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 | ||
16 | #define MAX_SPANS 512 | |
17 | #define MAX_BLOCKS 64 | |
18 | #define MAX_BLOCKS_PER_ROW 128 | |
19 | ||
f0931e56 | 20 | #define RENDER_STATE_MASK_EVALUATE 0x20 |
21 | #define RENDER_FLAGS_MODULATE_TEXELS 0x1 | |
22 | #define RENDER_FLAGS_BLEND 0x2 | |
d5c08ed3 | 23 | #define RENDER_INTERLACE_ENABLED 0x1 |
f0931e56 | 24 | |
cb88320b | 25 | #include "psx_gpu_offsets.h" |
75e28f62 | 26 | |
cb88320b | 27 | #define psx_gpu_b_dx_offset (psx_gpu_b_block_span_offset + 4) |
75e28f62 | 28 | |
75e28f62 E |
29 | #define edge_data_left_x_offset 0 |
30 | #define edge_data_num_blocks_offset 2 | |
31 | #define edge_data_right_mask_offset 4 | |
32 | #define edge_data_y_offset 6 | |
33 | ||
ed0fd81d | 34 | .syntax unified |
35 | .text | |
75e28f62 | 36 | |
3f0189c6 | 37 | #if 0 |
38 | #define save_abi_regs() \ | |
39 | vpush {q4-q7} | |
40 | #define restore_abi_regs() \ | |
41 | vpop {q4-q7} | |
42 | #else | |
43 | #define save_abi_regs() | |
44 | #define restore_abi_regs() | |
45 | #endif | |
46 | ||
75e28f62 E |
47 | #define psx_gpu r0 |
48 | #define v_a r1 | |
49 | #define v_b r2 | |
50 | #define v_c r3 | |
51 | ||
52 | #define x0 r4 | |
53 | #define x1 r5 | |
54 | #define x2 r6 | |
55 | #define x0_x1 r5 | |
56 | #define x1_x2 r6 | |
57 | #define y0 r7 | |
58 | #define y1 r8 | |
59 | #define y2 r9 | |
60 | #define y0_y1 r7 | |
61 | #define y1_y2 r8 | |
62 | #define b0 r9 | |
63 | #define b1 r10 | |
64 | #define b2 r11 | |
65 | #define b0_b1 r10 | |
66 | #define b1_b2 r11 | |
67 | ||
68 | ||
69 | #define area_r_s r5 | |
70 | ||
71 | #define g_bx0 r2 | |
72 | #define g_bx r3 | |
73 | #define g_bx2 r4 | |
74 | #define g_bx3 r5 | |
75 | #define b_base r6 | |
76 | #define g_by r8 | |
77 | ||
78 | #define gs_bx r7 | |
79 | #define gs_by r10 | |
80 | ||
81 | #define ga_bx g_bx | |
82 | #define ga_by g_by | |
83 | ||
84 | #define gw_bx_h g_bx | |
85 | #define gw_by_h g_by | |
86 | ||
87 | #define gw_bx_l r11 | |
88 | #define gw_by_l gw_bx_l | |
89 | ||
90 | #define store_a r0 | |
91 | #define store_b r1 | |
92 | #define store_inc r5 | |
93 | ||
94 | ||
95 | #define v0 q0 | |
96 | #define uvrgb0 d0 | |
97 | #define x0_y0 d1 | |
98 | ||
99 | #define v1 q1 | |
100 | #define uvrgb1 d2 | |
101 | #define x1_y1 d3 | |
102 | ||
103 | #define v2 q2 | |
104 | #define uvrgb2 d4 | |
105 | #define x2_y2 d5 | |
106 | ||
107 | #define x0_ab q3 | |
108 | #define uvrg_xxxx0 q3 | |
109 | #define uvrg0 d6 | |
110 | #define xxxx0 d7 | |
111 | ||
112 | #define x1_ab q4 | |
113 | #define uvrg_xxxx1 q4 | |
114 | #define uvrg1 d8 | |
115 | #define xxxx1 d9 | |
116 | ||
117 | #define x2_ab q5 | |
118 | #define uvrg_xxxx2 q5 | |
119 | #define uvrg2 d10 | |
120 | #define xxxx2 d11 | |
121 | ||
122 | #define y0_ab q6 | |
123 | #define yyyy_uvrg0 q6 | |
124 | #define yyyy0 d12 | |
125 | #define uvrg0b d13 | |
126 | ||
127 | #define y1_ab q7 | |
128 | #define yyyy_uvrg1 q7 | |
129 | #define yyyy1 d14 | |
130 | #define uvrg1b d15 | |
131 | ||
132 | #define y2_ab q8 | |
133 | #define yyyy_uvrg2 q8 | |
134 | #define yyyy2 d16 | |
135 | #define uvrg2b d17 | |
136 | ||
137 | #define d0_ab q9 | |
138 | #define d0_a d18 | |
139 | #define d0_b d19 | |
140 | ||
141 | #define d1_ab q10 | |
142 | #define d1_a d20 | |
143 | #define d1_b d21 | |
144 | ||
145 | #define d2_ab q11 | |
146 | #define d2_a d22 | |
147 | #define d2_b d23 | |
148 | ||
149 | #define d3_ab q12 | |
150 | #define d3_a d24 | |
151 | #define d3_b d25 | |
152 | ||
153 | #define ga_uvrg_x q1 | |
154 | #define ga_uvrg_y q4 | |
155 | ||
156 | #define dx x0_x1 | |
157 | #define dy y0_y1 | |
158 | #define db b0_b1 | |
159 | ||
160 | #define uvrg_base q11 | |
161 | ||
162 | #define gs_uvrg_x q5 | |
163 | #define gs_uvrg_y q6 | |
164 | ||
165 | #define g_uvrg_x q1 | |
166 | #define ga_uv_x d2 | |
167 | #define g_uv_x d2 | |
168 | #define ga_rg_x d3 | |
169 | #define g_rg_x d3 | |
170 | ||
171 | #define g_uvrg_y q4 | |
172 | #define ga_uv_y d8 | |
173 | #define g_uv_y d8 | |
174 | #define ga_rg_y d9 | |
175 | #define g_rg_y d9 | |
176 | ||
177 | #define gw_uv_x q1 | |
178 | #define gw_rg_x q2 | |
179 | #define gw_uv_y q4 | |
180 | #define gw_rg_y q3 | |
181 | ||
182 | #define w_mask q9 | |
183 | #define w_mask_l d18 | |
184 | ||
185 | #define r_shift q10 | |
186 | ||
187 | #define uvrg_dx0 q0 | |
188 | #define uvrg_dx0l d0 | |
189 | #define uvrg_dx0h d1 | |
190 | ||
191 | #define uvrg_dx1 q1 | |
192 | #define uvrg_dx1l d2 | |
193 | #define uvrg_dx1h d3 | |
194 | ||
195 | #define uvrg_dx2 q2 | |
196 | #define uvrg_dx2l d4 | |
197 | #define uvrg_dx2h d5 | |
198 | ||
199 | #define uvrg_dx3 q3 | |
200 | #define uvrg_dx3l d6 | |
201 | #define uvrg_dx3h d7 | |
202 | ||
c6063f89 | 203 | #define uvrgb_phase q13 |
75e28f62 E |
204 | |
205 | .align 4 | |
206 | ||
0e4ad319 | 207 | #include "arm_features.h" |
8184d7c5 | 208 | |
0e4ad319 | 209 | #define function(name) FUNCTION(name): |
210 | ||
211 | #ifndef TEXRELS_FORBIDDEN | |
75e28f62 | 212 | |
8184d7c5 | 213 | #define JT_OP_REL(table_label, index_reg, temp) |
214 | #define JT_OP(x...) x | |
215 | #define JTE(start, target) target | |
216 | ||
217 | #else | |
218 | ||
8184d7c5 | 219 | #define JT_OP_REL(table_label, index_reg, temp) \ |
220 | adr temp, table_label; \ | |
e1f6de8f | 221 | ldr temp, [temp, index_reg, lsl #2]; \ |
8184d7c5 | 222 | add pc, pc, temp \ |
223 | ||
224 | #define JT_OP(x...) | |
225 | #define JTE(start, target) (target - start) | |
226 | ||
0e4ad319 | 227 | #endif |
4d646738 | 228 | |
0e4ad319 | 229 | #ifdef __MACH__ |
8184d7c5 | 230 | #define flush_render_block_buffer _flush_render_block_buffer |
231 | #define setup_sprite_untextured_simple _setup_sprite_untextured_simple | |
232 | #define update_texture_8bpp_cache _update_texture_8bpp_cache | |
8184d7c5 | 233 | #endif |
234 | ||
75e28f62 E |
235 | @ r0: psx_gpu |
236 | @ r1: v_a | |
237 | @ r2: v_b | |
238 | @ r3: v_c | |
239 | ||
240 | function(compute_all_gradients) | |
241 | // First compute the triangle area reciprocal and shift. The division will | |
242 | // happen concurrently with much of the work which follows. | |
243 | @ r12 = psx_gpu->triangle_area | |
e1f6de8f | 244 | ldr r12, [psx_gpu, #psx_gpu_triangle_area_offset] |
75e28f62 | 245 | stmdb sp!, { r4 - r11, lr } |
3f0189c6 | 246 | save_abi_regs() |
75e28f62 E |
247 | |
248 | @ load exponent of 62 into upper half of double | |
249 | movw r4, #0 | |
250 | clz r14, r12 @ r14 = shift | |
251 | ||
252 | movt r4, #((62 + 1023) << 4) | |
253 | mov r12, r12, lsl r14 @ r12 = triangle_area_normalized | |
254 | ||
255 | @ load area normalized into lower half of double | |
256 | mov r5, r12, lsr #10 | |
257 | vmov.f64 d30, r5, r4 @ d30 = (1 << 62) + ta_n | |
258 | ||
259 | movt r4, #((1022 + 31) << 4) | |
260 | mov r5, r12, lsl #20 | |
261 | ||
262 | add r4, r4, r12, lsr #11 | |
263 | vmov.f64 d31, r5, r4 | |
264 | ||
265 | vdiv.f64 d30, d30, d31 @ d30 = ((1 << 62) + ta_n) / ta_n | |
266 | ||
267 | // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) = | |
268 | // ( d0 * d1 ) - ( d2 * d3 ) = | |
269 | // ( m0 ) - ( m1 ) = gradient | |
270 | ||
271 | // This is split to do 12 elements at a time over three sets: a, b, and c. | |
272 | // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so | |
273 | // two of the slots are unused. | |
274 | ||
275 | // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as | |
276 | // is g. | |
277 | ||
278 | // First type is: uvrg bxxx xxxx | |
279 | // Second type is: yyyy ybyy uvrg | |
280 | // Since x_a and y_c are the same the same variable is used for both. | |
281 | ||
e1f6de8f | 282 | vld1.u32 { v0 }, [v_a, :128] @ v0 = { uvrg0, b0, x0, y0 } |
283 | ldrsh x0, [v_a, #8] @ load x0 | |
75e28f62 | 284 | |
e1f6de8f | 285 | vld1.u32 { v1 }, [v_b, :128] @ v1 = { uvrg1, b1, x1, y1} |
286 | ldrh x1, [v_b, #8] @ load x1 | |
75e28f62 | 287 | |
e1f6de8f | 288 | vld1.u32 { v2 }, [v_c, :128] @ v2 = { uvrg2, b2, x2, y2 } |
289 | ldrh x2, [v_c, #8] @ load x2 | |
75e28f62 E |
290 | |
291 | vmovl.u8 uvrg_xxxx0, uvrgb0 @ uvrg_xxxx0 = { uv0, rg0, b0-, -- } | |
e1f6de8f | 292 | ldrh y0, [v_a, #10] @ load y0 |
75e28f62 E |
293 | |
294 | vmovl.u8 uvrg_xxxx1, uvrgb1 @ uvrg_xxxx1 = { uv1, rg1, b1-, -- } | |
e1f6de8f | 295 | ldrh y1, [v_b, #10] @ load y1 |
75e28f62 E |
296 | |
297 | vmovl.u8 uvrg_xxxx2, uvrgb2 @ uvrg_xxxx2 = { uv2, rg2, b2-, -- } | |
e1f6de8f | 298 | ldrh y2, [v_c, #10] @ load y2 |
75e28f62 E |
299 | |
300 | vmov.u8 uvrg0b, uvrg0 @ uvrg0b = { uv0, rg0 } | |
301 | vdup.u16 xxxx0, x0_y0[0] @ xxxx0 = { xx0, xx0 } | |
302 | ||
303 | orr x1_x2, x1, x2, lsl #16 @ x1_x2 = { x1, x2 } | |
304 | pkhbt x0_x1, x0, x1, lsl #16 @ x0_x1 = { x0, x1 } | |
305 | ||
306 | vmov.u8 uvrg1b, uvrg1 @ uvrg1b = { uv1, rg1 } | |
307 | vdup.u16 xxxx1, x1_y1[0] @ xxxx1 = { xx1, xx1 } | |
308 | ||
309 | vmov.u8 uvrg2b, uvrg2 @ uvrg2b = { uv2, rg2 } | |
310 | vdup.u16 xxxx2, x2_y2[0] @ xxxx2 = { xx2, xx2 } | |
311 | ||
e1f6de8f | 312 | ldrb b2, [v_c, #4] @ load b2 |
75e28f62 E |
313 | orr y0_y1, y0, y1, lsl #16 @ y0_y1 = { y0, y1 } |
314 | ||
e1f6de8f | 315 | ldrb b1, [v_b, #4] @ load b1 |
75e28f62 E |
316 | orr y1_y2, y1, y2, lsl #16 @ y1_y2 = { y1, y2 } |
317 | ||
318 | vdup.u16 yyyy0, x0_y0[1] @ yyyy0 = { yy0, yy0 } | |
319 | vsub.s16 d0_ab, x1_ab, x0_ab | |
320 | ||
e1f6de8f | 321 | ldrb b0, [v_a, #4] @ load b0 |
75e28f62 E |
322 | orr b1_b2, b1, b2, lsl #16 @ b1_b2 = { b1, b2 } |
323 | ||
324 | vdup.u16 yyyy1, x1_y1[1] @ yyyy1 = { yy1, yy1 } | |
325 | vsub.s16 d2_ab, x2_ab, x1_ab | |
326 | ||
327 | vdup.u16 yyyy2, x2_y2[1] @ yyyy2 = { yy2, yy2 } | |
328 | vsub.s16 d1_ab, y2_ab, y1_ab | |
329 | ||
330 | orr b0_b1, b0, b1, lsl #16 @ b1_b2 = { b1, b2 } | |
331 | ssub16 dx, x1_x2, x0_x1 @ dx = { x1 - x0, x2 - x1 } | |
332 | ||
333 | ssub16 dy, y1_y2, y0_y1 @ dy = { y1 - y0, y2 - y1 } | |
334 | ssub16 db, b1_b2, b0_b1 @ db = { b1 - b0, b2 - b1 } | |
335 | ||
336 | vsub.s16 d3_ab, y1_ab, y0_ab | |
337 | smusdx ga_by, dx, db @ ga_by = ((x1 - x0) * (b2 - b1)) - | |
338 | @ ((x2 - X1) * (b1 - b0)) | |
339 | vmull.s16 ga_uvrg_x, d0_a, d1_a | |
340 | smusdx ga_bx, db, dy @ ga_bx = ((b1 - b0) * (y2 - y1)) - | |
341 | @ ((b2 - b1) * (y1 - y0)) | |
342 | vmlsl.s16 ga_uvrg_x, d2_a, d3_a | |
343 | movs gs_bx, ga_bx, asr #31 | |
344 | ||
345 | vmull.s16 ga_uvrg_y, d0_b, d1_b | |
346 | rsbmi ga_bx, ga_bx, #0 | |
347 | ||
c6063f89 | 348 | @ r12 = psx_gpu->uvrgb_phase |
e1f6de8f | 349 | ldr r12, [psx_gpu, #psx_gpu_uvrgb_phase_offset] |
c6063f89 | 350 | |
75e28f62 E |
351 | vmlsl.s16 ga_uvrg_y, d2_b, d3_b |
352 | movs gs_by, ga_by, asr #31 | |
353 | ||
354 | vshr.u64 d0, d30, #22 | |
c6063f89 | 355 | add b_base, r12, b0, lsl #16 |
356 | ||
357 | vdup.u32 uvrgb_phase, r12 | |
75e28f62 E |
358 | |
359 | rsbmi ga_by, ga_by, #0 | |
360 | vclt.s32 gs_uvrg_x, ga_uvrg_x, #0 @ gs_uvrg_x = ga_uvrg_x < 0 | |
361 | ||
362 | @ r12 = psx_gpu->triangle_winding_offset | |
e1f6de8f | 363 | ldrb r12, [psx_gpu, #psx_gpu_triangle_winding_offset] |
75e28f62 E |
364 | vclt.s32 gs_uvrg_y, ga_uvrg_y, #0 @ gs_uvrg_y = ga_uvrg_y < 0 |
365 | ||
75e28f62 E |
366 | rsb r12, r12, #0 @ r12 = -(triangle->winding) |
367 | ||
368 | vdup.u32 w_mask, r12 @ w_mask = { -w, -w, -w, -w } | |
369 | sub r14, r14, #(62 - 12) @ r14 = shift - (62 - FIXED_BITS) | |
370 | ||
371 | vshll.u16 uvrg_base, uvrg0, #16 @ uvrg_base = uvrg0 << 16 | |
372 | vdup.u32 r_shift, r14 @ r_shift = { shift, shift, shift, shift } | |
373 | ||
c6063f89 | 374 | vadd.u32 uvrg_base, uvrgb_phase |
75e28f62 E |
375 | vabs.s32 ga_uvrg_x, ga_uvrg_x @ ga_uvrg_x = abs(ga_uvrg_x) |
376 | ||
377 | vmov area_r_s, s0 @ area_r_s = triangle_reciprocal | |
378 | vabs.s32 ga_uvrg_y, ga_uvrg_y @ ga_uvrg_y = abs(ga_uvrg_y) | |
379 | ||
380 | vmull.u32 gw_rg_x, ga_rg_x, d0[0] | |
381 | vmull.u32 gw_uv_x, ga_uv_x, d0[0] | |
382 | vmull.u32 gw_rg_y, ga_rg_y, d0[0] | |
383 | vmull.u32 gw_uv_y, ga_uv_y, d0[0] | |
384 | ||
385 | vshl.u64 gw_rg_x, gw_rg_x, r_shift | |
386 | vshl.u64 gw_uv_x, gw_uv_x, r_shift | |
387 | vshl.u64 gw_rg_y, gw_rg_y, r_shift | |
388 | vshl.u64 gw_uv_y, gw_uv_y, r_shift | |
389 | ||
390 | veor.u32 gs_uvrg_x, gs_uvrg_x, w_mask | |
391 | vmovn.u64 g_uv_x, gw_uv_x | |
392 | ||
393 | veor.u32 gs_uvrg_y, gs_uvrg_y, w_mask | |
394 | vmovn.u64 g_rg_x, gw_rg_x | |
395 | ||
396 | veor.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
397 | vmovn.u64 g_uv_y, gw_uv_y | |
398 | ||
399 | vsub.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
400 | vmovn.u64 g_rg_y, gw_rg_y | |
401 | ||
402 | veor.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
403 | mov ga_bx, ga_bx, lsl #13 | |
404 | ||
405 | vsub.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
406 | mov ga_by, ga_by, lsl #13 | |
407 | ||
408 | vdup.u32 x0_y0, x0 | |
409 | umull gw_bx_l, gw_bx_h, ga_bx, area_r_s | |
410 | ||
411 | vshl.u32 g_uvrg_x, g_uvrg_x, #4 | |
412 | vshl.u32 g_uvrg_y, g_uvrg_y, #4 | |
413 | ||
414 | umull gw_by_l, gw_by_h, ga_by, area_r_s | |
415 | vmls.s32 uvrg_base, ga_uvrg_x, x0_y0[0] | |
416 | ||
417 | eor gs_bx, gs_bx, r12 | |
418 | vadd.u32 uvrg_dx2, uvrg_dx1, uvrg_dx1 | |
419 | ||
420 | veor.u32 uvrg_dx0, uvrg_dx0, uvrg_dx0 | |
421 | eor gs_by, gs_by, r12 | |
422 | ||
423 | rsb r11, r14, #0 @ r11 = negative shift for scalar lsr | |
424 | add store_a, psx_gpu, #psx_gpu_uvrg_offset | |
425 | ||
426 | sub r11, r11, #(32 - 13) | |
427 | ||
