ||
| 373 | vmull.u32 gw_rg_x, ga_rg_x, d0[0] | |
| 374 | vmull.u32 gw_uv_x, ga_uv_x, d0[0] | |
| 375 | vmull.u32 gw_rg_y, ga_rg_y, d0[0] | |
| 376 | vmull.u32 gw_uv_y, ga_uv_y, d0[0] | |
| 377 | ||
| 378 | vshl.u64 gw_rg_x, gw_rg_x, r_shift | |
| 379 | vshl.u64 gw_uv_x, gw_uv_x, r_shift | |
| 380 | vshl.u64 gw_rg_y, gw_rg_y, r_shift | |
| 381 | vshl.u64 gw_uv_y, gw_uv_y, r_shift | |
| 382 | ||
| 383 | veor.u32 gs_uvrg_x, gs_uvrg_x, w_mask | |
| 384 | vmovn.u64 g_uv_x, gw_uv_x | |
| 385 | ||
| 386 | veor.u32 gs_uvrg_y, gs_uvrg_y, w_mask | |
| 387 | vmovn.u64 g_rg_x, gw_rg_x | |
| 388 | ||
| 389 | veor.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
| 390 | vmovn.u64 g_uv_y, gw_uv_y | |
| 391 | ||
| 392 | vsub.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
| 393 | vmovn.u64 g_rg_y, gw_rg_y | |
| 394 | ||
| 395 | veor.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
| 396 | mov ga_bx, ga_bx, lsl #13 | |
| 397 | ||
| 398 | vsub.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
| 399 | mov ga_by, ga_by, lsl #13 | |
| 400 | ||
| 401 | vdup.u32 x0_y0, x0 | |
| 402 | umull gw_bx_l, gw_bx_h, ga_bx, area_r_s | |
| 403 | ||
| 404 | vshl.u32 g_uvrg_x, g_uvrg_x, #4 | |
| 405 | vshl.u32 g_uvrg_y, g_uvrg_y, #4 | |
| 406 | ||
| 407 | umull gw_by_l, gw_by_h, ga_by, area_r_s | |
| 408 | vmls.s32 uvrg_base, ga_uvrg_x, x0_y0[0] | |
| 409 | ||
| 410 | eor gs_bx, gs_bx, r12 | |
| 411 | vadd.u32 uvrg_dx2, uvrg_dx1, uvrg_dx1 | |
| 412 | ||
| 413 | veor.u32 uvrg_dx0, uvrg_dx0, uvrg_dx0 | |
| 414 | eor gs_by, gs_by, r12 | |
| 415 | ||
| 416 | rsb r11, r14, #0 @ r11 = negative shift for scalar lsr | |
| 417 | add store_a, psx_gpu, #psx_gpu_uvrg_offset | |
| 418 | ||
| 419 | sub r11, r11, #(32 - 13) | |
| 420 | ||
| 421 | add store_b, store_a, #16 | |
| 422 | mov store_inc, #32 | |
| 423 | ||
| 424 | vadd.u32 uvrg_dx3, uvrg_dx2, uvrg_dx1 | |
| e1f6de8f | 425 | vst1.u32 { uvrg_base }, [store_a, :128], store_inc |
| 75e28f62 | 426 | |
| e1f6de8f | 427 | vst1.u32 { uvrg_dx1 }, [store_b, :128], store_inc |
| 75e28f62 E |
428 | mov g_bx, gw_bx_h, lsr r11 |
| 429 | ||
| e1f6de8f | 430 | vst1.u32 { g_uvrg_y }, [store_a, :128], store_inc |
| 75e28f62 E |
431 | mov g_by, gw_by_h, lsr r11 |
| 432 | ||
| 433 | vst4.u32 { uvrg_dx0l, uvrg_dx1l, uvrg_dx2l, uvrg_dx3l }, \ | |
| e1f6de8f | 434 | [store_b, :128], store_inc |
| 75e28f62 E |
435 | eor g_bx, g_bx, gs_bx |
| 436 | ||
| 437 | vst4.u32 { uvrg_dx0h, uvrg_dx1h, uvrg_dx2h, uvrg_dx3h }, \ | |
| e1f6de8f | 438 | [store_b, :128], store_inc |
| 75e28f62 E |
439 | sub g_bx, g_bx, gs_bx |
| 440 | ||
| 441 | lsl g_bx, g_bx, #4 | |
| 442 | eor g_by, g_by, gs_by | |
| 443 | ||
| 444 | mls b_base, g_bx, x0, b_base | |
| 445 | sub g_by, g_by, gs_by | |
| 446 | ||
| 447 | lsl g_by, g_by, #4 | |
| 448 | mov g_bx0, #0 | |
| 449 | ||
| 450 | add g_bx2, g_bx, g_bx | |
| 451 | add g_bx3, g_bx, g_bx2 | |
| 452 | ||
| 453 | stmia store_b, { g_bx0, g_bx, g_bx2, g_bx3, b_base, g_by } | |
| 454 | ||
| 3f0189c6 | 455 | restore_abi_regs() |
| 75e28f62 E |
456 | ldmia sp!, { r4 - r11, pc } |
| 457 | ||
| 458 | ||
| 459 | #define psx_gpu r0 | |
| 460 | #define v_a r1 | |
| 461 | #define v_b r2 | |
| 462 | #define v_c r3 | |
| 463 | ||
| 464 | #define temp r14 | |
| 465 | ||
| 466 | #define x_a r4 | |
| 467 | #define x_b r5 | |
| 468 | #define x_c r6 | |
| 469 | #define y_a r1 | |
| 470 | #define y_b r2 | |
| 471 | #define y_c r3 | |
| 472 | ||
| 473 | #define height_minor_a r7 | |
| 474 | #define height_minor_b r8 | |
| 475 | #define height_major r9 | |
| 476 | #define height r9 | |
| 477 | ||
| 478 | #define reciprocal_table_ptr r10 | |
| 479 | ||
| 480 | #define edge_alt_low r4 | |
| 481 | #define edge_alt_high r5 | |
| 482 | #define edge_dx_dy_alt r6 | |
| 483 | #define edge_shift_alt r10 | |
| 484 | ||
| 485 | #define edge_dx_dy_alt_low r4 | |
| 486 | #define edge_dx_dy_alt_high r5 | |
| 487 | ||
| 488 | #define span_edge_data r4 | |
| 489 | #define span_uvrg_offset r5 | |
| 490 | #define span_b_offset r6 | |
| 491 | ||
| 492 | #define clip r14 | |
| 493 | ||
| 494 | #define b r11 | |
| 495 | #define b_dy r12 | |
| 496 | ||
| 497 | ||
| 498 | #define alternate_x q0 | |
| 499 | #define alternate_dx_dy q1 | |
| 500 | #define alternate_x_32 q2 | |
| 501 | ||
| 502 | #define alternate_x_low d0 | |
| 503 | #define alternate_x_high d1 | |
| 504 | #define alternate_dx_dy_low d2 | |
| 505 | #define alternate_dx_dy_high d3 | |
| 506 | #define alternate_x_32_low d4 | |
| 507 | #define alternate_x_32_high d5 | |
| 508 | ||
| 509 | #define left_x q3 | |
| 510 | #define right_x q4 | |
| 511 | #define left_dx_dy q5 | |
| 512 | #define right_dx_dy q6 | |
| 513 | #define left_edge q7 | |
| 514 | #define right_edge q8 | |
| 515 | ||
| 516 | #define left_x_low d6 | |
| 517 | #define left_x_high d7 | |
| 518 | #define right_x_low d8 | |
| 519 | #define right_x_high d9 | |
| 520 | #define left_dx_dy_low d10 | |
| 521 | #define left_dx_dy_high d11 | |
| 522 | #define right_dx_dy_low d12 | |
| 523 | #define right_dx_dy_high d13 | |
| 524 | #define left_edge_low d14 | |
| 525 | #define left_edge_high d15 | |
| 526 | #define right_edge_low d16 | |
| 527 | #define right_edge_high d17 | |
| 528 | ||
| 529 | #define y_mid_point d18 | |
| 530 | #define c_0x0004 d19 | |
| 531 | ||
| 532 | #define left_right_x_16 q11 | |
| 533 | #define span_shifts_y q12 | |
| 534 | #define c_0x0001 q13 | |
| 535 | ||
| 536 | #define span_shifts d24 | |
| 537 | #define y_x4 d25 | |
| 538 | #define c_0xFFFE d26 | |
| 539 | #define c_0x0007 d27 | |
| 540 | ||
| 541 | #define left_right_x_16_low d22 | |
| 542 | #define left_right_x_16_high d23 | |
| 543 | ||
| 544 | #define uvrg q14 | |
| 545 | #define uvrg_dy q15 | |
| 546 | ||
| 547 | #define alternate_x_16 d4 | |
| 548 | ||
| 549 | #define v_clip q3 | |
| 550 | #define v_clip_low d6 | |
| 551 | ||
| 552 | #define right_x_32 q10 | |
| 553 | #define left_x_32 q11 | |
| 554 | #define alternate_select d24 | |
| 555 | ||
| 556 | #define right_x_32_low d20 | |
| 557 | #define right_x_32_high d21 | |
| 558 | #define left_x_32_low d22 | |
| 559 | #define left_x_32_high d23 | |
| 560 | ||
| 2d658c89 | 561 | #define tmp_max_blocks d20 |
| 562 | ||
| 75e28f62 E |
563 | #define edges_xy q0 |
| 564 | #define edges_dx_dy d2 | |
| 565 | #define edge_shifts d3 | |
| 566 | #define edge_shifts_64 q2 | |
| 567 | ||
| 568 | #define edges_xy_left d0 | |
| 569 | #define edges_xy_right d1 | |
| 570 | ||
| 571 | #define height_reciprocals d6 | |
| 572 | #define heights d7 | |
| 573 | ||
| 574 | #define widths d8 | |
| 575 | #define c_0x01 d9 | |
| 576 | #define x_starts d10 | |
| 577 | #define x_ends d11 | |
| 578 | ||
| 579 | #define heights_b d12 | |
| 580 | #define edges_dx_dy_64 q10 | |
| 581 | ||
| 582 | #define edges_dx_dy_64_left d20 | |
| 583 | #define edges_dx_dy_64_right d21 | |
| 584 | ||
| 585 | ||
| 586 | #define setup_spans_prologue() \ | |
| 587 | stmdb sp!, { r4 - r11, lr }; \ | |
| 3f0189c6 | 588 | save_abi_regs(); \ |
| 75e28f62 | 589 | \ |
| e1f6de8f | 590 | ldrsh x_a, [v_a, #8]; \ |
| 591 | ldrsh x_b, [v_b, #8]; \ | |
| 592 | ldrsh x_c, [v_c, #8]; \ | |
| 593 | ldrsh y_a, [v_a, #10]; \ | |
| 594 | ldrsh y_b, [v_b, #10]; \ | |
| 595 | ldrsh y_c, [v_c, #10]; \ | |
| 75e28f62 E |
596 | \ |
| 597 | add temp, psx_gpu, #psx_gpu_uvrg_offset; \ | |
| e1f6de8f | 598 | vld1.32 { uvrg }, [temp]; \ |
| 75e28f62 | 599 | add temp, psx_gpu, #psx_gpu_uvrg_dy_offset; \ |
| e1f6de8f | 600 | vld1.32 { uvrg_dy }, [temp]; \ |
| 601 | ldr reciprocal_table_ptr, [psx_gpu, #psx_gpu_reciprocal_table_ptr_offset]; \ | |
| 75e28f62 E |
602 | \ |
| 603 | vmov.u32 c_0x01, #0x01 \ | |
| 604 | ||
| 605 | #define setup_spans_load_b() \ | |
| e1f6de8f | 606 | ldr b, [psx_gpu, #psx_gpu_b_offset]; \ |
| 607 | ldr b_dy, [psx_gpu, #psx_gpu_b_dy_offset] \ | |
| 75e28f62 E |
608 | |
| 609 | #define setup_spans_prologue_b() \ | |
| 610 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 611 | add temp, psx_gpu, #psx_gpu_viewport_start_x_offset; \ | |
| 612 | \ | |
| 613 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 614 | vmov.u16 c_0x0004, #0x0004; \ | |
| 615 | \ | |
| 616 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 617 | vmov.u16 c_0x0001, #0x0001; \ | |
| 618 | \ | |
| e1f6de8f | 619 | vld1.u16 { left_edge_low[], left_edge_high[] }, [temp]; \ |
| 75e28f62 E |
620 | add temp, psx_gpu, #psx_gpu_viewport_end_x_offset; \ |
| 621 | \ | |
| e1f6de8f | 622 | vld1.u16 { right_edge_low[], right_edge_high[] }, [temp]; \ |
| 75e28f62 E |
623 | vadd.u16 right_edge, right_edge, c_0x0001; \ |
| 624 | \ | |
| 625 | vmov.u16 c_0x0007, #0x0007; \ | |
| 626 | vmvn.u16 c_0xFFFE, #0x0001 \ | |
| 627 | ||
| 628 | ||
| 629 | #define compute_edge_delta_x2() \ | |
| e1f6de8f | 630 | ldr temp, [reciprocal_table_ptr, height, lsl #2]; \ |
| 75e28f62 E |
631 | \ |
| 632 | vdup.u32 heights, height; \ | |
| 633 | vsub.u32 widths, x_ends, x_starts; \ | |
| 634 | \ | |
| 635 | vdup.u32 edge_shifts, temp; \ | |
| 636 | vsub.u32 heights_b, heights, c_0x01; \ | |
| 7d5140f5 | 637 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
| 75e28f62 E |
638 | \ |
| 639 | vmla.s32 heights_b, x_starts, heights; \ | |
| 640 | vbic.u16 edge_shifts, #0xE0; \ | |
| 641 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
| 642 | vmull.s32 edges_xy, heights_b, height_reciprocals \ | |
| 643 | ||
| 644 | #define width_alt r6 | |
| 645 | #define height_reciprocal_alt r11 | |
| 646 | #define height_b_alt r12 | |
| 647 | ||
| 648 | #define compute_edge_delta_x3(start_c, height_a, height_b) \ | |
| ed0fd81d | 649 | vmov heights, height_a, height_b; \ |
| e1f6de8f | 650 | ldr temp, [reciprocal_table_ptr, height_a, lsl #2]; \ |
| 75e28f62 | 651 | vmov.u32 edge_shifts[0], temp; \ |
| e1f6de8f | 652 | ldr temp, [reciprocal_table_ptr, height_b, lsl #2]; \ |
| 75e28f62 | 653 | vmov.u32 edge_shifts[1], temp; \ |
| e1f6de8f | 654 | ldr edge_shift_alt, [reciprocal_table_ptr, height_minor_b, lsl #2]; \ |
| 75e28f62 E |
655 | \ |
| 656 | vsub.u32 widths, x_ends, x_starts; \ | |
| 657 | sub width_alt, x_c, start_c; \ | |
| 658 | \ | |
| 659 | vsub.u32 heights_b, heights, c_0x01; \ | |
| 660 | sub height_b_alt, height_minor_b, #1; \ | |
| 661 | \ | |
| 7d5140f5 E |
662 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
| 663 | lsr height_reciprocal_alt, edge_shift_alt, #10; \ | |
| 75e28f62 E |
664 | \ |
| 665 | vmla.s32 heights_b, x_starts, heights; \ | |
| 666 | mla height_b_alt, height_minor_b, start_c, height_b_alt; \ | |
| 667 | \ | |
| 668 | vbic.u16 edge_shifts, #0xE0; \ | |
| 669 | and edge_shift_alt, edge_shift_alt, #0x1F; \ | |
| 670 | \ | |
| 671 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
| 672 | mul edge_dx_dy_alt, width_alt, height_reciprocal_alt; \ | |
| 673 | \ | |
| 674 | vmull.s32 edges_xy, heights_b, height_reciprocals; \ | |
| 675 | smull edge_alt_low, edge_alt_high, height_b_alt, height_reciprocal_alt \ | |
| 676 | ||
| 677 | ||
| 678 | #define setup_spans_adjust_y_up() \ | |
| 679 | vsub.u32 y_x4, y_x4, c_0x0004 \ | |
| 680 | ||
| 681 | #define setup_spans_adjust_y_down() \ | |
| 682 | vadd.u32 y_x4, y_x4, c_0x0004 \ | |
| 683 | ||
| 684 | #define setup_spans_adjust_interpolants_up() \ | |
| 685 | vsub.u32 uvrg, uvrg, uvrg_dy; \ | |
| 686 | sub b, b, b_dy \ | |
| 687 | ||
| 688 | #define setup_spans_adjust_interpolants_down() \ | |
| 689 | vadd.u32 uvrg, uvrg, uvrg_dy; \ | |
| 690 | add b, b, b_dy \ | |
| 691 | ||
| 692 | ||
| 693 | #define setup_spans_clip_interpolants_increment() \ | |
| 694 | mla b, b_dy, clip, b; \ | |
| 695 | vmla.s32 uvrg, uvrg_dy, v_clip \ | |
| 696 | ||
| 697 | #define setup_spans_clip_interpolants_decrement() \ | |
| 698 | mls b, b_dy, clip, b; \ | |
| 699 | vmls.s32 uvrg, uvrg_dy, v_clip \ | |
| 700 | ||
| 701 | #define setup_spans_clip_alternate_yes() \ | |
| 702 | smlal edge_alt_low, edge_alt_high, edge_dx_dy_alt, clip \ | |
| 703 | ||
| 704 | #define setup_spans_clip_alternate_no() \ | |
| 705 | ||
| 706 | #define setup_spans_clip(direction, alternate_active) \ | |
| 707 | vdup.u32 v_clip, clip; \ | |
| 708 | setup_spans_clip_alternate_##alternate_active(); \ | |
| 709 | setup_spans_clip_interpolants_##direction(); \ | |
| 710 | vmlal.s32 edges_xy, edges_dx_dy, v_clip_low \ | |
| 711 | ||
| 712 | ||
| 713 | #define setup_spans_adjust_edges_alternate_no(left_index, right_index) \ | |
| 714 | vmovl.s32 edge_shifts_64, edge_shifts; \ | |
| 715 | vmovl.s32 edges_dx_dy_64, edges_dx_dy; \ | |
| 716 | \ | |
| 717 | vshl.s64 edges_xy, edges_xy, edge_shifts_64; \ | |
| 718 | vshl.s64 edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64; \ | |
| 719 | \ | |
| 720 | vmov left_x_low, edges_xy_##left_index; \ | |
| 721 | vmov right_x_low, edges_xy_##right_index; \ | |
| 722 | \ | |
| 723 | vmov left_dx_dy_low, edges_dx_dy_64_##left_index; \ | |
| 724 | vmov left_dx_dy_high, edges_dx_dy_64_##left_index; \ | |
| 725 | vmov right_dx_dy_low, edges_dx_dy_64_##right_index; \ | |
| 726 | vmov right_dx_dy_high, edges_dx_dy_64_##right_index; \ | |
| 727 | \ | |
| 728 | vadd.u64 left_x_high, left_x_low, left_dx_dy_low; \ | |
| 729 | vadd.u64 right_x_high, right_x_low, right_dx_dy_low; \ | |
| 730 | \ | |
| 731 | vadd.u64 left_dx_dy, left_dx_dy, left_dx_dy; \ | |
| 732 | vadd.u64 right_dx_dy, right_dx_dy, right_dx_dy \ | |
| 733 | ||
| 734 | ||
| 735 | #define setup_spans_adjust_edges_alternate_yes(left_index, right_index) \ | |
| 736 | setup_spans_adjust_edges_alternate_no(left_index, right_index); \ | |
| 737 | \ | |
| 738 | vdup.u16 y_mid_point, y_b; \ | |
| 739 | rsb temp, edge_shift_alt, #32; \ | |
| 740 | \ | |
| 741 | lsl edge_alt_high, edge_alt_high, edge_shift_alt; \ | |
| 742 | orr edge_alt_high, edge_alt_high, edge_alt_low, lsr temp; \ | |
| 743 | lsl edge_alt_low, edge_alt_low, edge_shift_alt; \ | |
| 744 | vmov alternate_x_low, edge_alt_low, edge_alt_high; \ | |
| 745 | \ | |
| 746 | asr edge_dx_dy_alt_high, edge_dx_dy_alt, temp; \ | |
| 747 | lsl edge_dx_dy_alt_low, edge_dx_dy_alt, edge_shift_alt; \ | |
| 748 | vmov alternate_dx_dy_low, edge_dx_dy_alt_low, edge_dx_dy_alt_high; \ | |
| 749 | vmov alternate_dx_dy_high, alternate_dx_dy_low; \ | |
| 750 | \ | |
| 751 | vadd.u64 alternate_x_high, alternate_x_low, alternate_dx_dy_low; \ | |
| 752 | vadd.u64 alternate_dx_dy, alternate_dx_dy, alternate_dx_dy \ | |
| 753 | ||
| 754 | ||
| 755 | #define setup_spans_y_select_up() \ | |
| 756 | vclt.s16 alternate_select, y_x4, y_mid_point \ | |
| 757 | ||
| 758 | #define setup_spans_y_select_down() \ | |
| 759 | vcgt.s16 alternate_select, y_x4, y_mid_point \ | |
| 760 | ||
| 761 | ||
| 762 | #define setup_spans_alternate_select_left() \ | |
| 763 | vbit.u16 left_right_x_16_low, alternate_x_16, alternate_select \ | |
| 764 | ||
| 765 | #define setup_spans_alternate_select_right() \ | |
| 766 | vbit.u16 left_right_x_16_high, alternate_x_16, alternate_select \ | |
| 767 | ||
| 768 | ||
| 769 | #define setup_spans_set_x4_alternate_yes(alternate, direction) \ | |
| 770 | vshrn.s64 alternate_x_32_low, alternate_x, #32; \ | |
| 771 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
| 772 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
| 773 | \ | |
| 774 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
| 775 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 776 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 777 | \ | |
| 778 | vshrn.s64 alternate_x_32_high, alternate_x, #32; \ | |
| 779 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
| 780 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
| 781 | \ | |
| 782 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
| 783 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 784 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 785 | \ | |
| 786 | vmovn.u32 alternate_x_16, alternate_x_32; \ | |
| 787 | setup_spans_y_select_##direction(); \ | |
| 788 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
| 789 | \ | |
| 790 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
| 791 | setup_spans_alternate_select_##alternate(); \ | |
| 792 | \ | |
| e1f6de8f | 793 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 794 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
795 | setup_spans_adjust_interpolants_##direction(); \ |
| 796 | \ | |
| 797 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
| 798 | \ | |
| e1f6de8f | 799 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 800 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
801 | setup_spans_adjust_interpolants_##direction(); \ |
| 802 | \ | |
| 803 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
| 804 | \ | |
| e1f6de8f | 805 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 806 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
807 | setup_spans_adjust_interpolants_##direction(); \ |
| 808 | \ | |
| 809 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
| 810 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
| 811 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
| 812 | \ | |
| e1f6de8f | 813 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 814 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
815 | setup_spans_adjust_interpolants_##direction(); \ |
| 816 | \ | |
| 2d658c89 | 817 | vmov.u16 tmp_max_blocks, #MAX_BLOCKS_PER_ROW; \ |
| 75e28f62 E |
818 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ |
| 819 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
| 2d658c89 | 820 | vmin.u16 left_right_x_16_high, left_right_x_16_high, tmp_max_blocks; \ |
| 75e28f62 | 821 | \ |
| e1f6de8f | 822 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
| 75e28f62 E |
823 | \ |
| 824 | setup_spans_adjust_y_##direction() \ | |
| 825 | ||
| 826 | ||
| 827 | #define setup_spans_set_x4_alternate_no(alternate, direction) \ | |
| 828 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
| 829 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
| 830 | \ | |
| 831 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 832 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 833 | \ | |
| 834 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
| 835 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
| 836 | \ | |
| 837 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
| 838 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
| 839 | \ | |
| 840 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
| 841 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
| 842 | \ | |
| e1f6de8f | 843 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 844 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
845 | setup_spans_adjust_interpolants_##direction(); \ |
| 846 | \ | |
| 847 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
| 848 | \ | |
| e1f6de8f | 849 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 850 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
851 | setup_spans_adjust_interpolants_##direction(); \ |
| 852 | \ | |
| 853 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
| 854 | \ | |
| e1f6de8f | 855 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 856 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
857 | setup_spans_adjust_interpolants_##direction(); \ |
| 858 | \ | |
| 859 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
| 860 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
| 861 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
| 862 | \ | |
| e1f6de8f | 863 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
| 864 | str b, [span_b_offset], #4; \ | |
| 75e28f62 E |
865 | setup_spans_adjust_interpolants_##direction(); \ |
| 866 | \ | |
| 2d658c89 | 867 | vmov.u16 tmp_max_blocks, #MAX_BLOCKS_PER_ROW; \ |
| 75e28f62 | 868 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ |
| 2d658c89 | 869 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ |
| 870 | vmin.u16 left_right_x_16_high, left_right_x_16_high, tmp_max_blocks; \ | |
| 75e28f62 | 871 | \ |
| e1f6de8f | 872 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
| 75e28f62 E |
873 | \ |
| 874 | setup_spans_adjust_y_##direction() \ | |
| 875 | ||
| 876 | ||
| 877 | #define edge_adjust_low r11 | |
| 878 | #define edge_adjust_high r12 | |
| 879 | ||
| 880 | #define setup_spans_alternate_adjust_yes() \ | |
| 881 | smull edge_adjust_low, edge_adjust_high, edge_dx_dy_alt, height_minor_a; \ | |
| 882 | subs edge_alt_low, edge_alt_low, edge_adjust_low; \ | |
| 883 | sbc edge_alt_high, edge_alt_high, edge_adjust_high \ | |
| 884 | ||
| 885 | #define setup_spans_alternate_adjust_no() \ | |
| 886 | ||
| 887 | ||
| 888 | #define setup_spans_down(left_index, right_index, alternate, alternate_active) \ | |
| 889 | setup_spans_alternate_adjust_##alternate_active(); \ | |
| 890 | setup_spans_load_b(); \ | |
| 891 | \ | |
| e1f6de8f | 892 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
| 75e28f62 E |
893 | subs y_c, y_c, temp; \ |
| 894 | subgt height, height, y_c; \ | |
| 895 | addgt height, height, #1; \ | |
| 896 | \ | |
| e1f6de8f | 897 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
| 75e28f62 E |
898 | subs clip, temp, y_a; \ |
| 899 | ble 0f; \ | |
| 900 | \ | |
| 901 | sub height, height, clip; \ | |
| 902 | add y_a, y_a, clip; \ | |
| 903 | setup_spans_clip(increment, alternate_active); \ | |
| 904 | \ | |
| 905 | 0: \ | |
| 906 | cmp height, #0; \ | |
| 907 | ble 1f; \ | |
| 908 | \ | |
| 909 | orr temp, y_a, y_a, lsl #16; \ | |
| 2d658c89 | 910 | cmp height, #512; \ |
| 75e28f62 | 911 | add temp, temp, #(1 << 16); \ |
| 2d658c89 | 912 | movgt height, #512; \ |
| 75e28f62 E |
913 | add y_a, temp, #2; \ |
| 914 | add y_a, y_a, #(2 << 16); \ | |
| ed0fd81d | 915 | vmov y_x4, temp, y_a; \ |
| 75e28f62 E |
916 | \ |
| 917 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
| 918 | right_index); \ | |
| 919 | setup_spans_prologue_b(); \ | |
| 920 | \ | |
| e1f6de8f | 921 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
922 | \ |
| 923 | 2: \ | |
| 924 | setup_spans_set_x4_alternate_##alternate_active(alternate, down); \ | |
| 925 | subs height, height, #4; \ | |
| 926 | bhi 2b; \ | |
| 927 | \ | |
| 928 | 1: \ | |
| 929 | ||
| 930 | ||
| 931 | #define setup_spans_alternate_pre_increment_yes() \ | |
| 932 | adds edge_alt_low, edge_alt_low, edge_dx_dy_alt; \ | |
| 933 | adc edge_alt_high, edge_alt_high, edge_dx_dy_alt, asr #31 \ | |
| 934 | ||
| 935 | #define setup_spans_alternate_pre_increment_no() \ | |
| 936 | ||
| 937 | ||
| 938 | #define setup_spans_up_decrement_yes() \ | |
| 939 | suble height, height, #1 \ | |
| 940 | ||
| 941 | #define setup_spans_up_decrement_no() \ | |
| 942 | ||
| 943 | ||
| 944 | #define setup_spans_up(left_index, right_index, alternate, alternate_active) \ | |
| 945 | setup_spans_alternate_adjust_##alternate_active(); \ | |
| 946 | setup_spans_load_b(); \ | |
| 947 | sub y_a, y_a, #1; \ | |
| 948 | \ | |
| e1f6de8f | 949 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
| 75e28f62 E |
950 | subs temp, temp, y_c; \ |
| 951 | subgt height, height, temp; \ | |
| 952 | setup_spans_up_decrement_##alternate_active(); \ | |
| 953 | \ | |
| e1f6de8f | 954 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
| 75e28f62 E |
955 | subs clip, y_a, temp; \ |
| 956 | ble 0f; \ | |
| 957 | \ | |
| 958 | sub height, height, clip; \ | |
| 959 | sub y_a, y_a, clip; \ | |
| 960 | setup_spans_clip(decrement, alternate_active); \ | |
| 961 | \ | |
| 962 | 0: \ | |
| 963 | cmp height, #0; \ | |
| 964 | ble 1f; \ | |
| 965 | \ | |
| 966 | orr temp, y_a, y_a, lsl #16; \ | |
| 2d658c89 | 967 | cmp height, #512; \ |
| 75e28f62 | 968 | sub temp, temp, #(1 << 16); \ |
| 2d658c89 | 969 | movgt height, #512; \ |
| 75e28f62 E |
970 | sub y_a, temp, #2; \ |