428 | add store_b, store_a, #16 | |
429 | mov store_inc, #32 | |
430 | ||
431 | vadd.u32 uvrg_dx3, uvrg_dx2, uvrg_dx1 | |
e1f6de8f | 432 | vst1.u32 { uvrg_base }, [store_a, :128], store_inc |
75e28f62 | 433 | |
e1f6de8f | 434 | vst1.u32 { uvrg_dx1 }, [store_b, :128], store_inc |
75e28f62 E |
435 | mov g_bx, gw_bx_h, lsr r11 |
436 | ||
e1f6de8f | 437 | vst1.u32 { g_uvrg_y }, [store_a, :128], store_inc |
75e28f62 E |
438 | mov g_by, gw_by_h, lsr r11 |
439 | ||
440 | vst4.u32 { uvrg_dx0l, uvrg_dx1l, uvrg_dx2l, uvrg_dx3l }, \ | |
e1f6de8f | 441 | [store_b, :128], store_inc |
75e28f62 E |
442 | eor g_bx, g_bx, gs_bx |
443 | ||
444 | vst4.u32 { uvrg_dx0h, uvrg_dx1h, uvrg_dx2h, uvrg_dx3h }, \ | |
e1f6de8f | 445 | [store_b, :128], store_inc |
75e28f62 E |
446 | sub g_bx, g_bx, gs_bx |
447 | ||
448 | lsl g_bx, g_bx, #4 | |
449 | eor g_by, g_by, gs_by | |
450 | ||
451 | mls b_base, g_bx, x0, b_base | |
452 | sub g_by, g_by, gs_by | |
453 | ||
454 | lsl g_by, g_by, #4 | |
455 | mov g_bx0, #0 | |
456 | ||
457 | add g_bx2, g_bx, g_bx | |
458 | add g_bx3, g_bx, g_bx2 | |
459 | ||
460 | stmia store_b, { g_bx0, g_bx, g_bx2, g_bx3, b_base, g_by } | |
461 | ||
3f0189c6 | 462 | restore_abi_regs() |
75e28f62 E |
463 | ldmia sp!, { r4 - r11, pc } |
464 | ||
465 | ||
466 | #define psx_gpu r0 | |
467 | #define v_a r1 | |
468 | #define v_b r2 | |
469 | #define v_c r3 | |
470 | ||
471 | #define temp r14 | |
472 | ||
473 | #define x_a r4 | |
474 | #define x_b r5 | |
475 | #define x_c r6 | |
476 | #define y_a r1 | |
477 | #define y_b r2 | |
478 | #define y_c r3 | |
479 | ||
480 | #define height_minor_a r7 | |
481 | #define height_minor_b r8 | |
482 | #define height_major r9 | |
483 | #define height r9 | |
484 | ||
485 | #define reciprocal_table_ptr r10 | |
486 | ||
487 | #define edge_alt_low r4 | |
488 | #define edge_alt_high r5 | |
489 | #define edge_dx_dy_alt r6 | |
490 | #define edge_shift_alt r10 | |
491 | ||
492 | #define edge_dx_dy_alt_low r4 | |
493 | #define edge_dx_dy_alt_high r5 | |
494 | ||
495 | #define span_edge_data r4 | |
496 | #define span_uvrg_offset r5 | |
497 | #define span_b_offset r6 | |
498 | ||
499 | #define clip r14 | |
500 | ||
501 | #define b r11 | |
502 | #define b_dy r12 | |
503 | ||
504 | ||
505 | #define alternate_x q0 | |
506 | #define alternate_dx_dy q1 | |
507 | #define alternate_x_32 q2 | |
508 | ||
509 | #define alternate_x_low d0 | |
510 | #define alternate_x_high d1 | |
511 | #define alternate_dx_dy_low d2 | |
512 | #define alternate_dx_dy_high d3 | |
513 | #define alternate_x_32_low d4 | |
514 | #define alternate_x_32_high d5 | |
515 | ||
516 | #define left_x q3 | |
517 | #define right_x q4 | |
518 | #define left_dx_dy q5 | |
519 | #define right_dx_dy q6 | |
520 | #define left_edge q7 | |
521 | #define right_edge q8 | |
522 | ||
523 | #define left_x_low d6 | |
524 | #define left_x_high d7 | |
525 | #define right_x_low d8 | |
526 | #define right_x_high d9 | |
527 | #define left_dx_dy_low d10 | |
528 | #define left_dx_dy_high d11 | |
529 | #define right_dx_dy_low d12 | |
530 | #define right_dx_dy_high d13 | |
531 | #define left_edge_low d14 | |
532 | #define left_edge_high d15 | |
533 | #define right_edge_low d16 | |
534 | #define right_edge_high d17 | |
535 | ||
536 | #define y_mid_point d18 | |
537 | #define c_0x0004 d19 | |
538 | ||
539 | #define left_right_x_16 q11 | |
540 | #define span_shifts_y q12 | |
541 | #define c_0x0001 q13 | |
542 | ||
543 | #define span_shifts d24 | |
544 | #define y_x4 d25 | |
545 | #define c_0xFFFE d26 | |
546 | #define c_0x0007 d27 | |
547 | ||
548 | #define left_right_x_16_low d22 | |
549 | #define left_right_x_16_high d23 | |
550 | ||
551 | #define uvrg q14 | |
552 | #define uvrg_dy q15 | |
553 | ||
554 | #define alternate_x_16 d4 | |
555 | ||
556 | #define v_clip q3 | |
557 | #define v_clip_low d6 | |
558 | ||
559 | #define right_x_32 q10 | |
560 | #define left_x_32 q11 | |
561 | #define alternate_select d24 | |
562 | ||
563 | #define right_x_32_low d20 | |
564 | #define right_x_32_high d21 | |
565 | #define left_x_32_low d22 | |
566 | #define left_x_32_high d23 | |
567 | ||
568 | #define edges_xy q0 | |
569 | #define edges_dx_dy d2 | |
570 | #define edge_shifts d3 | |
571 | #define edge_shifts_64 q2 | |
572 | ||
573 | #define edges_xy_left d0 | |
574 | #define edges_xy_right d1 | |
575 | ||
576 | #define height_reciprocals d6 | |
577 | #define heights d7 | |
578 | ||
579 | #define widths d8 | |
580 | #define c_0x01 d9 | |
581 | #define x_starts d10 | |
582 | #define x_ends d11 | |
583 | ||
584 | #define heights_b d12 | |
585 | #define edges_dx_dy_64 q10 | |
586 | ||
587 | #define edges_dx_dy_64_left d20 | |
588 | #define edges_dx_dy_64_right d21 | |
589 | ||
590 | ||
591 | #define setup_spans_prologue() \ | |
592 | stmdb sp!, { r4 - r11, lr }; \ | |
3f0189c6 | 593 | save_abi_regs(); \ |
75e28f62 | 594 | \ |
e1f6de8f | 595 | ldrsh x_a, [v_a, #8]; \ |
596 | ldrsh x_b, [v_b, #8]; \ | |
597 | ldrsh x_c, [v_c, #8]; \ | |
598 | ldrsh y_a, [v_a, #10]; \ | |
599 | ldrsh y_b, [v_b, #10]; \ | |
600 | ldrsh y_c, [v_c, #10]; \ | |
75e28f62 E |
601 | \ |
602 | add temp, psx_gpu, #psx_gpu_uvrg_offset; \ | |
e1f6de8f | 603 | vld1.32 { uvrg }, [temp]; \ |
75e28f62 | 604 | add temp, psx_gpu, #psx_gpu_uvrg_dy_offset; \ |
e1f6de8f | 605 | vld1.32 { uvrg_dy }, [temp]; \ |
606 | ldr reciprocal_table_ptr, [psx_gpu, #psx_gpu_reciprocal_table_ptr_offset]; \ | |
75e28f62 E |
607 | \ |
608 | vmov.u32 c_0x01, #0x01 \ | |
609 | ||
610 | #define setup_spans_load_b() \ | |
e1f6de8f | 611 | ldr b, [psx_gpu, #psx_gpu_b_offset]; \ |
612 | ldr b_dy, [psx_gpu, #psx_gpu_b_dy_offset] \ | |
75e28f62 E |
613 | |
614 | #define setup_spans_prologue_b() \ | |
615 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
616 | add temp, psx_gpu, #psx_gpu_viewport_start_x_offset; \ | |
617 | \ | |
618 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
619 | vmov.u16 c_0x0004, #0x0004; \ | |
620 | \ | |
621 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
622 | vmov.u16 c_0x0001, #0x0001; \ | |
623 | \ | |
e1f6de8f | 624 | vld1.u16 { left_edge_low[], left_edge_high[] }, [temp]; \ |
75e28f62 E |
625 | add temp, psx_gpu, #psx_gpu_viewport_end_x_offset; \ |
626 | \ | |
e1f6de8f | 627 | vld1.u16 { right_edge_low[], right_edge_high[] }, [temp]; \ |
75e28f62 E |
628 | vadd.u16 right_edge, right_edge, c_0x0001; \ |
629 | \ | |
630 | vmov.u16 c_0x0007, #0x0007; \ | |
631 | vmvn.u16 c_0xFFFE, #0x0001 \ | |
632 | ||
633 | ||
634 | #define compute_edge_delta_x2() \ | |
e1f6de8f | 635 | ldr temp, [reciprocal_table_ptr, height, lsl #2]; \ |
75e28f62 E |
636 | \ |
637 | vdup.u32 heights, height; \ | |
638 | vsub.u32 widths, x_ends, x_starts; \ | |
639 | \ | |
640 | vdup.u32 edge_shifts, temp; \ | |
641 | vsub.u32 heights_b, heights, c_0x01; \ | |
7d5140f5 | 642 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
75e28f62 E |
643 | \ |
644 | vmla.s32 heights_b, x_starts, heights; \ | |
645 | vbic.u16 edge_shifts, #0xE0; \ | |
646 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
647 | vmull.s32 edges_xy, heights_b, height_reciprocals \ | |
648 | ||
649 | #define width_alt r6 | |
650 | #define height_reciprocal_alt r11 | |
651 | #define height_b_alt r12 | |
652 | ||
653 | #define compute_edge_delta_x3(start_c, height_a, height_b) \ | |
ed0fd81d | 654 | vmov heights, height_a, height_b; \ |
e1f6de8f | 655 | ldr temp, [reciprocal_table_ptr, height_a, lsl #2]; \ |
75e28f62 | 656 | vmov.u32 edge_shifts[0], temp; \ |
e1f6de8f | 657 | ldr temp, [reciprocal_table_ptr, height_b, lsl #2]; \ |
75e28f62 | 658 | vmov.u32 edge_shifts[1], temp; \ |
e1f6de8f | 659 | ldr edge_shift_alt, [reciprocal_table_ptr, height_minor_b, lsl #2]; \ |
75e28f62 E |
660 | \ |
661 | vsub.u32 widths, x_ends, x_starts; \ | |
662 | sub width_alt, x_c, start_c; \ | |
663 | \ | |
664 | vsub.u32 heights_b, heights, c_0x01; \ | |
665 | sub height_b_alt, height_minor_b, #1; \ | |
666 | \ | |
7d5140f5 E |
667 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
668 | lsr height_reciprocal_alt, edge_shift_alt, #10; \ | |
75e28f62 E |
669 | \ |
670 | vmla.s32 heights_b, x_starts, heights; \ | |
671 | mla height_b_alt, height_minor_b, start_c, height_b_alt; \ | |
672 | \ | |
673 | vbic.u16 edge_shifts, #0xE0; \ | |
674 | and edge_shift_alt, edge_shift_alt, #0x1F; \ | |
675 | \ | |
676 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
677 | mul edge_dx_dy_alt, width_alt, height_reciprocal_alt; \ | |
678 | \ | |
679 | vmull.s32 edges_xy, heights_b, height_reciprocals; \ | |
680 | smull edge_alt_low, edge_alt_high, height_b_alt, height_reciprocal_alt \ | |
681 | ||
682 | ||
683 | #define setup_spans_adjust_y_up() \ | |
684 | vsub.u32 y_x4, y_x4, c_0x0004 \ | |
685 | ||
686 | #define setup_spans_adjust_y_down() \ | |
687 | vadd.u32 y_x4, y_x4, c_0x0004 \ | |
688 | ||
689 | #define setup_spans_adjust_interpolants_up() \ | |
690 | vsub.u32 uvrg, uvrg, uvrg_dy; \ | |
691 | sub b, b, b_dy \ | |
692 | ||
693 | #define setup_spans_adjust_interpolants_down() \ | |
694 | vadd.u32 uvrg, uvrg, uvrg_dy; \ | |
695 | add b, b, b_dy \ | |
696 | ||
697 | ||
698 | #define setup_spans_clip_interpolants_increment() \ | |
699 | mla b, b_dy, clip, b; \ | |
700 | vmla.s32 uvrg, uvrg_dy, v_clip \ | |
701 | ||
702 | #define setup_spans_clip_interpolants_decrement() \ | |
703 | mls b, b_dy, clip, b; \ | |
704 | vmls.s32 uvrg, uvrg_dy, v_clip \ | |
705 | ||
706 | #define setup_spans_clip_alternate_yes() \ | |
707 | smlal edge_alt_low, edge_alt_high, edge_dx_dy_alt, clip \ | |
708 | ||
709 | #define setup_spans_clip_alternate_no() \ | |
710 | ||
711 | #define setup_spans_clip(direction, alternate_active) \ | |
712 | vdup.u32 v_clip, clip; \ | |
713 | setup_spans_clip_alternate_##alternate_active(); \ | |
714 | setup_spans_clip_interpolants_##direction(); \ | |
715 | vmlal.s32 edges_xy, edges_dx_dy, v_clip_low \ | |
716 | ||
717 | ||
718 | #define setup_spans_adjust_edges_alternate_no(left_index, right_index) \ | |
719 | vmovl.s32 edge_shifts_64, edge_shifts; \ | |
720 | vmovl.s32 edges_dx_dy_64, edges_dx_dy; \ | |
721 | \ | |
722 | vshl.s64 edges_xy, edges_xy, edge_shifts_64; \ | |
723 | vshl.s64 edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64; \ | |
724 | \ | |
725 | vmov left_x_low, edges_xy_##left_index; \ | |
726 | vmov right_x_low, edges_xy_##right_index; \ | |
727 | \ | |
728 | vmov left_dx_dy_low, edges_dx_dy_64_##left_index; \ | |
729 | vmov left_dx_dy_high, edges_dx_dy_64_##left_index; \ | |
730 | vmov right_dx_dy_low, edges_dx_dy_64_##right_index; \ | |
731 | vmov right_dx_dy_high, edges_dx_dy_64_##right_index; \ | |
732 | \ | |
733 | vadd.u64 left_x_high, left_x_low, left_dx_dy_low; \ | |
734 | vadd.u64 right_x_high, right_x_low, right_dx_dy_low; \ | |
735 | \ | |
736 | vadd.u64 left_dx_dy, left_dx_dy, left_dx_dy; \ | |
737 | vadd.u64 right_dx_dy, right_dx_dy, right_dx_dy \ | |
738 | ||
739 | ||
740 | #define setup_spans_adjust_edges_alternate_yes(left_index, right_index) \ | |
741 | setup_spans_adjust_edges_alternate_no(left_index, right_index); \ | |
742 | \ | |
743 | vdup.u16 y_mid_point, y_b; \ | |
744 | rsb temp, edge_shift_alt, #32; \ | |
745 | \ | |
746 | lsl edge_alt_high, edge_alt_high, edge_shift_alt; \ | |
747 | orr edge_alt_high, edge_alt_high, edge_alt_low, lsr temp; \ | |
748 | lsl edge_alt_low, edge_alt_low, edge_shift_alt; \ | |
749 | vmov alternate_x_low, edge_alt_low, edge_alt_high; \ | |
750 | \ | |
751 | asr edge_dx_dy_alt_high, edge_dx_dy_alt, temp; \ | |
752 | lsl edge_dx_dy_alt_low, edge_dx_dy_alt, edge_shift_alt; \ | |
753 | vmov alternate_dx_dy_low, edge_dx_dy_alt_low, edge_dx_dy_alt_high; \ | |
754 | vmov alternate_dx_dy_high, alternate_dx_dy_low; \ | |
755 | \ | |
756 | vadd.u64 alternate_x_high, alternate_x_low, alternate_dx_dy_low; \ | |
757 | vadd.u64 alternate_dx_dy, alternate_dx_dy, alternate_dx_dy \ | |
758 | ||
759 | ||
760 | #define setup_spans_y_select_up() \ | |
761 | vclt.s16 alternate_select, y_x4, y_mid_point \ | |
762 | ||
763 | #define setup_spans_y_select_down() \ | |
764 | vcgt.s16 alternate_select, y_x4, y_mid_point \ | |
765 | ||
766 | ||
767 | #define setup_spans_alternate_select_left() \ | |
768 | vbit.u16 left_right_x_16_low, alternate_x_16, alternate_select \ | |
769 | ||
770 | #define setup_spans_alternate_select_right() \ | |
771 | vbit.u16 left_right_x_16_high, alternate_x_16, alternate_select \ | |
772 | ||
773 | ||
774 | #define setup_spans_set_x4_alternate_yes(alternate, direction) \ | |
775 | vshrn.s64 alternate_x_32_low, alternate_x, #32; \ | |
776 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
777 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
778 | \ | |
779 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
780 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
781 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
782 | \ | |
783 | vshrn.s64 alternate_x_32_high, alternate_x, #32; \ | |
784 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
785 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
786 | \ | |
787 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
788 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
789 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
790 | \ | |
791 | vmovn.u32 alternate_x_16, alternate_x_32; \ | |
792 | setup_spans_y_select_##direction(); \ | |
793 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
794 | \ | |
795 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
796 | setup_spans_alternate_select_##alternate(); \ | |
797 | \ | |
e1f6de8f | 798 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
799 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
800 | setup_spans_adjust_interpolants_##direction(); \ |
801 | \ | |
802 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
803 | \ | |
e1f6de8f | 804 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
805 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
806 | setup_spans_adjust_interpolants_##direction(); \ |
807 | \ | |
808 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
809 | \ | |
e1f6de8f | 810 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
811 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
812 | setup_spans_adjust_interpolants_##direction(); \ |
813 | \ | |
814 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
815 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
816 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
817 | \ | |
e1f6de8f | 818 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
819 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
820 | setup_spans_adjust_interpolants_##direction(); \ |
821 | \ | |
822 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
823 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
824 | \ | |
e1f6de8f | 825 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
75e28f62 E |
826 | \ |
827 | setup_spans_adjust_y_##direction() \ | |
828 | ||
829 | ||
830 | #define setup_spans_set_x4_alternate_no(alternate, direction) \ | |
831 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
832 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
833 | \ | |
834 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
835 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
836 | \ | |
837 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
838 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
839 | \ | |
840 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
841 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
842 | \ | |
843 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
844 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
845 | \ | |
e1f6de8f | 846 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
847 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
848 | setup_spans_adjust_interpolants_##direction(); \ |
849 | \ | |
850 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
851 | \ | |
e1f6de8f | 852 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
853 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
854 | setup_spans_adjust_interpolants_##direction(); \ |
855 | \ | |
856 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
857 | \ | |
e1f6de8f | 858 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
859 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
860 | setup_spans_adjust_interpolants_##direction(); \ |
861 | \ | |
862 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
863 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
864 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
865 | \ | |
e1f6de8f | 866 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
867 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
868 | setup_spans_adjust_interpolants_##direction(); \ |
869 | \ | |
870 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
871 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
872 | \ | |
e1f6de8f | 873 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
75e28f62 E |
874 | \ |
875 | setup_spans_adjust_y_##direction() \ | |
876 | ||
877 | ||
878 | #define edge_adjust_low r11 | |
879 | #define edge_adjust_high r12 | |
880 | ||
881 | #define setup_spans_alternate_adjust_yes() \ | |
882 | smull edge_adjust_low, edge_adjust_high, edge_dx_dy_alt, height_minor_a; \ | |
883 | subs edge_alt_low, edge_alt_low, edge_adjust_low; \ | |
884 | sbc edge_alt_high, edge_alt_high, edge_adjust_high \ | |
885 | ||
886 | #define setup_spans_alternate_adjust_no() \ | |
887 | ||
888 | ||
889 | #define setup_spans_down(left_index, right_index, alternate, alternate_active) \ | |
890 | setup_spans_alternate_adjust_##alternate_active(); \ | |
891 | setup_spans_load_b(); \ | |
892 | \ | |
e1f6de8f | 893 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
75e28f62 E |
894 | subs y_c, y_c, temp; \ |
895 | subgt height, height, y_c; \ | |
896 | addgt height, height, #1; \ | |
897 | \ | |
e1f6de8f | 898 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
75e28f62 E |
899 | subs clip, temp, y_a; \ |
900 | ble 0f; \ | |
901 | \ | |
902 | sub height, height, clip; \ | |
903 | add y_a, y_a, clip; \ | |
904 | setup_spans_clip(increment, alternate_active); \ | |
905 | \ | |
906 | 0: \ | |
907 | cmp height, #0; \ | |
908 | ble 1f; \ | |
909 | \ | |
910 | orr temp, y_a, y_a, lsl #16; \ | |
911 | add temp, temp, #(1 << 16); \ | |
912 | add y_a, temp, #2; \ | |
913 | add y_a, y_a, #(2 << 16); \ | |
ed0fd81d | 914 | vmov y_x4, temp, y_a; \ |
75e28f62 E |
915 | \ |
916 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
917 | right_index); \ | |
918 | setup_spans_prologue_b(); \ | |
919 | \ | |
e1f6de8f | 920 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
921 | \ |
922 | 2: \ | |
923 | setup_spans_set_x4_alternate_##alternate_active(alternate, down); \ | |
924 | subs height, height, #4; \ | |
925 | bhi 2b; \ | |
926 | \ | |
927 | 1: \ | |
928 | ||
929 | ||
930 | #define setup_spans_alternate_pre_increment_yes() \ | |
931 | adds edge_alt_low, edge_alt_low, edge_dx_dy_alt; \ | |
932 | adc edge_alt_high, edge_alt_high, edge_dx_dy_alt, asr #31 \ | |
933 | ||
934 | #define setup_spans_alternate_pre_increment_no() \ | |
935 | ||
936 | ||
937 | #define setup_spans_up_decrement_yes() \ | |
938 | suble height, height, #1 \ | |
939 | ||
940 | #define setup_spans_up_decrement_no() \ | |
941 | ||
942 | ||
943 | #define setup_spans_up(left_index, right_index, alternate, alternate_active) \ | |
944 | setup_spans_alternate_adjust_##alternate_active(); \ | |
945 | setup_spans_load_b(); \ | |
946 | sub y_a, y_a, #1; \ | |
947 | \ | |
e1f6de8f | 948 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
75e28f62 E |
949 | subs temp, temp, y_c; \ |
950 | subgt height, height, temp; \ | |
951 | setup_spans_up_decrement_##alternate_active(); \ | |
952 | \ | |
e1f6de8f | 953 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
75e28f62 E |
954 | subs clip, y_a, temp; \ |
955 | ble 0f; \ | |
956 | \ | |
957 | sub height, height, clip; \ | |
958 | sub y_a, y_a, clip; \ | |
959 | setup_spans_clip(decrement, alternate_active); \ | |
960 | \ | |
961 | 0: \ | |
962 | cmp height, #0; \ | |
963 | ble 1f; \ | |
964 | \ | |
965 | orr temp, y_a, y_a, lsl #16; \ | |
966 | sub temp, temp, #(1 << 16); \ | |
967 | sub y_a, temp, #2; \ | |
968 | sub y_a, y_a, #(2 << 16); \ | |
ed0fd81d | 969 | vmov y_x4, temp, y_a; \ |
75e28f62 E |
970 | \ |
971 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy; \ | |
972 | \ | |
973 | setup_spans_alternate_pre_increment_##alternate_active(); \ | |
974 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
975 | right_index); \ | |
976 | setup_spans_adjust_interpolants_up(); \ | |
977 | setup_spans_prologue_b(); \ | |
978 | \ | |
e1f6de8f | 979 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
980 | \ |
981 | 2: \ | |
982 | setup_spans_set_x4_alternate_##alternate_active(alternate, up); \ | |
983 | subs height, height, #4; \ | |
984 | bhi 2b; \ | |
985 | \ | |
986 | 1: \ | |
987 | ||
988 | ||
989 | #define setup_spans_epilogue() \ | |
3f0189c6 | 990 | restore_abi_regs(); \ |
75e28f62 E |
991 | ldmia sp!, { r4 - r11, pc } \ |
992 | ||
993 | ||
994 | #define setup_spans_up_up(minor, major) \ | |
995 | setup_spans_prologue(); \ | |
996 | sub height_minor_a, y_a, y_b; \ | |
997 | sub height_minor_b, y_b, y_c; \ | |
998 | sub height, y_a, y_c; \ | |
999 | \ | |
1000 | vdup.u32 x_starts, x_a; \ | |
ed0fd81d | 1001 | vmov x_ends, x_c, x_b; \ |
75e28f62 E |
1002 | \ |
1003 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
1004 | setup_spans_up(major, minor, minor, yes); \ | |
1005 | setup_spans_epilogue() \ | |
1006 | ||
1007 | function(setup_spans_up_left) | |
1008 | setup_spans_up_up(left, right) | |
1009 | ||
1010 | function(setup_spans_up_right) | |
1011 | setup_spans_up_up(right, left) | |
1012 | ||
75e28f62 E |
1013 | #define setup_spans_down_down(minor, major) \ |
1014 | setup_spans_prologue(); \ | |
1015 | sub height_minor_a, y_b, y_a; \ | |
1016 | sub height_minor_b, y_c, y_b; \ | |
1017 | sub height, y_c, y_a; \ | |
1018 | \ | |
1019 | vdup.u32 x_starts, x_a; \ | |
ed0fd81d | 1020 | vmov x_ends, x_c, x_b; \ |
75e28f62 E |
1021 | \ |
1022 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
1023 | setup_spans_down(major, minor, minor, yes); \ | |
1024 | setup_spans_epilogue() \ | |
1025 | ||
1026 | function(setup_spans_down_left) | |
1027 | setup_spans_down_down(left, right) | |
1028 | ||
1029 | function(setup_spans_down_right) | |
1030 | setup_spans_down_down(right, left) | |
1031 | ||
1032 | ||
1033 | #define setup_spans_up_flat() \ | |
1034 | sub height, y_a, y_c; \ | |
1035 | \ | |
1036 | compute_edge_delta_x2(); \ | |
1037 | setup_spans_up(left, right, none, no); \ | |
1038 | setup_spans_epilogue() \ | |
1039 | ||
1040 | function(setup_spans_up_a) | |
1041 | setup_spans_prologue() | |
1042 | ||
ed0fd81d | 1043 | vmov x_starts, x_a, x_b |
75e28f62 E |
1044 | vdup.u32 x_ends, x_c |
1045 | ||
1046 | setup_spans_up_flat() | |
1047 | ||
1048 | function(setup_spans_up_b) | |
1049 | setup_spans_prologue() | |
1050 | ||
1051 | vdup.u32 x_starts, x_a | |
ed0fd81d | 1052 | vmov x_ends, x_b, x_c |
75e28f62 E |
1053 | |
1054 | setup_spans_up_flat() | |
1055 | ||
1056 | #define setup_spans_down_flat() \ | |
1057 | sub height, y_c, y_a; \ | |
1058 | \ | |
1059 | compute_edge_delta_x2(); \ | |
1060 | setup_spans_down(left, right, none, no); \ | |
1061 | setup_spans_epilogue() \ | |
1062 | ||
1063 | function(setup_spans_down_a) | |
1064 | setup_spans_prologue() | |
1065 | ||
ed0fd81d | 1066 | vmov x_starts, x_a, x_b |
75e28f62 E |
1067 | vdup.u32 x_ends, x_c |
1068 | ||
1069 | setup_spans_down_flat() | |
1070 | ||
1071 | function(setup_spans_down_b) | |
1072 | setup_spans_prologue() | |
1073 | ||
1074 | vdup.u32 x_starts, x_a | |
ed0fd81d | 1075 | vmov x_ends, x_b, x_c |
75e28f62 E |
1076 | |
1077 | setup_spans_down_flat() | |
1078 | ||
1079 | ||
1080 | #define middle_y r9 | |
1081 | ||
1082 | #define edges_xy_b q11 | |
1083 | #define edges_dx_dy_b d26 | |
1084 | #define edge_shifts_b d27 | |
1085 | #define edges_dx_dy_and_shifts_b q13 | |
1086 | #define height_increment d20 | |
1087 | ||
1088 | #define edges_dx_dy_and_shifts q1 | |
1089 | ||
1090 | #define edges_xy_b_left d22 | |
1091 | #define edges_xy_b_right d23 | |
1092 | ||
1093 | #define setup_spans_up_down_load_edge_set_b() \ | |
1094 | vmov edges_xy, edges_xy_b; \ | |
1095 | vmov edges_dx_dy_and_shifts, edges_dx_dy_and_shifts_b \ | |
1096 | ||
1097 | ||
1098 | function(setup_spans_up_down) | |
1099 | setup_spans_prologue() | |
1100 | ||
1101 | // s32 middle_y = y_a; | |
1102 | sub height_minor_a, y_a, y_b | |
1103 | sub height_minor_b, y_c, y_a | |
1104 | sub height_major, y_c, y_b | |
1105 | ||
ed0fd81d | 1106 | vmov x_starts, x_a, x_c |
75e28f62 E |
1107 | vdup.u32 x_ends, x_b |
1108 | ||
1109 | compute_edge_delta_x3(x_a, height_minor_a, height_major) | |
1110 | ||
1111 | mov temp, #0 | |
ed0fd81d | 1112 | vmov height_increment, temp, height_minor_b |
75e28f62 E |
1113 | vmlal.s32 edges_xy, edges_dx_dy, height_increment |
1114 | ||
1115 | vmov edges_xy_b_left, edge_alt_low, edge_alt_high | |
1116 | vmov edges_xy_b_right, edges_xy_right | |
1117 | ||
1118 | vmov edge_shifts_b, edge_shifts | |
1119 | vmov.u32 edge_shifts_b[0], edge_shift_alt | |
1120 | ||
1121 | vneg.s32 edges_dx_dy_b, edges_dx_dy | |
1122 | vmov.u32 edges_dx_dy_b[0], edge_dx_dy_alt | |
1123 | ||
1124 | mov middle_y, y_a | |
1125 | ||
1126 | setup_spans_load_b() | |
1127 | sub y_a, y_a, #1 | |
1128 | ||
e1f6de8f | 1129 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
75e28f62 E |
1130 | subs temp, temp, y_b |
1131 | subgt height_minor_a, height_minor_a, temp | |
1132 | ||
e1f6de8f | 1133 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
75e28f62 E |
1134 | subs clip, y_a, temp |
1135 | ble 0f | |
1136 | ||
1137 | sub height_minor_a, height_minor_a, clip | |
1138 | sub y_a, y_a, clip | |
1139 | setup_spans_clip(decrement, no) | |
1140 | ||
1141 | 0: | |
1142 | cmp height_minor_a, #0 | |
1143 | ble 3f | |
1144 | ||
1145 | orr temp, y_a, y_a, lsl #16 | |
1146 | sub temp, temp, #(1 << 16) | |
1147 | sub y_a, temp, #2 | |
1148 | sub y_a, y_a, #(2 << 16) | |
ed0fd81d | 1149 | vmov y_x4, temp, y_a |
75e28f62 E |
1150 | |
1151 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy | |
1152 | ||
e1f6de8f | 1153 | strh height_minor_a, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1154 | |
1155 | setup_spans_adjust_edges_alternate_no(left, right); | |
1156 | setup_spans_adjust_interpolants_up() | |
1157 | setup_spans_up_down_load_edge_set_b() | |
1158 | ||
1159 | setup_spans_prologue_b() | |
1160 | ||
1161 | ||
1162 | 2: | |
1163 | setup_spans_set_x4_alternate_no(none, up) | |
1164 | subs height_minor_a, height_minor_a, #4 | |
1165 | bhi 2b | |
1166 | ||