| 971 | sub y_a, y_a, #(2 << 16); \ | |
| ed0fd81d | 972 | vmov y_x4, temp, y_a; \ |
| 75e28f62 E |
973 | \ |
| 974 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy; \ | |
| 975 | \ | |
| 976 | setup_spans_alternate_pre_increment_##alternate_active(); \ | |
| 977 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
| 978 | right_index); \ | |
| 979 | setup_spans_adjust_interpolants_up(); \ | |
| 980 | setup_spans_prologue_b(); \ | |
| 981 | \ | |
| e1f6de8f | 982 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
983 | \ |
| 984 | 2: \ | |
| 985 | setup_spans_set_x4_alternate_##alternate_active(alternate, up); \ | |
| 986 | subs height, height, #4; \ | |
| 987 | bhi 2b; \ | |
| 988 | \ | |
| 989 | 1: \ | |
| 990 | ||
| 991 | ||
| 992 | #define setup_spans_epilogue() \ | |
| 3f0189c6 | 993 | restore_abi_regs(); \ |
| 75e28f62 E |
994 | ldmia sp!, { r4 - r11, pc } \ |
| 995 | ||
| 996 | ||
| 997 | #define setup_spans_up_up(minor, major) \ | |
| 998 | setup_spans_prologue(); \ | |
| 999 | sub height_minor_a, y_a, y_b; \ | |
| 1000 | sub height_minor_b, y_b, y_c; \ | |
| 1001 | sub height, y_a, y_c; \ | |
| 1002 | \ | |
| 1003 | vdup.u32 x_starts, x_a; \ | |
| ed0fd81d | 1004 | vmov x_ends, x_c, x_b; \ |
| 75e28f62 E |
1005 | \ |
| 1006 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
| 1007 | setup_spans_up(major, minor, minor, yes); \ | |
| 1008 | setup_spans_epilogue() \ | |
| 1009 | ||
| 1010 | function(setup_spans_up_left) | |
| 1011 | setup_spans_up_up(left, right) | |
| 1012 | ||
| 1013 | function(setup_spans_up_right) | |
| 1014 | setup_spans_up_up(right, left) | |
| 1015 | ||
| 75e28f62 E |
1016 | #define setup_spans_down_down(minor, major) \ |
| 1017 | setup_spans_prologue(); \ | |
| 1018 | sub height_minor_a, y_b, y_a; \ | |
| 1019 | sub height_minor_b, y_c, y_b; \ | |
| 1020 | sub height, y_c, y_a; \ | |
| 1021 | \ | |
| 1022 | vdup.u32 x_starts, x_a; \ | |
| ed0fd81d | 1023 | vmov x_ends, x_c, x_b; \ |
| 75e28f62 E |
1024 | \ |
| 1025 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
| 1026 | setup_spans_down(major, minor, minor, yes); \ | |
| 1027 | setup_spans_epilogue() \ | |
| 1028 | ||
| 1029 | function(setup_spans_down_left) | |
| 1030 | setup_spans_down_down(left, right) | |
| 1031 | ||
| 1032 | function(setup_spans_down_right) | |
| 1033 | setup_spans_down_down(right, left) | |
| 1034 | ||
| 1035 | ||
| 1036 | #define setup_spans_up_flat() \ | |
| 1037 | sub height, y_a, y_c; \ | |
| 1038 | \ | |
| 1039 | compute_edge_delta_x2(); \ | |
| 1040 | setup_spans_up(left, right, none, no); \ | |
| 1041 | setup_spans_epilogue() \ | |
| 1042 | ||
| 1043 | function(setup_spans_up_a) | |
| 1044 | setup_spans_prologue() | |
| 1045 | ||
| ed0fd81d | 1046 | vmov x_starts, x_a, x_b |
| 75e28f62 E |
1047 | vdup.u32 x_ends, x_c |
| 1048 | ||
| 1049 | setup_spans_up_flat() | |
| 1050 | ||
| 1051 | function(setup_spans_up_b) | |
| 1052 | setup_spans_prologue() | |
| 1053 | ||
| 1054 | vdup.u32 x_starts, x_a | |
| ed0fd81d | 1055 | vmov x_ends, x_b, x_c |
| 75e28f62 E |
1056 | |
| 1057 | setup_spans_up_flat() | |
| 1058 | ||
| 1059 | #define setup_spans_down_flat() \ | |
| 1060 | sub height, y_c, y_a; \ | |
| 1061 | \ | |
| 1062 | compute_edge_delta_x2(); \ | |
| 1063 | setup_spans_down(left, right, none, no); \ | |
| 1064 | setup_spans_epilogue() \ | |
| 1065 | ||
| 1066 | function(setup_spans_down_a) | |
| 1067 | setup_spans_prologue() | |
| 1068 | ||
| ed0fd81d | 1069 | vmov x_starts, x_a, x_b |
| 75e28f62 E |
1070 | vdup.u32 x_ends, x_c |
| 1071 | ||
| 1072 | setup_spans_down_flat() | |
| 1073 | ||
| 1074 | function(setup_spans_down_b) | |
| 1075 | setup_spans_prologue() | |
| 1076 | ||
| 1077 | vdup.u32 x_starts, x_a | |
| ed0fd81d | 1078 | vmov x_ends, x_b, x_c |
| 75e28f62 E |
1079 | |
| 1080 | setup_spans_down_flat() | |
| 1081 | ||
| 1082 | ||
| 1083 | #define middle_y r9 | |
| 1084 | ||
| 1085 | #define edges_xy_b q11 | |
| 1086 | #define edges_dx_dy_b d26 | |
| 1087 | #define edge_shifts_b d27 | |
| 1088 | #define edges_dx_dy_and_shifts_b q13 | |
| 1089 | #define height_increment d20 | |
| 1090 | ||
| 1091 | #define edges_dx_dy_and_shifts q1 | |
| 1092 | ||
| 1093 | #define edges_xy_b_left d22 | |
| 1094 | #define edges_xy_b_right d23 | |
| 1095 | ||
| 1096 | #define setup_spans_up_down_load_edge_set_b() \ | |
| 1097 | vmov edges_xy, edges_xy_b; \ | |
| 1098 | vmov edges_dx_dy_and_shifts, edges_dx_dy_and_shifts_b \ | |
| 1099 | ||
| 1100 | ||
| 1101 | function(setup_spans_up_down) | |
| 1102 | setup_spans_prologue() | |
| 1103 | ||
| 1104 | // s32 middle_y = y_a; | |
| 1105 | sub height_minor_a, y_a, y_b | |
| 1106 | sub height_minor_b, y_c, y_a | |
| 1107 | sub height_major, y_c, y_b | |
| 1108 | ||
| ed0fd81d | 1109 | vmov x_starts, x_a, x_c |
| 75e28f62 E |
1110 | vdup.u32 x_ends, x_b |
| 1111 | ||
| 1112 | compute_edge_delta_x3(x_a, height_minor_a, height_major) | |
| 1113 | ||
| 1114 | mov temp, #0 | |
| ed0fd81d | 1115 | vmov height_increment, temp, height_minor_b |
| 75e28f62 E |
1116 | vmlal.s32 edges_xy, edges_dx_dy, height_increment |
| 1117 | ||
| 1118 | vmov edges_xy_b_left, edge_alt_low, edge_alt_high | |
| 1119 | vmov edges_xy_b_right, edges_xy_right | |
| 1120 | ||
| 1121 | vmov edge_shifts_b, edge_shifts | |
| 1122 | vmov.u32 edge_shifts_b[0], edge_shift_alt | |
| 1123 | ||
| 1124 | vneg.s32 edges_dx_dy_b, edges_dx_dy | |
| 1125 | vmov.u32 edges_dx_dy_b[0], edge_dx_dy_alt | |
| 1126 | ||
| 1127 | mov middle_y, y_a | |
| 1128 | ||
| 1129 | setup_spans_load_b() | |
| 1130 | sub y_a, y_a, #1 | |
| 1131 | ||
| e1f6de8f | 1132 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
| 75e28f62 E |
1133 | subs temp, temp, y_b |
| 1134 | subgt height_minor_a, height_minor_a, temp | |
| 1135 | ||
| e1f6de8f | 1136 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
| 75e28f62 E |
1137 | subs clip, y_a, temp |
| 1138 | ble 0f | |
| 1139 | ||
| 1140 | sub height_minor_a, height_minor_a, clip | |
| 1141 | sub y_a, y_a, clip | |
| 1142 | setup_spans_clip(decrement, no) | |
| 1143 | ||
| 1144 | 0: | |
| 1145 | cmp height_minor_a, #0 | |
| 1146 | ble 3f | |
| 1147 | ||
| 1148 | orr temp, y_a, y_a, lsl #16 | |
| 1149 | sub temp, temp, #(1 << 16) | |
| 1150 | sub y_a, temp, #2 | |
| 1151 | sub y_a, y_a, #(2 << 16) | |
| ed0fd81d | 1152 | vmov y_x4, temp, y_a |
| 75e28f62 E |
1153 | |
| 1154 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy | |
| 1155 | ||
| e1f6de8f | 1156 | strh height_minor_a, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
1157 | |
| 1158 | setup_spans_adjust_edges_alternate_no(left, right); | |
| 1159 | setup_spans_adjust_interpolants_up() | |
| 1160 | setup_spans_up_down_load_edge_set_b() | |
| 1161 | ||
| 1162 | setup_spans_prologue_b() | |
| 1163 | ||
| 1164 | ||
| 1165 | 2: | |
| 1166 | setup_spans_set_x4_alternate_no(none, up) | |
| 1167 | subs height_minor_a, height_minor_a, #4 | |
| 1168 | bhi 2b | |
| 1169 | ||
| 1170 | add span_edge_data, span_edge_data, height_minor_a, lsl #3 | |
| 1171 | add span_uvrg_offset, span_uvrg_offset, height_minor_a, lsl #4 | |
| 1172 | add span_b_offset, span_b_offset, height_minor_a, lsl #2 | |
| 1173 | ||
| 1174 | 4: | |
| 1175 | add temp, psx_gpu, #psx_gpu_uvrg_offset | |
| e1f6de8f | 1176 | vld1.32 { uvrg }, [temp] |
| 75e28f62 E |
1177 | mov y_a, middle_y |
| 1178 | ||
| 1179 | setup_spans_load_b() | |
| 1180 | ||
| e1f6de8f | 1181 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
| 75e28f62 E |
1182 | subs y_c, y_c, temp |
| 1183 | subgt height_minor_b, height_minor_b, y_c | |
| 1184 | addgt height_minor_b, height_minor_b, #1 | |
| 1185 | ||
| e1f6de8f | 1186 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
| 75e28f62 E |
1187 | subs clip, temp, y_a |
| 1188 | ble 0f | |
| 1189 | ||
| 1190 | sub height_minor_b, height_minor_b, clip | |
| 1191 | add y_a, y_a, clip | |
| 1192 | setup_spans_clip(increment, no) | |
| 1193 | ||
| 1194 | 0: | |
| 1195 | cmp height_minor_b, #0 | |
| 1196 | ble 1f | |
| 1197 | ||
| 1198 | orr temp, y_a, y_a, lsl #16 | |
| 1199 | add temp, temp, #(1 << 16) | |
| 1200 | add y_a, temp, #2 | |
| 1201 | add y_a, y_a, #(2 << 16) | |
| ed0fd81d | 1202 | vmov y_x4, temp, y_a |
| 75e28f62 E |
1203 | |
| 1204 | setup_spans_adjust_edges_alternate_no(left, right) | |
| 1205 | ||
| e1f6de8f | 1206 | ldrh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 | 1207 | add temp, temp, height_minor_b |
| b7569147 | 1208 | |
| 1209 | cmp temp, #MAX_SPANS | |
| 1210 | beq 5f | |
| 1211 | ||
| e1f6de8f | 1212 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
1213 | |
| 1214 | 2: | |
| 1215 | setup_spans_set_x4_alternate_no(none, down) | |
| 1216 | subs height_minor_b, height_minor_b, #4 | |
| 1217 | bhi 2b | |
| 1218 | ||
| 1219 | 1: | |
| 1220 | setup_spans_epilogue() | |
| 1221 | ||
| 1222 | 3: | |
| 1223 | setup_spans_up_down_load_edge_set_b() | |
| 1224 | setup_spans_prologue_b() | |
| 1225 | bal 4b | |
| 1226 | ||
| b7569147 | 1227 | 5: |
| 1228 | // FIXME: overflow corner case | |
| 1229 | sub temp, temp, height_minor_b | |
| 1230 | bics height_minor_b, #3 | |
| 1231 | add temp, temp, height_minor_b | |
| e1f6de8f | 1232 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
| b7569147 | 1233 | bne 2b |
| 1234 | bal 1b | |
| 1235 | ||
| 75e28f62 E |
1236 | #undef span_uvrg_offset |
| 1237 | #undef span_edge_data | |
| 1238 | #undef span_b_offset | |
| 1239 | #undef left_x | |
| 1240 | #undef b | |
| 1241 | ||
| 1242 | #define psx_gpu r0 | |
| 1243 | #define num_spans r1 | |
| 1244 | #define span_uvrg_offset r2 | |
| 1245 | #define span_edge_data r3 | |
| 1246 | #define span_b_offset r4 | |
| 1247 | #define b_dx r5 | |
| 1248 | #define span_num_blocks r6 | |
| 1249 | #define y r7 | |
| 1250 | #define left_x r8 | |
| 1251 | #define b r9 | |
| 1252 | #define dither_offset_ptr r10 | |
| 1253 | #define block_ptr_a r11 | |
| 1254 | #define fb_ptr r12 | |
| 1255 | #define num_blocks r14 | |
| 1256 | ||
| 1257 | #define uvrg_dx_ptr r2 | |
| 1258 | #define texture_mask_ptr r3 | |
| 1259 | #define dither_shift r8 | |
| 1260 | #define dither_row r10 | |
| 1261 | ||
| 1262 | #define c_32 r7 | |
| 1263 | #define b_dx4 r8 | |
| 1264 | #define b_dx8 r9 | |
| 1265 | #define block_ptr_b r10 | |
| 1266 | ||
| 1267 | #define block_span_ptr r10 | |
| 1268 | #define right_mask r8 | |
| 1269 | ||
| 1270 | #define color r2 | |
| 1271 | #define color_r r3 | |
| 1272 | #define color_g r4 | |
| 1273 | #define color_b r5 | |
| 1274 | ||
| 1275 | #undef uvrg | |
| 1276 | ||
| 1277 | #define u_block q0 | |
| 1278 | #define v_block q1 | |
| 1279 | #define r_block q2 | |
| 1280 | #define g_block q3 | |
| 1281 | #define b_block q4 | |
| 1282 | ||
| 1283 | #define uv_dx4 d10 | |
| 1284 | #define rg_dx4 d11 | |
| 1285 | #define uv_dx8 d12 | |
| 1286 | #define rg_dx8 d13 | |
| 1287 | #define b_whole_8 d14 | |
| 1288 | #define fb_mask_ptrs d15 | |
| 1289 | ||
| 1290 | #define uvrg_dx4 q5 | |
| 1291 | #define uvrg_dx8 q6 | |
| 1292 | #define uv_dx8 d12 | |
| 1293 | #define rg_dx8 d13 | |
| 1294 | ||
| 1295 | #define u_whole q8 | |
| 1296 | #define v_whole q9 | |
| 1297 | #define r_whole q10 | |
| 1298 | #define g_whole q11 | |
| 1299 | #define b_whole q12 | |
| 1300 | ||
| 1301 | #define u_whole_low d16 | |
| 1302 | #define u_whole_high d17 | |
| 1303 | #define v_whole_low d18 | |
| 1304 | #define v_whole_high d19 | |
| 1305 | #define r_whole_low d20 | |
| 1306 | #define r_whole_high d21 | |
| 1307 | #define g_whole_low d22 | |
| 1308 | #define g_whole_high d23 | |
| 1309 | #define b_whole_low d24 | |
| 1310 | #define b_whole_high d25 | |
| 1311 | ||
| 1312 | #define dx4 q13 | |
| 1313 | #define dx8 q13 | |
| 1314 | ||
| 1315 | #define u_whole_8 d26 | |
| 1316 | #define v_whole_8 d27 | |
| 1317 | #define u_whole_8b d24 | |
| 1318 | #define r_whole_8 d24 | |
| 1319 | #define g_whole_8 d25 | |
| 1320 | ||
| 1321 | #define uv_whole_8 q13 | |
| 1322 | #define uv_whole_8b q14 | |
| 1323 | ||
| 1324 | #define dither_offsets q14 | |
| 1325 | #define texture_mask q15 | |
| 1326 | #define texture_mask_u d30 | |
| 1327 | #define texture_mask_v d31 | |
| 1328 | ||
| 1329 | #define dither_offsets_short d28 | |
| 1330 | ||
| 1331 | #define v_left_x q8 | |
| 1332 | #define uvrg q9 | |
| 1333 | #define block_span q10 | |
| 1334 | ||
| 1335 | #define uv d18 | |
| 1336 | #define rg d19 | |
| 1337 | ||
| 1338 | #define draw_mask q1 | |
| 1339 | #define draw_mask_edge q13 | |
| 1340 | #define test_mask q0 | |
| 1341 | ||
| 1342 | #define uvrg_dx q3 | |
| 1343 | ||
| 1344 | #define colors q2 | |
| 1345 | ||
| 1346 | #define setup_blocks_texture_swizzled() \ | |
| 1347 | vand.u8 u_whole_8b, u_whole_8, texture_mask_u; \ | |
| 1348 | vsli.u8 u_whole_8, v_whole_8, #4; \ | |
| 1349 | vsri.u8 v_whole_8, u_whole_8b, #4 \ | |
| 1350 | ||
| 1351 | #define setup_blocks_texture_unswizzled() \ | |
| 1352 | ||
| 1353 | ||
| 1354 | #define setup_blocks_shaded_textured_builder(swizzling) \ | |
| 1355 | .align 3; \ | |
| 1356 | \ | |
| 1357 | function(setup_blocks_shaded_textured_dithered_##swizzling##_indirect) \ | |
| e1f6de8f | 1358 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
1359 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
| 1360 | \ | |
| e1f6de8f | 1361 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
| 75e28f62 E |
1362 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
| 1363 | \ | |
| 1364 | cmp num_spans, #0; \ | |
| 1365 | bxeq lr; \ | |
| 1366 | \ | |
| 1367 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 1368 | save_abi_regs(); \ |
| 75e28f62 E |
1369 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
| 1370 | \ | |
| e1f6de8f | 1371 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
| 75e28f62 E |
1372 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ |
| 1373 | \ | |
| e1f6de8f | 1374 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
| 75e28f62 E |
1375 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
| 1376 | \ | |
| e1f6de8f | 1377 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1378 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
| 1379 | \ | |
| 1380 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 1381 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1382 | \ | |
| 1383 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 1384 | \ | |
| 1385 | 0: \ | |
| 1386 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1387 | \ | |
| e1f6de8f | 1388 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
1389 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 1390 | \ | |
| e1f6de8f | 1391 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 1392 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
1393 | \ |
| 1394 | cmp span_num_blocks, #0; \ | |
| 1395 | beq 1f; \ | |
| 1396 | \ | |
| e1f6de8f | 1397 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
1398 | add num_blocks, span_num_blocks, num_blocks; \ |
| 1399 | \ | |
| 1400 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 1401 | bgt 2f; \ | |
| 1402 | \ | |
| 1403 | 3: \ | |
| e1f6de8f | 1404 | ldr b, [span_b_offset]; \ |
| 75e28f62 E |
1405 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 1406 | \ | |
| 1407 | vdup.u32 v_left_x, left_x; \ | |
| 1408 | and y, y, #0x3; \ | |
| 1409 | \ | |
| e1f6de8f | 1410 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
1411 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 1412 | \ | |
| 1413 | mla b, b_dx, left_x, b; \ | |
| 1414 | and dither_shift, left_x, #0x03; \ | |
| 1415 | \ | |
| e1f6de8f | 1416 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
1417 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
| 1418 | \ | |
| 1419 | mov dither_shift, dither_shift, lsl #3; \ | |
| 1420 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
| 1421 | \ | |
| 1422 | mov c_32, #32; \ | |
| 1423 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1424 | \ | |
| 1425 | mov dither_row, dither_row, ror dither_shift; \ | |
| 1426 | mov b_dx4, b_dx, lsl #2; \ | |
| 1427 | \ | |
| 1428 | vdup.u32 dither_offsets_short, dither_row; \ | |
| 1429 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
| 1430 | \ | |
| 1431 | vdup.u32 b_block, b; \ | |
| 1432 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
| 1433 | \ | |
| 1434 | vdup.u32 u_block, uv[0]; \ | |
| 1435 | mov b_dx8, b_dx, lsl #3; \ | |
| 1436 | \ | |
| 1437 | vdup.u32 v_block, uv[1]; \ | |
| 1438 | vdup.u32 r_block, rg[0]; \ | |
| 1439 | vdup.u32 g_block, rg[1]; \ | |
| 1440 | \ | |
| e1f6de8f | 1441 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1442 | \ |
| 1443 | vadd.u32 u_block, u_block, block_span; \ | |
| e1f6de8f | 1444 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1445 | \ |
| 1446 | vadd.u32 v_block, v_block, block_span; \ | |
| e1f6de8f | 1447 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1448 | \ |
| 1449 | vadd.u32 r_block, r_block, block_span; \ | |
| e1f6de8f | 1450 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1451 | \ |
| 1452 | vadd.u32 g_block, g_block, block_span; \ | |
| e1f6de8f | 1453 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
| 75e28f62 E |
1454 | \ |
| 1455 | vadd.u32 b_block, b_block, block_span; \ | |
| 1456 | add block_ptr_b, block_ptr_a, #16; \ | |
| 1457 | \ | |
| 1458 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1459 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1460 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 1461 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 1462 | \ | |
| 1463 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1464 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 1465 | \ | |
| 1466 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1467 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1468 | \ | |
| 1469 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1470 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 1471 | \ | |
| 1472 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 1473 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 1474 | \ | |
| 1475 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 1476 | vdup.u32 dx4, b_dx4; \ | |
| 1477 | \ | |
| 1478 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 1479 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1480 | \ | |
| 1481 | vadd.u32 u_block, u_block, dx8; \ | |
| 1482 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1483 | \ | |
| 1484 | vadd.u32 v_block, v_block, dx8; \ | |
| 1485 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 1486 | \ | |
| 1487 | vadd.u32 r_block, r_block, dx8; \ | |
| 1488 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 1489 | \ | |
| 1490 | vadd.u32 g_block, g_block, dx8; \ | |
| 1491 | vdup.u32 dx8, b_dx8; \ | |
| 1492 | \ | |
| 1493 | vadd.u32 b_block, b_block, dx8; \ | |
| 1494 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1495 | \ | |
| 1496 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1497 | \ | |
| 1498 | vmovn.u16 b_whole_8, b_whole; \ | |
| e1f6de8f | 1499 | pld [fb_ptr]; \ |
| 75e28f62 E |
1500 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
| 1501 | \ | |
| 1502 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1503 | setup_blocks_texture_##swizzling(); \ | |
| 1504 | \ | |
| 1505 | vmovn.u16 r_whole_8, r_whole; \ | |
| 1506 | beq 5f; \ | |
| 1507 | \ | |
| 1508 | 4: \ | |
| 1509 | vmovn.u16 g_whole_8, g_whole; \ | |
| 1510 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1511 | \ | |
| e1f6de8f | 1512 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1513 | vshrn.u32 v_whole_low, v_block, #16; \ |
| 1514 | \ | |
| e1f6de8f | 1515 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
| 75e28f62 E |
1516 | vshrn.u32 r_whole_low, r_block, #16; \ |
| 1517 | \ | |
| e1f6de8f | 1518 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1519 | vshrn.u32 g_whole_low, g_block, #16; \ |
| 1520 | \ | |
| 1521 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1522 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 1523 | \ | |
| 1524 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1525 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1526 | \ | |
| 1527 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1528 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 1529 | \ | |
| 1530 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 1531 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 1532 | \ | |
| 1533 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 1534 | vdup.u32 dx4, b_dx4; \ | |
| 1535 | \ | |
| 1536 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 1537 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1538 | \ | |
| 1539 | vadd.u32 u_block, u_block, dx8; \ | |
| 1540 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1541 | \ | |
| 1542 | vadd.u32 v_block, v_block, dx8; \ | |
| 1543 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 1544 | \ | |
| 1545 | vadd.u32 r_block, r_block, dx8; \ | |
| 1546 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 1547 | \ | |
| 1548 | vadd.u32 g_block, g_block, dx8; \ | |
| 1549 | vdup.u32 dx8, b_dx8; \ | |
| 1550 | \ | |
| 1551 | vadd.u32 b_block, b_block, dx8; \ | |
| 1552 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1553 | \ | |
| 1554 | add fb_ptr, fb_ptr, #16; \ | |
| 1555 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1556 | \ | |
| e1f6de8f | 1557 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
| 75e28f62 E |
1558 | vmovn.u16 b_whole_8, b_whole; \ |
| 1559 | \ | |
| e1f6de8f | 1560 | pld [fb_ptr]; \ |
| 75e28f62 E |
1561 | \ |
| 1562 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1563 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1564 | \ | |
| 1565 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1566 | setup_blocks_texture_##swizzling(); \ | |
| 1567 | \ | |
| 1568 | vmovn.u16 r_whole_8, r_whole; \ | |
| 1569 | bne 4b; \ | |
| 1570 | \ | |
| 1571 | 5: \ | |
| 1572 | vmovn.u16 g_whole_8, g_whole; \ | |
| e1f6de8f | 1573 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 | 1574 | \ |
| e1f6de8f | 1575 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
1576 | vdup.u8 draw_mask, right_mask; \ |
| 1577 | \ | |
| 1578 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
| 1579 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 1580 | vzip.u8 u_whole_8, v_whole_8; \ | |
| 1581 | \ | |
| 1582 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
| e1f6de8f | 1583 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
| 1584 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ | |
| 1585 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
| 1586 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
| 75e28f62 E |
1587 | \ |
| 1588 | 1: \ | |
| 1589 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 1590 | add span_b_offset, span_b_offset, #4; \ | |
| 1591 | \ | |
| 1592 | add span_edge_data, span_edge_data, #8; \ | |
| 1593 | subs num_spans, num_spans, #1; \ | |
| 1594 | \ | |
| e1f6de8f | 1595 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1596 | bne 0b; \ |
| 1597 | \ | |
| 3f0189c6 | 1598 | restore_abi_regs(); \ |
| 75e28f62 E |
1599 | ldmia sp!, { r4 - r11, pc }; \ |
| 1600 | \ | |
| 1601 | 2: \ | |
| 1602 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
| 1603 | vpush { texture_mask }; \ | |
| 1604 | vpush { uvrg_dx4 }; \ | |
| 1605 | \ | |
| 26e3e2aa | 1606 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; /* r14=num_blocks */ \ |
| 75e28f62 | 1607 | bl flush_render_block_buffer; \ |
| 26e3e2aa | 1608 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ |
| 75e28f62 E |
1609 | \ |
| 1610 | vpop { uvrg_dx4 }; \ | |
| 1611 | vpop { texture_mask }; \ | |
| 1612 | \ | |
| 1613 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
| 1614 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1615 | \ | |
| 1616 | mov num_blocks, span_num_blocks; \ | |
| 1617 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1618 | bal 3b \ | |
| 1619 | ||
| 1620 | ||
| 1621 | setup_blocks_shaded_textured_builder(swizzled) | |
| 1622 | setup_blocks_shaded_textured_builder(unswizzled) | |
| 1623 | ||
| 1624 | ||
| 1625 | #define setup_blocks_unshaded_textured_builder(swizzling) \ | |
| 1626 | .align 3; \ | |
| 1627 | \ | |
| 1628 | function(setup_blocks_unshaded_textured_dithered_##swizzling##_indirect) \ | |
| e1f6de8f | 1629 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
1630 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
| 1631 | \ | |
| e1f6de8f | 1632 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
| 75e28f62 E |
1633 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
| 1634 | \ | |
| 1635 | cmp num_spans, #0; \ | |
| 1636 | bxeq lr; \ | |
| 1637 | \ | |
| 1638 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 1639 | save_abi_regs(); \ |
| 75e28f62 E |
1640 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
| 1641 | \ | |
| 1642 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
| 1643 | \ | |
| e1f6de8f | 1644 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
| 75e28f62 E |
1645 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
| 1646 | \ | |
| e1f6de8f | 1647 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1648 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
| 1649 | \ | |
| 1650 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1651 | \ | |
| 1652 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 1653 | \ | |
| 1654 | 0: \ | |
| 1655 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1656 | \ | |
| e1f6de8f | 1657 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
1658 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 1659 | \ | |
| e1f6de8f | 1660 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 1661 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
1662 | \ |
| 1663 | cmp span_num_blocks, #0; \ | |
| 1664 | beq 1f; \ | |
| 1665 | \ | |
| e1f6de8f | 1666 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
1667 | add num_blocks, span_num_blocks, num_blocks; \ |
| 1668 | \ | |
| 1669 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 1670 | bgt 2f; \ | |
| 1671 | \ | |
| 1672 | 3: \ | |
| 1673 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
| 1674 | \ | |
| 1675 | vdup.u32 v_left_x, left_x; \ | |
| 1676 | and y, y, #0x3; \ | |
| 1677 | \ | |
| e1f6de8f | 1678 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
1679 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 1680 | \ | |
| 1681 | and dither_shift, left_x, #0x03; \ | |
| 1682 | \ | |
| e1f6de8f | 1683 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
1684 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
| 1685 | \ | |
| 1686 | mov dither_shift, dither_shift, lsl #3; \ | |
| 1687 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
| 1688 | \ | |
| 1689 | mov c_32, #32; \ | |
| 1690 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1691 | \ | |
| 1692 | mov dither_row, dither_row, ror dither_shift; \ | |
| 1693 | \ | |
| 1694 | vdup.u32 dither_offsets_short, dither_row; \ | |
| 1695 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
| 1696 | \ | |
| 1697 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
| 1698 | \ | |
| 1699 | vdup.u32 u_block, uv[0]; \ | |
| 1700 | \ | |
| 1701 | vdup.u32 v_block, uv[1]; \ | |
| e1f6de8f | 1702 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1703 | \ |
| 1704 | vadd.u32 u_block, u_block, block_span; \ | |
| e1f6de8f | 1705 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
1706 | \ |
| 1707 | vadd.u32 v_block, v_block, block_span; \ | |
| 1708 | add block_ptr_b, block_ptr_a, #16; \ | |
| 1709 | \ | |
| 1710 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1711 | vshrn.u32 v_whole_low, v_block, #16; \ | |
| 1712 | \ | |
| 1713 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1714 | \ | |
| 1715 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1716 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1717 | \ | |
| 1718 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1719 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1720 | \ | |
| 1721 | vadd.u32 u_block, u_block, dx8; \ | |
| 1722 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1723 | \ | |
| 1724 | vadd.u32 v_block, v_block, dx8; \ | |
| 1725 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1726 | \ | |
| 1727 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1728 | \ | |
| e1f6de8f | 1729 | pld [fb_ptr]; \ |
| 75e28f62 E |
1730 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
| 1731 | \ | |
| 1732 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1733 | setup_blocks_texture_##swizzling(); \ | |
| 1734 | \ | |
| 1735 | beq 5f; \ | |
| 1736 | \ | |
| 1737 | 4: \ | |
| 1738 | vshrn.u32 u_whole_low, u_block, #16; \ | |
| 1739 | \ | |
| e1f6de8f | 1740 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1741 | vshrn.u32 v_whole_low, v_block, #16; \ |
| 1742 | \ | |
| 1743 | add block_ptr_b, block_ptr_b, #32; \ | |
| e1f6de8f | 1744 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
| 75e28f62 E |
1745 | \ |
| 1746 | vdup.u32 dx4, uv_dx4[0]; \ | |
| 1747 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
| 1748 | vdup.u32 dx4, uv_dx4[1]; \ | |
| 1749 | \ | |
| 1750 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
| 1751 | vdup.u32 dx8, uv_dx8[0]; \ | |
| 1752 | \ | |
| 1753 | vadd.u32 u_block, u_block, dx8; \ | |
| 1754 | vdup.u32 dx8, uv_dx8[1]; \ | |
| 1755 | \ | |
| 1756 | vadd.u32 v_block, v_block, dx8; \ | |
| 1757 | vmovn.u16 u_whole_8, u_whole; \ | |
| 1758 | \ | |
| 1759 | add fb_ptr, fb_ptr, #16; \ | |
| 1760 | vmovn.u16 v_whole_8, v_whole; \ | |
| 1761 | \ | |
| e1f6de8f | 1762 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
| 1763 | pld [fb_ptr]; \ | |
| 75e28f62 E |
1764 | \ |
| 1765 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
| 1766 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 1767 | \ | |
| 1768 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
| 1769 | setup_blocks_texture_##swizzling(); \ | |
| 1770 | \ | |
| 1771 | bne 4b; \ | |
| 1772 | \ | |
| 1773 | 5: \ | |
| e1f6de8f | 1774 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 | 1775 | \ |
| e1f6de8f | 1776 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
1777 | vdup.u8 draw_mask, right_mask; \ |
| 1778 | \ | |
| 1779 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
| 1780 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 1781 | vzip.u8 u_whole_8, v_whole_8; \ | |
| 1782 | \ | |
| 1783 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
| 1784 | add block_ptr_b, block_ptr_b, #32; \ | |
| e1f6de8f | 1785 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ |
| 1786 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
| 1787 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
| 75e28f62 E |
1788 | \ |
| 1789 | 1: \ | |
| 1790 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 1791 | add span_edge_data, span_edge_data, #8; \ | |
| 1792 | subs num_spans, num_spans, #1; \ | |
| 1793 | \ | |
| e1f6de8f | 1794 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
1795 | bne 0b; \ |
| 1796 | \ | |
| 3f0189c6 | 1797 | restore_abi_regs(); \ |
| 75e28f62 E |
1798 | ldmia sp!, { r4 - r11, pc }; \ |
| 1799 | \ | |
| 1800 | 2: \ | |
| 1801 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
| 1802 | vpush { texture_mask }; \ | |
| 1803 | vpush { uvrg_dx4 }; \ | |
| 1804 | \ | |
| 26e3e2aa | 1805 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; /* r14=num_blocks */ \ |
| 75e28f62 | 1806 | bl flush_render_block_buffer; \ |
| 26e3e2aa | 1807 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ |
| 75e28f62 E |
1808 | \ |
| 1809 | vpop { uvrg_dx4 }; \ | |
| 1810 | vpop { texture_mask }; \ | |
| 1811 | \ | |
| 1812 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
| 1813 | vmov.u8 fb_mask_ptrs, #0; \ | |
| 1814 | \ | |
| 1815 | mov num_blocks, span_num_blocks; \ | |
| 1816 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 1817 | bal 3b \ | |
| 1818 | ||
| 1819 | ||
| 1820 | setup_blocks_unshaded_textured_builder(swizzled) | |
| 1821 | setup_blocks_unshaded_textured_builder(unswizzled) | |
| 1822 | ||
| 1823 | ||
| 1824 | .align 3 | |
| 1825 | ||
| 1826 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_indirect) | |
| e1f6de8f | 1827 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
1828 | veor.u32 draw_mask, draw_mask, draw_mask |
| 1829 | ||
| 1830 | cmp num_spans, #0 | |
| 1831 | bxeq lr | |
| 1832 | ||
| 1833 | stmdb sp!, { r4 - r11, r14 } | |
| 3f0189c6 | 1834 | save_abi_regs() |
| e1f6de8f | 1835 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 | 1836 | |
| e1f6de8f | 1837 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
| 75e28f62 E |
1838 | |
| 1839 | ubfx color_r, color, #3, #5 | |
| 1840 | ubfx color_g, color, #11, #5 | |
| 1841 | ubfx color_b, color, #19, #5 | |
| 1842 | ||
| 1843 | orr color, color_r, color_b, lsl #10 | |
| 1844 | orr color, color, color_g, lsl #5 | |
| 1845 | ||
| 1846 | vdup.u16 colors, color | |
| 1847 | ||
| e1f6de8f | 1848 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
1849 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset |
| 1850 | ||
| 1851 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
| 1852 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6 | |
| 1853 | ||
| 1854 | 0: | |
| e1f6de8f | 1855 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
| 1856 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
| 75e28f62 | 1857 | |
| e1f6de8f | 1858 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 75e28f62 E |
1859 | |
| 1860 | cmp span_num_blocks, #0 | |
| 1861 | beq 1f | |
| 1862 | ||
| e1f6de8f | 1863 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
| 75e28f62 E |
1864 | add num_blocks, span_num_blocks, num_blocks |
| 1865 | ||
| 1866 | cmp num_blocks, #MAX_BLOCKS | |
| 1867 | bgt 2f | |
| 1868 | ||
| 1869 | 3: | |
| 1870 | add fb_ptr, fb_ptr, y, lsl #11 | |
| 1871 | and y, y, #0x3 | |
| 1872 | ||
| 1873 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
| 1874 | mov c_32, #32 | |
| 1875 | ||
| 1876 | subs span_num_blocks, span_num_blocks, #1 | |
| 1877 | ||
| 1878 | add block_ptr_b, block_ptr_a, #16 | |
| e1f6de8f | 1879 | pld [fb_ptr] |
| 75e28f62 E |
1880 | |
| 1881 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
| 1882 | beq 5f | |
| 1883 | ||
| 1884 | 4: | |
| e1f6de8f | 1885 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_32 |
| 1886 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 | |
| 1887 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 | |
| 75e28f62 E |
1888 | |
| 1889 | add fb_ptr, fb_ptr, #16 | |
| 1890 | add block_ptr_b, block_ptr_b, #32 | |
| 1891 | ||
| e1f6de8f | 1892 | pld [fb_ptr] |
| 75e28f62 E |
1893 | |
| 1894 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
| 1895 | subs span_num_blocks, span_num_blocks, #1 | |
| 1896 | ||
| 1897 | bne 4b | |
| 1898 | ||
| 1899 | 5: | |
| e1f6de8f | 1900 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset] |
| 75e28f62 E |
1901 | |
| 1902 | vdup.u8 draw_mask_edge, right_mask | |
| 1903 | vtst.u16 draw_mask_edge, draw_mask_edge, test_mask | |
| 1904 | ||
| e1f6de8f | 1905 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 |
| 1906 | vst1.u32 { draw_mask_edge }, [block_ptr_a, :128], c_32 | |
| 75e28f62 | 1907 | add block_ptr_b, block_ptr_b, #32 |
| e1f6de8f | 1908 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 |
| 75e28f62 E |
1909 | |
| 1910 | 1: | |
| 1911 | add span_edge_data, span_edge_data, #8 | |
| 1912 | subs num_spans, num_spans, #1 | |
| 1913 | ||
| e1f6de8f | 1914 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
1915 | bne 0b |
| 1916 | ||
| 3f0189c6 | 1917 | restore_abi_regs() |
| 75e28f62 E |
1918 | ldmia sp!, { r4 - r11, pc } |
| 1919 | ||
| 1920 | 2: | |
| 1921 | vpush { colors } | |
| 1922 | ||
| 4d646738 | 1923 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
| 75e28f62 | 1924 | bl flush_render_block_buffer |
| 4d646738 | 1925 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
| 75e28f62 E |
1926 | |
| 1927 | vpop { colors } | |
| 1928 | ||
| e1f6de8f | 1929 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 E |
1930 | veor.u32 draw_mask, draw_mask, draw_mask |
| 1931 | ||
| 1932 | mov num_blocks, span_num_blocks | |
| 1933 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
| 1934 | bal 3b | |
| 1935 | ||
| 1936 | ||
| 1937 | #define mask_msb_scalar r14 | |
| 1938 | ||
| 1939 | #define msb_mask q15 | |
| 1940 | ||
| 1941 | #define pixels_low d16 | |
| 1942 | ||
| 1943 | #define msb_mask_low d30 | |
| 1944 | #define msb_mask_high d31 | |
| 1945 | ||
| 1946 | ||
| 1947 | .align 3 | |
| 1948 | ||
| 1949 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_direct) | |
| e1f6de8f | 1950 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
| 75e28f62 E |
1951 | |
| 1952 | cmp num_spans, #0 | |
| 1953 | bxeq lr | |
| 1954 | ||
| 1955 | stmdb sp!, { r4 - r11, r14 } | |
| 1956 | ||
| e1f6de8f | 1957 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
| 75e28f62 E |
1958 | |
| 1959 | ubfx color_r, color, #3, #5 | |
| 1960 | ubfx color_g, color, #11, #5 | |
| 1961 | ||
| e1f6de8f | 1962 | ldrh mask_msb_scalar, [psx_gpu, #psx_gpu_mask_msb_offset] |
| 75e28f62 E |
1963 | ubfx color_b, color, #19, #5 |
| 1964 | ||
| 1965 | orr color, color_r, color_b, lsl #10 | |
| 1966 | orr color, color, color_g, lsl #5 | |
| 1967 | orr color, color, mask_msb_scalar | |
| 1968 | ||
| 1969 | vdup.u16 colors, color | |
| 1970 | ||
| 1971 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
| ed0fd81d | 1972 | orr color, color, color, lsl #16 |
| 3867c6ef | 1973 | |
| 75e28f62 E |
1974 | |
| 1975 | 0: | |
| e1f6de8f | 1976 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
| 1977 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
| 75e28f62 | 1978 | |
| e1f6de8f | 1979 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 75e28f62 E |
1980 | |
| 1981 | cmp span_num_blocks, #0 | |
| 1982 | beq 1f | |
| 1983 | ||
| e1f6de8f | 1984 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
| 75e28f62 E |
1985 | |
| 1986 | add fb_ptr, fb_ptr, y, lsl #11 | |
| 1987 | subs span_num_blocks, span_num_blocks, #1 | |
| 1988 | ||
| 1989 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
| 1990 | beq 3f | |
| 1991 | ||
| 1992 | 2: | |
| e1f6de8f | 1993 | vst1.u32 { colors }, [fb_ptr]! |
| 75e28f62 E |
1994 | subs span_num_blocks, span_num_blocks, #1 |
| 1995 | ||
| 1996 | bne 2b | |
| 1997 | ||
| 1998 | 3: | |
| e1f6de8f | 1999 | ldrb right_mask, [span_edge_data, #edge_data_right_mask_offset] |
| 75e28f62 | 2000 | |
| 3867c6ef E |
2001 | cmp right_mask, #0x0 |
| 2002 | beq 5f | |
| 2003 | ||
| 2004 | tst right_mask, #0xF | |
| e1f6de8f | 2005 | streq color, [fb_ptr], #4 |
| 3867c6ef | 2006 | moveq right_mask, right_mask, lsr #4 |
| e1f6de8f | 2007 | streq color, [fb_ptr], #4 |
| 3867c6ef E |
2008 | |
| 2009 | tst right_mask, #0x3 | |
| e1f6de8f | 2010 | streq color, [fb_ptr], #4 |
| 3867c6ef E |
2011 | moveq right_mask, right_mask, lsr #2 |
| 2012 | ||
| 2013 | tst right_mask, #0x1 | |
| e1f6de8f | 2014 | strheq color, [fb_ptr] |
| 75e28f62 E |
2015 | |
| 2016 | 1: | |
| 2017 | add span_edge_data, span_edge_data, #8 | |
| 2018 | subs num_spans, num_spans, #1 | |
| 75e28f62 E |
2019 | bne 0b |
| 2020 | ||
| 2021 | ldmia sp!, { r4 - r11, pc } | |
| 2022 | ||
| 3867c6ef | 2023 | 5: |
| e1f6de8f | 2024 | vst1.u32 { colors }, [fb_ptr] |
| 3867c6ef | 2025 | bal 1b |
| 75e28f62 E |
2026 | |
| 2027 | ||
| 2028 | #undef c_64 | |
| 2029 | ||
| 2030 | #define c_64 r7 | |
| 2031 | #define rg_dx_ptr r2 | |
| 2032 | ||
| 2033 | ||
| 2034 | #undef r_block | |
| 2035 | #undef g_block | |
| 2036 | #undef b_block | |
| 2037 | #undef r_whole | |
| 2038 | #undef g_whole | |
| 2039 | #undef b_whole | |
| 2040 | #undef r_whole_low | |
| 2041 | #undef r_whole_high | |
| 2042 | #undef g_whole_low | |
| 2043 | #undef g_whole_high | |
| 2044 | #undef b_whole_low | |
| 2045 | #undef b_whole_high | |
| 2046 | #undef r_whole_8 | |
| 2047 | #undef g_whole_8 | |
| 2048 | #undef b_whole_8 | |
| 2049 | #undef dither_offsets | |
| 2050 | #undef rg_dx4 | |
| 2051 | #undef rg_dx8 | |
| 2052 | #undef dx4 | |
| 2053 | #undef dx8 | |
| 2054 | #undef v_left_x | |
| 2055 | #undef uvrg | |
| 2056 | #undef block_span | |
| 2057 | #undef rg | |
| 2058 | #undef draw_mask | |
| 2059 | #undef test_mask | |
| 2060 | ||
| 2061 | #define r_block q0 | |
| 2062 | #define g_block q1 | |
| 2063 | #define b_block q2 | |
| 2064 | ||
| 2065 | #define r_whole q3 | |
| 2066 | #define g_whole q4 | |
| 2067 | #define b_whole q5 | |
| 2068 | ||
| 2069 | #define r_whole_low d6 | |
| 2070 | #define r_whole_high d7 | |
| 2071 | #define g_whole_low d8 | |
| 2072 | #define g_whole_high d9 | |
| 2073 | #define b_whole_low d10 | |
| 2074 | #define b_whole_high d11 | |
| 2075 | ||
| 2076 | #define gb_whole_8 q6 | |
| 2077 | ||
| 2078 | #define g_whole_8 d12 | |
| 2079 | #define b_whole_8 d13 | |
| 2080 | ||
| 2081 | #define r_whole_8 d14 | |
| 2082 | ||
| 2083 | #define pixels q8 | |
| 2084 | ||
| 2085 | #define rg_dx4 d18 | |
| 2086 | #define rg_dx8 d19 | |
| 2087 | ||
| 2088 | #define dx4 q10 | |
| 2089 | #define dx8 q10 | |
| 2090 | ||
| 2091 | #define v_left_x d6 | |
| 2092 | #define uvrg q4 | |
| 2093 | #define block_span q5 | |
| 2094 | ||
| 2095 | #define rg d9 | |
| 2096 | ||
| 2097 | #define d64_1 d22 | |
| 2098 | #define d64_128 d23 | |
| 2099 | ||
| 2100 | #define d128_4 q12 | |
| 2101 | #define d128_0x7 q13 | |
| 2102 | ||
| 2103 | #define d64_4 d24 | |
| 2104 | ||
| 2105 | #define dither_offsets q14 | |
| 2106 | #define draw_mask q15 | |
| 2107 | ||
| 2108 | #define dither_offsets_low d28 | |
| 2109 | ||
| 2110 | #define rg_dx d0 | |
| 2111 | #define test_mask q10 | |
| 2112 | ||
| 2113 | ||
| 2114 | #define setup_blocks_shaded_untextured_dither_a_dithered() \ | |
| 2115 | vqadd.u8 r_whole_8, r_whole_8, dither_offsets_low; \ | |
| 2116 | vqadd.u8 gb_whole_8, gb_whole_8, dither_offsets; \ | |
| 2117 | ||
| 2118 | #define setup_blocks_shaded_untextured_dither_b_dithered() \ | |
| 2119 | vqsub.u8 r_whole_8, r_whole_8, d64_4; \ | |
| 2120 | vqsub.u8 gb_whole_8, gb_whole_8, d128_4 \ | |
| 2121 | ||
| 2122 | #define setup_blocks_shaded_untextured_dither_a_undithered() \ | |
| 2123 | ||
| 2124 | #define setup_blocks_shaded_untextured_dither_b_undithered() \ | |
| 2125 | ||
| 2126 | ||
| 2127 | #define setup_blocks_shaded_untextured_indirect_builder(dithering) \ | |
| 2128 | .align 3; \ | |
| 2129 | \ | |
| 2130 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_indirect) \ | |
| e1f6de8f | 2131 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
2132 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
| 2133 | \ | |
| e1f6de8f | 2134 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
| 75e28f62 E |
2135 | \ |
| 2136 | cmp num_spans, #0; \ | |
| 2137 | bxeq lr; \ | |
| 2138 | \ | |
| 2139 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 2140 | save_abi_regs(); \ |
| 75e28f62 E |
2141 | vshl.u32 rg_dx4, rg_dx, #2; \ |
| 2142 | \ | |
| e1f6de8f | 2143 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
| 75e28f62 E |
2144 | vshl.u32 rg_dx8, rg_dx, #3; \ |
| 2145 | \ | |
| 2146 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 2147 | \ | |
| e1f6de8f | 2148 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
2149 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
| 2150 | \ | |
| 2151 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 2152 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 2153 | \ | |
| 2154 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
| 2155 | vmov.u8 d64_1, #1; \ | |
| 2156 | \ | |
| 2157 | vmov.u8 d128_4, #4; \ | |
| 2158 | vmov.u8 d64_128, #128; \ | |
| 2159 | \ | |
| 2160 | vmov.u8 d128_0x7, #0x7; \ | |
| 2161 | \ | |
| 2162 | 0: \ | |
| e1f6de8f | 2163 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
2164 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 2165 | \ | |
| e1f6de8f | 2166 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 2167 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
2168 | \ |
| 2169 | cmp span_num_blocks, #0; \ | |
| 2170 | beq 1f; \ | |
| 2171 | \ | |
| e1f6de8f | 2172 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
2173 | add num_blocks, span_num_blocks, num_blocks; \ |
| 2174 | \ | |
| 2175 | cmp num_blocks, #MAX_BLOCKS; \ | |
| 2176 | bgt 2f; \ | |
| 2177 | \ | |
| 2178 | 3: \ | |
| e1f6de8f | 2179 | ldr b, [span_b_offset]; \ |
| 75e28f62 E |
2180 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 2181 | \ | |
| 2182 | vdup.u32 v_left_x, left_x; \ | |
| 2183 | and y, y, #0x3; \ | |
| 2184 | \ | |
| e1f6de8f | 2185 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
2186 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 2187 | \ | |
| 2188 | mla b, b_dx, left_x, b; \ | |
| 2189 | and dither_shift, left_x, #0x03; \ | |
| 2190 | \ | |
| e1f6de8f | 2191 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
2192 | vshr.u32 rg_dx, rg_dx4, #2; \ |
| 2193 | \ | |
| 2194 | mov dither_shift, dither_shift, lsl #3; \ | |
| 2195 | vmla.u32 rg, rg_dx, v_left_x; \ | |
| 2196 | \ | |
| 2197 | mov c_64, #64; \ | |
| 2198 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2199 | \ | |
| 2200 | mov dither_row, dither_row, ror dither_shift; \ | |
| 2201 | mov b_dx4, b_dx, lsl #2; \ | |
| 2202 | \ | |
| 2203 | vdup.u32 dither_offsets, dither_row; \ | |
| 2204 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
| 2205 | \ | |
| 2206 | vdup.u32 b_block, b; \ | |
| 2207 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
| 2208 | \ | |
| 2209 | mov b_dx8, b_dx, lsl #3; \ | |
| 2210 | vdup.u32 r_block, rg[0]; \ | |
| 2211 | vdup.u32 g_block, rg[1]; \ | |
| 2212 | \ | |
| e1f6de8f | 2213 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2214 | \ |
| 2215 | vadd.u32 r_block, r_block, block_span; \ | |
| e1f6de8f | 2216 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2217 | \ |
| 2218 | vadd.u32 g_block, g_block, block_span; \ | |
| e1f6de8f | 2219 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
| 75e28f62 E |
2220 | \ |
| 2221 | vadd.u32 b_block, b_block, block_span; \ | |
| 2222 | add block_ptr_b, block_ptr_a, #16; \ | |
| 2223 | \ | |
| 2224 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2225 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2226 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2227 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2228 | \ | |
| 2229 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2230 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2231 | \ | |
| 2232 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2233 | vdup.u32 dx4, b_dx4; \ | |
| 2234 | \ | |
| 2235 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2236 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2237 | \ | |
| 2238 | vadd.u32 r_block, r_block, dx8; \ | |
| 2239 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2240 | \ | |
| 2241 | vadd.u32 g_block, g_block, dx8; \ | |
| 2242 | vdup.u32 dx8, b_dx8; \ | |
| 2243 | \ | |
| 2244 | vadd.u32 b_block, b_block, dx8; \ | |
| 2245 | \ | |
| 2246 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2247 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2248 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2249 | \ | |
| 2250 | beq 5f; \ | |
| 2251 | veor.u32 draw_mask, draw_mask, draw_mask; \ | |
| 2252 | \ | |
| 2253 | 4: \ | |