1167 | add span_edge_data, span_edge_data, height_minor_a, lsl #3 | |
1168 | add span_uvrg_offset, span_uvrg_offset, height_minor_a, lsl #4 | |
1169 | add span_b_offset, span_b_offset, height_minor_a, lsl #2 | |
1170 | ||
1171 | 4: | |
1172 | add temp, psx_gpu, #psx_gpu_uvrg_offset | |
e1f6de8f | 1173 | vld1.32 { uvrg }, [temp] |
75e28f62 E |
1174 | mov y_a, middle_y |
1175 | ||
1176 | setup_spans_load_b() | |
1177 | ||
e1f6de8f | 1178 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
75e28f62 E |
1179 | subs y_c, y_c, temp |
1180 | subgt height_minor_b, height_minor_b, y_c | |
1181 | addgt height_minor_b, height_minor_b, #1 | |
1182 | ||
e1f6de8f | 1183 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
75e28f62 E |
1184 | subs clip, temp, y_a |
1185 | ble 0f | |
1186 | ||
1187 | sub height_minor_b, height_minor_b, clip | |
1188 | add y_a, y_a, clip | |
1189 | setup_spans_clip(increment, no) | |
1190 | ||
1191 | 0: | |
1192 | cmp height_minor_b, #0 | |
1193 | ble 1f | |
1194 | ||
1195 | orr temp, y_a, y_a, lsl #16 | |
1196 | add temp, temp, #(1 << 16) | |
1197 | add y_a, temp, #2 | |
1198 | add y_a, y_a, #(2 << 16) | |
ed0fd81d | 1199 | vmov y_x4, temp, y_a |
75e28f62 E |
1200 | |
1201 | setup_spans_adjust_edges_alternate_no(left, right) | |
1202 | ||
e1f6de8f | 1203 | ldrh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 | 1204 | add temp, temp, height_minor_b |
b7569147 | 1205 | |
1206 | cmp temp, #MAX_SPANS | |
1207 | beq 5f | |
1208 | ||
e1f6de8f | 1209 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1210 | |
1211 | 2: | |
1212 | setup_spans_set_x4_alternate_no(none, down) | |
1213 | subs height_minor_b, height_minor_b, #4 | |
1214 | bhi 2b | |
1215 | ||
1216 | 1: | |
1217 | setup_spans_epilogue() | |
1218 | ||
1219 | 3: | |
1220 | setup_spans_up_down_load_edge_set_b() | |
1221 | setup_spans_prologue_b() | |
1222 | bal 4b | |
1223 | ||
b7569147 | 1224 | 5: |
1225 | // FIXME: overflow corner case | |
1226 | sub temp, temp, height_minor_b | |
1227 | bics height_minor_b, #3 | |
1228 | add temp, temp, height_minor_b | |
e1f6de8f | 1229 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
b7569147 | 1230 | bne 2b |
1231 | bal 1b | |
1232 | ||
75e28f62 E |
1233 | #undef span_uvrg_offset |
1234 | #undef span_edge_data | |
1235 | #undef span_b_offset | |
1236 | #undef left_x | |
1237 | #undef b | |
1238 | ||
1239 | #define psx_gpu r0 | |
1240 | #define num_spans r1 | |
1241 | #define span_uvrg_offset r2 | |
1242 | #define span_edge_data r3 | |
1243 | #define span_b_offset r4 | |
1244 | #define b_dx r5 | |
1245 | #define span_num_blocks r6 | |
1246 | #define y r7 | |
1247 | #define left_x r8 | |
1248 | #define b r9 | |
1249 | #define dither_offset_ptr r10 | |
1250 | #define block_ptr_a r11 | |
1251 | #define fb_ptr r12 | |
1252 | #define num_blocks r14 | |
1253 | ||
1254 | #define uvrg_dx_ptr r2 | |
1255 | #define texture_mask_ptr r3 | |
1256 | #define dither_shift r8 | |
1257 | #define dither_row r10 | |
1258 | ||
1259 | #define c_32 r7 | |
1260 | #define b_dx4 r8 | |
1261 | #define b_dx8 r9 | |
1262 | #define block_ptr_b r10 | |
1263 | ||
1264 | #define block_span_ptr r10 | |
1265 | #define right_mask r8 | |
1266 | ||
1267 | #define color r2 | |
1268 | #define color_r r3 | |
1269 | #define color_g r4 | |
1270 | #define color_b r5 | |
1271 | ||
1272 | #undef uvrg | |
1273 | ||
1274 | #define u_block q0 | |
1275 | #define v_block q1 | |
1276 | #define r_block q2 | |
1277 | #define g_block q3 | |
1278 | #define b_block q4 | |
1279 | ||
1280 | #define uv_dx4 d10 | |
1281 | #define rg_dx4 d11 | |
1282 | #define uv_dx8 d12 | |
1283 | #define rg_dx8 d13 | |
1284 | #define b_whole_8 d14 | |
1285 | #define fb_mask_ptrs d15 | |
1286 | ||
1287 | #define uvrg_dx4 q5 | |
1288 | #define uvrg_dx8 q6 | |
1289 | #define uv_dx8 d12 | |
1290 | #define rg_dx8 d13 | |
1291 | ||
1292 | #define u_whole q8 | |
1293 | #define v_whole q9 | |
1294 | #define r_whole q10 | |
1295 | #define g_whole q11 | |
1296 | #define b_whole q12 | |
1297 | ||
1298 | #define u_whole_low d16 | |
1299 | #define u_whole_high d17 | |
1300 | #define v_whole_low d18 | |
1301 | #define v_whole_high d19 | |
1302 | #define r_whole_low d20 | |
1303 | #define r_whole_high d21 | |
1304 | #define g_whole_low d22 | |
1305 | #define g_whole_high d23 | |
1306 | #define b_whole_low d24 | |
1307 | #define b_whole_high d25 | |
1308 | ||
1309 | #define dx4 q13 | |
1310 | #define dx8 q13 | |
1311 | ||
1312 | #define u_whole_8 d26 | |
1313 | #define v_whole_8 d27 | |
1314 | #define u_whole_8b d24 | |
1315 | #define r_whole_8 d24 | |
1316 | #define g_whole_8 d25 | |
1317 | ||
1318 | #define uv_whole_8 q13 | |
1319 | #define uv_whole_8b q14 | |
1320 | ||
1321 | #define dither_offsets q14 | |
1322 | #define texture_mask q15 | |
1323 | #define texture_mask_u d30 | |
1324 | #define texture_mask_v d31 | |
1325 | ||
1326 | #define dither_offsets_short d28 | |
1327 | ||
1328 | #define v_left_x q8 | |
1329 | #define uvrg q9 | |
1330 | #define block_span q10 | |
1331 | ||
1332 | #define uv d18 | |
1333 | #define rg d19 | |
1334 | ||
1335 | #define draw_mask q1 | |
1336 | #define draw_mask_edge q13 | |
1337 | #define test_mask q0 | |
1338 | ||
1339 | #define uvrg_dx q3 | |
1340 | ||
1341 | #define colors q2 | |
1342 | ||
1343 | #define setup_blocks_texture_swizzled() \ | |
1344 | vand.u8 u_whole_8b, u_whole_8, texture_mask_u; \ | |
1345 | vsli.u8 u_whole_8, v_whole_8, #4; \ | |
1346 | vsri.u8 v_whole_8, u_whole_8b, #4 \ | |
1347 | ||
1348 | #define setup_blocks_texture_unswizzled() \ | |
1349 | ||
1350 | ||
1351 | #define setup_blocks_shaded_textured_builder(swizzling) \ | |
1352 | .align 3; \ | |
1353 | \ | |
1354 | function(setup_blocks_shaded_textured_dithered_##swizzling##_indirect) \ | |
e1f6de8f | 1355 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
1356 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
1357 | \ | |
e1f6de8f | 1358 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
75e28f62 E |
1359 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
1360 | \ | |
1361 | cmp num_spans, #0; \ | |
1362 | bxeq lr; \ | |
1363 | \ | |
1364 | stmdb sp!, { r4 - r11, r14 }; \ | |
3f0189c6 | 1365 | save_abi_regs(); \ |
75e28f62 E |
1366 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
1367 | \ | |
e1f6de8f | 1368 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
75e28f62 E |
1369 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ |
1370 | \ | |
e1f6de8f | 1371 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
75e28f62 E |
1372 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
1373 | \ | |
e1f6de8f | 1374 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1375 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
1376 | \ | |
1377 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
1378 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1379 | \ | |
1380 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
1381 | \ | |
1382 | 0: \ | |
1383 | vmov.u8 fb_mask_ptrs, #0; \ | |
1384 | \ | |
e1f6de8f | 1385 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
1386 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
1387 | \ | |
e1f6de8f | 1388 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
1389 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
1390 | \ |
1391 | cmp span_num_blocks, #0; \ | |
1392 | beq 1f; \ | |
1393 | \ | |
e1f6de8f | 1394 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
1395 | add num_blocks, span_num_blocks, num_blocks; \ |
1396 | \ | |
1397 | cmp num_blocks, #MAX_BLOCKS; \ | |
1398 | bgt 2f; \ | |
1399 | \ | |
1400 | 3: \ | |
e1f6de8f | 1401 | ldr b, [span_b_offset]; \ |
75e28f62 E |
1402 | add fb_ptr, fb_ptr, y, lsl #11; \ |
1403 | \ | |
1404 | vdup.u32 v_left_x, left_x; \ | |
1405 | and y, y, #0x3; \ | |
1406 | \ | |
e1f6de8f | 1407 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
1408 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
1409 | \ | |
1410 | mla b, b_dx, left_x, b; \ | |
1411 | and dither_shift, left_x, #0x03; \ | |
1412 | \ | |
e1f6de8f | 1413 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
1414 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
1415 | \ | |
1416 | mov dither_shift, dither_shift, lsl #3; \ | |
1417 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
1418 | \ | |
1419 | mov c_32, #32; \ | |
1420 | subs span_num_blocks, span_num_blocks, #1; \ | |
1421 | \ | |
1422 | mov dither_row, dither_row, ror dither_shift; \ | |
1423 | mov b_dx4, b_dx, lsl #2; \ | |
1424 | \ | |
1425 | vdup.u32 dither_offsets_short, dither_row; \ | |
1426 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
1427 | \ | |
1428 | vdup.u32 b_block, b; \ | |
1429 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
1430 | \ | |
1431 | vdup.u32 u_block, uv[0]; \ | |
1432 | mov b_dx8, b_dx, lsl #3; \ | |
1433 | \ | |
1434 | vdup.u32 v_block, uv[1]; \ | |
1435 | vdup.u32 r_block, rg[0]; \ | |
1436 | vdup.u32 g_block, rg[1]; \ | |
1437 | \ | |
e1f6de8f | 1438 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1439 | \ |
1440 | vadd.u32 u_block, u_block, block_span; \ | |
e1f6de8f | 1441 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1442 | \ |
1443 | vadd.u32 v_block, v_block, block_span; \ | |
e1f6de8f | 1444 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1445 | \ |
1446 | vadd.u32 r_block, r_block, block_span; \ | |
e1f6de8f | 1447 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1448 | \ |
1449 | vadd.u32 g_block, g_block, block_span; \ | |
e1f6de8f | 1450 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
75e28f62 E |
1451 | \ |
1452 | vadd.u32 b_block, b_block, block_span; \ | |
1453 | add block_ptr_b, block_ptr_a, #16; \ | |
1454 | \ | |
1455 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1456 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1457 | vshrn.u32 r_whole_low, r_block, #16; \ | |
1458 | vshrn.u32 g_whole_low, g_block, #16; \ | |
1459 | \ | |
1460 | vdup.u32 dx4, uv_dx4[0]; \ | |
1461 | vshrn.u32 b_whole_low, b_block, #16; \ | |
1462 | \ | |
1463 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1464 | vdup.u32 dx4, uv_dx4[1]; \ | |
1465 | \ | |
1466 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1467 | vdup.u32 dx4, rg_dx4[0]; \ | |
1468 | \ | |
1469 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
1470 | vdup.u32 dx4, rg_dx4[1]; \ | |
1471 | \ | |
1472 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
1473 | vdup.u32 dx4, b_dx4; \ | |
1474 | \ | |
1475 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
1476 | vdup.u32 dx8, uv_dx8[0]; \ | |
1477 | \ | |
1478 | vadd.u32 u_block, u_block, dx8; \ | |
1479 | vdup.u32 dx8, uv_dx8[1]; \ | |
1480 | \ | |
1481 | vadd.u32 v_block, v_block, dx8; \ | |
1482 | vdup.u32 dx8, rg_dx8[0]; \ | |
1483 | \ | |
1484 | vadd.u32 r_block, r_block, dx8; \ | |
1485 | vdup.u32 dx8, rg_dx8[1]; \ | |
1486 | \ | |
1487 | vadd.u32 g_block, g_block, dx8; \ | |
1488 | vdup.u32 dx8, b_dx8; \ | |
1489 | \ | |
1490 | vadd.u32 b_block, b_block, dx8; \ | |
1491 | vmovn.u16 u_whole_8, u_whole; \ | |
1492 | \ | |
1493 | vmovn.u16 v_whole_8, v_whole; \ | |
1494 | \ | |
1495 | vmovn.u16 b_whole_8, b_whole; \ | |
e1f6de8f | 1496 | pld [fb_ptr]; \ |
75e28f62 E |
1497 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
1498 | \ | |
1499 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1500 | setup_blocks_texture_##swizzling(); \ | |
1501 | \ | |
1502 | vmovn.u16 r_whole_8, r_whole; \ | |
1503 | beq 5f; \ | |
1504 | \ | |
1505 | 4: \ | |
1506 | vmovn.u16 g_whole_8, g_whole; \ | |
1507 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1508 | \ | |
e1f6de8f | 1509 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1510 | vshrn.u32 v_whole_low, v_block, #16; \ |
1511 | \ | |
e1f6de8f | 1512 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
75e28f62 E |
1513 | vshrn.u32 r_whole_low, r_block, #16; \ |
1514 | \ | |
e1f6de8f | 1515 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1516 | vshrn.u32 g_whole_low, g_block, #16; \ |
1517 | \ | |
1518 | vdup.u32 dx4, uv_dx4[0]; \ | |
1519 | vshrn.u32 b_whole_low, b_block, #16; \ | |
1520 | \ | |
1521 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1522 | vdup.u32 dx4, uv_dx4[1]; \ | |
1523 | \ | |
1524 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1525 | vdup.u32 dx4, rg_dx4[0]; \ | |
1526 | \ | |
1527 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
1528 | vdup.u32 dx4, rg_dx4[1]; \ | |
1529 | \ | |
1530 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
1531 | vdup.u32 dx4, b_dx4; \ | |
1532 | \ | |
1533 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
1534 | vdup.u32 dx8, uv_dx8[0]; \ | |
1535 | \ | |
1536 | vadd.u32 u_block, u_block, dx8; \ | |
1537 | vdup.u32 dx8, uv_dx8[1]; \ | |
1538 | \ | |
1539 | vadd.u32 v_block, v_block, dx8; \ | |
1540 | vdup.u32 dx8, rg_dx8[0]; \ | |
1541 | \ | |
1542 | vadd.u32 r_block, r_block, dx8; \ | |
1543 | vdup.u32 dx8, rg_dx8[1]; \ | |
1544 | \ | |
1545 | vadd.u32 g_block, g_block, dx8; \ | |
1546 | vdup.u32 dx8, b_dx8; \ | |
1547 | \ | |
1548 | vadd.u32 b_block, b_block, dx8; \ | |
1549 | vmovn.u16 u_whole_8, u_whole; \ | |
1550 | \ | |
1551 | add fb_ptr, fb_ptr, #16; \ | |
1552 | vmovn.u16 v_whole_8, v_whole; \ | |
1553 | \ | |
e1f6de8f | 1554 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
75e28f62 E |
1555 | vmovn.u16 b_whole_8, b_whole; \ |
1556 | \ | |
e1f6de8f | 1557 | pld [fb_ptr]; \ |
75e28f62 E |