| 2254 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2255 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2256 | \ | |
| 2257 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2258 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2259 | \ | |
| 2260 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| e1f6de8f | 2261 | str fb_ptr, [block_ptr_a, #44]; \ |
| 75e28f62 E |
2262 | \ |
| 2263 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2264 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2265 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2266 | \ | |
| 2267 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2268 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2269 | \ | |
| 2270 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2271 | vdup.u32 dx4, b_dx4; \ | |
| 2272 | \ | |
| 2273 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2274 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2275 | \ | |
| 2276 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
| 2277 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2278 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2279 | \ | |
| 2280 | vadd.u32 r_block, r_block, dx8; \ | |
| 2281 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2282 | \ | |
| 2283 | vadd.u32 g_block, g_block, dx8; \ | |
| 2284 | vdup.u32 dx8, b_dx8; \ | |
| 2285 | \ | |
| 2286 | vadd.u32 b_block, b_block, dx8; \ | |
| 2287 | add fb_ptr, fb_ptr, #16; \ | |
| 2288 | \ | |
| 2289 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2290 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2291 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2292 | \ | |
| e1f6de8f | 2293 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
| 2294 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
| 75e28f62 | 2295 | \ |
| e1f6de8f | 2296 | pld [fb_ptr]; \ |
| 75e28f62 E |
2297 | \ |
| 2298 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2299 | bne 4b; \ | |
| 2300 | \ | |
| 2301 | 5: \ | |
| e1f6de8f | 2302 | str fb_ptr, [block_ptr_a, #44]; \ |
| 75e28f62 E |
2303 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ |
| 2304 | \ | |
| e1f6de8f | 2305 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 E |
2306 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
| 2307 | \ | |
| 2308 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2309 | vdup.u8 draw_mask, right_mask; \ | |
| 2310 | \ | |
| 2311 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| e1f6de8f | 2312 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
2313 | \ |
| 2314 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 2315 | \ | |
| 2316 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
| 2317 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2318 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2319 | \ | |
| e1f6de8f | 2320 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
| 2321 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
| 75e28f62 E |
2322 | \ |
| 2323 | 1: \ | |
| 2324 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 2325 | add span_b_offset, span_b_offset, #4; \ | |
| 2326 | \ | |
| 2327 | add span_edge_data, span_edge_data, #8; \ | |
| 2328 | subs num_spans, num_spans, #1; \ | |
| 2329 | \ | |
| e1f6de8f | 2330 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
2331 | bne 0b; \ |
| 2332 | \ | |
| 3f0189c6 | 2333 | restore_abi_regs(); \ |
| 75e28f62 E |
2334 | ldmia sp!, { r4 - r11, pc }; \ |
| 2335 | \ | |
| 2336 | 2: \ | |
| 2337 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
| 2338 | vpush { rg_dx4 }; \ | |
| 2339 | \ | |
| 4d646738 | 2340 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
| 75e28f62 | 2341 | bl flush_render_block_buffer; \ |
| 4d646738 | 2342 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
| 75e28f62 E |
2343 | \ |
| 2344 | vpop { rg_dx4 }; \ | |
| 2345 | \ | |
| 2346 | vmov.u8 d64_1, #1; \ | |
| 2347 | vmov.u8 d128_4, #4; \ | |
| 2348 | vmov.u8 d64_128, #128; \ | |
| 2349 | vmov.u8 d128_0x7, #0x7; \ | |
| 2350 | \ | |
| 2351 | vadd.u32 rg_dx8, rg_dx4, rg_dx4; \ | |
| 2352 | \ | |
| 2353 | mov num_blocks, span_num_blocks; \ | |
| 2354 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 2355 | bal 3b \ | |
| 2356 | ||
| 2357 | ||
| 2358 | setup_blocks_shaded_untextured_indirect_builder(undithered) | |
| 2359 | setup_blocks_shaded_untextured_indirect_builder(dithered) | |
| 2360 | ||
| 2361 | ||
| 2362 | #undef draw_mask | |
| 2363 | ||
| 2364 | #define mask_msb_ptr r14 | |
| 2365 | ||
| 2366 | #define draw_mask q0 | |
| 2367 | #define pixels_low d16 | |
| 3867c6ef | 2368 | #define pixels_high d17 |
| 75e28f62 E |
2369 | |
| 2370 | ||
| 2371 | ||
| 2372 | #define setup_blocks_shaded_untextured_direct_builder(dithering) \ | |
| 2373 | .align 3; \ | |
| 2374 | \ | |
| 2375 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_direct) \ | |
| e1f6de8f | 2376 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
| 75e28f62 E |
2377 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
| 2378 | \ | |
| e1f6de8f | 2379 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
| 75e28f62 E |
2380 | \ |
| 2381 | cmp num_spans, #0; \ | |
| 2382 | bxeq lr; \ | |
| 2383 | \ | |
| 2384 | stmdb sp!, { r4 - r11, r14 }; \ | |
| 3f0189c6 | 2385 | save_abi_regs(); \ |
| 75e28f62 E |
2386 | vshl.u32 rg_dx4, rg_dx, #2; \ |
| 2387 | \ | |
| e1f6de8f | 2388 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
| 75e28f62 E |
2389 | vshl.u32 rg_dx8, rg_dx, #3; \ |
| 2390 | \ | |
| 2391 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
| 2392 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
| 2393 | \ | |
| 2394 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
| 2395 | vmov.u8 d64_1, #1; \ | |
| 2396 | \ | |
| 2397 | vmov.u8 d128_4, #4; \ | |
| 2398 | vmov.u8 d64_128, #128; \ | |
| 2399 | \ | |
| 2400 | vmov.u8 d128_0x7, #0x7; \ | |
| 2401 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| e1f6de8f | 2402 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
2403 | \ |
| 2404 | 0: \ | |
| e1f6de8f | 2405 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
| 75e28f62 E |
2406 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
| 2407 | \ | |
| e1f6de8f | 2408 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
| 2409 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 E |
2410 | \ |
| 2411 | cmp span_num_blocks, #0; \ | |
| 2412 | beq 1f; \ | |
| 2413 | \ | |
| e1f6de8f | 2414 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
| 75e28f62 E |
2415 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 2416 | \ | |
| e1f6de8f | 2417 | ldr b, [span_b_offset]; \ |
| 75e28f62 E |
2418 | vdup.u32 v_left_x, left_x; \ |
| 2419 | and y, y, #0x3; \ | |
| 2420 | \ | |
| e1f6de8f | 2421 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
| 75e28f62 E |
2422 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
| 2423 | \ | |
| 2424 | mla b, b_dx, left_x, b; \ | |
| 2425 | and dither_shift, left_x, #0x03; \ | |
| 2426 | \ | |
| e1f6de8f | 2427 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
| 75e28f62 E |
2428 | vshr.u32 rg_dx, rg_dx4, #2; \ |
| 2429 | \ | |
| 2430 | mov dither_shift, dither_shift, lsl #3; \ | |
| 2431 | vmla.u32 rg, rg_dx, v_left_x; \ | |
| 2432 | \ | |
| 2433 | subs span_num_blocks, span_num_blocks, #1; \ | |
| 2434 | \ | |
| 2435 | mov dither_row, dither_row, ror dither_shift; \ | |
| 2436 | mov b_dx4, b_dx, lsl #2; \ | |
| 2437 | \ | |
| 2438 | vdup.u32 dither_offsets, dither_row; \ | |
| 2439 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
| 2440 | \ | |
| 2441 | vdup.u32 b_block, b; \ | |
| 2442 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
| 2443 | \ | |
| 2444 | mov b_dx8, b_dx, lsl #3; \ | |
| 2445 | vdup.u32 r_block, rg[0]; \ | |
| 2446 | vdup.u32 g_block, rg[1]; \ | |
| 2447 | \ | |
| e1f6de8f | 2448 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2449 | \ |
| 2450 | vadd.u32 r_block, r_block, block_span; \ | |
| e1f6de8f | 2451 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
| 75e28f62 E |
2452 | \ |
| 2453 | vadd.u32 g_block, g_block, block_span; \ | |
| e1f6de8f | 2454 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
| 75e28f62 E |
2455 | \ |
| 2456 | vadd.u32 b_block, b_block, block_span; \ | |
| 2457 | add block_ptr_b, block_ptr_a, #16; \ | |
| 2458 | \ | |
| 2459 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2460 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2461 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2462 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2463 | \ | |
| 2464 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2465 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2466 | \ | |
| 2467 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2468 | vdup.u32 dx4, b_dx4; \ | |
| 2469 | \ | |
| 2470 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2471 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2472 | \ | |
| 2473 | vadd.u32 r_block, r_block, dx8; \ | |
| 2474 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2475 | \ | |
| 2476 | vadd.u32 g_block, g_block, dx8; \ | |
| 2477 | vdup.u32 dx8, b_dx8; \ | |
| 2478 | \ | |
| 2479 | vadd.u32 b_block, b_block, dx8; \ | |
| 2480 | \ | |
| 2481 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2482 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2483 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2484 | \ | |
| 2485 | beq 3f; \ | |
| 2486 | \ | |
| 2487 | 2: \ | |
| 2488 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2489 | vshrn.u32 r_whole_low, r_block, #16; \ | |
| 2490 | \ | |
| 2491 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
| 2492 | vshrn.u32 g_whole_low, g_block, #16; \ | |
| 2493 | \ | |
| 2494 | vshrn.u32 b_whole_low, b_block, #16; \ | |
| 2495 | \ | |
| 2496 | vdup.u32 dx4, rg_dx4[0]; \ | |
| 2497 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 2498 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 2499 | \ | |
| 2500 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
| 2501 | vdup.u32 dx4, rg_dx4[1]; \ | |
| 2502 | \ | |
| 2503 | vmov pixels, msb_mask; \ | |
| 2504 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
| 2505 | vdup.u32 dx4, b_dx4; \ | |
| 2506 | \ | |
| 2507 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
| 2508 | vdup.u32 dx8, rg_dx8[0]; \ | |
| 2509 | \ | |
| 2510 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
| 2511 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2512 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2513 | \ | |
| 2514 | vadd.u32 r_block, r_block, dx8; \ | |
| 2515 | vdup.u32 dx8, rg_dx8[1]; \ | |
| 2516 | \ | |
| 2517 | vadd.u32 g_block, g_block, dx8; \ | |
| 2518 | vdup.u32 dx8, b_dx8; \ | |
| 2519 | \ | |
| 2520 | vadd.u32 b_block, b_block, dx8; \ | |
| 2521 | \ | |
| 2522 | vmovn.u16 r_whole_8, r_whole; \ | |
| 2523 | vmovn.u16 g_whole_8, g_whole; \ | |
| 2524 | vmovn.u16 b_whole_8, b_whole; \ | |
| 2525 | \ | |
| e1f6de8f | 2526 | vst1.u32 { pixels }, [fb_ptr]!; \ |
| 75e28f62 E |
2527 | subs span_num_blocks, span_num_blocks, #1; \ |
| 2528 | bne 2b; \ | |
| 2529 | \ | |
| 2530 | 3: \ | |
| 2531 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
| 2532 | \ | |
| e1f6de8f | 2533 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
| 75e28f62 E |
2534 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
| 2535 | \ | |
| 2536 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
| 3867c6ef | 2537 | rbit right_mask, right_mask; \ |
| 75e28f62 E |
2538 | vmov pixels, msb_mask; \ |
| 2539 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
| 3867c6ef | 2540 | clz right_mask, right_mask; \ |
| 75e28f62 E |
2541 | \ |
| 2542 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
| 2543 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
| 2544 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
| 2545 | \ | |
| 8184d7c5 | 2546 | JT_OP_REL(100f, right_mask, temp); \ |
| e1f6de8f | 2547 | JT_OP(ldr pc, [pc, right_mask, lsl #2]); \ |
| 3867c6ef | 2548 | nop; \ |
| 8184d7c5 | 2549 | 100: \ |
| 3867c6ef | 2550 | nop; \ |
| 8184d7c5 | 2551 | .word JTE(100b, 4f); \ |
| 2552 | .word JTE(100b, 5f); \ | |
| 2553 | .word JTE(100b, 6f); \ | |
| 2554 | .word JTE(100b, 7f); \ | |
| 2555 | .word JTE(100b, 8f); \ | |
| 2556 | .word JTE(100b, 9f); \ | |
| 2557 | .word JTE(100b, 10f); \ | |
| 2558 | .word JTE(100b, 11f); \ | |
| 3867c6ef | 2559 | \ |
| 75e28f62 | 2560 | 4: \ |
| e1f6de8f | 2561 | vst1.u16 { pixels_low[0] }, [fb_ptr]; \ |
| 3867c6ef E |
2562 | bal 1f; \ |
| 2563 | \ | |
| 2564 | 5: \ | |
| e1f6de8f | 2565 | vst1.u32 { pixels_low[0] }, [fb_ptr]; \ |
| 3867c6ef E |
2566 | bal 1f; \ |
| 2567 | \ | |
| 2568 | 6: \ | |
| e1f6de8f | 2569 | vst1.u32 { pixels_low[0] }, [fb_ptr]!; \ |
| 2570 | vst1.u16 { pixels_low[2] }, [fb_ptr]; \ | |
| 3867c6ef E |
2571 | bal 1f; \ |
| 2572 | \ | |
| 2573 | 7: \ | |
| e1f6de8f | 2574 | vst1.u32 { pixels_low }, [fb_ptr]; \ |
| 3867c6ef E |
2575 | bal 1f; \ |
| 2576 | \ | |
| 2577 | 8: \ | |
| e1f6de8f | 2578 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
| 2579 | vst1.u16 { pixels_high[0] }, [fb_ptr]; \ | |
| 3867c6ef E |
2580 | bal 1f; \ |
| 2581 | \ | |
| 2582 | 9: \ | |
| e1f6de8f | 2583 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
| 2584 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
| 3867c6ef E |
2585 | bal 1f; \ |
| 2586 | \ | |
| 2587 | 10: \ | |
| e1f6de8f | 2588 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
| 2589 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
| 2590 | vst1.u16 { pixels_high[2] }, [fb_ptr]; \ | |
| 3867c6ef E |
2591 | bal 1f; \ |
| 2592 | \ | |
| 2593 | 11: \ | |
| e1f6de8f | 2594 | vst1.u32 { pixels }, [fb_ptr]; \ |
| 3867c6ef | 2595 | bal 1f; \ |
| 75e28f62 E |
2596 | \ |
| 2597 | 1: \ | |
| 2598 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
| 2599 | add span_b_offset, span_b_offset, #4; \ | |
| 2600 | \ | |
| 2601 | add span_edge_data, span_edge_data, #8; \ | |
| 2602 | subs num_spans, num_spans, #1; \ | |
| 2603 | \ | |
| 2604 | bne 0b; \ | |
| 2605 | \ | |
| 3f0189c6 | 2606 | restore_abi_regs(); \ |
| 75e28f62 E |
2607 | ldmia sp!, { r4 - r11, pc } \ |
| 2608 | ||
| 2609 | setup_blocks_shaded_untextured_direct_builder(undithered) | |
| 2610 | setup_blocks_shaded_untextured_direct_builder(dithered) | |
| 2611 | ||
| 2612 | ||
| 2613 | #undef psx_gpu | |
| 2614 | #undef num_blocks | |
| 2615 | #undef triangle | |
| 2616 | #undef c_64 | |
| 2617 | ||
| 2618 | #define psx_gpu r0 | |
| 2619 | #define block_ptr r1 | |
| 2620 | #define num_blocks r2 | |
| 2621 | #define uv_01 r3 | |
| 2622 | #define uv_23 r4 | |
| 2623 | #define uv_45 r5 | |
| 2624 | #define uv_67 r6 | |
| 2625 | #define uv_0 r7 | |
| 2626 | #define uv_1 r3 | |
| 2627 | #define uv_2 r8 | |
| 2628 | #define uv_3 r4 | |
| 2629 | #define uv_4 r9 | |
| 2630 | #define uv_5 r5 | |
| 2631 | #define uv_6 r10 | |
| 2632 | #define uv_7 r6 | |
| 2633 | #define texture_ptr r11 | |
| 2634 | ||
| 2635 | #define pixel_0 r7 | |
| 2636 | #define pixel_1 r3 | |
| 2637 | #define pixel_2 r8 | |
| 2638 | #define pixel_3 r4 | |
| 2639 | #define pixel_4 r9 | |
| 2640 | #define pixel_5 r5 | |
| 2641 | #define pixel_6 r10 | |
| 2642 | #define pixel_7 r6 | |
| 2643 | ||
| 2644 | #define pixels_a r7 | |
| 2645 | #define pixels_b r9 | |
| 2646 | #define pixels_c r8 | |
| 2647 | #define pixels_d r10 | |
| 2648 | ||
| 2649 | #define c_64 r0 | |
| 2650 | ||
| 2651 | #define clut_ptr r12 | |
| 2652 | #define current_texture_mask r5 | |
| 2653 | #define dirty_textures_mask r6 | |
| 2654 | ||
| 2655 | #define texels d0 | |
| 2656 | ||
| 2657 | #define clut_low_a d2 | |
| 2658 | #define clut_low_b d3 | |
| 2659 | #define clut_high_a d4 | |
| 2660 | #define clut_high_b d5 | |
| 2661 | ||
| 2662 | #define clut_a q1 | |
| 2663 | #define clut_b q2 | |
| 2664 | ||
| 2665 | #define texels_low d6 | |
| 2666 | #define texels_high d7 | |
| 2667 | ||
| 2668 | .align 3 | |
| 2669 | ||
| 2670 | function(texture_blocks_untextured) | |
| 2671 | bx lr | |
| 2672 | ||
| 2673 | ||
| 2674 | .align 3 | |
| 2675 | ||
| 2676 | function(texture_blocks_4bpp) | |
| 2677 | stmdb sp!, { r3 - r11, r14 } | |
| 2678 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2679 | ||
| e1f6de8f | 2680 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 2681 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
| 75e28f62 | 2682 | |
| e1f6de8f | 2683 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
| 2684 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128] | |
| 75e28f62 | 2685 | |
| e1f6de8f | 2686 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
| 75e28f62 E |
2687 | vuzp.u8 clut_a, clut_b |
| 2688 | ||
| e1f6de8f | 2689 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
| 75e28f62 E |
2690 | tst dirty_textures_mask, current_texture_mask |
| 2691 | ||
| 2692 | bne 1f | |
| 2693 | mov c_64, #64 | |
| 2694 | ||
| 2695 | 0: | |
| 2696 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
| 2697 | ||
| 2698 | uxtah uv_0, texture_ptr, uv_01 | |
| 2699 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
| 2700 | ||
| 2701 | uxtah uv_2, texture_ptr, uv_23 | |
| 2702 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
| 2703 | ||
| 2704 | uxtah uv_4, texture_ptr, uv_45 | |
| e1f6de8f | 2705 | ldrb pixel_0, [uv_0] |
| 75e28f62 E |
2706 | |
| 2707 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
| e1f6de8f | 2708 | ldrb pixel_1, [uv_1] |
| 75e28f62 E |
2709 | |
| 2710 | uxtah uv_6, texture_ptr, uv_67 | |
| e1f6de8f | 2711 | ldrb pixel_2, [uv_2] |
| 75e28f62 E |
2712 | |
| 2713 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
| e1f6de8f | 2714 | ldrb pixel_3, [uv_3] |
| 75e28f62 | 2715 | |
| e1f6de8f | 2716 | ldrb pixel_4, [uv_4] |
| 75e28f62 E |
2717 | subs num_blocks, num_blocks, #1 |
| 2718 | ||
| e1f6de8f | 2719 | ldrb pixel_5, [uv_5] |
| 75e28f62 E |
2720 | orr pixels_a, pixel_0, pixel_1, lsl #8 |
| 2721 | ||
| e1f6de8f | 2722 | ldrb pixel_6, [uv_6] |
| 75e28f62 E |
2723 | orr pixels_b, pixel_4, pixel_5, lsl #8 |
| 2724 | ||
| e1f6de8f | 2725 | ldrb pixel_7, [uv_7] |
| 75e28f62 E |
2726 | orr pixels_a, pixels_a, pixel_2, lsl #16 |
| 2727 | ||
| 2728 | orr pixels_b, pixels_b, pixel_6, lsl #16 | |
| 2729 | orr pixels_a, pixels_a, pixel_3, lsl #24 | |
| 2730 | ||
| 2731 | orr pixels_b, pixels_b, pixel_7, lsl #24 | |
| ed0fd81d | 2732 | vmov texels, pixels_a, pixels_b |
| 75e28f62 E |
2733 | |
| 2734 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels | |
| 2735 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels | |
| 2736 | ||
| e1f6de8f | 2737 | vst2.u8 { texels_low, texels_high }, [block_ptr, :128], c_64 |
| 75e28f62 E |
2738 | bne 0b |
| 2739 | ||
| 2740 | ldmia sp!, { r3 - r11, pc } | |
| 2741 | ||
| 2742 | 1: | |
| 2743 | stmdb sp!, { r1 - r2 } | |
| 2744 | bl update_texture_4bpp_cache | |
| 2745 | ||
| 2746 | mov c_64, #64 | |
| 2747 | ldmia sp!, { r1 - r2 } | |
| 2748 | bal 0b | |
| 2749 | ||
| 2750 | ||
| 2751 | .align 3 | |
| 2752 | ||
| 2753 | function(texture_blocks_8bpp) | |
| 2754 | stmdb sp!, { r3 - r11, r14 } | |
| 2755 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2756 | ||
| e1f6de8f | 2757 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 2758 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
| 75e28f62 | 2759 | |
| e1f6de8f | 2760 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
| 2761 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] | |
| 75e28f62 | 2762 | |
| e1f6de8f | 2763 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset] |
| 75e28f62 E |
2764 | tst dirty_textures_mask, current_texture_mask |
| 2765 | ||
| 2766 | bne 1f | |
| 2767 | nop | |
| 2768 | ||
| 2769 | 0: | |
| 2770 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
| 2771 | ||
| 2772 | uxtah uv_0, texture_ptr, uv_01 | |
| 2773 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
| 2774 | ||
| 2775 | uxtah uv_2, texture_ptr, uv_23 | |
| 2776 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
| 2777 | ||
| 2778 | uxtah uv_4, texture_ptr, uv_45 | |
| e1f6de8f | 2779 | ldrb pixel_0, [uv_0] |
| 75e28f62 E |
2780 | |
| 2781 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
| e1f6de8f | 2782 | ldrb pixel_1, [uv_1] |
| 75e28f62 E |
2783 | |
| 2784 | uxtah uv_6, texture_ptr, uv_67 | |
| e1f6de8f | 2785 | ldrb pixel_2, [uv_2] |
| 75e28f62 E |
2786 | |
| 2787 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
| e1f6de8f | 2788 | ldrb pixel_3, [uv_3] |
| 75e28f62 | 2789 | |
| e1f6de8f | 2790 | ldrb pixel_4, [uv_4] |
| 75e28f62 E |
2791 | add pixel_0, pixel_0, pixel_0 |
| 2792 | ||
| e1f6de8f | 2793 | ldrb pixel_5, [uv_5] |
| 75e28f62 E |
2794 | add pixel_1, pixel_1, pixel_1 |
| 2795 | ||
| e1f6de8f | 2796 | ldrb pixel_6, [uv_6] |
| 75e28f62 E |
2797 | add pixel_2, pixel_2, pixel_2 |
| 2798 | ||
| e1f6de8f | 2799 | ldrb pixel_7, [uv_7] |
| 75e28f62 E |
2800 | add pixel_3, pixel_3, pixel_3 |
| 2801 | ||
| e1f6de8f | 2802 | ldrh pixel_0, [clut_ptr, pixel_0] |
| 75e28f62 E |
2803 | add pixel_4, pixel_4, pixel_4 |
| 2804 | ||
| e1f6de8f | 2805 | ldrh pixel_1, [clut_ptr, pixel_1] |
| 75e28f62 E |
2806 | add pixel_5, pixel_5, pixel_5 |
| 2807 | ||
| e1f6de8f | 2808 | ldrh pixel_2, [clut_ptr, pixel_2] |
| 75e28f62 E |
2809 | add pixel_6, pixel_6, pixel_6 |
| 2810 | ||
| e1f6de8f | 2811 | ldrh pixel_3, [clut_ptr, pixel_3] |
| 75e28f62 E |
2812 | add pixel_7, pixel_7, pixel_7 |
| 2813 | ||
| e1f6de8f | 2814 | ldrh pixel_4, [clut_ptr, pixel_4] |
| 75e28f62 E |
2815 | orr pixels_a, pixel_0, pixel_1, lsl #16 |
| 2816 | ||
| e1f6de8f | 2817 | ldrh pixel_5, [clut_ptr, pixel_5] |
| 75e28f62 E |
2818 | orr pixels_c, pixel_2, pixel_3, lsl #16 |
| 2819 | ||
| e1f6de8f | 2820 | ldrh pixel_6, [clut_ptr, pixel_6] |
| 75e28f62 E |
2821 | subs num_blocks, num_blocks, #1 |
| 2822 | ||
| e1f6de8f | 2823 | ldrh pixel_7, [clut_ptr, pixel_7] |
| 75e28f62 E |
2824 | orr pixels_b, pixel_4, pixel_5, lsl #16 |
| 2825 | ||
| 2826 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
| 2827 | stm block_ptr, { pixels_a, pixels_c, pixels_b, pixels_d } | |
| 2828 | ||
| 2829 | add block_ptr, block_ptr, #64 | |
| 2830 | bne 0b | |
| 2831 | ||
| 2832 | ldmia sp!, { r3 - r11, pc } | |
| 2833 | ||
| 2834 | 1: | |
| 4d646738 | 2835 | stmdb sp!, { r1 - r2, EXTRA_UNSAVED_REGS r12 } |
| 75e28f62 E |
2836 | |
| 2837 | bl update_texture_8bpp_cache | |
| 2838 | ||
| 4d646738 | 2839 | ldmia sp!, { r1 - r2, EXTRA_UNSAVED_REGS r12 } |
| 75e28f62 E |
2840 | bal 0b |
| 2841 | ||
| 2842 | ||
| 2843 | #undef uv_0 | |
| 2844 | #undef uv_1 | |
| 2845 | #undef uv_2 | |
| 2846 | #undef uv_3 | |
| 2847 | #undef uv_4 | |
| 2848 | #undef uv_5 | |
| 2849 | #undef uv_6 | |
| 2850 | #undef uv_7 | |
| 2851 | ||
| 2852 | #undef pixel_0 | |
| 2853 | #undef pixel_1 | |
| 2854 | #undef pixel_2 | |
| 2855 | #undef pixel_3 | |
| 2856 | #undef pixel_4 | |
| 2857 | #undef pixel_5 | |
| 2858 | #undef pixel_6 | |
| 2859 | #undef pixel_7 | |
| 2860 | ||
| 2861 | #undef texture_ptr | |
| 2862 | ||
| 2863 | #undef pixels_a | |
| 2864 | #undef pixels_b | |
| 2865 | #undef pixels_c | |
| 2866 | #undef pixels_d | |
| 2867 | ||
| 2868 | #define psx_gpu r0 | |
| 2869 | #define block_ptr r1 | |
| 2870 | #define num_blocks r2 | |
| 2871 | ||
| 2872 | #define uv_0 r3 | |
| 2873 | #define uv_1 r4 | |
| 2874 | #define u_0 r3 | |
| 2875 | #define u_1 r4 | |
| 2876 | #define v_0 r5 | |
| 2877 | #define v_1 r6 | |
| 2878 | ||
| 2879 | #define uv_2 r5 | |
| 2880 | #define uv_3 r6 | |
| 2881 | #define u_2 r5 | |
| 2882 | #define u_3 r6 | |
| 2883 | #define v_2 r7 | |
| 2884 | #define v_3 r8 | |
| 2885 | ||
| 2886 | #define uv_4 r7 | |
| 2887 | #define uv_5 r8 | |
| 2888 | #define u_4 r7 | |
| 2889 | #define u_5 r8 | |
| 2890 | #define v_4 r9 | |
| 2891 | #define v_5 r10 | |
| 2892 | ||
| 2893 | #define uv_6 r9 | |
| 2894 | #define uv_7 r10 | |
| 2895 | #define u_6 r9 | |
| 2896 | #define u_7 r10 | |
| 2897 | #define v_6 r11 | |
| 2898 | #define v_7 r0 | |
| 2899 | ||
| 2900 | #define pixel_0 r3 | |
| 2901 | #define pixel_1 r4 | |
| 2902 | #define pixel_2 r5 | |
| 2903 | #define pixel_3 r6 | |
| 2904 | #define pixel_4 r7 | |
| 2905 | #define pixel_5 r8 | |
| 2906 | #define pixel_6 r9 | |
| 2907 | #define pixel_7 r10 | |
| 2908 | ||
| 2909 | #define pixels_a r3 | |
| 2910 | #define pixels_b r5 | |
| 2911 | #define pixels_c r7 | |
| 2912 | #define pixels_d r9 | |
| 2913 | ||
| 2914 | #define texture_ptr r12 | |
| 2915 | ||
| 2916 | ||
| 2917 | .align 3 | |
| 2918 | ||
| 2919 | function(texture_blocks_16bpp) | |
| 2920 | stmdb sp!, { r3 - r11, r14 } | |
| 2921 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 2922 | ||
| e1f6de8f | 2923 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 2924 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] | |
| 75e28f62 E |
2925 | |
| 2926 | 0: | |
| e1f6de8f | 2927 | ldrh uv_0, [block_ptr] |
| 75e28f62 E |
2928 | subs num_blocks, num_blocks, #1 |
| 2929 | ||
| e1f6de8f | 2930 | ldrh uv_1, [block_ptr, #2] |
| 75e28f62 E |
2931 | |
| 2932 | and v_0, uv_0, #0xFF00 | |
| 2933 | and v_1, uv_1, #0xFF00 | |
| 2934 | ||
| 2935 | and u_0, uv_0, #0xFF | |
| 2936 | and u_1, uv_1, #0xFF | |
| 2937 | ||
| 2938 | add uv_0, u_0, v_0, lsl #2 | |
| e1f6de8f | 2939 | ldrh uv_2, [block_ptr, #4] |
| 75e28f62 E |
2940 | |
| 2941 | add uv_1, u_1, v_1, lsl #2 | |
| e1f6de8f | 2942 | ldrh uv_3, [block_ptr, #6] |
| 75e28f62 E |
2943 | |
| 2944 | add uv_0, uv_0, uv_0 | |
| 2945 | add uv_1, uv_1, uv_1 | |
| 2946 | ||
| 2947 | and v_2, uv_2, #0xFF00 | |
| 2948 | and v_3, uv_3, #0xFF00 | |
| 2949 | ||
| 2950 | and u_2, uv_2, #0xFF | |
| 2951 | and u_3, uv_3, #0xFF | |
| 2952 | ||
| 2953 | add uv_2, u_2, v_2, lsl #2 | |
| e1f6de8f | 2954 | ldrh uv_4, [block_ptr, #8] |
| 75e28f62 E |
2955 | |
| 2956 | add uv_3, u_3, v_3, lsl #2 | |
| e1f6de8f | 2957 | ldrh uv_5, [block_ptr, #10] |
| 75e28f62 E |
2958 | |
| 2959 | add uv_2, uv_2, uv_2 | |
| 2960 | add uv_3, uv_3, uv_3 | |
| 2961 | ||
| 2962 | and v_4, uv_4, #0xFF00 | |
| 2963 | and v_5, uv_5, #0xFF00 | |
| 2964 | ||
| 2965 | and u_4, uv_4, #0xFF | |
| 2966 | and u_5, uv_5, #0xFF | |
| 2967 | ||
| 2968 | add uv_4, u_4, v_4, lsl #2 | |
| e1f6de8f | 2969 | ldrh uv_6, [block_ptr, #12] |
| 75e28f62 E |
2970 | |
| 2971 | add uv_5, u_5, v_5, lsl #2 | |
| e1f6de8f | 2972 | ldrh uv_7, [block_ptr, #14] |
| 75e28f62 E |
2973 | |
| 2974 | add uv_4, uv_4, uv_4 | |
| e1f6de8f | 2975 | ldrh pixel_0, [texture_ptr, uv_0] |
| 75e28f62 E |
2976 | |
| 2977 | add uv_5, uv_5, uv_5 | |
| e1f6de8f | 2978 | ldrh pixel_1, [texture_ptr, uv_1] |
| 75e28f62 E |
2979 | |
| 2980 | and v_6, uv_6, #0xFF00 | |
| e1f6de8f | 2981 | ldrh pixel_2, [texture_ptr, uv_2] |
| 75e28f62 E |
2982 | |
| 2983 | and v_7, uv_7, #0xFF00 | |
| e1f6de8f | 2984 | ldrh pixel_3, [texture_ptr, uv_3] |
| 75e28f62 E |
2985 | |
| 2986 | and u_6, uv_6, #0xFF | |
| e1f6de8f | 2987 | ldrh pixel_4, [texture_ptr, uv_4] |
| 75e28f62 E |
2988 | |
| 2989 | and u_7, uv_7, #0xFF | |
| e1f6de8f | 2990 | ldrh pixel_5, [texture_ptr, uv_5] |
| 75e28f62 E |
2991 | |
| 2992 | add uv_6, u_6, v_6, lsl #2 | |
| 2993 | add uv_7, u_7, v_7, lsl #2 | |
| 2994 | ||
| 2995 | add uv_6, uv_6, uv_6 | |
| 2996 | add uv_7, uv_7, uv_7 | |
| 2997 | ||
| 2998 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
| 2999 | orr pixels_b, pixel_2, pixel_3, lsl #16 | |
| 3000 | ||
| e1f6de8f | 3001 | ldrh pixel_6, [texture_ptr, uv_6] |
| 75e28f62 E |
3002 | orr pixels_c, pixel_4, pixel_5, lsl #16 |
| 3003 | ||
| e1f6de8f | 3004 | ldrh pixel_7, [texture_ptr, uv_7] |
| 75e28f62 E |
3005 | orr pixels_d, pixel_6, pixel_7, lsl #16 |
| 3006 | ||
| 3007 | stm block_ptr, { pixels_a, pixels_b, pixels_c, pixels_d } | |
| 3008 | add block_ptr, block_ptr, #64 | |
| 3009 | ||
| 3010 | bne 0b | |
| 3011 | ||
| 3012 | ldmia sp!, { r3 - r11, pc } | |
| 3013 | ||
| 3014 | ||
| 3015 | #undef num_blocks | |
| 3016 | ||
| 3017 | #undef test_mask | |
| 3018 | #undef texels | |
| 3019 | #undef pixels_b | |
| 3020 | #undef pixels | |
| 3021 | #undef d64_1 | |
| 3022 | #undef d64_4 | |
| 3023 | #undef d64_128 | |
| 3024 | #undef draw_mask | |
| 3025 | #undef msb_mask | |
| 3026 | #undef msb_mask_low | |
| 3027 | #undef msb_mask_high | |
| 3028 | #undef fb_pixels | |
| 3029 | ||
| 3030 | #undef c_32 | |
| 3031 | #undef fb_ptr | |
| 3032 | #undef mask_msb_ptr | |
| 3033 | ||
| 3034 | #define psx_gpu r0 | |
| 3035 | #define num_blocks r1 | |
| 3036 | #define color_ptr r2 | |
| 3867c6ef E |
3037 | #define colors_scalar r2 |
| 3038 | #define colors_scalar_compare r3 | |
| 75e28f62 E |
3039 | #define mask_msb_ptr r2 |
| 3040 | ||
| 3041 | #define block_ptr_load_a r0 | |
| 3042 | #define block_ptr_store r3 | |
| 3043 | #define block_ptr_load_b r12 | |
| 3044 | #define c_32 r2 | |
| 3045 | ||
| 3046 | #define c_48 r4 | |
| 3047 | #define fb_ptr r14 | |
| 3048 | #define draw_mask_bits_scalar r5 | |
| 3049 | ||
| 3050 | #define d128_0x07 q0 | |
| 3051 | #define d128_0x1F q1 | |
| 3052 | #define d128_0x8000 q2 | |
| 3053 | #define test_mask q3 | |
| 3054 | #define texels q4 | |
| 3055 | #define colors_rg q5 | |
| 3056 | #define colors_b_dm_bits q6 | |
| 3057 | #define texels_rg q7 | |
| 3058 | #define pixels_r q8 | |
| 3059 | #define pixels_g q9 | |
| 3060 | #define pixels_b q10 | |
| 3061 | #define pixels q11 | |
| 3062 | #define zero_mask q4 | |
| 3063 | #define draw_mask q12 | |
| 3064 | #define msb_mask q13 | |
| 3065 | ||
| 3066 | #define fb_pixels q8 | |
| 3067 | ||
| 3068 | #define pixels_gb_low q9 | |
| 3069 | ||
| 3070 | #define colors_r d10 | |
| 3071 | #define colors_g d11 | |
| 3072 | #define colors_b d12 | |
| 3073 | #define draw_mask_bits d13 | |
| 3074 | #define texels_r d14 | |
| 3075 | #define texels_g d15 | |
| 3076 | #define pixels_r_low d16 | |
| 3077 | #define pixels_g_low d18 | |
| 3078 | #define pixels_b_low d19 | |
| 3079 | #define msb_mask_low d26 | |
| 3080 | #define msb_mask_high d27 | |
| 3081 | ||
| 3082 | #define d64_1 d28 | |
| 3083 | #define d64_4 d29 | |
| 3084 | #define d64_128 d30 | |
| 3085 | #define texels_b d31 | |
| 3086 | ||
| 3087 | #define shade_blocks_textured_modulated_prologue_indirect() \ | |
| 3088 | mov c_48, #48; \ | |
| 3089 | add block_ptr_store, psx_gpu, #psx_gpu_blocks_offset \ | |
| 3090 | ||
| 3091 | #define shade_blocks_textured_modulated_prologue_direct() \ | |
| 3092 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
| e1f6de8f | 3093 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] \ |
| 75e28f62 | 3094 | |
| 75e28f62 | 3095 | |
| 3867c6ef E |
3096 | #define shade_blocks_textured_modulated_prologue_shaded(dithering, target) \ |
| 3097 | ||
| 3098 | #define shade_blocks_textured_false_modulation_check_undithered(target) \ | |
| e1f6de8f | 3099 | ldr colors_scalar, [psx_gpu, #psx_gpu_triangle_color_offset]; \ |
| 3867c6ef E |
3100 | movw colors_scalar_compare, #0x8080; \ |
| 3101 | \ | |
| 3102 | movt colors_scalar_compare, #0x80; \ | |
| 3103 | cmp colors_scalar, colors_scalar_compare; \ | |
| 3104 | beq shade_blocks_textured_unmodulated_##target \ | |
| 3105 | ||
| 3106 | #define shade_blocks_textured_false_modulation_check_dithered(target) \ | |
| 3107 | ||
| 3108 | #define shade_blocks_textured_modulated_prologue_unshaded(dithering, target) \ | |
| 3109 | shade_blocks_textured_false_modulation_check_##dithering(target); \ | |
| 75e28f62 | 3110 | add color_ptr, psx_gpu, #psx_gpu_triangle_color_offset; \ |
| e1f6de8f | 3111 | vld1.u32 { colors_r[] }, [color_ptr, :32]; \ |
| 75e28f62 E |
3112 | vdup.u8 colors_g, colors_r[1]; \ |
| 3113 | vdup.u8 colors_b, colors_r[2]; \ | |
| 3114 | vdup.u8 colors_r, colors_r[0] \ | |
| 3115 | ||
| 3116 | ||
| 3117 | #define shade_blocks_textured_modulated_load_dithered(target) \ | |
| e1f6de8f | 3118 | vld1.u32 { target }, [block_ptr_load_b, :128] \ |
| 75e28f62 E |
3119 | |
| 3120 | #define shade_blocks_textured_modulated_load_last_dithered(target) \ | |
| e1f6de8f | 3121 | vld1.u32 { target }, [block_ptr_load_b, :128], c_32 \ |
| 75e28f62 E |
3122 | |
| 3123 | #define shade_blocks_textured_modulated_load_undithered(target) \ | |
| 3124 | ||
| 3125 | #define shade_blocks_textured_modulated_load_last_undithered(target) \ | |
| 3126 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
| 3127 | ||
| 3128 | #define shade_blocks_textured_modulate_dithered(channel) \ | |
| 3129 | vmlal.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
| 3130 | ||
| 3131 | #define shade_blocks_textured_modulate_undithered(channel) \ | |
| 3132 | vmull.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
| 3133 | ||
| 3134 | ||
| 3135 | #define shade_blocks_textured_modulated_store_draw_mask_indirect(offset) \ | |
| e1f6de8f | 3136 | vst1.u32 { draw_mask }, [block_ptr_store, :128]! \ |
| 75e28f62 E |
3137 | |
| 3138 | #define shade_blocks_textured_modulated_store_draw_mask_direct(offset) \ | |
| e1f6de8f | 3139 | ldr fb_ptr, [block_ptr_load_b, #(offset - 64)]; \ |
| 3140 | vld1.u32 { fb_pixels }, [fb_ptr]; \ | |
| 75e28f62 E |
3141 | vbit.u16 pixels, fb_pixels, draw_mask \ |
| 3142 | ||
| 3143 | #define shade_blocks_textured_modulated_store_pixels_indirect() \ | |
| e1f6de8f | 3144 | vst1.u32 { pixels }, [block_ptr_store, :128], c_48 \ |
| 75e28f62 E |
3145 | |
| 3146 | #define shade_blocks_textured_modulated_store_pixels_direct() \ | |
| e1f6de8f | 3147 | vst1.u32 { pixels }, [fb_ptr] \ |
| 75e28f62 E |
3148 | |
| 3149 | ||
| 3150 | #define shade_blocks_textured_modulated_load_rg_shaded() \ | |
| e1f6de8f | 3151 | vld1.u32 { colors_r, colors_g }, [block_ptr_load_b, :128], c_32 \ |
| 75e28f62 E |
3152 | |
| 3153 | #define shade_blocks_textured_modulated_load_rg_unshaded() \ | |
| 3154 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
| 3155 | ||
| 3156 | #define shade_blocks_textured_modulated_load_bdm_shaded() \ | |
| e1f6de8f | 3157 | vld1.u32 { colors_b, draw_mask_bits }, [block_ptr_load_a, :128], c_32 \ |
| 75e28f62 E |
3158 | |
| 3159 | #define shade_blocks_textured_modulated_load_bdm_unshaded() \ | |
| e1f6de8f | 3160 | ldr draw_mask_bits_scalar, [block_ptr_load_a, #8]; \ |
| 75e28f62 E |
3161 | add block_ptr_load_a, block_ptr_load_a, #32 \ |
| 3162 | ||
| 3163 | #define shade_blocks_textured_modulated_expand_draw_mask_shaded() \ | |
| 3164 | vdup.u16 draw_mask, draw_mask_bits[0] \ | |
| 3165 | ||
| 3166 | #define shade_blocks_textured_modulated_expand_draw_mask_unshaded() \ | |
| 3167 | vdup.u16 draw_mask, draw_mask_bits_scalar \ | |
| 3168 | ||
| 3169 | ||
| 3170 | #define shade_blocks_textured_modulated_apply_msb_mask_indirect() \ | |
| 3171 | ||
| 3172 | #define shade_blocks_textured_modulated_apply_msb_mask_direct() \ | |
| 3173 | vorr.u16 pixels, pixels, msb_mask \ | |
| 3174 | ||
| 3175 | ||
| 3176 | #define shade_blocks_textured_modulated_builder(shading, dithering, target) \ | |
| 3177 | .align 3; \ | |
| 3178 | \ | |
| 3179 | function(shade_blocks_##shading##_textured_modulated_##dithering##_##target) \ | |
| fab27ba2 | 3180 | save_abi_regs(); \ |
| 3867c6ef | 3181 | shade_blocks_textured_modulated_prologue_##shading(dithering, target); \ |
| 75e28f62 | 3182 | stmdb sp!, { r4 - r5, lr }; \ |
| e1f6de8f | 3183 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 | 3184 | \ |
| e1f6de8f | 3185 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
| 75e28f62 E |
3186 | \ |
| 3187 | shade_blocks_textured_modulated_prologue_##target(); \ | |
| 75e28f62 E |
3188 | \ |
| 3189 | add block_ptr_load_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3190 | mov c_32, #32; \ | |
| 3191 | \ | |
| 3192 | add block_ptr_load_b, block_ptr_load_a, #16; \ | |
| 3193 | vmov.u8 d64_1, #1; \ | |
| 3194 | vmov.u8 d64_4, #4; \ | |
| 3195 | vmov.u8 d64_128, #128; \ | |
| 3196 | \ | |
| e1f6de8f | 3197 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
| 75e28f62 E |
3198 | vmov.u8 d128_0x07, #0x07; \ |
| 3199 | \ | |
| 3200 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
| 3201 | vmov.u8 d128_0x1F, #0x1F; \ | |
| 3202 | \ | |
| 3203 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
| 3204 | vmov.u16 d128_0x8000, #0x8000; \ | |
| 3205 | \ | |
| 3206 | vmovn.u16 texels_r, texels; \ | |
| 3207 | vshrn.u16 texels_g, texels, #5; \ | |
| 3208 | \ | |
| 3209 | vshrn.u16 texels_b, texels, #7; \ | |
| 3210 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
| 3211 | \ | |
| 3212 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
| 3213 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 3214 | \ | |
| 3215 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
| 3216 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
| 3217 | \ | |
| 3218 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
| 3219 | vshr.u8 texels_b, texels_b, #3; \ | |
| 3220 | \ | |
| 3221 | shade_blocks_textured_modulate_##dithering(r); \ | |
| 3222 | shade_blocks_textured_modulate_##dithering(g); \ | |
| 3223 | shade_blocks_textured_modulate_##dithering(b); \ | |
| 3224 | \ | |
| 3225 | vand.u16 pixels, texels, d128_0x8000; \ | |
| 3226 | vceq.u16 zero_mask, texels, #0; \ | |
| 3227 | \ | |
| 3228 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
| 3229 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
| 3230 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
| 3231 | \ | |
| 3232 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
| 3233 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
| 3234 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
| 3235 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
| 3236 | \ | |
| 3237 | subs num_blocks, num_blocks, #1; \ | |
| 3238 | beq 1f; \ | |
| 3239 | \ | |
| 3240 | .align 3; \ | |
| 3241 | \ | |
| 3242 | 0: \ | |
| e1f6de8f | 3243 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
| 75e28f62 E |
3244 | shade_blocks_textured_modulated_load_rg_##shading(); \ |
| 3245 | vshrn.u16 texels_g, texels, #5; \ | |
| 3246 | \ | |
| 3247 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
| 3248 | vshrn.u16 texels_b, texels, #7; \ | |
| 3249 | \ | |
| e1f6de8f | 3250 | pld [block_ptr_load_a]; \ |
| 75e28f62 E |
3251 | vmovn.u16 texels_r, texels; \ |
| 3252 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
| 3253 | \ | |
| 3254 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
| 3255 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
| 3256 | shade_blocks_textured_modulated_store_draw_mask_##target(-4); \ | |
| 3257 | \ | |
| 3258 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
| 3259 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
| 3260 | \ | |
| 3261 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
| 3262 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
| 3263 | \ | |
| 3264 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
| 3265 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
| 3266 | \ | |
| 3267 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
| 3268 | vshr.u8 texels_b, texels_b, #3; \ | |
| 3269 | \ | |
| 3270 | shade_blocks_textured_modulate_##dithering(r); \ | |
| 3271 | shade_blocks_textured_modulate_##dithering(g); \ | |
| 3272 | shade_blocks_textured_modulate_##dithering(b); \ | |
| 3273 | \ | |
| 3274 | vand.u16 pixels, texels, d128_0x8000; \ | |
| 3275 | vceq.u16 zero_mask, texels, #0; \ | |
| 3276 | \ | |
| 3277 | subs num_blocks, num_blocks, #1; \ | |
| 3278 | \ | |
| 3279 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
| 3280 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
| 3281 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
| 3282 | \ | |
| 3283 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
| 3284 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
| 3285 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
| 3286 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
| 3287 | \ | |
| 3288 | bne 0b; \ | |
| 3289 | \ | |
| 3290 | 1: \ | |
| 3291 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
| 3292 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
| 3293 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
| 3294 | \ | |
| 3295 | shade_blocks_textured_modulated_store_draw_mask_##target(28); \ | |
| 3296 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
| 3297 | \ | |
| fab27ba2 | 3298 | ldmia sp!, { r4 - r5, lr }; \ |
| 3f0189c6 | 3299 | restore_abi_regs(); \ |
| fab27ba2 | 3300 | bx lr \ |
| 75e28f62 E |
3301 | |
| 3302 | ||
| 3303 | shade_blocks_textured_modulated_builder(shaded, dithered, direct); | |
| 3304 | shade_blocks_textured_modulated_builder(shaded, undithered, direct); | |
| 3305 | shade_blocks_textured_modulated_builder(unshaded, dithered, direct); | |
| 3306 | shade_blocks_textured_modulated_builder(unshaded, undithered, direct); | |
| 3307 | ||
| 3308 | shade_blocks_textured_modulated_builder(shaded, dithered, indirect); | |
| 3309 | shade_blocks_textured_modulated_builder(shaded, undithered, indirect); | |
| 3310 | shade_blocks_textured_modulated_builder(unshaded, dithered, indirect); | |
| 3311 | shade_blocks_textured_modulated_builder(unshaded, undithered, indirect); | |
| 3312 | ||
| 3313 | ||
| 3314 | #undef c_64 | |
| 3315 | #undef fb_ptr | |
| 3316 | #undef color_ptr | |
| 3317 | ||
| 3318 | #undef color_r | |
| 3319 | #undef color_g | |
| 3320 | #undef color_b | |
| 3321 | ||
| 3322 | #undef test_mask | |
| 3323 | #undef pixels | |
| 3324 | #undef draw_mask | |
| 3325 | #undef zero_mask | |
| 3326 | #undef fb_pixels | |
| 3327 | #undef msb_mask | |
| 3328 | #undef msb_mask_low | |
| 3329 | #undef msb_mask_high | |
| 3330 | ||
| 3331 | #define psx_gpu r0 | |
| 3332 | #define num_blocks r1 | |
| 3333 | #define mask_msb_ptr r2 | |
| 3334 | #define color_ptr r3 | |
| 3335 | ||
| 3336 | #define block_ptr_load r0 | |
| 3337 | #define draw_mask_store_ptr r3 | |
| 3338 | #define draw_mask_bits_ptr r12 | |
| 3339 | #define draw_mask_ptr r12 | |
| 3340 | #define pixel_store_ptr r14 | |
| 3341 | ||
| 3342 | #define fb_ptr_cmp r4 | |
| 3343 | ||
| 3344 | #define fb_ptr r3 | |
| 3345 | #define fb_ptr_next r14 | |
| 3346 | ||
| 3347 | #define c_64 r2 | |
| 3348 | ||
| 3349 | #define test_mask q0 | |
| 3350 | #define pixels q1 | |
| 3351 | #define draw_mask q2 | |
| 3352 | #define zero_mask q3 | |
| 3353 | #define draw_mask_combined q4 | |
| 3354 | #define fb_pixels q5 | |
| 3355 | #define fb_pixels_next q6 | |
| 3356 | #define msb_mask q7 | |
| 3357 | ||
| 3358 | #define draw_mask_low d4 | |
| 3359 | #define draw_mask_high d5 | |
| 3360 | #define msb_mask_low d14 | |
| 3361 | #define msb_mask_high d15 | |
| 3362 | ||
| 3363 | .align 3 | |
| 3364 | function(shade_blocks_textured_unmodulated_indirect) | |
| 3f0189c6 | 3365 | stmdb sp!, { r4, r14 } |
| 3366 | save_abi_regs() | |
| 75e28f62 E |
3367 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
| 3368 | ||
| e1f6de8f | 3369 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
3370 | add pixel_store_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
| 3371 | ||
| e1f6de8f | 3372 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 E |
3373 | add draw_mask_store_ptr, psx_gpu, #psx_gpu_blocks_offset |
| 3374 | ||
| 3375 | mov c_64, #64 | |
| 3376 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
| 3377 | ||
| e1f6de8f | 3378 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 75e28f62 | 3379 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ |
| e1f6de8f | 3380 | [draw_mask_bits_ptr, :16], c_64 |
| 75e28f62 E |
3381 | vceq.u16 zero_mask, pixels, #0 |
| 3382 | ||
| 3383 | vtst.u16 draw_mask, draw_mask, test_mask | |
| e1f6de8f | 3384 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
| 75e28f62 E |
3385 | |
| 3386 | subs num_blocks, num_blocks, #1 | |
| 3387 | beq 1f | |
| 3388 | ||
| 3389 | 0: | |
| e1f6de8f | 3390 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 75e28f62 E |
3391 | vorr.u16 draw_mask_combined, draw_mask, zero_mask |
| 3392 | ||
| 3393 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| e1f6de8f | 3394 | [draw_mask_bits_ptr, :16], c_64 |
| 75e28f62 E |
3395 | vceq.u16 zero_mask, pixels, #0 |
| 3396 | ||
| 3397 | vtst.u16 draw_mask, draw_mask, test_mask | |
| e1f6de8f | 3398 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
| 75e28f62 | 3399 | |
| e1f6de8f | 3400 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
| 75e28f62 E |
3401 | subs num_blocks, num_blocks, #1 |
| 3402 | ||
| 3403 | bne 0b | |
| 3404 | ||
| 3405 | 1: | |
| 3406 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| e1f6de8f | 3407 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
| 75e28f62 | 3408 | |
| 3f0189c6 | 3409 | restore_abi_regs() |
| 3410 | ldmia sp!, { r4, pc } | |
| 75e28f62 E |
3411 | |
| 3412 | ||
| 3413 | .align 3 | |
| 3414 | ||
| 3415 | function(shade_blocks_textured_unmodulated_direct) | |
| 3416 | stmdb sp!, { r4, r14 } | |
| 3f0189c6 | 3417 | save_abi_regs() |
| 75e28f62 E |
3418 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
| 3419 | ||
| e1f6de8f | 3420 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
3421 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
| 3422 | ||
| e1f6de8f | 3423 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
| 75e28f62 E |
3424 | mov c_64, #64 |
| 3425 | ||
| e1f6de8f | 3426 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| 75e28f62 E |
3427 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset |
| 3428 | ||
| 3429 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| e1f6de8f | 3430 | [draw_mask_bits_ptr, :16], c_64 |
| 3431 | ldr fb_ptr_next, [block_ptr_load, #44] | |
| 75e28f62 | 3432 | |
| e1f6de8f | 3433 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 3434 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
| 75e28f62 E |
3435 | vceq.u16 zero_mask, pixels, #0 |
| 3436 | vtst.u16 draw_mask, draw_mask, test_mask | |
| 3437 | ||
| 3438 | subs num_blocks, num_blocks, #1 | |
| 3439 | beq 1f | |
| 3440 | ||
| 3441 | 0: | |
| 3442 | mov fb_ptr, fb_ptr_next | |
| e1f6de8f | 3443 | ldr fb_ptr_next, [block_ptr_load, #44] |
| 75e28f62 E |
3444 | |
| 3445 | vorr.u16 pixels, pixels, msb_mask | |
| 3446 | ||
| 3447 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3448 | vmov fb_pixels, fb_pixels_next | |
| 3449 | ||
| 3450 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
| e1f6de8f | 3451 | [draw_mask_bits_ptr, :16], c_64 |
| 75e28f62 E |
3452 | vbif.u16 fb_pixels, pixels, draw_mask_combined |
| 3453 | ||
| 75e28f62 | 3454 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr |
| e1f6de8f | 3455 | pld [fb_ptr_next, #64] |
| 8438c3c7 | 3456 | |
| 75e28f62 | 3457 | add fb_ptr_cmp, fb_ptr_cmp, #14 |
| e1f6de8f | 3458 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
| 8438c3c7 | 3459 | |
| 75e28f62 E |
3460 | cmp fb_ptr_cmp, #28 |
| 3461 | bls 4f | |
| 3462 | ||
| e1f6de8f | 3463 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 E |
3464 | vceq.u16 zero_mask, pixels, #0 |
| 3465 | ||
| e1f6de8f | 3466 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 E |
3467 | vtst.u16 draw_mask, draw_mask, test_mask |
| 3468 | ||
| 3469 | 3: | |
| 3470 | subs num_blocks, num_blocks, #1 | |
| 3471 | bne 0b | |
| 3472 | ||
| 3473 | 1: | |
| 3474 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
| 3475 | vbif.u16 fb_pixels_next, pixels, draw_mask_combined | |
| 3476 | ||
| e1f6de8f | 3477 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 | 3478 | |
| 3f0189c6 | 3479 | restore_abi_regs() |
| 75e28f62 E |
3480 | ldmia sp!, { r4, pc } |
| 3481 | ||
| 3482 | 4: | |