1558 | \ |
1559 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1560 | subs span_num_blocks, span_num_blocks, #1; \ | |
1561 | \ | |
1562 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1563 | setup_blocks_texture_##swizzling(); \ | |
1564 | \ | |
1565 | vmovn.u16 r_whole_8, r_whole; \ | |
1566 | bne 4b; \ | |
1567 | \ | |
1568 | 5: \ | |
1569 | vmovn.u16 g_whole_8, g_whole; \ | |
e1f6de8f | 1570 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 | 1571 | \ |
e1f6de8f | 1572 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
1573 | vdup.u8 draw_mask, right_mask; \ |
1574 | \ | |
1575 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
1576 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
1577 | vzip.u8 u_whole_8, v_whole_8; \ | |
1578 | \ | |
1579 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
e1f6de8f | 1580 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
1581 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ | |
1582 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
1583 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
75e28f62 E |
1584 | \ |
1585 | 1: \ | |
1586 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
1587 | add span_b_offset, span_b_offset, #4; \ | |
1588 | \ | |
1589 | add span_edge_data, span_edge_data, #8; \ | |
1590 | subs num_spans, num_spans, #1; \ | |
1591 | \ | |
e1f6de8f | 1592 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1593 | bne 0b; \ |
1594 | \ | |
3f0189c6 | 1595 | restore_abi_regs(); \ |
75e28f62 E |
1596 | ldmia sp!, { r4 - r11, pc }; \ |
1597 | \ | |
1598 | 2: \ | |
1599 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
1600 | vpush { texture_mask }; \ | |
1601 | vpush { uvrg_dx4 }; \ | |
1602 | \ | |
4d646738 | 1603 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 | 1604 | bl flush_render_block_buffer; \ |
4d646738 | 1605 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 E |
1606 | \ |
1607 | vpop { uvrg_dx4 }; \ | |
1608 | vpop { texture_mask }; \ | |
1609 | \ | |
1610 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
1611 | vmov.u8 fb_mask_ptrs, #0; \ | |
1612 | \ | |
1613 | mov num_blocks, span_num_blocks; \ | |
1614 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1615 | bal 3b \ | |
1616 | ||
1617 | ||
1618 | setup_blocks_shaded_textured_builder(swizzled) | |
1619 | setup_blocks_shaded_textured_builder(unswizzled) | |
1620 | ||
1621 | ||
1622 | #define setup_blocks_unshaded_textured_builder(swizzling) \ | |
1623 | .align 3; \ | |
1624 | \ | |
1625 | function(setup_blocks_unshaded_textured_dithered_##swizzling##_indirect) \ | |
e1f6de8f | 1626 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
1627 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
1628 | \ | |
e1f6de8f | 1629 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
75e28f62 E |
1630 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
1631 | \ | |
1632 | cmp num_spans, #0; \ | |
1633 | bxeq lr; \ | |
1634 | \ | |
1635 | stmdb sp!, { r4 - r11, r14 }; \ | |
3f0189c6 | 1636 | save_abi_regs(); \ |
75e28f62 E |
1637 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
1638 | \ | |
1639 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
1640 | \ | |
e1f6de8f | 1641 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
75e28f62 E |
1642 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
1643 | \ | |
e1f6de8f | 1644 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1645 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
1646 | \ | |
1647 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1648 | \ | |
1649 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
1650 | \ | |
1651 | 0: \ | |
1652 | vmov.u8 fb_mask_ptrs, #0; \ | |
1653 | \ | |
e1f6de8f | 1654 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
1655 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
1656 | \ | |
e1f6de8f | 1657 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
1658 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
1659 | \ |
1660 | cmp span_num_blocks, #0; \ | |
1661 | beq 1f; \ | |
1662 | \ | |
e1f6de8f | 1663 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
1664 | add num_blocks, span_num_blocks, num_blocks; \ |
1665 | \ | |
1666 | cmp num_blocks, #MAX_BLOCKS; \ | |
1667 | bgt 2f; \ | |
1668 | \ | |
1669 | 3: \ | |
1670 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
1671 | \ | |
1672 | vdup.u32 v_left_x, left_x; \ | |
1673 | and y, y, #0x3; \ | |
1674 | \ | |
e1f6de8f | 1675 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
1676 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
1677 | \ | |
1678 | and dither_shift, left_x, #0x03; \ | |
1679 | \ | |
e1f6de8f | 1680 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
1681 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
1682 | \ | |
1683 | mov dither_shift, dither_shift, lsl #3; \ | |
1684 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
1685 | \ | |
1686 | mov c_32, #32; \ | |
1687 | subs span_num_blocks, span_num_blocks, #1; \ | |
1688 | \ | |
1689 | mov dither_row, dither_row, ror dither_shift; \ | |
1690 | \ | |
1691 | vdup.u32 dither_offsets_short, dither_row; \ | |
1692 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
1693 | \ | |
1694 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
1695 | \ | |
1696 | vdup.u32 u_block, uv[0]; \ | |
1697 | \ | |
1698 | vdup.u32 v_block, uv[1]; \ | |
e1f6de8f | 1699 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1700 | \ |
1701 | vadd.u32 u_block, u_block, block_span; \ | |
e1f6de8f | 1702 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1703 | \ |
1704 | vadd.u32 v_block, v_block, block_span; \ | |
1705 | add block_ptr_b, block_ptr_a, #16; \ | |
1706 | \ | |
1707 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1708 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1709 | \ | |
1710 | vdup.u32 dx4, uv_dx4[0]; \ | |
1711 | \ | |
1712 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1713 | vdup.u32 dx4, uv_dx4[1]; \ | |
1714 | \ | |
1715 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1716 | vdup.u32 dx8, uv_dx8[0]; \ | |
1717 | \ | |
1718 | vadd.u32 u_block, u_block, dx8; \ | |
1719 | vdup.u32 dx8, uv_dx8[1]; \ | |
1720 | \ | |
1721 | vadd.u32 v_block, v_block, dx8; \ | |
1722 | vmovn.u16 u_whole_8, u_whole; \ | |
1723 | \ | |
1724 | vmovn.u16 v_whole_8, v_whole; \ | |
1725 | \ | |
e1f6de8f | 1726 | pld [fb_ptr]; \ |
75e28f62 E |
1727 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
1728 | \ | |
1729 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1730 | setup_blocks_texture_##swizzling(); \ | |
1731 | \ | |
1732 | beq 5f; \ | |
1733 | \ | |
1734 | 4: \ | |
1735 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1736 | \ | |
e1f6de8f | 1737 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1738 | vshrn.u32 v_whole_low, v_block, #16; \ |
1739 | \ | |
1740 | add block_ptr_b, block_ptr_b, #32; \ | |
e1f6de8f | 1741 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1742 | \ |
1743 | vdup.u32 dx4, uv_dx4[0]; \ | |
1744 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1745 | vdup.u32 dx4, uv_dx4[1]; \ | |
1746 | \ | |
1747 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1748 | vdup.u32 dx8, uv_dx8[0]; \ | |
1749 | \ | |
1750 | vadd.u32 u_block, u_block, dx8; \ | |
1751 | vdup.u32 dx8, uv_dx8[1]; \ | |
1752 | \ | |
1753 | vadd.u32 v_block, v_block, dx8; \ | |
1754 | vmovn.u16 u_whole_8, u_whole; \ | |
1755 | \ | |
1756 | add fb_ptr, fb_ptr, #16; \ | |
1757 | vmovn.u16 v_whole_8, v_whole; \ | |
1758 | \ | |
e1f6de8f | 1759 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
1760 | pld [fb_ptr]; \ | |
75e28f62 E |
1761 | \ |
1762 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1763 | subs span_num_blocks, span_num_blocks, #1; \ | |
1764 | \ | |
1765 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1766 | setup_blocks_texture_##swizzling(); \ | |
1767 | \ | |
1768 | bne 4b; \ | |
1769 | \ | |
1770 | 5: \ | |
e1f6de8f | 1771 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 | 1772 | \ |
e1f6de8f | 1773 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
1774 | vdup.u8 draw_mask, right_mask; \ |
1775 | \ | |
1776 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
1777 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
1778 | vzip.u8 u_whole_8, v_whole_8; \ | |
1779 | \ | |
1780 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
1781 | add block_ptr_b, block_ptr_b, #32; \ | |
e1f6de8f | 1782 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ |
1783 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
1784 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
75e28f62 E |
1785 | \ |
1786 | 1: \ | |
1787 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
1788 | add span_edge_data, span_edge_data, #8; \ | |
1789 | subs num_spans, num_spans, #1; \ | |
1790 | \ | |
e1f6de8f | 1791 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1792 | bne 0b; \ |
1793 | \ | |
3f0189c6 | 1794 | restore_abi_regs(); \ |
75e28f62 E |
1795 | ldmia sp!, { r4 - r11, pc }; \ |
1796 | \ | |
1797 | 2: \ | |
1798 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
1799 | vpush { texture_mask }; \ | |
1800 | vpush { uvrg_dx4 }; \ | |
1801 | \ | |
4d646738 | 1802 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 | 1803 | bl flush_render_block_buffer; \ |
4d646738 | 1804 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 E |
1805 | \ |
1806 | vpop { uvrg_dx4 }; \ | |
1807 | vpop { texture_mask }; \ | |
1808 | \ | |
1809 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
1810 | vmov.u8 fb_mask_ptrs, #0; \ | |
1811 | \ | |
1812 | mov num_blocks, span_num_blocks; \ | |
1813 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1814 | bal 3b \ | |
1815 | ||
1816 | ||
1817 | setup_blocks_unshaded_textured_builder(swizzled) | |
1818 | setup_blocks_unshaded_textured_builder(unswizzled) | |
1819 | ||
1820 | ||
1821 | .align 3 | |
1822 | ||
1823 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_indirect) | |
e1f6de8f | 1824 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1825 | veor.u32 draw_mask, draw_mask, draw_mask |
1826 | ||
1827 | cmp num_spans, #0 | |
1828 | bxeq lr | |
1829 | ||
1830 | stmdb sp!, { r4 - r11, r14 } | |
3f0189c6 | 1831 | save_abi_regs() |
e1f6de8f | 1832 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 | 1833 | |
e1f6de8f | 1834 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
75e28f62 E |
1835 | |
1836 | ubfx color_r, color, #3, #5 | |
1837 | ubfx color_g, color, #11, #5 | |
1838 | ubfx color_b, color, #19, #5 | |
1839 | ||
1840 | orr color, color_r, color_b, lsl #10 | |
1841 | orr color, color, color_g, lsl #5 | |
1842 | ||
1843 | vdup.u16 colors, color | |
1844 | ||
e1f6de8f | 1845 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
1846 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset |
1847 | ||
1848 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
1849 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6 | |
1850 | ||
1851 | 0: | |
e1f6de8f | 1852 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
1853 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
75e28f62 | 1854 | |
e1f6de8f | 1855 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
75e28f62 E |
1856 | |
1857 | cmp span_num_blocks, #0 | |
1858 | beq 1f | |
1859 | ||
e1f6de8f | 1860 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
75e28f62 E |
1861 | add num_blocks, span_num_blocks, num_blocks |
1862 | ||
1863 | cmp num_blocks, #MAX_BLOCKS | |
1864 | bgt 2f | |
1865 | ||
1866 | 3: | |
1867 | add fb_ptr, fb_ptr, y, lsl #11 | |
1868 | and y, y, #0x3 | |
1869 | ||
1870 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
1871 | mov c_32, #32 | |
1872 | ||
1873 | subs span_num_blocks, span_num_blocks, #1 | |
1874 | ||
1875 | add block_ptr_b, block_ptr_a, #16 | |
e1f6de8f | 1876 | pld [fb_ptr] |
75e28f62 E |
1877 | |
1878 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
1879 | beq 5f | |
1880 | ||
1881 | 4: | |
e1f6de8f | 1882 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_32 |
1883 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 | |
1884 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 | |
75e28f62 E |
1885 | |
1886 | add fb_ptr, fb_ptr, #16 | |
1887 | add block_ptr_b, block_ptr_b, #32 | |
1888 | ||
e1f6de8f | 1889 | pld [fb_ptr] |
75e28f62 E |
1890 | |
1891 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
1892 | subs span_num_blocks, span_num_blocks, #1 | |
1893 | ||
1894 | bne 4b | |
1895 | ||
1896 | 5: | |
e1f6de8f | 1897 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset] |
75e28f62 E |
1898 | |
1899 | vdup.u8 draw_mask_edge, right_mask | |
1900 | vtst.u16 draw_mask_edge, draw_mask_edge, test_mask | |
1901 | ||
e1f6de8f | 1902 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 |
1903 | vst1.u32 { draw_mask_edge }, [block_ptr_a, :128], c_32 | |
75e28f62 | 1904 | add block_ptr_b, block_ptr_b, #32 |
e1f6de8f | 1905 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 |
75e28f62 E |
1906 | |
1907 | 1: | |
1908 | add span_edge_data, span_edge_data, #8 | |
1909 | subs num_spans, num_spans, #1 | |
1910 | ||
e1f6de8f | 1911 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
1912 | bne 0b |
1913 | ||
3f0189c6 | 1914 | restore_abi_regs() |
75e28f62 E |
1915 | ldmia sp!, { r4 - r11, pc } |
1916 | ||
1917 | 2: | |
1918 | vpush { colors } | |
1919 | ||