| e1f6de8f | 3483 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 E |
3484 | vceq.u16 zero_mask, pixels, #0 |
| 3485 | ||
| e1f6de8f | 3486 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 E |
3487 | vtst.u16 draw_mask, draw_mask, test_mask |
| 3488 | ||
| 3489 | bal 3b | |
| 3490 | ||
| 3491 | ||
| 3492 | function(shade_blocks_unshaded_untextured_indirect) | |
| 3493 | bx lr | |
| 3494 | ||
| 3495 | .align 3 | |
| 3496 | ||
| 3497 | function(shade_blocks_unshaded_untextured_direct) | |
| 3498 | stmdb sp!, { r4, r14 } | |
| 3f0189c6 | 3499 | save_abi_regs() |
| 75e28f62 E |
3500 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset |
| 3501 | ||
| e1f6de8f | 3502 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
3503 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
| 3504 | ||
| e1f6de8f | 3505 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
| 75e28f62 E |
3506 | add color_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
| 3507 | ||
| 3508 | add block_ptr_load, psx_gpu, #(psx_gpu_blocks_offset + 44) | |
| e1f6de8f | 3509 | vld1.u16 { pixels }, [color_ptr, :128] |
| 75e28f62 E |
3510 | |
| 3511 | mov c_64, #64 | |
| e1f6de8f | 3512 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
| 75e28f62 E |
3513 | |
| 3514 | vorr.u16 pixels, pixels, msb_mask | |
| 3515 | subs num_blocks, num_blocks, #1 | |
| 3516 | ||
| e1f6de8f | 3517 | ldr fb_ptr_next, [block_ptr_load], #64 |
| 75e28f62 | 3518 | |
| e1f6de8f | 3519 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 E |
3520 | beq 1f |
| 3521 | ||
| 3522 | 0: | |
| 3523 | vmov fb_pixels, fb_pixels_next | |
| 3524 | mov fb_ptr, fb_ptr_next | |
| e1f6de8f | 3525 | ldr fb_ptr_next, [block_ptr_load], #64 |
| 75e28f62 E |
3526 | |
| 3527 | vbif.u16 fb_pixels, pixels, draw_mask | |
| e1f6de8f | 3528 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
| 75e28f62 E |
3529 | |
| 3530 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr | |
| 3531 | add fb_ptr_cmp, fb_ptr_cmp, #14 | |
| 3532 | cmp fb_ptr_cmp, #28 | |
| 3533 | bls 4f | |
| 3534 | ||
| e1f6de8f | 3535 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 3536 | vst1.u16 { fb_pixels }, [fb_ptr] | |
| 75e28f62 E |
3537 | |
| 3538 | 3: | |
| 3539 | subs num_blocks, num_blocks, #1 | |
| 3540 | bne 0b | |
| 3541 | ||
| 3542 | 1: | |
| 3543 | vbif.u16 fb_pixels_next, pixels, draw_mask | |
| e1f6de8f | 3544 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
| 75e28f62 | 3545 | |
| 3f0189c6 | 3546 | restore_abi_regs() |
| 75e28f62 E |
3547 | ldmia sp!, { r4, pc } |
| 3548 | ||
| 3549 | 4: | |
| e1f6de8f | 3550 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 3551 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
| 75e28f62 E |
3552 | bal 3b |
| 3553 | ||
| 3554 | ||
| 3555 | #undef draw_mask_ptr | |
| 3556 | #undef c_64 | |
| 3557 | #undef fb_ptr | |
| 3558 | #undef fb_ptr_next | |
| 3559 | #undef fb_ptr_cmp | |
| 3560 | ||
| 3561 | #define psx_gpu r0 | |
| 3562 | #define num_blocks r1 | |
| 3563 | #define msb_mask_ptr r2 | |
| 3564 | #define pixel_ptr r3 | |
| 3565 | #define draw_mask_ptr r0 | |
| 3566 | #define c_64 r2 | |
| 3567 | #define fb_ptr r12 | |
| 3568 | #define fb_ptr_next r14 | |
| 3569 | #define fb_ptr_cmp r4 | |
| 3570 | ||
| 3571 | #undef msb_mask | |
| 3572 | #undef draw_mask | |
| 3573 | #undef pixels | |
| 3574 | #undef fb_pixels | |
| 3575 | #undef d128_0x8000 | |
| 3576 | #undef msb_mask_low | |
| 3577 | #undef msb_mask_high | |
| 3578 | #undef draw_mask_next | |
| 3579 | #undef pixels_g | |
| 3580 | #undef blend_pixels | |
| 3581 | #undef fb_pixels_next | |
| 3582 | ||
| 3583 | #define msb_mask q0 | |
| 3584 | #define draw_mask q1 | |
| 3585 | #define pixels q2 | |
| 3586 | #define fb_pixels q3 | |
| 3587 | #define blend_pixels q4 | |
| 3588 | #define pixels_no_msb q5 | |
| 3589 | #define blend_mask q6 | |
| 3590 | #define fb_pixels_no_msb q7 | |
| 3591 | #define d128_0x8000 q8 | |
| 3592 | #define d128_0x0421 q9 | |
| 3593 | #define fb_pixels_next q10 | |
| 3594 | #define blend_pixels_next q11 | |
| 3595 | #define pixels_next q12 | |
| 3596 | #define draw_mask_next q13 | |
| 3597 | #define write_mask q14 | |
| 3598 | ||
| 3599 | #define pixels_rb q5 | |
| 3600 | #define pixels_mg q7 | |
| 3601 | #define pixels_g q7 | |
| 3602 | #define d128_0x7C1F q8 | |
| 3603 | #define d128_0x03E0 q9 | |
| 3604 | #define fb_pixels_rb q10 | |
| 3605 | #define fb_pixels_g q11 | |
| 3606 | #define fb_pixels_masked q11 | |
| 3607 | #define d128_0x83E0 q15 | |
| 3608 | #define pixels_fourth q7 | |
| 3609 | #define d128_0x1C07 q12 | |
| 3610 | #define d128_0x00E0 q13 | |
| 3611 | #define d128_0x80E0 q13 | |
| 3612 | ||
| 3613 | #define msb_mask_low d0 | |
| 3614 | #define msb_mask_high d1 | |
| 3615 | ||
| 3616 | #define blend_blocks_average_set_blend_mask_textured(source) \ | |
| 3617 | vclt.s16 blend_mask, source, #0 \ | |
| 3618 | ||
| 3619 | #define blend_blocks_average_set_stp_bit_textured() \ | |
| 3620 | vorr.u16 blend_pixels, #0x8000 \ | |
| 3621 | ||
| 3622 | #define blend_blocks_average_combine_textured(source) \ | |
| 3623 | vbif.u16 blend_pixels, source, blend_mask \ | |
| 3624 | ||
| 3625 | #define blend_blocks_average_set_blend_mask_untextured(source) \ | |
| 3626 | ||
| 3627 | #define blend_blocks_average_set_stp_bit_untextured() \ | |
| 3628 | ||
| 3629 | #define blend_blocks_average_combine_untextured(source) \ | |
| 3630 | ||
| 3631 | #define blend_blocks_average_mask_set_on() \ | |
| 3632 | vclt.s16 write_mask, fb_pixels_next, #0 \ | |
| 3633 | ||
| 3634 | #define blend_blocks_average_mask_copy_on() \ | |
| 3635 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
| 3636 | ||
| 3637 | #define blend_blocks_average_mask_copy_b_on() \ | |
| 3638 | vorr.u16 draw_mask_next, draw_mask_next, write_mask \ | |
| 3639 | ||
| 3640 | #define blend_blocks_average_mask_set_off() \ | |
| 3641 | ||
| 3642 | #define blend_blocks_average_mask_copy_off() \ | |
| 3643 | vmov draw_mask, draw_mask_next \ | |
| 3644 | ||
| 3645 | #define blend_blocks_average_mask_copy_b_off() \ | |
| 3646 | ||
| 3647 | #define blend_blocks_average_builder(texturing, mask_evaluate) \ | |
| 3648 | .align 3; \ | |
| 3649 | \ | |
| 3650 | function(blend_blocks_##texturing##_average_##mask_evaluate) \ | |
| 3651 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 3652 | save_abi_regs(); \ |
| 75e28f62 | 3653 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 3654 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
3655 | \ |
| 3656 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 3657 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
3658 | \ |
| 3659 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3660 | mov c_64, #64; \ | |
| 3661 | \ | |
| 3662 | vmov.u16 d128_0x8000, #0x8000; \ | |
| e1f6de8f | 3663 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 3664 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 75e28f62 E |
3665 | \ |
| 3666 | vmov.u16 d128_0x0421, #0x0400; \ | |
| e1f6de8f | 3667 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3668 | \ |
| 3669 | vorr.u16 d128_0x0421, #0x0021; \ | |
| e1f6de8f | 3670 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 E |
3671 | \ |
| 3672 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
| 3673 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3674 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3675 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3676 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
| 3677 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3678 | \ | |
| 3679 | subs num_blocks, num_blocks, #1; \ | |
| 3680 | beq 1f; \ | |
| 3681 | \ | |
| 3682 | 0: \ | |
| 3683 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 3684 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 E |
3685 | \ |
| 3686 | vmov pixels, pixels_next; \ | |
| e1f6de8f | 3687 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3688 | \ |
| 3689 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
| 3690 | \ | |
| 3691 | blend_blocks_average_mask_copy_##mask_evaluate(); \ | |
| e1f6de8f | 3692 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
3693 | \ |
| 3694 | blend_blocks_average_set_blend_mask_##texturing(pixels); \ | |
| 3695 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
| 3696 | vmov fb_pixels, fb_pixels_next; \ | |
| 3697 | blend_blocks_average_combine_##texturing(pixels); \ | |
| 3698 | \ | |
| 3699 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3700 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3701 | cmp fb_ptr_cmp, #28; \ | |
| 3702 | bls 2f; \ | |
| 3703 | \ | |
| e1f6de8f | 3704 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 E |
3705 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
| 3706 | \ | |
| 3707 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3708 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3709 | \ | |
| 3710 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3711 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
| 3712 | \ | |
| 3713 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3714 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3715 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
| e1f6de8f | 3716 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3717 | \ |
| 3718 | 3: \ | |
| 3719 | subs num_blocks, num_blocks, #1; \ | |
| 3720 | bne 0b; \ | |
| 3721 | \ | |
| 3722 | 1: \ | |
| 3723 | blend_blocks_average_mask_copy_b_##mask_evaluate(); \ | |
| 3724 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
| 3725 | \ | |
| 3726 | blend_blocks_average_set_blend_mask_##texturing(pixels_next); \ | |
| 3727 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
| 3728 | blend_blocks_average_combine_##texturing(pixels_next); \ | |
| 3729 | \ | |
| 3730 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3731 | vbif.u16 fb_pixels_next, blend_pixels, draw_mask_next; \ | |
| e1f6de8f | 3732 | vst1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 | 3733 | \ |
| 3f0189c6 | 3734 | restore_abi_regs(); \ |
| 75e28f62 E |
3735 | ldmia sp!, { r4, pc }; \ |
| 3736 | \ | |
| 3737 | 2: \ | |
| 3738 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3739 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
| e1f6de8f | 3740 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
| 75e28f62 | 3741 | \ |
| e1f6de8f | 3742 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
| 75e28f62 E |
3743 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
| 3744 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
| 3745 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
| 3746 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
| 3747 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
| 3748 | \ | |
| 3749 | bal 3b \ | |
| 3750 | ||
| 3751 | blend_blocks_average_builder(textured, off) | |
| 3752 | blend_blocks_average_builder(untextured, off) | |
| 3753 | blend_blocks_average_builder(textured, on) | |
| 3754 | blend_blocks_average_builder(untextured, on) | |
| 3755 | ||
| 3756 | ||
| 3757 | #define blend_blocks_add_mask_set_on() \ | |
| 3758 | vclt.s16 write_mask, fb_pixels, #0 \ | |
| 3759 | ||
| 3760 | #define blend_blocks_add_mask_copy_on() \ | |
| 3761 | vorr.u16 draw_mask, draw_mask, write_mask \ | |
| 3762 | ||
| 3763 | #define blend_blocks_add_mask_set_off() \ | |
| 3764 | ||
| 3765 | #define blend_blocks_add_mask_copy_off() \ | |
| 3766 | ||
| 3767 | ||
| 3768 | #define blend_blocks_add_textured_builder(mask_evaluate) \ | |
| 3769 | .align 3; \ | |
| 3770 | \ | |
| 3771 | function(blend_blocks_textured_add_##mask_evaluate) \ | |
| 3772 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 3773 | save_abi_regs(); \ |
| 75e28f62 | 3774 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 3775 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
3776 | \ |
| 3777 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 3778 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
3779 | \ |
| 3780 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3781 | mov c_64, #64; \ | |
| 3782 | \ | |
| 3783 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 3784 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 3785 | vmov.u16 d128_0x83E0, #0x8000; \ | |
| 3786 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 3787 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 3788 | vorr.u16 d128_0x83E0, d128_0x83E0, d128_0x03E0; \ | |
| 3789 | \ | |
| e1f6de8f | 3790 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 3791 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 3792 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 75e28f62 | 3793 | vclt.s16 blend_mask, pixels, #0; \ |
| e1f6de8f | 3794 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3795 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3796 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3797 | \ | |
| 3798 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3799 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 3800 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3801 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
| 3802 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 3803 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 3804 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3805 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 3806 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3807 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 3808 | \ | |
| 3809 | subs num_blocks, num_blocks, #1; \ | |
| 3810 | beq 1f; \ | |
| 3811 | \ | |
| 3812 | 0: \ | |
| 3813 | mov fb_ptr, fb_ptr_next; \ | |
| 3814 | \ | |
| e1f6de8f | 3815 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 3816 | \ |
| e1f6de8f | 3817 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3818 | vclt.s16 blend_mask, pixels, #0; \ |
| 3819 | \ | |
| 3820 | vorr.u16 pixels, pixels, msb_mask; \ | |
| 3821 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3822 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
| 3823 | \ | |
| 8438c3c7 | 3824 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ |
| e1f6de8f | 3825 | pld [fb_ptr_next, #64]; \ |
| 75e28f62 E |
3826 | \ |
| 3827 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 8438c3c7 | 3828 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
| 3829 | \ | |
| 75e28f62 | 3830 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ |
| e1f6de8f | 3831 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 8438c3c7 | 3832 | \ |
| 75e28f62 E |
3833 | cmp fb_ptr_cmp, #28; \ |
| 3834 | bls 2f; \ | |
| 3835 | \ | |
| e1f6de8f | 3836 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3837 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3838 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3839 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3840 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3841 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| e1f6de8f | 3842 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3843 | \ |
| 3844 | 3: \ | |
| 3845 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
| 3846 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3847 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
| 3848 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3849 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
| 3850 | \ | |
| 3851 | subs num_blocks, num_blocks, #1; \ | |
| 3852 | bne 0b; \ | |
| 3853 | \ | |
| 3854 | 1: \ | |
| 3855 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3856 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 3857 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 3858 | \ |
| 3f0189c6 | 3859 | restore_abi_regs(); \ |
| 75e28f62 E |
3860 | ldmia sp!, { r4, pc }; \ |
| 3861 | \ | |
| 3862 | 2: \ | |
| e1f6de8f | 3863 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3864 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
| 3865 | \ | |
| e1f6de8f | 3866 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3867 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3868 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
| 3869 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3870 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
| 3871 | bal 3b \ | |
| 3872 | ||
| 3873 | ||
| 3874 | #define blend_blocks_add_untextured_builder(mask_evaluate) \ | |
| 3875 | .align 3; \ | |
| 3876 | \ | |
| 3877 | function(blend_blocks_untextured_add_##mask_evaluate) \ | |
| 3878 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 3879 | save_abi_regs(); \ |
| 75e28f62 | 3880 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 3881 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
3882 | \ |
| 3883 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 3884 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
3885 | \ |
| 3886 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 3887 | mov c_64, #64; \ | |
| 3888 | \ | |
| 3889 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 3890 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 3891 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 3892 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 3893 | \ | |
| e1f6de8f | 3894 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 3895 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 3896 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 3897 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
| 75e28f62 E |
3898 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3899 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3900 | \ | |
| 3901 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3902 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
| 3903 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3904 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3905 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3906 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3907 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3908 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 3909 | \ | |
| 3910 | subs num_blocks, num_blocks, #1; \ | |
| 3911 | beq 1f; \ | |
| 3912 | \ | |
| 3913 | 0: \ | |
| 3914 | mov fb_ptr, fb_ptr_next; \ | |
| 3915 | \ | |
| e1f6de8f | 3916 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 3917 | \ |
| e1f6de8f | 3918 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
3919 | \ |
| 3920 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3921 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3922 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
| 3923 | \ | |
| 3924 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 3925 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
3926 | \ |
| 3927 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 3928 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 3929 | cmp fb_ptr_cmp, #28; \ | |
| 3930 | bls 2f; \ | |
| 3931 | \ | |
| e1f6de8f | 3932 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3933 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3934 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3935 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
| 3936 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| e1f6de8f | 3937 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3938 | \ |
| 3939 | 3: \ | |
| 3940 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 3941 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 3942 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 3943 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 3944 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 3945 | \ | |
| 3946 | subs num_blocks, num_blocks, #1; \ | |
| 3947 | bne 0b; \ | |
| 3948 | \ | |
| 3949 | 1: \ | |
| 3950 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 3951 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 3952 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 3953 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 3954 | \ |
| 3f0189c6 | 3955 | restore_abi_regs(); \ |
| 75e28f62 E |
3956 | ldmia sp!, { r4, pc }; \ |
| 3957 | \ | |
| 3958 | 2: \ | |
| e1f6de8f | 3959 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
3960 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
| 3961 | \ | |
| e1f6de8f | 3962 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
3963 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 3964 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 3965 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 3966 | bal 3b \ | |
| 3967 | ||
| 3968 | ||
| 3969 | blend_blocks_add_textured_builder(off) | |
| 3970 | blend_blocks_add_textured_builder(on) | |
| 3971 | blend_blocks_add_untextured_builder(off) | |
| 3972 | blend_blocks_add_untextured_builder(on) | |
| 3973 | ||
| 3974 | #define blend_blocks_subtract_set_blend_mask_textured() \ | |
| 3975 | vclt.s16 blend_mask, pixels_next, #0 \ | |
| 3976 | ||
| 3977 | #define blend_blocks_subtract_combine_textured() \ | |
| 3978 | vbif.u16 blend_pixels, pixels, blend_mask \ | |
| 3979 | ||
| 718a9e58 | 3980 | #define blend_blocks_subtract_set_stp_textured() \ |
| 75e28f62 E |
3981 | vorr.u16 blend_pixels, #0x8000 \ |
| 3982 | ||
| 3983 | #define blend_blocks_subtract_msb_mask_textured() \ | |
| 3984 | vorr.u16 pixels, pixels_next, msb_mask \ | |
| 3985 | ||
| 3986 | #define blend_blocks_subtract_set_blend_mask_untextured() \ | |
| 3987 | ||
| 3988 | #define blend_blocks_subtract_combine_untextured() \ | |
| 3989 | ||
| 718a9e58 | 3990 | #define blend_blocks_subtract_set_stp_untextured() \ |
| 75e28f62 E |
3991 | vorr.u16 blend_pixels, blend_pixels, msb_mask \ |
| 3992 | ||
| 3993 | #define blend_blocks_subtract_msb_mask_untextured() \ | |
| 3994 | ||
| 3995 | ||
| 3996 | #define blend_blocks_subtract_mask_set_on() \ | |
| 3997 | vclt.s16 write_mask, fb_pixels, #0 \ | |
| 3998 | ||
| 3999 | #define blend_blocks_subtract_mask_copy_on() \ | |
| 4000 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
| 4001 | ||
| 4002 | #define blend_blocks_subtract_mask_set_off() \ | |
| 4003 | ||
| 4004 | #define blend_blocks_subtract_mask_copy_off() \ | |
| 4005 | vmov draw_mask, draw_mask_next \ | |
| 4006 | ||
| 4007 | ||
| 4008 | #define blend_blocks_subtract_builder(texturing, mask_evaluate) \ | |
| 4009 | .align 3; \ | |
| 4010 | \ | |
| 4011 | function(blend_blocks_##texturing##_subtract_##mask_evaluate) \ | |
| 4012 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 4013 | save_abi_regs(); \ |
| 75e28f62 | 4014 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 4015 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
4016 | \ |
| 4017 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 4018 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
4019 | \ |
| 4020 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4021 | mov c_64, #64; \ | |
| 4022 | \ | |
| 4023 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4024 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 4025 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 4026 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 4027 | \ | |
| e1f6de8f | 4028 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 4029 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 4030 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ | |
| 75e28f62 | 4031 | blend_blocks_subtract_set_blend_mask_##texturing(); \ |
| e1f6de8f | 4032 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4033 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
| 4034 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
| 4035 | \ | |
| 4036 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