4d646738 | 1920 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 1921 | bl flush_render_block_buffer |
4d646738 | 1922 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 E |
1923 | |
1924 | vpop { colors } | |
1925 | ||
e1f6de8f | 1926 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 E |
1927 | veor.u32 draw_mask, draw_mask, draw_mask |
1928 | ||
1929 | mov num_blocks, span_num_blocks | |
1930 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
1931 | bal 3b | |
1932 | ||
1933 | ||
1934 | #define mask_msb_scalar r14 | |
1935 | ||
1936 | #define msb_mask q15 | |
1937 | ||
1938 | #define pixels_low d16 | |
1939 | ||
1940 | #define msb_mask_low d30 | |
1941 | #define msb_mask_high d31 | |
1942 | ||
1943 | ||
1944 | .align 3 | |
1945 | ||
1946 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_direct) | |
e1f6de8f | 1947 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1948 | |
1949 | cmp num_spans, #0 | |
1950 | bxeq lr | |
1951 | ||
1952 | stmdb sp!, { r4 - r11, r14 } | |
1953 | ||
e1f6de8f | 1954 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
75e28f62 E |
1955 | |
1956 | ubfx color_r, color, #3, #5 | |
1957 | ubfx color_g, color, #11, #5 | |
1958 | ||
e1f6de8f | 1959 | ldrh mask_msb_scalar, [psx_gpu, #psx_gpu_mask_msb_offset] |
75e28f62 E |
1960 | ubfx color_b, color, #19, #5 |
1961 | ||
1962 | orr color, color_r, color_b, lsl #10 | |
1963 | orr color, color, color_g, lsl #5 | |
1964 | orr color, color, mask_msb_scalar | |
1965 | ||
1966 | vdup.u16 colors, color | |
1967 | ||
1968 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
ed0fd81d | 1969 | orr color, color, color, lsl #16 |
3867c6ef | 1970 | |
75e28f62 E |
1971 | |
1972 | 0: | |
e1f6de8f | 1973 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
1974 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
75e28f62 | 1975 | |
e1f6de8f | 1976 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
75e28f62 E |
1977 | |
1978 | cmp span_num_blocks, #0 | |
1979 | beq 1f | |
1980 | ||
e1f6de8f | 1981 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
75e28f62 E |
1982 | |
1983 | add fb_ptr, fb_ptr, y, lsl #11 | |
1984 | subs span_num_blocks, span_num_blocks, #1 | |
1985 | ||
1986 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
1987 | beq 3f | |
1988 | ||
1989 | 2: | |
e1f6de8f | 1990 | vst1.u32 { colors }, [fb_ptr]! |
75e28f62 E |
1991 | subs span_num_blocks, span_num_blocks, #1 |
1992 | ||
1993 | bne 2b | |
1994 | ||
1995 | 3: | |
e1f6de8f | 1996 | ldrb right_mask, [span_edge_data, #edge_data_right_mask_offset] |
75e28f62 | 1997 | |
3867c6ef E |
1998 | cmp right_mask, #0x0 |
1999 | beq 5f | |
2000 | ||
2001 | tst right_mask, #0xF | |
e1f6de8f | 2002 | streq color, [fb_ptr], #4 |
3867c6ef | 2003 | moveq right_mask, right_mask, lsr #4 |
e1f6de8f | 2004 | streq color, [fb_ptr], #4 |
3867c6ef E |
2005 | |
2006 | tst right_mask, #0x3 | |
e1f6de8f | 2007 | streq color, [fb_ptr], #4 |
3867c6ef E |
2008 | moveq right_mask, right_mask, lsr #2 |
2009 | ||
2010 | tst right_mask, #0x1 | |
e1f6de8f | 2011 | strheq color, [fb_ptr] |
75e28f62 E |
2012 | |
2013 | 1: | |
2014 | add span_edge_data, span_edge_data, #8 | |
2015 | subs num_spans, num_spans, #1 | |
75e28f62 E |
2016 | bne 0b |
2017 | ||
2018 | ldmia sp!, { r4 - r11, pc } | |
2019 | ||
3867c6ef | 2020 | 5: |
e1f6de8f | 2021 | vst1.u32 { colors }, [fb_ptr] |
3867c6ef | 2022 | bal 1b |
75e28f62 E |
2023 | |
2024 | ||
2025 | #undef c_64 | |
2026 | ||
2027 | #define c_64 r7 | |
2028 | #define rg_dx_ptr r2 | |
2029 | ||
2030 | ||
2031 | #undef r_block | |
2032 | #undef g_block | |
2033 | #undef b_block | |
2034 | #undef r_whole | |
2035 | #undef g_whole | |
2036 | #undef b_whole | |
2037 | #undef r_whole_low | |
2038 | #undef r_whole_high | |
2039 | #undef g_whole_low | |
2040 | #undef g_whole_high | |
2041 | #undef b_whole_low | |
2042 | #undef b_whole_high | |
2043 | #undef r_whole_8 | |
2044 | #undef g_whole_8 | |
2045 | #undef b_whole_8 | |
2046 | #undef dither_offsets | |
2047 | #undef rg_dx4 | |
2048 | #undef rg_dx8 | |
2049 | #undef dx4 | |
2050 | #undef dx8 | |
2051 | #undef v_left_x | |
2052 | #undef uvrg | |
2053 | #undef block_span | |
2054 | #undef rg | |
2055 | #undef draw_mask | |
2056 | #undef test_mask | |
2057 | ||
2058 | #define r_block q0 | |
2059 | #define g_block q1 | |
2060 | #define b_block q2 | |
2061 | ||
2062 | #define r_whole q3 | |
2063 | #define g_whole q4 | |
2064 | #define b_whole q5 | |
2065 | ||
2066 | #define r_whole_low d6 | |
2067 | #define r_whole_high d7 | |
2068 | #define g_whole_low d8 | |
2069 | #define g_whole_high d9 | |
2070 | #define b_whole_low d10 | |
2071 | #define b_whole_high d11 | |
2072 | ||
2073 | #define gb_whole_8 q6 | |
2074 | ||
2075 | #define g_whole_8 d12 | |
2076 | #define b_whole_8 d13 | |
2077 | ||
2078 | #define r_whole_8 d14 | |
2079 | ||
2080 | #define pixels q8 | |
2081 | ||
2082 | #define rg_dx4 d18 | |
2083 | #define rg_dx8 d19 | |
2084 | ||
2085 | #define dx4 q10 | |
2086 | #define dx8 q10 | |
2087 | ||
2088 | #define v_left_x d6 | |
2089 | #define uvrg q4 | |
2090 | #define block_span q5 | |
2091 | ||
2092 | #define rg d9 | |
2093 | ||
2094 | #define d64_1 d22 | |
2095 | #define d64_128 d23 | |
2096 | ||
2097 | #define d128_4 q12 | |
2098 | #define d128_0x7 q13 | |
2099 | ||
2100 | #define d64_4 d24 | |
2101 | ||
2102 | #define dither_offsets q14 | |
2103 | #define draw_mask q15 | |
2104 | ||
2105 | #define dither_offsets_low d28 | |
2106 | ||
2107 | #define rg_dx d0 | |
2108 | #define test_mask q10 | |
2109 | ||
2110 | ||
2111 | #define setup_blocks_shaded_untextured_dither_a_dithered() \ | |
2112 | vqadd.u8 r_whole_8, r_whole_8, dither_offsets_low; \ | |
2113 | vqadd.u8 gb_whole_8, gb_whole_8, dither_offsets; \ | |
2114 | ||
2115 | #define setup_blocks_shaded_untextured_dither_b_dithered() \ | |
2116 | vqsub.u8 r_whole_8, r_whole_8, d64_4; \ | |
2117 | vqsub.u8 gb_whole_8, gb_whole_8, d128_4 \ | |
2118 | ||
2119 | #define setup_blocks_shaded_untextured_dither_a_undithered() \ | |
2120 | ||
2121 | #define setup_blocks_shaded_untextured_dither_b_undithered() \ | |
2122 | ||
2123 | ||
2124 | #define setup_blocks_shaded_untextured_indirect_builder(dithering) \ | |
2125 | .align 3; \ | |
2126 | \ | |
2127 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_indirect) \ | |
e1f6de8f | 2128 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
2129 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
2130 | \ | |
e1f6de8f | 2131 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
75e28f62 E |
2132 | \ |
2133 | cmp num_spans, #0; \ | |
2134 | bxeq lr; \ | |
2135 | \ | |
2136 | stmdb sp!, { r4 - r11, r14 }; \ | |
3f0189c6 | 2137 | save_abi_regs(); \ |
75e28f62 E |
2138 | vshl.u32 rg_dx4, rg_dx, #2; \ |
2139 | \ | |
e1f6de8f | 2140 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
75e28f62 E |
2141 | vshl.u32 rg_dx8, rg_dx, #3; \ |
2142 | \ | |
2143 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
2144 | \ | |
e1f6de8f | 2145 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
2146 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
2147 | \ | |
2148 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
2149 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
2150 | \ | |
2151 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
2152 | vmov.u8 d64_1, #1; \ | |
2153 | \ | |
2154 | vmov.u8 d128_4, #4; \ | |
2155 | vmov.u8 d64_128, #128; \ | |
2156 | \ | |
2157 | vmov.u8 d128_0x7, #0x7; \ | |
2158 | \ | |
2159 | 0: \ | |
e1f6de8f | 2160 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
2161 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
2162 | \ | |
e1f6de8f | 2163 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
2164 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
2165 | \ |
2166 | cmp span_num_blocks, #0; \ | |
2167 | beq 1f; \ | |
2168 | \ | |
e1f6de8f | 2169 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
2170 | add num_blocks, span_num_blocks, num_blocks; \ |
2171 | \ | |
2172 | cmp num_blocks, #MAX_BLOCKS; \ | |
2173 | bgt 2f; \ | |
2174 | \ | |
2175 | 3: \ | |
e1f6de8f | 2176 | ldr b, [span_b_offset]; \ |
75e28f62 E |
2177 | add fb_ptr, fb_ptr, y, lsl #11; \ |
2178 | \ | |
2179 | vdup.u32 v_left_x, left_x; \ | |
2180 | and y, y, #0x3; \ | |
2181 | \ | |
e1f6de8f | 2182 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
2183 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
2184 | \ | |
2185 | mla b, b_dx, left_x, b; \ | |
2186 | and dither_shift, left_x, #0x03; \ | |
2187 | \ | |
e1f6de8f | 2188 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
2189 | vshr.u32 rg_dx, rg_dx4, #2; \ |
2190 | \ | |
2191 | mov dither_shift, dither_shift, lsl #3; \ | |
2192 | vmla.u32 rg, rg_dx, v_left_x; \ | |
2193 | \ | |
2194 | mov c_64, #64; \ | |
2195 | subs span_num_blocks, span_num_blocks, #1; \ | |
2196 | \ | |
2197 | mov dither_row, dither_row, ror dither_shift; \ | |
2198 | mov b_dx4, b_dx, lsl #2; \ | |
2199 | \ | |
2200 | vdup.u32 dither_offsets, dither_row; \ | |
2201 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
2202 | \ | |
2203 | vdup.u32 b_block, b; \ | |
2204 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
2205 | \ | |
2206 | mov b_dx8, b_dx, lsl #3; \ | |
2207 | vdup.u32 r_block, rg[0]; \ | |
2208 | vdup.u32 g_block, rg[1]; \ | |
2209 | \ | |
e1f6de8f | 2210 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2211 | \ |
2212 | vadd.u32 r_block, r_block, block_span; \ | |
e1f6de8f | 2213 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2214 | \ |
2215 | vadd.u32 g_block, g_block, block_span; \ | |
e1f6de8f | 2216 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
75e28f62 E |
2217 | \ |
2218 | vadd.u32 b_block, b_block, block_span; \ | |
2219 | add block_ptr_b, block_ptr_a, #16; \ | |
2220 | \ | |
2221 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2222 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2223 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2224 | vdup.u32 dx4, rg_dx4[0]; \ | |
2225 | \ | |
2226 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2227 | vdup.u32 dx4, rg_dx4[1]; \ | |
2228 | \ | |
2229 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2230 | vdup.u32 dx4, b_dx4; \ | |
2231 | \ | |
2232 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2233 | vdup.u32 dx8, rg_dx8[0]; \ | |
2234 | \ | |
2235 | vadd.u32 r_block, r_block, dx8; \ | |
2236 | vdup.u32 dx8, rg_dx8[1]; \ | |
2237 | \ | |
2238 | vadd.u32 g_block, g_block, dx8; \ | |
2239 | vdup.u32 dx8, b_dx8; \ | |
2240 | \ | |
2241 | vadd.u32 b_block, b_block, dx8; \ | |
2242 | \ | |
2243 | vmovn.u16 r_whole_8, r_whole; \ | |
2244 | vmovn.u16 g_whole_8, g_whole; \ | |
2245 | vmovn.u16 b_whole_8, b_whole; \ | |
2246 | \ | |
2247 | beq 5f; \ | |
2248 | veor.u32 draw_mask, draw_mask, draw_mask; \ | |
2249 | \ | |
2250 | 4: \ | |
2251 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2252 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2253 | \ | |
2254 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
2255 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2256 | \ | |
2257 | vshrn.u32 b_whole_low, b_block, #16; \ | |
e1f6de8f | 2258 | str fb_ptr, [block_ptr_a, #44]; \ |
75e28f62 E |
2259 | \ |
2260 | vdup.u32 dx4, rg_dx4[0]; \ | |
2261 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2262 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
2263 | \ | |
2264 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2265 | vdup.u32 dx4, rg_dx4[1]; \ | |
2266 | \ | |
2267 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2268 | vdup.u32 dx4, b_dx4; \ | |
2269 | \ | |
2270 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2271 | vdup.u32 dx8, rg_dx8[0]; \ | |
2272 | \ | |
2273 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
2274 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2275 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2276 | \ | |
2277 | vadd.u32 r_block, r_block, dx8; \ | |
2278 | vdup.u32 dx8, rg_dx8[1]; \ | |
2279 | \ | |
2280 | vadd.u32 g_block, g_block, dx8; \ | |
2281 | vdup.u32 dx8, b_dx8; \ | |
2282 | \ | |
2283 | vadd.u32 b_block, b_block, dx8; \ | |
2284 | add fb_ptr, fb_ptr, #16; \ | |
2285 | \ | |
2286 | vmovn.u16 r_whole_8, r_whole; \ | |
2287 | vmovn.u16 g_whole_8, g_whole; \ | |
2288 | vmovn.u16 b_whole_8, b_whole; \ | |
2289 | \ | |
e1f6de8f | 2290 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
2291 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
75e28f62 | 2292 | \ |
e1f6de8f | 2293 | pld [fb_ptr]; \ |
75e28f62 E |
2294 | \ |
2295 | subs span_num_blocks, span_num_blocks, #1; \ | |
2296 | bne 4b; \ | |
2297 | \ | |
2298 | 5: \ | |
e1f6de8f | 2299 | str fb_ptr, [block_ptr_a, #44]; \ |
75e28f62 E |
2300 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ |
2301 | \ | |
e1f6de8f | 2302 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 E |
2303 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
2304 | \ | |
2305 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2306 | vdup.u8 draw_mask, right_mask; \ | |
2307 | \ | |
2308 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
e1f6de8f | 2309 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
2310 | \ |
2311 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
2312 | \ | |
2313 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
2314 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2315 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2316 | \ | |