| 4037 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4038 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4039 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4040 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4041 | \ | |
| 4042 | subs num_blocks, num_blocks, #1; \ | |
| 4043 | beq 1f; \ | |
| 4044 | \ | |
| 4045 | 0: \ | |
| 4046 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
| 4047 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 4048 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 4049 | \ |
| e1f6de8f | 4050 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
4051 | blend_blocks_subtract_msb_mask_##texturing(); \ |
| 4052 | \ | |
| e1f6de8f | 4053 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
4054 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
| 4055 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
| 718a9e58 | 4056 | blend_blocks_subtract_set_stp_##texturing(); \ |
| 75e28f62 E |
4057 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ |
| 4058 | blend_blocks_subtract_combine_##texturing(); \ | |
| 4059 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
| 4060 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| 4061 | \ | |
| 4062 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4063 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4064 | cmp fb_ptr_cmp, #28; \ | |
| 4065 | bls 2f; \ | |
| 4066 | \ | |
| e1f6de8f | 4067 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4068 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
| 4069 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4070 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4071 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| e1f6de8f | 4072 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4073 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
| 4074 | \ | |
| 4075 | 3: \ | |
| 4076 | subs num_blocks, num_blocks, #1; \ | |
| 4077 | bne 0b; \ | |
| 4078 | \ | |
| 4079 | 1: \ | |
| 4080 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
| 4081 | \ | |
| 4082 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
| 4083 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 718a9e58 | 4084 | blend_blocks_subtract_set_stp_##texturing(); \ |
| 75e28f62 E |
4085 | blend_blocks_subtract_combine_##texturing(); \ |
| 4086 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4087 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4088 | \ |
| 3f0189c6 | 4089 | restore_abi_regs(); \ |
| 75e28f62 E |
4090 | ldmia sp!, { r4, pc }; \ |
| 4091 | \ | |
| 4092 | 2: \ | |
| e1f6de8f | 4093 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 4094 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
| 75e28f62 E |
4095 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
| 4096 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4097 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4098 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4099 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4100 | bal 3b \ | |
| 4101 | ||
| 4102 | ||
| 4103 | blend_blocks_subtract_builder(textured, off) | |
| 4104 | blend_blocks_subtract_builder(textured, on) | |
| 4105 | blend_blocks_subtract_builder(untextured, off) | |
| 4106 | blend_blocks_subtract_builder(untextured, on) | |
| 4107 | ||
| 4108 | ||
| 4109 | #define blend_blocks_add_fourth_textured_builder(mask_evaluate) \ | |
| 4110 | .align 3; \ | |
| 4111 | \ | |
| 4112 | function(blend_blocks_textured_add_fourth_##mask_evaluate) \ | |
| 4113 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 4114 | save_abi_regs(); \ |
| 75e28f62 | 4115 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 4116 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
4117 | \ |
| 4118 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 4119 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
4120 | \ |
| 4121 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4122 | mov c_64, #64; \ | |
| 4123 | \ | |
| 4124 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4125 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 75e28f62 | 4126 | vmov.u16 d128_0x1C07, #0x1C00; \ |
| d1c75d1e | 4127 | vmov.u16 d128_0x00E0, #0x00E0; \ |
| 75e28f62 E |
4128 | vorr.u16 d128_0x7C1F, #0x001F; \ |
| 4129 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 75e28f62 | 4130 | vorr.u16 d128_0x1C07, #0x0007; \ |
| 75e28f62 | 4131 | \ |
| e1f6de8f | 4132 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 4133 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 4134 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 75e28f62 | 4135 | vclt.s16 blend_mask, pixels, #0; \ |
| e1f6de8f | 4136 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4137 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4138 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| d1c75d1e | 4139 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
| 75e28f62 E |
4140 | \ |
| 4141 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| d1c75d1e E |
4142 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 4143 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4144 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 75e28f62 | 4145 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
| d1c75d1e | 4146 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
| 75e28f62 | 4147 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
| d1c75d1e | 4148 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
| 75e28f62 E |
4149 | \ |
| 4150 | subs num_blocks, num_blocks, #1; \ | |
| 4151 | beq 1f; \ | |
| 4152 | \ | |
| 4153 | 0: \ | |
| 4154 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 4155 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 4156 | \ |
| d1c75d1e | 4157 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
| 718a9e58 | 4158 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
| d1c75d1e E |
4159 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
| 4160 | \ | |
| e1f6de8f | 4161 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 | 4162 | vclt.s16 blend_mask, pixels, #0; \ |
| 75e28f62 | 4163 | vshr.s16 pixels_fourth, pixels, #2; \ |
| d1c75d1e | 4164 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
| 75e28f62 E |
4165 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
| 4166 | \ | |
| 4167 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4168 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
4169 | \ |
| 4170 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4171 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4172 | cmp fb_ptr_cmp, #28; \ | |
| 4173 | bls 2f; \ | |
| 4174 | \ | |
| e1f6de8f | 4175 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4176 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 75e28f62 | 4177 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
| d1c75d1e E |
4178 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 4179 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| e1f6de8f | 4180 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4181 | \ |
| 4182 | 3: \ | |
| d1c75d1e | 4183 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ |
| 75e28f62 | 4184 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
| d1c75d1e | 4185 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
| 75e28f62 | 4186 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
| d1c75d1e | 4187 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
| 75e28f62 E |
4188 | \ |
| 4189 | subs num_blocks, num_blocks, #1; \ | |
| 4190 | bne 0b; \ | |
| 4191 | \ | |
| 4192 | 1: \ | |
| 4193 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 718a9e58 | 4194 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
| d1c75d1e | 4195 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
| 718a9e58 | 4196 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
| 75e28f62 | 4197 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
| e1f6de8f | 4198 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4199 | \ |
| 3f0189c6 | 4200 | restore_abi_regs(); \ |
| 75e28f62 E |
4201 | ldmia sp!, { r4, pc }; \ |
| 4202 | \ | |
| 4203 | 2: \ | |
| e1f6de8f | 4204 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| d1c75d1e | 4205 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 75e28f62 | 4206 | \ |
| e1f6de8f | 4207 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4208 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 75e28f62 | 4209 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
| d1c75d1e | 4210 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ |
| 75e28f62 E |
4211 | bal 3b \ |
| 4212 | ||
| 4213 | ||
| d1c75d1e | 4214 | |
| 75e28f62 E |
4215 | #define blend_blocks_add_fourth_untextured_builder(mask_evaluate) \ |
| 4216 | .align 3; \ | |
| 4217 | \ | |
| 4218 | function(blend_blocks_untextured_add_fourth_##mask_evaluate) \ | |
| 4219 | stmdb sp!, { r4, r14 }; \ | |
| 3f0189c6 | 4220 | save_abi_regs(); \ |
| 75e28f62 | 4221 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
| e1f6de8f | 4222 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
| 75e28f62 E |
4223 | \ |
| 4224 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
| e1f6de8f | 4225 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
| 75e28f62 E |
4226 | \ |
| 4227 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
| 4228 | mov c_64, #64; \ | |
| 4229 | \ | |
| 4230 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
| 4231 | vmov.u16 d128_0x03E0, #0x0300; \ | |
| 75e28f62 E |
4232 | vmov.u16 d128_0x1C07, #0x1C00; \ |
| 4233 | vmov.u16 d128_0x00E0, #0x00E0; \ | |
| 4234 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
| 4235 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
| 75e28f62 E |
4236 | vorr.u16 d128_0x1C07, #0x0007; \ |
| 4237 | \ | |
| e1f6de8f | 4238 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 4239 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
| 4240 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
| 4241 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
| 75e28f62 E |
4242 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4243 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4244 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4245 | \ | |
| 4246 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4247 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
| 4248 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4249 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4250 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4251 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4252 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4253 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 4254 | \ | |
| 4255 | subs num_blocks, num_blocks, #1; \ | |
| 4256 | beq 1f; \ | |
| 4257 | \ | |
| 4258 | 0: \ | |
| 4259 | mov fb_ptr, fb_ptr_next; \ | |
| e1f6de8f | 4260 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
| 75e28f62 | 4261 | \ |
| e1f6de8f | 4262 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
| 75e28f62 E |
4263 | \ |
| 4264 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4265 | vshr.s16 pixels_fourth, pixels, #2; \ | |
| 4266 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 4267 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
| 4268 | \ | |
| 4269 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4270 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
| 75e28f62 E |
4271 | \ |
| 4272 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
| 4273 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
| 4274 | cmp fb_ptr_cmp, #28; \ | |
| 4275 | bls 2f; \ | |
| 4276 | \ | |
| e1f6de8f | 4277 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4278 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4279 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4280 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
| 4281 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| e1f6de8f | 4282 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4283 | \ |
| 4284 | 3: \ | |
| 4285 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
| 4286 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
| 4287 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
| 4288 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
| 4289 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
| 4290 | \ | |
| 4291 | subs num_blocks, num_blocks, #1; \ | |
| 4292 | bne 0b; \ | |
| 4293 | \ | |
| 4294 | 1: \ | |
| 4295 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
| 4296 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
| 4297 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
| e1f6de8f | 4298 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
| 75e28f62 | 4299 | \ |
| 3f0189c6 | 4300 | restore_abi_regs(); \ |
| 75e28f62 E |
4301 | ldmia sp!, { r4, pc }; \ |
| 4302 | \ | |
| 4303 | 2: \ | |
| e1f6de8f | 4304 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
| 75e28f62 E |
4305 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
| 4306 | \ | |
| e1f6de8f | 4307 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
| 75e28f62 E |
4308 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
| 4309 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
| 4310 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
| 4311 | bal 3b \ | |
| 4312 | ||
| 4313 | ||
| 4314 | blend_blocks_add_fourth_textured_builder(off) | |
| 4315 | blend_blocks_add_fourth_textured_builder(on) | |
| 4316 | blend_blocks_add_fourth_untextured_builder(off) | |
| 4317 | blend_blocks_add_fourth_untextured_builder(on) | |
| 4318 | ||
| 4319 | // TODO: Optimize this more. Need a scene that actually uses it for | |
| 4320 | // confirmation.. | |
| 4321 | ||
| 4322 | .align 3 | |
| 4323 | ||
| 4324 | function(blend_blocks_textured_unblended_on) | |
| 4325 | stmdb sp!, { r4, r14 } | |
| 3f0189c6 | 4326 | save_abi_regs() |
| 75e28f62 | 4327 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
| e1f6de8f | 4328 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
4329 | |
| 4330 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
| e1f6de8f | 4331 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
| 75e28f62 E |
4332 | |
| 4333 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
| 4334 | mov c_64, #64 | |
| 4335 | ||
| e1f6de8f | 4336 | ldr fb_ptr, [pixel_ptr, #28] |
| 4337 | vld1.u16 { fb_pixels }, [fb_ptr] | |
| 4338 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
| 75e28f62 | 4339 | vclt.s16 write_mask, fb_pixels, #0 |
| e1f6de8f | 4340 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
| 75e28f62 E |
4341 | |
| 4342 | subs num_blocks, num_blocks, #1 | |
| 4343 | beq 1f | |
| 4344 | ||
| 4345 | 0: | |
| 134f81ec | 4346 | vorr.u16 pixels, pixels, msb_mask |
| 75e28f62 E |
4347 | vorr.u16 draw_mask, draw_mask, write_mask |
| 4348 | vbif.u16 fb_pixels, pixels, draw_mask | |
| e1f6de8f | 4349 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 | 4350 | |
| e1f6de8f | 4351 | ldr fb_ptr, [pixel_ptr, #28] |
| 4352 | vld1.u16 { fb_pixels }, [fb_ptr] | |
| 4353 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
| 75e28f62 | 4354 | vclt.s16 write_mask, fb_pixels, #0 |
| e1f6de8f | 4355 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
| 75e28f62 E |
4356 | |
| 4357 | subs num_blocks, num_blocks, #1 | |
| 4358 | bne 0b | |
| 4359 | ||
| 4360 | 1: | |
| 134f81ec | 4361 | vorr.u16 pixels, pixels, msb_mask |
| 75e28f62 E |
4362 | vorr.u16 draw_mask, draw_mask, write_mask |
| 4363 | vbif.u16 fb_pixels, pixels, draw_mask | |
| e1f6de8f | 4364 | vst1.u16 { fb_pixels }, [fb_ptr] |
| 75e28f62 | 4365 | |
| 3f0189c6 | 4366 | restore_abi_regs() |
| 75e28f62 E |
4367 | ldmia sp!, { r4, pc } |
| 4368 | ||
| 4369 | ||
| 4370 | function(blend_blocks_textured_unblended_off) | |
| 4371 | bx lr | |
| 4372 | ||
| 4373 | ||
| 4374 | function(warmup) | |
| 4375 | mov r3, #64 | |
| 4376 | cmp r0, #0 | |
| 4377 | bxeq lr | |
| 4378 | ||
| 4379 | 0: | |
| e1f6de8f | 4380 | vld1.u32 { u_whole_8, v_whole_8 }, [r1, :128], r3 |
| 75e28f62 E |
4381 | |
| 4382 | subs r0, r0, #1 | |
| 4383 | bne 0b | |
| 4384 | ||
| 4385 | bx lr | |
| 4386 | ||
| 6c4a10c4 | 4387 | #undef vram_ptr |
| 75e28f62 | 4388 | #undef color |
| 6c4a10c4 | 4389 | #undef width |
| 75e28f62 | 4390 | #undef height |
| 6c4a10c4 | 4391 | #undef pitch |
| 75e28f62 E |
4392 | |
| 4393 | #define vram_ptr r0 | |
| 6c4a10c4 E |
4394 | #define color r1 |
| 4395 | #define width r2 | |
| 4396 | #define height r3 | |
| 75e28f62 | 4397 | |
| 6c4a10c4 | 4398 | #define pitch r1 |
| 75e28f62 | 4399 | |
| 6c4a10c4 | 4400 | #define num_width r12 |
| 75e28f62 | 4401 | |
| 87c45ad1 E |
4402 | #undef colors_a |
| 4403 | #undef colors_b | |
| 75e28f62 | 4404 | |
| 87c45ad1 E |
4405 | #define colors_a q0 |
| 4406 | #define colors_b q1 | |
| 75e28f62 E |
4407 | |
| 4408 | .align 3 | |
| 4409 | ||
| 4410 | function(render_block_fill_body) | |
| 87c45ad1 | 4411 | vdup.u16 colors_a, color |
| 6c4a10c4 | 4412 | mov pitch, #2048 |
| 75e28f62 | 4413 | |
| 87c45ad1 | 4414 | vmov colors_b, colors_a |
| 75e28f62 | 4415 | sub pitch, pitch, width, lsl #1 |
| 75e28f62 | 4416 | |
| 6c4a10c4 | 4417 | mov num_width, width |
| 75e28f62 | 4418 | |
| 6c4a10c4 | 4419 | 0: |
| e1f6de8f | 4420 | vst1.u32 { colors_a, colors_b }, [vram_ptr, :256]! |
| 75e28f62 | 4421 | |
| d1c75d1e | 4422 | subs num_width, num_width, #16 |
| 6c4a10c4 | 4423 | bne 0b |
| 75e28f62 | 4424 | |
| 75e28f62 | 4425 | add vram_ptr, vram_ptr, pitch |
| 6c4a10c4 E |
4426 | mov num_width, width |
| 4427 | ||
| 75e28f62 E |
4428 | subs height, height, #1 |
| 4429 | bne 0b | |
| 75e28f62 | 4430 | |
| 6c4a10c4 E |
4431 | bx lr |
| 4432 | ||
| 75e28f62 E |
4433 | |
| 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 E |
4525 | |
| 4526 | ||
| 59d15d23 | 4527 | setup_sprite_flush_blocks: |
| 4528 | vpush { q1 - q5 } | |
| 75e28f62 | 4529 | |
| 4d646738 | 4530 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
| 75e28f62 | 4531 | bl flush_render_block_buffer |
| 4d646738 | 4532 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
| 75e28f62 | 4533 | |
| 59d15d23 | 4534 | vpop { q1 - q5 } |
| 75e28f62 E |
4535 | |
| 4536 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 75e28f62 E |
4537 | bx lr |
| 4538 | ||
| 4539 | ||
| 4540 | setup_sprite_update_texture_4bpp_cache: | |
| 4541 | stmdb sp!, { r0 - r3, r14 } | |
| 4542 | bl update_texture_4bpp_cache | |
| 4543 | ldmia sp!, { r0 - r3, pc } | |
| 4544 | ||
| 4545 | ||
| 4546 | setup_sprite_update_texture_8bpp_cache: | |
| 4d646738 | 4547 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r14 } |
| 75e28f62 | 4548 | bl update_texture_8bpp_cache |
| 4d646738 | 4549 | ldmia sp!, { r0 - r3, 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); \ |
| 3f0189c6 | 4836 | restore_abi_regs(); \ |
| 75e28f62 E |
4837 | ldmia sp!, { r4 - r11, pc } \ |
| 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);\ |
| 3f0189c6 | 4873 | restore_abi_regs(); \ |
| 75e28f62 E |
4874 | ldmia sp!, { r4 - r11, pc } \ |
| 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) \ |
| 75e28f62 | 5220 | stmdb sp!, { r4 - r11, r14 }; \ |
| 59d15d23 | 5221 | setup_sprite_tiled_initialize_##texture_mode##x4mode(); \ |
| 75e28f62 | 5222 | \ |
| e1f6de8f | 5223 | ldr v, [sp, #36]; \ |
| 75e28f62 E |
5224 | and offset_u, u, #0xF; \ |
| 5225 | \ | |
| e1f6de8f | 5226 | ldr width, [sp, #40]; \ |
| 5227 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
| 75e28f62 | 5228 | \ |
| e1f6de8f | 5229 | ldr height, [sp, #44]; \ |
| 75e28f62 E |
5230 | add fb_ptr, fb_ptr, y, lsl #11; \ |
| 5231 | \ | |
| 3f0189c6 | 5232 | save_abi_regs(); \ |
| 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) | |
| 5356 | stmdb sp!, { r4 - r11, r14 } | |
| 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 | |
| 5409 | ||
| 5410 | ldmia sp!, { r4 - r11, pc } | |
| 5411 | ||
| 5412 | ||
| 5413 | #undef width_rounded | |
| 5414 | #undef texture_mask | |
| 5415 | #undef num_blocks | |
| 5416 | #undef texture_offset | |
| 59d15d23 | 5417 | #undef texels_low |
| 5418 | #undef texels_high | |
| 5419 | #undef texels_wide_low | |
| 5420 | #undef texels_wide_high | |
| 5421 | #undef texels_wide | |
| 5422 | #undef fb_ptr2 | |
| 8184d7c5 | 5423 | #undef temp |
| 75e28f62 E |
5424 | |
| 5425 | #define psx_gpu r0 | |
| 5426 | #define x r1 | |
| 5427 | #define y r2 | |
| 5428 | #define u r3 | |
| 5429 | #define v r4 | |
| 5430 | #define width r5 | |
| 5431 | #define height r6 | |
| 5432 | #define left_offset r8 | |
| 5433 | #define width_rounded r9 | |
| 5434 | #define right_width r10 | |
| 59d15d23 | 5435 | |
| 75e28f62 E |
5436 | #define block_width r11 |
| 5437 | ||
| 5438 | #define texture_offset_base r1 | |
| 5439 | #define texture_mask r2 | |
| 5440 | #define texture_page_ptr r3 | |
| 5441 | #define num_blocks r4 | |
| 5442 | #define block r5 | |
| 5443 | #define fb_ptr r7 | |
| 5444 | #define texture_offset r8 | |
| 5445 | #define blocks_remaining r9 | |
| 59d15d23 | 5446 | #define fb_ptr2 r10 |
| 75e28f62 E |
5447 | #define fb_ptr_pitch r12 |
| 5448 | #define texture_block_ptr r14 | |
| 5449 | ||
| 5450 | #define texture_mask_width r2 | |
| 5451 | #define texture_mask_height r3 | |
| 5452 | #define left_mask_bits r4 | |
| 5453 | #define right_mask_bits r5 | |
| 5454 | ||
| 5455 | ||
| 5456 | #undef block_masks | |
| 5457 | #undef block_masks_shifted | |
| 5458 | #undef texels | |
| 5459 | ||
| 5460 | #define block_masks d0 | |
| 5461 | #define block_masks_shifted d1 | |
| 5462 | #define draw_mask_fb_ptr d2 | |
| 5463 | #define texels q2 | |
| 5464 | ||
| 59d15d23 | 5465 | #define draw_mask_fb_ptr_a d2 |
| 5466 | #define draw_mask_fb_ptr_b d3 | |
| 5467 | #define texels_low d4 | |
| 5468 | #define texels_high d5 | |
| 5469 | #define texels_wide_low d6 | |
| 5470 | #define texels_wide_high d7 | |
| 5471 | #define texels_wide q3 | |
| 75e28f62 | 5472 | |
| 75e28f62 | 5473 | |
| 59d15d23 | 5474 | setup_sprites_16bpp_flush: |
| 5475 | vpush { d0 - d3 } | |
| 75e28f62 | 5476 | |
| 4d646738 | 5477 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
| 75e28f62 | 5478 | bl flush_render_block_buffer |
| 4d646738 | 5479 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
| 75e28f62 | 5480 | |
| 59d15d23 | 5481 | vpop { d0 - d3 } |
| 75e28f62 E |
5482 | |
| 5483 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 5484 | mov num_blocks, block_width | |
| 5485 | ||
| 5486 | bx lr | |
| 5487 | ||
| 5488 | function(setup_sprite_16bpp) | |
| 5489 | stmdb sp!, { r4 - r11, r14 } | |
| e1f6de8f | 5490 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 75e28f62 | 5491 | |
| e1f6de8f | 5492 | ldr v, [sp, #36] |
| 75e28f62 E |
5493 | add fb_ptr, fb_ptr, y, lsl #11 |
| 5494 | ||
| e1f6de8f | 5495 | ldr width, [sp, #40] |
| 75e28f62 E |
5496 | add fb_ptr, fb_ptr, x, lsl #1 |
| 5497 | ||
| e1f6de8f | 5498 | ldr height, [sp, #44] |
| 75e28f62 E |
5499 | and left_offset, u, #0x7 |
| 5500 | ||
| 5501 | add texture_offset_base, u, u | |
| 5502 | add width_rounded, width, #7 | |
| 5503 | ||
| ed0fd81d | 5504 | add texture_offset_base, texture_offset_base, v, lsl #11 |
| 75e28f62 E |
5505 | mov left_mask_bits, #0xFF |
| 5506 | ||
| e1f6de8f | 5507 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
| 75e28f62 E |
5508 | add width_rounded, width_rounded, left_offset |
| 5509 | ||
| e1f6de8f | 5510 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
| 75e28f62 E |
5511 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
| 5512 | ||
| 5513 | add texture_mask, texture_mask_width, texture_mask_width | |
| 5514 | mov right_mask_bits, #0xFE | |
| 5515 | ||
| 5516 | and right_width, width_rounded, #0x7 | |
| 5517 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
| 5518 | ||
| ed0fd81d | 5519 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
| 75e28f62 E |
5520 | mov block_width, width_rounded, lsr #3 |
| 5521 | ||
| 5522 | mov right_mask_bits, right_mask_bits, lsl right_width | |
| 5523 | movw fb_ptr_pitch, #(2048 + 16) | |
| 5524 | ||
| 5525 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4 | |
| 5526 | vmov block_masks, left_mask_bits, right_mask_bits | |
| 5527 | ||
| e1f6de8f | 5528 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
5529 | add block, psx_gpu, #psx_gpu_blocks_offset |
| 5530 | ||
| 6ea0f7bf | 5531 | bic texture_offset_base, texture_offset_base, #0xF |
| 75e28f62 E |
5532 | cmp block_width, #1 |
| 5533 | ||
| e1f6de8f | 5534 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 75e28f62 E |
5535 | add block, block, num_blocks, lsl #6 |
| 5536 | ||
| 5537 | bne 0f | |
| 5538 | ||
| 5539 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
| 5540 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
| 5541 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
| 5542 | ||
| 5543 | 1: | |
| 5544 | add num_blocks, num_blocks, #1 | |
| 5545 | cmp num_blocks, #MAX_BLOCKS | |
| 59d15d23 | 5546 | blgt setup_sprites_16bpp_flush |
| 75e28f62 E |
5547 | |
| 5548 | and texture_block_ptr, texture_offset_base, texture_mask | |
| 5549 | subs height, height, #1 | |
| 5550 | ||
| 5551 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5552 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 | 5553 | |
| e1f6de8f | 5554 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5555 | add block, block, #40 |
| 5556 | ||
| 5557 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5558 | pld [fb_ptr] |
| 75e28f62 | 5559 | |
| e1f6de8f | 5560 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5561 | |
| 5562 | add block, block, #24 | |
| 5563 | add texture_offset_base, texture_offset_base, #2048 | |
| 5564 | add fb_ptr, fb_ptr, #2048 | |
| e1f6de8f | 5565 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
5566 | bne 1b |
| 5567 | ||
| 5568 | ldmia sp!, { r4 - r11, pc } | |
| 5569 | ||
| 5570 | 0: | |
| 5571 | add num_blocks, num_blocks, block_width | |
| 5572 | mov texture_offset, texture_offset_base | |
| 5573 | ||
| 5574 | cmp num_blocks, #MAX_BLOCKS | |
| 59d15d23 | 5575 | blgt setup_sprites_16bpp_flush |
| 75e28f62 E |
5576 | |
| 5577 | add texture_offset_base, texture_offset_base, #2048 | |
| 5578 | and texture_block_ptr, texture_offset, texture_mask | |
| 5579 | ||
| 5580 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5581 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 | 5582 | |
| e1f6de8f | 5583 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5584 | add block, block, #40 |
| 5585 | ||
| 5586 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
| 5587 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5588 | pld [fb_ptr] |
| 75e28f62 | 5589 | |
| e1f6de8f | 5590 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5591 | subs blocks_remaining, block_width, #2 |
| 5592 | ||
| 5593 | add texture_offset, texture_offset, #16 | |
| 5594 | add fb_ptr, fb_ptr, #16 | |
| 5595 | ||
| 5596 | vmov.u8 draw_mask_fb_ptr, #0 | |
| 5597 | ||
| 5598 | add block, block, #24 | |
| 5599 | beq 2f | |
| 5600 | ||
| 5601 | 1: | |
| 5602 | and texture_block_ptr, texture_offset, texture_mask | |
| 5603 | subs blocks_remaining, blocks_remaining, #1 | |
| 5604 | ||
| 5605 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5606 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 | 5607 | |
| e1f6de8f | 5608 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5609 | add block, block, #40 |
| 5610 | ||
| 5611 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5612 | pld [fb_ptr] |
| 75e28f62 | 5613 | |
| e1f6de8f | 5614 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5615 | |
| 5616 | add texture_offset, texture_offset, #16 | |
| 5617 | add fb_ptr, fb_ptr, #16 | |
| 5618 | ||
| 5619 | add block, block, #24 | |
| 5620 | bne 1b | |
| 5621 | ||
| 5622 | 2: | |
| 5623 | and texture_block_ptr, texture_offset, texture_mask | |
| 5624 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5625 | ||
| e1f6de8f | 5626 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 75e28f62 E |
5627 | vdup.u8 draw_mask_fb_ptr, block_masks[4] |
| 5628 | ||
| e1f6de8f | 5629 | vst1.u32 { texels }, [block, :128] |
| 75e28f62 E |
5630 | add block, block, #40 |
| 5631 | ||
| 5632 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5633 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
| 75e28f62 E |
5634 | |
| 5635 | add block, block, #24 | |
| 5636 | subs height, height, #1 | |
| 5637 | ||
| 5638 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
| e1f6de8f | 5639 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 75e28f62 E |
5640 | |
| 5641 | bne 0b | |
| 5642 | ||
| 5643 | ldmia sp!, { r4 - r11, pc } | |
| 5644 | ||
| 5645 | ||
| 59d15d23 | 5646 | // 4x version |
| 5647 | // FIXME: duplicate code with normal version :( | |
| 5648 | #undef draw_mask_fb_ptr | |
| 5649 | ||
| 5650 | function(setup_sprite_16bpp_4x) | |
| 5651 | stmdb sp!, { r4 - r11, r14 } | |
| e1f6de8f | 5652 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
| 59d15d23 | 5653 | |
| e1f6de8f | 5654 | ldr v, [sp, #36] |
| 59d15d23 | 5655 | add fb_ptr, fb_ptr, y, lsl #11 |
| 5656 | ||
| e1f6de8f | 5657 | ldr width, [sp, #40] |
| 59d15d23 | 5658 | add fb_ptr, fb_ptr, x, lsl #1 |
| 5659 | ||
| e1f6de8f | 5660 | ldr height, [sp, #44] |
| 59d15d23 | 5661 | and left_offset, u, #0x7 |
| 5662 | ||
| 5663 | add texture_offset_base, u, u | |
| 5664 | add width_rounded, width, #7 | |
| 5665 | ||
| ed0fd81d | 5666 | add texture_offset_base, texture_offset_base, v, lsl #11 |
| 59d15d23 | 5667 | movw left_mask_bits, #0xFFFF |
| 5668 | ||
| e1f6de8f | 5669 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
| 59d15d23 | 5670 | add width_rounded, width_rounded, left_offset |
| 5671 | ||
| 5672 | lsl left_offset, #1 | |
| 5673 | ||
| e1f6de8f | 5674 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
| 59d15d23 | 5675 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
| 5676 | ||
| 5677 | add texture_mask, texture_mask_width, texture_mask_width | |
| 5678 | movw right_mask_bits, #0xFFFC | |
| 5679 | ||
| 5680 | and right_width, width_rounded, #0x7 | |
| 5681 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
| 5682 | ||
| 5683 | lsl right_width, #1 | |
| 5684 | ||
| ed0fd81d | 5685 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
| 59d15d23 | 5686 | mov block_width, width_rounded, lsr #3 |
| 5687 | ||
| 5688 | mov right_mask_bits, right_mask_bits, lsl right_width | |
| 5689 | movw fb_ptr_pitch, #(2048 + 16) * 2 | |
| 5690 | ||
| 5691 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4+1 | |
| 5692 | vmov block_masks, left_mask_bits, right_mask_bits | |
| 5693 | ||
| e1f6de8f | 5694 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 59d15d23 | 5695 | add block, psx_gpu, #psx_gpu_blocks_offset |
| 5696 | ||
| 5697 | bic texture_offset_base, texture_offset_base, #0xF | |
| 5698 | cmp block_width, #1 | |
| 5699 | ||
| e1f6de8f | 5700 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
| 59d15d23 | 5701 | add block, block, num_blocks, lsl #6 |
| 5702 | ||
| 5703 | lsl block_width, #2 | |
| 5704 | bne 0f | |
| 5705 | ||
| 5706 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
| 5707 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
| 5708 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] | |
| 5709 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
| 5710 | ||
| 5711 | 1: | |
| 5712 | add num_blocks, num_blocks, block_width | |
| 5713 | cmp num_blocks, #MAX_BLOCKS | |
| 5714 | blgt setup_sprites_16bpp_flush | |
| 5715 | ||
| 5716 | and texture_block_ptr, texture_offset_base, texture_mask | |
| 5717 | subs height, height, #1 | |
| 5718 | ||
| 5719 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5720 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5721 | |
| 5722 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5723 | ||
| 5724 | add texture_offset_base, texture_offset_base, #2048 | |
| 5725 | add fb_ptr, fb_ptr, #2048*2 | |
| e1f6de8f | 5726 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 59d15d23 | 5727 | bne 1b |
| 5728 | ||
| 5729 | ldmia sp!, { r4 - r11, pc } | |
| 5730 | ||
| 5731 | 0: | |
| 5732 | add num_blocks, num_blocks, block_width | |
| 5733 | mov texture_offset, texture_offset_base | |
| 5734 | ||
| 5735 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] // left_mask_bits | |
| 5736 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
| 5737 | ||
| 5738 | cmp num_blocks, #MAX_BLOCKS | |
| 5739 | blgt setup_sprites_16bpp_flush | |
| 5740 | ||
| 5741 | add texture_offset_base, texture_offset_base, #2048 | |
| 5742 | and texture_block_ptr, texture_offset, texture_mask | |
| 5743 | ||
| 5744 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5745 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5746 | |
| 5747 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5748 | ||
| 5749 | subs blocks_remaining, block_width, #2*4 | |
| 5750 | add texture_offset, texture_offset, #16 | |
| 5751 | ||
| 5752 | vmov.u8 draw_mask_fb_ptr_a, #0 | |
| 5753 | vmov.u8 draw_mask_fb_ptr_b, #0 | |
| 5754 | ||
| 5755 | add fb_ptr, fb_ptr, #16*2 | |
| 5756 | beq 2f | |
| 5757 | ||
| 5758 | 1: | |
| 5759 | and texture_block_ptr, texture_offset, texture_mask | |
| 5760 | subs blocks_remaining, blocks_remaining, #4 | |
| 5761 | ||
| 5762 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| e1f6de8f | 5763 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5764 | |
| 5765 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5766 | add texture_offset, texture_offset, #16 | |
| 5767 | ||
| 5768 | add fb_ptr, fb_ptr, #16*2 | |
| 5769 | bgt 1b | |
| 5770 | ||
| 5771 | 2: | |
| 5772 | vdup.u8 draw_mask_fb_ptr_a, block_masks[4] // right_mask_bits | |
| 5773 | vdup.u8 draw_mask_fb_ptr_b, block_masks[5] | |
| 5774 | ||
| 5775 | and texture_block_ptr, texture_offset, texture_mask | |
| 5776 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
| 5777 | ||
| e1f6de8f | 5778 | vld1.u32 { texels }, [texture_block_ptr, :128] |
| 59d15d23 | 5779 | |
| 5780 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
| 5781 | subs height, height, #1 | |
| 5782 | ||
| 5783 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
| e1f6de8f | 5784 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| 59d15d23 | 5785 | |
| 5786 | bne 0b | |
| 5787 | ||
| 5788 | ldmia sp!, { r4 - r11, pc } | |
| 5789 | ||
| 5790 | ||
| f0931e56 | 5791 | #undef width |
| 5792 | #undef right_width | |
| 5793 | #undef right_mask_bits | |
| 5794 | #undef color | |
| 5795 | #undef height | |
| 5796 | #undef blocks_remaining | |
| 5797 | #undef colors | |
| 5798 | #undef right_mask | |
| 5799 | #undef test_mask | |
| 5800 | #undef draw_mask | |
| 5801 | ||
| 5802 | #define psx_gpu r0 | |
| 5803 | #define x r1 | |
| 5804 | #define y r2 | |
| 5805 | #define width r3 | |
| 5806 | #define right_width r5 | |
| 5807 | #define right_mask_bits r6 | |
| 5808 | #define fb_ptr r7 | |
| 5809 | #define color r8 | |
| 5810 | #define height r9 | |
| 5811 | #define fb_ptr_pitch r12 | |
| 5812 | ||
| 5813 | // referenced by setup_sprites_16bpp_flush | |
| 5814 | #define num_blocks r4 | |
| 5815 | #define block r5 | |
| 5816 | #define block_width r11 | |
| 5817 | ||
| 5818 | #define color_r r1 | |
| 5819 | #define color_g r2 | |
| 5820 | #define color_b r8 | |
| 5821 | #define blocks_remaining r6 | |
| 5822 | ||
| 5823 | #define colors q0 | |
| 5824 | #define right_mask q1 | |
| 5825 | #define test_mask q2 | |
| 5826 | #define draw_mask q2 | |
| 5827 | #define draw_mask_bits_fb_ptr d6 | |
| 5828 | ||
| 5829 | ||
| 5830 | .align 3 | |
| 5831 | ||
| 2d658c89 | 5832 | function(setup_sprite_untextured_512) |
| f0931e56 | 5833 | stmdb sp!, { r4 - r11, r14 } |
| 5834 | ||
| e1f6de8f | 5835 | ldr width, [sp, #40] |
| 5836 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] | |
| f0931e56 | 5837 | |
| e1f6de8f | 5838 | ldr height, [sp, #44] |
| f0931e56 | 5839 | add fb_ptr, fb_ptr, y, lsl #11 |
| 5840 | ||
| 5841 | add fb_ptr, fb_ptr, x, lsl #1 | |
| 5842 | sub right_width, width, #1 | |
| 5843 | ||
| e1f6de8f | 5844 | ldr color, [sp, #48] |
| f0931e56 | 5845 | and right_width, #7 |
| 5846 | ||
| 5847 | add block_width, width, #7 | |
| 5848 | add right_width, #1 | |
| 5849 | ||
| 5850 | lsr block_width, #3 | |
| 5851 | mov right_mask_bits, #0xff | |
| 5852 | ||
| 5853 | sub fb_ptr_pitch, block_width, #1 | |
| 5854 | lsl right_mask_bits, right_width | |
| 5855 | ||
| 5856 | lsl fb_ptr_pitch, #3+1 | |
| 5857 | ubfx color_r, color, #3, #5 | |
| 5858 | ||
| 5859 | rsb fb_ptr_pitch, #1024*2 | |
| 5860 | ubfx color_g, color, #11, #5 | |
| 5861 | ||
| e1f6de8f | 5862 | vld1.u32 { test_mask }, [psx_gpu, :128] |
| f0931e56 | 5863 | ubfx color_b, color, #19, #5 |
| 5864 | ||
| 5865 | vdup.u16 right_mask, right_mask_bits | |
| 5866 | orr color, color_r, color_b, lsl #10 | |
| 5867 | ||
| e1f6de8f | 5868 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| f0931e56 | 5869 | orr color, color, color_g, lsl #5 |
| 5870 | ||
| 5871 | vtst.u16 right_mask, right_mask, test_mask | |
| 5872 | add block, psx_gpu, #psx_gpu_blocks_offset | |
| 5873 | ||
| 5874 | vdup.u16 colors, color | |
| 5875 | add block, block, num_blocks, lsl #6 | |
| 5876 | ||
| 5877 | ||
| 5878 | setup_sprite_untextured_height_loop: | |
| 5879 | add num_blocks, block_width | |
| 5880 | sub blocks_remaining, block_width, #1 | |
| 5881 | ||
| 5882 | cmp num_blocks, #MAX_BLOCKS | |
| 5883 | blgt setup_sprites_16bpp_flush | |
| 5884 | ||
| 5885 | cmp blocks_remaining, #0 | |
| 5886 | ble 1f | |
| 5887 | ||
| 5888 | vmov.u8 draw_mask, #0 /* zero_mask */ | |
| 5889 | vmov.u8 draw_mask_bits_fb_ptr, #0 | |
| 5890 | ||
| 5891 | 0: | |
| e1f6de8f | 5892 | vst1.u32 { draw_mask }, [block, :128]! |
| f0931e56 | 5893 | subs blocks_remaining, #1 |
| 5894 | ||
| e1f6de8f | 5895 | vst1.u32 { colors }, [block, :128] |
| f0931e56 | 5896 | add block, block, #24 |
| 5897 | ||
| 5898 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5899 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
| f0931e56 | 5900 | |
| 5901 | add block, block, #24 | |
| 5902 | add fb_ptr, #8*2 | |
| 5903 | bgt 0b | |
| 5904 | ||
| 5905 | 1: | |
| e1f6de8f | 5906 | vst1.u32 { right_mask }, [block, :128]! |
| f0931e56 | 5907 | subs height, #1 |
| 5908 | ||
| e1f6de8f | 5909 | vst1.u32 { colors }, [block, :128] |
| f0931e56 | 5910 | add block, block, #24 |
| 5911 | ||
| 5912 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
| e1f6de8f | 5913 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
| f0931e56 | 5914 | |
| 5915 | add block, block, #24 | |
| 5916 | add fb_ptr, fb_ptr_pitch | |
| 5917 | ||
| e1f6de8f | 5918 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
| f0931e56 | 5919 | bgt setup_sprite_untextured_height_loop |
| 5920 | ||
| 5921 | ldmia sp!, { r4 - r11, pc } | |
| 5922 | ||
| 5923 | ||
| 5924 | ||
| 75e28f62 E |
5925 | #undef texture_page_ptr |
| 5926 | #undef vram_ptr | |
| 5927 | #undef dirty_textures_mask | |
| 5928 | #undef current_texture_mask | |
| 5929 | ||
| 5930 | #define psx_gpu r0 | |
| 5931 | #define current_texture_page r1 | |
| 5932 | #define texture_page_ptr r2 | |
| 5933 | #define vram_ptr_a r3 | |
| 5934 | #define current_texture_page_x r12 | |
| 5935 | #define current_texture_page_y r4 | |
| 5936 | #define dirty_textures_mask r5 | |
| 5937 | #define tile_y r6 | |
| 5938 | #define tile_x r7 | |
| 5939 | #define sub_y r8 | |
| 5940 | #define current_texture_mask r9 | |
| 5941 | #define c_4096 r10 | |
| 5942 | #define vram_ptr_b r11 | |
| 5943 | ||
| 5944 | #define texel_block_a d0 | |
| 5945 | #define texel_block_b d1 | |
| 5946 | #define texel_block_expanded_a q1 | |
| 5947 | #define texel_block_expanded_b q2 | |
| 5948 | #define texel_block_expanded_ab q2 | |
| 5949 | #define texel_block_expanded_c q3 | |
| fab27ba2 | 5950 | #define texel_block_expanded_d q0 |
| 75e28f62 E |
5951 | #define texel_block_expanded_cd q3 |
| 5952 | ||
| 5953 | function(update_texture_4bpp_cache) | |
| 5954 | stmdb sp!, { r4 - r11, r14 } | |
| 5955 | vpush { q0 - q3 } | |
| 5956 | ||
| e1f6de8f | 5957 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
| 75e28f62 | 5958 | |
| e1f6de8f | 5959 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
| 5960 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
| 75e28f62 E |
5961 | |
| 5962 | and current_texture_page_x, current_texture_page, #0xF | |
| e1f6de8f | 5963 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
| 75e28f62 E |
5964 | |
| 5965 | mov current_texture_page_y, current_texture_page, lsr #4 | |
| e1f6de8f | 5966 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
| 75e28f62 E |
5967 | |
| 5968 | add vram_ptr_a, vram_ptr_a, current_texture_page_y, lsl #19 | |
| 5969 | mov tile_y, #16 | |
| 5970 | ||
| 5971 | add vram_ptr_a, vram_ptr_a, current_texture_page_x, lsl #7 | |
| 5972 | bic dirty_textures_mask, current_texture_mask | |
| 5973 | ||
| 5974 | mov tile_x, #16 | |
| e1f6de8f | 5975 | str dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
| 75e28f62 E |
5976 | |
| 5977 | mov sub_y, #8 | |
| 5978 | movw c_4096, #4096 | |
| 5979 | ||
| 5980 | add vram_ptr_b, vram_ptr_a, #2048 | |
| 5981 | ||
| 5982 | 0: | |
| e1f6de8f | 5983 | vld1.u32 { texel_block_a }, [vram_ptr_a, :64], c_4096 |
| 5984 | vld1.u32 { texel_block_b }, [vram_ptr_b, :64], c_4096 | |
| 75e28f62 E |
5985 | |
| 5986 | vmovl.u8 texel_block_expanded_a, texel_block_a | |
| 5987 | vshll.u8 texel_block_expanded_b, texel_block_a, #4 | |
| 5988 | vmovl.u8 texel_block_expanded_c, texel_block_b | |
| 5989 | vshll.u8 texel_block_expanded_d, texel_block_b, #4 | |
| 5990 | ||
| 5991 | vbic.u16 texel_block_expanded_a, #0x00F0 | |
| 5992 | vbic.u16 texel_block_expanded_b, #0x00F0 | |
| 5993 | vbic.u16 texel_block_expanded_c, #0x00F0 | |
| 5994 | vbic.u16 texel_block_expanded_d, #0x00F0 | |
| 5995 | ||
| 5996 | vorr.u16 texel_block_expanded_ab, texel_block_expanded_a, \ | |
| 5997 | texel_block_expanded_b | |
| 5998 | vorr.u16 texel_block_expanded_cd, texel_block_expanded_c, \ | |
| 5999 | texel_block_expanded_d | |
| 6000 | ||
| 6001 | vst1.u32 { texel_block_expanded_ab, texel_block_expanded_cd }, \ | |
| e1f6de8f | 6002 | [texture_page_ptr, :256]! |
| 75e28f62 E |
6003 | |
| 6004 | subs sub_y, sub_y, #1 | |
| 6005 | bne 0b | |
| 6006 | ||
| 6007 | mov sub_y, #8 | |
| 6008 | add vram_ptr_a, vram_ptr_a, #8 | |
| 6009 | add vram_ptr_b, vram_ptr_b, #8 | |
| 6010 | ||
| 6011 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6012 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6013 | ||
| 6014 | subs tile_x, tile_x, #1 | |
| 6015 | bne 0b | |
| 6016 | ||
| 6017 | mov tile_x, #16 | |
| 6018 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6019 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6020 | ||
| 6021 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
| 6022 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
| 6023 | ||
| 6024 | subs tile_y, tile_y, #1 | |
| 6025 | bne 0b | |
| 6026 | ||
| 6027 | vpop { q0 - q3 } | |
| 6028 | ldmia sp!, { r4 - r11, pc } | |
| 6029 | ||
| 6030 | ||
| 6031 | #undef current_texture_page | |
| 6032 | ||
| 6033 | #define psx_gpu r0 | |
| 6034 | #define texture_page r1 | |
| 6035 | #define texture_page_ptr r2 | |
| 6036 | #define vram_ptr_a r3 | |
| 6037 | #define texture_page_x r12 | |
| 6038 | #define texture_page_y r4 | |
| 6039 | #define current_texture_page r5 | |
| 6040 | #define tile_y r6 | |
| 6041 | #define tile_x r7 | |
| 6042 | #define sub_y r8 | |
| 6043 | #define c_4096 r10 | |
| 6044 | #define vram_ptr_b r11 | |
| 6045 | ||
| 6046 | ||
| 6047 | #undef texels_a | |
| 6048 | #undef texels_b | |
| 6049 | ||
| 6050 | #define texels_a q0 | |
| 6051 | #define texels_b q1 | |
| 6052 | #define texels_c q2 | |
| 6053 | #define texels_d q3 | |
| 6054 | ||
| 6055 | ||
| 6056 | function(update_texture_8bpp_cache_slice) | |
| 6057 | stmdb sp!, { r4 - r11, r14 } | |
| 6058 | vpush { q0 - q3 } | |
| 6059 | ||
| e1f6de8f | 6060 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
| 6061 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
| 75e28f62 | 6062 | |
| e1f6de8f | 6063 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
| 75e28f62 E |
6064 | mov tile_y, #16 |
| 6065 | ||
| 6066 | and texture_page_x, texture_page, #0xF | |
| 6067 | mov texture_page_y, texture_page, lsr #4 | |
| 6068 | ||
| 6069 | add vram_ptr_a, vram_ptr_a, texture_page_x, lsl #7 | |
| 6070 | mov tile_x, #8 | |
| 6071 | ||
| 6072 | add vram_ptr_a, vram_ptr_a, texture_page_y, lsl #19 | |
| 6073 | eor current_texture_page, current_texture_page, texture_page | |
| 6074 | ||
| 6075 | ands current_texture_page, current_texture_page, #0x1 | |
| 6076 | mov sub_y, #4 | |
| 6077 | ||
| 6078 | addne texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
| 6079 | movw c_4096, #4096 | |
| 6080 | ||
| 6081 | add vram_ptr_b, vram_ptr_a, #2048 | |
| 6082 | ||
| 6083 | 0: | |
| e1f6de8f | 6084 | vld1.u32 { texels_a }, [vram_ptr_a, :128], c_4096 |
| 6085 | vld1.u32 { texels_b }, [vram_ptr_b, :128], c_4096 | |
| 6086 | vld1.u32 { texels_c }, [vram_ptr_a, :128], c_4096 | |
| 6087 | vld1.u32 { texels_d }, [vram_ptr_b, :128], c_4096 | |
| 75e28f62 | 6088 | |
| e1f6de8f | 6089 | vst1.u32 { texels_a, texels_b }, [texture_page_ptr, :256]! |
| 6090 | vst1.u32 { texels_c, texels_d }, [texture_page_ptr, :256]! | |
| 75e28f62 E |
6091 | |
| 6092 | subs sub_y, sub_y, #1 | |
| 6093 | bne 0b | |
| 6094 | ||
| 6095 | mov sub_y, #4 | |
| 6096 | ||
| 6097 | add vram_ptr_a, vram_ptr_a, #16 | |
| 6098 | add vram_ptr_b, vram_ptr_b, #16 | |
| 6099 | ||
| 6100 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6101 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6102 | ||
| 6103 | subs tile_x, tile_x, #1 | |
| 6104 | bne 0b | |
| 6105 | ||
| 6106 | mov tile_x, #8 | |
| 6107 | ||
| 6108 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
| 6109 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
| 6110 | ||
| 6111 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
| 6112 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
| 6113 | ||
| 6114 | subs tile_y, tile_y, #1 | |
| 6115 | add texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
| 6116 | ||
| 6117 | bne 0b | |
| 6118 | ||
| 6119 | vpop { q0 - q3 } | |
| 6120 | ldmia sp!, { r4 - r11, pc } | |
| 6121 | ||
| 50f9355a | 6122 | |
| 6123 | /* void scale2x_tiles8(void *dst, const void *src, int w8, int h) */ | |
| 6124 | function(scale2x_tiles8) | |
| 6125 | push { r4, r14 } | |
| 6126 | ||
| 6127 | mov r4, r1 | |
| 6128 | add r12, r0, #1024*2 | |
| 6129 | mov r14, r2 | |
| 6130 | ||
| 6131 | 0: | |
| c2a25f67 | 6132 | pld [r1, #1024*2] |
| e1f6de8f | 6133 | vld1.u16 { q0 }, [r1, :128]! |
| 6134 | vld1.u16 { q2 }, [r1, :128]! | |
| 50f9355a | 6135 | vmov q1, q0 |
| 6136 | vmov q3, q2 | |
| 6137 | vzip.16 q0, q1 | |
| 6138 | vzip.16 q2, q3 | |
| 6139 | subs r14, #2 | |
| e1f6de8f | 6140 | vst1.u16 { q0, q1 }, [r0, :128]! |
| 6141 | vst1.u16 { q0, q1 }, [r12, :128]! | |
| 50f9355a | 6142 | blt 1f |
| e1f6de8f | 6143 | vst1.u16 { q2, q3 }, [r0, :128]! |
| 6144 | vst1.u16 { q2, q3 }, [r12, :128]! | |
| 50f9355a | 6145 | bgt 0b |
| 6146 | 1: | |
| 6147 | subs r3, #1 | |
| 6148 | mov r14, r2 | |
| 6149 | add r0, #1024*2*2 | |
| 6150 | add r4, #1024*2 | |
| ed0fd81d | 6151 | sub r0, r0, r2, lsl #4+1 |
| 50f9355a | 6152 | mov r1, r4 |
| 6153 | add r12, r0, #1024*2 | |
| 6154 | bgt 0b | |
| 6155 | nop | |
| 6156 | ||
| 6157 | pop { r4, pc } | |
| 59d15d23 | 6158 | |
| 6159 | // vim:filetype=armasm |