e1f6de8f | 2317 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
2318 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
75e28f62 E |
2319 | \ |
2320 | 1: \ | |
2321 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
2322 | add span_b_offset, span_b_offset, #4; \ | |
2323 | \ | |
2324 | add span_edge_data, span_edge_data, #8; \ | |
2325 | subs num_spans, num_spans, #1; \ | |
2326 | \ | |
e1f6de8f | 2327 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
2328 | bne 0b; \ |
2329 | \ | |
3f0189c6 | 2330 | restore_abi_regs(); \ |
75e28f62 E |
2331 | ldmia sp!, { r4 - r11, pc }; \ |
2332 | \ | |
2333 | 2: \ | |
2334 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
2335 | vpush { rg_dx4 }; \ | |
2336 | \ | |
4d646738 | 2337 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 | 2338 | bl flush_render_block_buffer; \ |
4d646738 | 2339 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 E |
2340 | \ |
2341 | vpop { rg_dx4 }; \ | |
2342 | \ | |
2343 | vmov.u8 d64_1, #1; \ | |
2344 | vmov.u8 d128_4, #4; \ | |
2345 | vmov.u8 d64_128, #128; \ | |
2346 | vmov.u8 d128_0x7, #0x7; \ | |
2347 | \ | |
2348 | vadd.u32 rg_dx8, rg_dx4, rg_dx4; \ | |
2349 | \ | |
2350 | mov num_blocks, span_num_blocks; \ | |
2351 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
2352 | bal 3b \ | |
2353 | ||
2354 | ||
2355 | setup_blocks_shaded_untextured_indirect_builder(undithered) | |
2356 | setup_blocks_shaded_untextured_indirect_builder(dithered) | |
2357 | ||
2358 | ||
2359 | #undef draw_mask | |
2360 | ||
2361 | #define mask_msb_ptr r14 | |
2362 | ||
2363 | #define draw_mask q0 | |
2364 | #define pixels_low d16 | |
3867c6ef | 2365 | #define pixels_high d17 |
75e28f62 E |
2366 | |
2367 | ||
2368 | ||
2369 | #define setup_blocks_shaded_untextured_direct_builder(dithering) \ | |
2370 | .align 3; \ | |
2371 | \ | |
2372 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_direct) \ | |
e1f6de8f | 2373 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
2374 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
2375 | \ | |
e1f6de8f | 2376 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
75e28f62 E |
2377 | \ |
2378 | cmp num_spans, #0; \ | |
2379 | bxeq lr; \ | |
2380 | \ | |
2381 | stmdb sp!, { r4 - r11, r14 }; \ | |
3f0189c6 | 2382 | save_abi_regs(); \ |
75e28f62 E |
2383 | vshl.u32 rg_dx4, rg_dx, #2; \ |
2384 | \ | |
e1f6de8f | 2385 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
75e28f62 E |
2386 | vshl.u32 rg_dx8, rg_dx, #3; \ |
2387 | \ | |
2388 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
2389 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
2390 | \ | |
2391 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
2392 | vmov.u8 d64_1, #1; \ | |
2393 | \ | |
2394 | vmov.u8 d128_4, #4; \ | |
2395 | vmov.u8 d64_128, #128; \ | |
2396 | \ | |
2397 | vmov.u8 d128_0x7, #0x7; \ | |
2398 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
e1f6de8f | 2399 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
2400 | \ |
2401 | 0: \ | |
e1f6de8f | 2402 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
2403 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
2404 | \ | |
e1f6de8f | 2405 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
2406 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
2407 | \ |
2408 | cmp span_num_blocks, #0; \ | |
2409 | beq 1f; \ | |
2410 | \ | |
e1f6de8f | 2411 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
2412 | add fb_ptr, fb_ptr, y, lsl #11; \ |
2413 | \ | |
e1f6de8f | 2414 | ldr b, [span_b_offset]; \ |
75e28f62 E |
2415 | vdup.u32 v_left_x, left_x; \ |
2416 | and y, y, #0x3; \ | |
2417 | \ | |
e1f6de8f | 2418 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
2419 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
2420 | \ | |
2421 | mla b, b_dx, left_x, b; \ | |
2422 | and dither_shift, left_x, #0x03; \ | |
2423 | \ | |
e1f6de8f | 2424 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
2425 | vshr.u32 rg_dx, rg_dx4, #2; \ |
2426 | \ | |
2427 | mov dither_shift, dither_shift, lsl #3; \ | |
2428 | vmla.u32 rg, rg_dx, v_left_x; \ | |
2429 | \ | |
2430 | subs span_num_blocks, span_num_blocks, #1; \ | |
2431 | \ | |
2432 | mov dither_row, dither_row, ror dither_shift; \ | |
2433 | mov b_dx4, b_dx, lsl #2; \ | |
2434 | \ | |
2435 | vdup.u32 dither_offsets, dither_row; \ | |
2436 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
2437 | \ | |
2438 | vdup.u32 b_block, b; \ | |
2439 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
2440 | \ | |
2441 | mov b_dx8, b_dx, lsl #3; \ | |
2442 | vdup.u32 r_block, rg[0]; \ | |
2443 | vdup.u32 g_block, rg[1]; \ | |
2444 | \ | |
e1f6de8f | 2445 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2446 | \ |
2447 | vadd.u32 r_block, r_block, block_span; \ | |
e1f6de8f | 2448 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2449 | \ |
2450 | vadd.u32 g_block, g_block, block_span; \ | |
e1f6de8f | 2451 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
75e28f62 E |
2452 | \ |
2453 | vadd.u32 b_block, b_block, block_span; \ | |
2454 | add block_ptr_b, block_ptr_a, #16; \ | |
2455 | \ | |
2456 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2457 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2458 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2459 | vdup.u32 dx4, rg_dx4[0]; \ | |
2460 | \ | |
2461 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2462 | vdup.u32 dx4, rg_dx4[1]; \ | |
2463 | \ | |
2464 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2465 | vdup.u32 dx4, b_dx4; \ | |
2466 | \ | |
2467 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2468 | vdup.u32 dx8, rg_dx8[0]; \ | |
2469 | \ | |
2470 | vadd.u32 r_block, r_block, dx8; \ | |
2471 | vdup.u32 dx8, rg_dx8[1]; \ | |
2472 | \ | |
2473 | vadd.u32 g_block, g_block, dx8; \ | |
2474 | vdup.u32 dx8, b_dx8; \ | |
2475 | \ | |
2476 | vadd.u32 b_block, b_block, dx8; \ | |
2477 | \ | |
2478 | vmovn.u16 r_whole_8, r_whole; \ | |
2479 | vmovn.u16 g_whole_8, g_whole; \ | |
2480 | vmovn.u16 b_whole_8, b_whole; \ | |
2481 | \ | |
2482 | beq 3f; \ | |
2483 | \ | |
2484 | 2: \ | |
2485 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2486 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2487 | \ | |
2488 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
2489 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2490 | \ | |
2491 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2492 | \ | |
2493 | vdup.u32 dx4, rg_dx4[0]; \ | |
2494 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2495 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
2496 | \ | |
2497 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2498 | vdup.u32 dx4, rg_dx4[1]; \ | |
2499 | \ | |
2500 | vmov pixels, msb_mask; \ | |
2501 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2502 | vdup.u32 dx4, b_dx4; \ | |
2503 | \ | |
2504 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2505 | vdup.u32 dx8, rg_dx8[0]; \ | |
2506 | \ | |
2507 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
2508 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2509 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2510 | \ | |
2511 | vadd.u32 r_block, r_block, dx8; \ | |
2512 | vdup.u32 dx8, rg_dx8[1]; \ | |
2513 | \ | |
2514 | vadd.u32 g_block, g_block, dx8; \ | |
2515 | vdup.u32 dx8, b_dx8; \ | |
2516 | \ | |
2517 | vadd.u32 b_block, b_block, dx8; \ | |
2518 | \ | |
2519 | vmovn.u16 r_whole_8, r_whole; \ | |
2520 | vmovn.u16 g_whole_8, g_whole; \ | |
2521 | vmovn.u16 b_whole_8, b_whole; \ | |
2522 | \ | |
e1f6de8f | 2523 | vst1.u32 { pixels }, [fb_ptr]!; \ |
75e28f62 E |
2524 | subs span_num_blocks, span_num_blocks, #1; \ |
2525 | bne 2b; \ | |
2526 | \ | |
2527 | 3: \ | |
2528 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2529 | \ | |
e1f6de8f | 2530 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 E |
2531 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
2532 | \ | |
2533 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
3867c6ef | 2534 | rbit right_mask, right_mask; \ |
75e28f62 E |
2535 | vmov pixels, msb_mask; \ |
2536 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
3867c6ef | 2537 | clz right_mask, right_mask; \ |
75e28f62 E |
2538 | \ |
2539 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
2540 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2541 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2542 | \ | |
8184d7c5 | 2543 | JT_OP_REL(100f, right_mask, temp); \ |
e1f6de8f | 2544 | JT_OP(ldr pc, [pc, right_mask, lsl #2]); \ |
3867c6ef | 2545 | nop; \ |
8184d7c5 | 2546 | 100: \ |
3867c6ef | 2547 | nop; \ |
8184d7c5 | 2548 | .word JTE(100b, 4f); \ |
2549 | .word JTE(100b, 5f); \ | |
2550 | .word JTE(100b, 6f); \ | |
2551 | .word JTE(100b, 7f); \ | |
2552 | .word JTE(100b, 8f); \ | |
2553 | .word JTE(100b, 9f); \ | |
2554 | .word JTE(100b, 10f); \ | |
2555 | .word JTE(100b, 11f); \ | |
3867c6ef | 2556 | \ |
75e28f62 | 2557 | 4: \ |
e1f6de8f | 2558 | vst1.u16 { pixels_low[0] }, [fb_ptr]; \ |
3867c6ef E |
2559 | bal 1f; \ |
2560 | \ | |
2561 | 5: \ | |
e1f6de8f | 2562 | vst1.u32 { pixels_low[0] }, [fb_ptr]; \ |
3867c6ef E |
2563 | bal 1f; \ |
2564 | \ | |
2565 | 6: \ | |
e1f6de8f | 2566 | vst1.u32 { pixels_low[0] }, [fb_ptr]!; \ |
2567 | vst1.u16 { pixels_low[2] }, [fb_ptr]; \ | |
3867c6ef E |
2568 | bal 1f; \ |
2569 | \ | |
2570 | 7: \ | |
e1f6de8f | 2571 | vst1.u32 { pixels_low }, [fb_ptr]; \ |
3867c6ef E |
2572 | bal 1f; \ |
2573 | \ | |
2574 | 8: \ | |
e1f6de8f | 2575 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
2576 | vst1.u16 { pixels_high[0] }, [fb_ptr]; \ | |
3867c6ef E |
2577 | bal 1f; \ |
2578 | \ | |
2579 | 9: \ | |
e1f6de8f | 2580 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
2581 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
3867c6ef E |
2582 | bal 1f; \ |
2583 | \ | |
2584 | 10: \ | |
e1f6de8f | 2585 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
2586 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
2587 | vst1.u16 { pixels_high[2] }, [fb_ptr]; \ | |
3867c6ef E |
2588 | bal 1f; \ |
2589 | \ | |
2590 | 11: \ | |
e1f6de8f | 2591 | vst1.u32 { pixels }, [fb_ptr]; \ |
3867c6ef | 2592 | bal 1f; \ |
75e28f62 E |
2593 | \ |
2594 | 1: \ | |
2595 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
2596 | add span_b_offset, span_b_offset, #4; \ | |
2597 | \ | |
2598 | add span_edge_data, span_edge_data, #8; \ | |
2599 | subs num_spans, num_spans, #1; \ | |
2600 | \ | |
2601 | bne 0b; \ | |
2602 | \ | |
3f0189c6 | 2603 | restore_abi_regs(); \ |
75e28f62 E |
2604 | ldmia sp!, { r4 - r11, pc } \ |
2605 | ||
2606 | setup_blocks_shaded_untextured_direct_builder(undithered) | |
2607 | setup_blocks_shaded_untextured_direct_builder(dithered) | |
2608 | ||
2609 | ||
2610 | #undef psx_gpu | |
2611 | #undef num_blocks | |
2612 | #undef triangle | |
2613 | #undef c_64 | |
2614 | ||
2615 | #define psx_gpu r0 | |
2616 | #define block_ptr r1 | |
2617 | #define num_blocks r2 | |
2618 | #define uv_01 r3 | |
2619 | #define uv_23 r4 | |
2620 | #define uv_45 r5 | |
2621 | #define uv_67 r6 | |
2622 | #define uv_0 r7 | |
2623 | #define uv_1 r3 | |
2624 | #define uv_2 r8 | |
2625 | #define uv_3 r4 | |
2626 | #define uv_4 r9 | |
2627 | #define uv_5 r5 | |
2628 | #define uv_6 r10 | |
2629 | #define uv_7 r6 | |
2630 | #define texture_ptr r11 | |
2631 | ||
2632 | #define pixel_0 r7 | |
2633 | #define pixel_1 r3 | |
2634 | #define pixel_2 r8 | |
2635 | #define pixel_3 r4 | |
2636 | #define pixel_4 r9 | |
2637 | #define pixel_5 r5 | |
2638 | #define pixel_6 r10 | |
2639 | #define pixel_7 r6 | |
2640 | ||
2641 | #define pixels_a r7 | |
2642 | #define pixels_b r9 | |
2643 | #define pixels_c r8 | |
2644 | #define pixels_d r10 | |
2645 | ||
2646 | #define c_64 r0 | |
2647 | ||
2648 | #define clut_ptr r12 | |
2649 | #define current_texture_mask r5 | |
2650 | #define dirty_textures_mask r6 | |
2651 | ||
2652 | #define texels d0 | |
2653 | ||
2654 | #define clut_low_a d2 | |
2655 | #define clut_low_b d3 | |
2656 | #define clut_high_a d4 | |
2657 | #define clut_high_b d5 | |
2658 | ||
2659 | #define clut_a q1 | |
2660 | #define clut_b q2 | |
2661 | ||
2662 | #define texels_low d6 | |
2663 | #define texels_high d7 | |
2664 | ||
2665 | .align 3 | |
2666 | ||
2667 | function(texture_blocks_untextured) | |
2668 | bx lr | |
2669 | ||
2670 | ||
2671 | .align 3 | |
2672 | ||
2673 | function(texture_blocks_4bpp) | |
2674 | stmdb sp!, { r3 - r11, r14 } | |
2675 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2676 | ||
e1f6de8f | 2677 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
2678 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
75e28f62 | 2679 | |
e1f6de8f | 2680 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
2681 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128] | |
75e28f62 | 2682 | |
e1f6de8f | 2683 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
75e28f62 E |
2684 | vuzp.u8 clut_a, clut_b |
2685 | ||
e1f6de8f | 2686 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
75e28f62 E |
2687 | tst dirty_textures_mask, current_texture_mask |
2688 | ||
2689 | bne 1f | |
2690 | mov c_64, #64 | |
2691 | ||
2692 | 0: | |
2693 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
2694 | ||
2695 | uxtah uv_0, texture_ptr, uv_01 | |
2696 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
2697 | ||
2698 | uxtah uv_2, texture_ptr, uv_23 | |
2699 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
2700 | ||
2701 | uxtah uv_4, texture_ptr, uv_45 | |
e1f6de8f | 2702 | ldrb pixel_0, [uv_0] |
75e28f62 E |
2703 | |
2704 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
e1f6de8f | 2705 | ldrb pixel_1, [uv_1] |
75e28f62 E |
2706 | |
2707 | uxtah uv_6, texture_ptr, uv_67 | |
e1f6de8f | 2708 | ldrb pixel_2, [uv_2] |
75e28f62 E |
2709 | |
2710 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
e1f6de8f | 2711 | ldrb pixel_3, [uv_3] |
75e28f62 | 2712 | |
e1f6de8f | 2713 | ldrb pixel_4, [uv_4] |
75e28f62 E |
2714 | subs num_blocks, num_blocks, #1 |
2715 | ||
e1f6de8f | 2716 | ldrb pixel_5, [uv_5] |
75e28f62 E |
2717 | orr pixels_a, pixel_0, pixel_1, lsl #8 |
2718 | ||
e1f6de8f | 2719 | ldrb pixel_6, [uv_6] |
75e28f62 E |
2720 | orr pixels_b, pixel_4, pixel_5, lsl #8 |
2721 | ||
e1f6de8f | 2722 | ldrb pixel_7, [uv_7] |
75e28f62 E |
2723 | orr pixels_a, pixels_a, pixel_2, lsl #16 |
2724 | ||
2725 | orr pixels_b, pixels_b, pixel_6, lsl #16 | |
2726 | orr pixels_a, pixels_a, pixel_3, lsl #24 | |
2727 | ||
2728 | orr pixels_b, pixels_b, pixel_7, lsl #24 | |
ed0fd81d | 2729 | vmov texels, pixels_a, pixels_b |
75e28f62 E |
2730 | |
2731 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels | |
2732 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels | |
2733 | ||
e1f6de8f | 2734 | vst2.u8 { texels_low, texels_high }, [block_ptr, :128], c_64 |
75e28f62 E |
2735 | bne 0b |
2736 | ||
2737 | ldmia sp!, { r3 - r11, pc } | |
2738 | ||
2739 | 1: | |
2740 | stmdb sp!, { r1 - r2 } | |
2741 | bl update_texture_4bpp_cache | |
2742 | ||
2743 | mov c_64, #64 | |
2744 | ldmia sp!, { r1 - r2 } | |
2745 | bal 0b | |
2746 | ||
2747 | ||
2748 | .align 3 | |
2749 | ||
2750 | function(texture_blocks_8bpp) | |
2751 | stmdb sp!, { r3 - r11, r14 } | |
2752 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2753 | ||
e1f6de8f | 2754 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
2755 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
75e28f62 | 2756 | |
e1f6de8f | 2757 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
2758 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] | |
75e28f62 | 2759 | |
e1f6de8f | 2760 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset] |
75e28f62 E |
2761 | tst dirty_textures_mask, current_texture_mask |
2762 | ||
2763 | bne 1f | |
2764 | nop | |
2765 | ||
2766 | 0: | |
2767 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
2768 | ||
2769 | uxtah uv_0, texture_ptr, uv_01 | |
2770 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
2771 | ||
2772 | uxtah uv_2, texture_ptr, uv_23 | |
2773 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
2774 | ||
2775 | uxtah uv_4, texture_ptr, uv_45 | |
e1f6de8f | 2776 | ldrb pixel_0, [uv_0] |
75e28f62 E |
2777 | |
2778 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
e1f6de8f | 2779 | ldrb pixel_1, [uv_1] |
75e28f62 E |
2780 | |
2781 | uxtah uv_6, texture_ptr, uv_67 | |
e1f6de8f | 2782 | ldrb pixel_2, [uv_2] |
75e28f62 E |
2783 | |
2784 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
e1f6de8f | 2785 | ldrb pixel_3, [uv_3] |
75e28f62 | 2786 | |
e1f6de8f | 2787 | ldrb pixel_4, [uv_4] |
75e28f62 E |
2788 | add pixel_0, pixel_0, pixel_0 |
2789 | ||
e1f6de8f | 2790 | ldrb pixel_5, [uv_5] |
75e28f62 E |
2791 | add pixel_1, pixel_1, pixel_1 |
2792 | ||
e1f6de8f | 2793 | ldrb pixel_6, [uv_6] |
75e28f62 E |
2794 | add pixel_2, pixel_2, pixel_2 |
2795 | ||
e1f6de8f | 2796 | ldrb pixel_7, [uv_7] |
75e28f62 E |
2797 | add pixel_3, pixel_3, pixel_3 |
2798 | ||
e1f6de8f | 2799 | ldrh pixel_0, [clut_ptr, pixel_0] |
75e28f62 E |
2800 | add pixel_4, pixel_4, pixel_4 |
2801 | ||
e1f6de8f | 2802 | ldrh pixel_1, [clut_ptr, pixel_1] |
75e28f62 E |
2803 | add pixel_5, pixel_5, pixel_5 |
2804 | ||
e1f6de8f | 2805 | ldrh pixel_2, [clut_ptr, pixel_2] |
75e28f62 E |
2806 | add pixel_6, pixel_6, pixel_6 |
2807 | ||
e1f6de8f | 2808 | ldrh pixel_3, [clut_ptr, pixel_3] |
75e28f62 E |
2809 | add pixel_7, pixel_7, pixel_7 |
2810 | ||
e1f6de8f | 2811 | ldrh pixel_4, [clut_ptr, pixel_4] |
75e28f62 E |
2812 | orr pixels_a, pixel_0, pixel_1, lsl #16 |
2813 | ||
e1f6de8f | 2814 | ldrh pixel_5, [clut_ptr, pixel_5] |
75e28f62 E |
2815 | orr pixels_c, pixel_2, pixel_3, lsl #16 |
2816 | ||
e1f6de8f | 2817 | ldrh pixel_6, [clut_ptr, pixel_6] |
75e28f62 E |
2818 | subs num_blocks, num_blocks, #1 |
2819 | ||
e1f6de8f | 2820 | ldrh pixel_7, [clut_ptr, pixel_7] |
75e28f62 E |
2821 | orr pixels_b, pixel_4, pixel_5, lsl #16 |
2822 | ||
2823 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
2824 | stm block_ptr, { pixels_a, pixels_c, pixels_b, pixels_d } | |
2825 | ||
2826 | add block_ptr, block_ptr, #64 | |
2827 | bne 0b | |
2828 | ||
2829 | ldmia sp!, { r3 - r11, pc } | |
2830 | ||
2831 | 1: | |
4d646738 | 2832 | stmdb sp!, { r1 - r2, EXTRA_UNSAVED_REGS r12 } |
75e28f62 E |
2833 | |
2834 | bl update_texture_8bpp_cache | |
2835 | ||
4d646738 | 2836 | ldmia sp!, { r1 - r2, EXTRA_UNSAVED_REGS r12 } |
75e28f62 E |
2837 | bal 0b |
2838 | ||
2839 | ||
2840 | #undef uv_0 | |
2841 | #undef uv_1 | |
2842 | #undef uv_2 | |
2843 | #undef uv_3 | |
2844 | #undef uv_4 | |
2845 | #undef uv_5 | |
2846 | #undef uv_6 | |
2847 | #undef uv_7 | |
2848 | ||
2849 | #undef pixel_0 | |
2850 | #undef pixel_1 | |
2851 | #undef pixel_2 | |
2852 | #undef pixel_3 | |
2853 | #undef pixel_4 | |
2854 | #undef pixel_5 | |
2855 | #undef pixel_6 | |
2856 | #undef pixel_7 | |
2857 | ||
2858 | #undef texture_ptr | |
2859 | ||
2860 | #undef pixels_a | |
2861 | #undef pixels_b | |
2862 | #undef pixels_c | |
2863 | #undef pixels_d | |
2864 | ||
2865 | #define psx_gpu r0 | |
2866 | #define block_ptr r1 | |
2867 | #define num_blocks r2 | |
2868 | ||
2869 | #define uv_0 r3 | |
2870 | #define uv_1 r4 | |
2871 | #define u_0 r3 | |
2872 | #define u_1 r4 | |
2873 | #define v_0 r5 | |
2874 | #define v_1 r6 | |
2875 | ||
2876 | #define uv_2 r5 | |
2877 | #define uv_3 r6 | |
2878 | #define u_2 r5 | |
2879 | #define u_3 r6 | |
2880 | #define v_2 r7 | |
2881 | #define v_3 r8 | |
2882 | ||
2883 | #define uv_4 r7 | |
2884 | #define uv_5 r8 | |
2885 | #define u_4 r7 | |
2886 | #define u_5 r8 | |
2887 | #define v_4 r9 | |
2888 | #define v_5 r10 | |
2889 | ||
2890 | #define uv_6 r9 | |
2891 | #define uv_7 r10 | |
2892 | #define u_6 r9 | |
2893 | #define u_7 r10 | |
2894 | #define v_6 r11 | |
2895 | #define v_7 r0 | |
2896 | ||
2897 | #define pixel_0 r3 | |
2898 | #define pixel_1 r4 | |
2899 | #define pixel_2 r5 | |
2900 | #define pixel_3 r6 | |
2901 | #define pixel_4 r7 | |
2902 | #define pixel_5 r8 | |
2903 | #define pixel_6 r9 | |
2904 | #define pixel_7 r10 | |
2905 | ||
2906 | #define pixels_a r3 | |
2907 | #define pixels_b r5 | |
2908 | #define pixels_c r7 | |
2909 | #define pixels_d r9 | |
2910 | ||
2911 | #define texture_ptr r12 | |
2912 | ||
2913 | ||
2914 | .align 3 | |
2915 | ||
2916 | function(texture_blocks_16bpp) | |
2917 | stmdb sp!, { r3 - r11, r14 } | |
2918 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2919 | ||
e1f6de8f | 2920 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
2921 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] | |
75e28f62 E |
2922 | |
2923 | 0: | |
e1f6de8f | 2924 | ldrh uv_0, [block_ptr] |
75e28f62 E |
2925 | subs num_blocks, num_blocks, #1 |
2926 | ||
e1f6de8f | 2927 | ldrh uv_1, [block_ptr, #2] |
75e28f62 E |
2928 | |
2929 | and v_0, uv_0, #0xFF00 | |
2930 | and v_1, uv_1, #0xFF00 | |
2931 | ||
2932 | and u_0, uv_0, #0xFF | |
2933 | and u_1, uv_1, #0xFF | |
2934 | ||
2935 | add uv_0, u_0, v_0, lsl #2 | |
e1f6de8f | 2936 | ldrh uv_2, [block_ptr, #4] |
75e28f62 E |
2937 | |
2938 | add uv_1, u_1, v_1, lsl #2 | |
e1f6de8f | 2939 | ldrh uv_3, [block_ptr, #6] |
75e28f62 E |
2940 | |
2941 | add uv_0, uv_0, uv_0 | |
2942 | add uv_1, uv_1, uv_1 | |
2943 | ||
2944 | and v_2, uv_2, #0xFF00 | |
2945 | and v_3, uv_3, #0xFF00 | |
2946 | ||
2947 | and u_2, uv_2, #0xFF | |
2948 | and u_3, uv_3, #0xFF | |
2949 | ||
2950 | add uv_2, u_2, v_2, lsl #2 | |
e1f6de8f | 2951 | ldrh uv_4, [block_ptr, #8] |
75e28f62 E |
2952 | |
2953 | add uv_3, u_3, v_3, lsl #2 | |
e1f6de8f | 2954 | ldrh uv_5, [block_ptr, #10] |
75e28f62 E |
2955 | |
2956 | add uv_2, uv_2, uv_2 | |
2957 | add uv_3, uv_3, uv_3 | |
2958 | ||
2959 | and v_4, uv_4, #0xFF00 | |
2960 | and v_5, uv_5, #0xFF00 | |
2961 | ||
2962 | and u_4, uv_4, #0xFF | |
2963 | and u_5, uv_5, #0xFF | |
2964 | ||
2965 | add uv_4, u_4, v_4, lsl #2 | |
e1f6de8f | 2966 | ldrh uv_6, [block_ptr, #12] |
75e28f62 E |
2967 | |
2968 | add uv_5, u_5, v_5, lsl #2 | |
e1f6de8f | 2969 | ldrh uv_7, [block_ptr, #14] |
75e28f62 E |
2970 | |
2971 | add uv_4, uv_4, uv_4 | |
e1f6de8f | 2972 | ldrh pixel_0, [texture_ptr, uv_0] |
75e28f62 E |
2973 | |
2974 | add uv_5, uv_5, uv_5 | |
e1f6de8f | 2975 | ldrh pixel_1, [texture_ptr, uv_1] |
75e28f62 E |
2976 | |
2977 | and v_6, uv_6, #0xFF00 | |
e1f6de8f | 2978 | ldrh pixel_2, [texture_ptr, uv_2] |
75e28f62 E |
2979 | |
2980 | and v_7, uv_7, #0xFF00 | |
e1f6de8f | 2981 | ldrh pixel_3, [texture_ptr, uv_3] |
75e28f62 E |
2982 | |
2983 | and u_6, uv_6, #0xFF | |
e1f6de8f | 2984 | ldrh pixel_4, [texture_ptr, uv_4] |
75e28f62 E |
2985 | |
2986 | and u_7, uv_7, #0xFF | |
e1f6de8f | 2987 | ldrh pixel_5, [texture_ptr, uv_5] |
75e28f62 E |
2988 | |
2989 | add uv_6, u_6, v_6, lsl #2 | |
2990 | add uv_7, u_7, v_7, lsl #2 | |
2991 | ||
2992 | add uv_6, uv_6, uv_6 | |
2993 | add uv_7, uv_7, uv_7 | |
2994 | ||
2995 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
2996 | orr pixels_b, pixel_2, pixel_3, lsl #16 | |
2997 | ||
e1f6de8f | 2998 | ldrh pixel_6, [texture_ptr, uv_6] |
75e28f62 E |
2999 | orr pixels_c, pixel_4, pixel_5, lsl #16 |
3000 | ||
e1f6de8f | 3001 | ldrh pixel_7, [texture_ptr, uv_7] |
75e28f62 E |
3002 | orr pixels_d, pixel_6, pixel_7, lsl #16 |
3003 | ||
3004 | stm block_ptr, { pixels_a, pixels_b, pixels_c, pixels_d } | |
3005 | add block_ptr, block_ptr, #64 | |
3006 | ||
3007 | bne 0b | |
3008 | ||
3009 | ldmia sp!, { r3 - r11, pc } | |
3010 | ||
3011 | ||
3012 | #undef num_blocks | |
3013 | ||
3014 | #undef test_mask | |
3015 | #undef texels | |
3016 | #undef pixels_b | |
3017 | #undef pixels | |
3018 | #undef d64_1 | |
3019 | #undef d64_4 | |
3020 | #undef d64_128 | |
3021 | #undef draw_mask | |
3022 | #undef msb_mask | |
3023 | #undef msb_mask_low | |
3024 | #undef msb_mask_high | |
3025 | #undef fb_pixels | |
3026 | ||
3027 | #undef c_32 | |
3028 | #undef fb_ptr | |
3029 | #undef mask_msb_ptr | |
3030 | ||
3031 | #define psx_gpu r0 | |
3032 | #define num_blocks r1 | |
3033 | #define color_ptr r2 | |
3867c6ef E |
3034 | #define colors_scalar r2 |
3035 | #define colors_scalar_compare r3 | |
75e28f62 E |
3036 | #define mask_msb_ptr r2 |
3037 | ||
3038 | #define block_ptr_load_a r0 | |
3039 | #define block_ptr_store r3 | |
3040 | #define block_ptr_load_b r12 | |
3041 | #define c_32 r2 | |
3042 | ||
3043 | #define c_48 r4 | |
3044 | #define fb_ptr r14 | |
3045 | #define draw_mask_bits_scalar r5 | |
3046 | ||
3047 | #define d128_0x07 q0 | |
3048 | #define d128_0x1F q1 | |
3049 | #define d128_0x8000 q2 | |
3050 | #define test_mask q3 | |
3051 | #define texels q4 | |
3052 | #define colors_rg q5 | |
3053 | #define colors_b_dm_bits q6 | |
3054 | #define texels_rg q7 | |
3055 | #define pixels_r q8 | |
3056 | #define pixels_g q9 | |
3057 | #define pixels_b q10 | |
3058 | #define pixels q11 | |
3059 | #define zero_mask q4 | |
3060 | #define draw_mask q12 | |
3061 | #define msb_mask q13 | |
3062 | ||
3063 | #define fb_pixels q8 | |
3064 | ||
3065 | #define pixels_gb_low q9 | |
3066 | ||
3067 | #define colors_r d10 | |
3068 | #define colors_g d11 | |
3069 | #define colors_b d12 | |
3070 | #define draw_mask_bits d13 | |
3071 | #define texels_r d14 | |
3072 | #define texels_g d15 | |
3073 | #define pixels_r_low d16 | |
3074 | #define pixels_g_low d18 | |
3075 | #define pixels_b_low d19 | |
3076 | #define msb_mask_low d26 | |
3077 | #define msb_mask_high d27 | |
3078 | ||
3079 | #define d64_1 d28 | |
3080 | #define d64_4 d29 | |
3081 | #define d64_128 d30 | |
3082 | #define texels_b d31 | |
3083 | ||
3084 | #define shade_blocks_textured_modulated_prologue_indirect() \ | |
3085 | mov c_48, #48; \ | |
3086 | add block_ptr_store, psx_gpu, #psx_gpu_blocks_offset \ | |
3087 | ||
3088 | #define shade_blocks_textured_modulated_prologue_direct() \ | |
3089 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
e1f6de8f | 3090 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] \ |
75e28f62 | 3091 | |
75e28f62 | 3092 | |
3867c6ef E |
3093 | #define shade_blocks_textured_modulated_prologue_shaded(dithering, target) \ |
3094 | ||
3095 | #define shade_blocks_textured_false_modulation_check_undithered(target) \ | |
e1f6de8f | 3096 | ldr colors_scalar, [psx_gpu, #psx_gpu_triangle_color_offset]; \ |
3867c6ef E |
3097 | movw colors_scalar_compare, #0x8080; \ |
3098 | \ | |
3099 | movt colors_scalar_compare, #0x80; \ | |
3100 | cmp colors_scalar, colors_scalar_compare; \ | |
3101 | beq shade_blocks_textured_unmodulated_##target \ | |
3102 | ||
3103 | #define shade_blocks_textured_false_modulation_check_dithered(target) \ | |
3104 | ||
3105 | #define shade_blocks_textured_modulated_prologue_unshaded(dithering, target) \ | |
3106 | shade_blocks_textured_false_modulation_check_##dithering(target); \ | |
75e28f62 | 3107 | add color_ptr, psx_gpu, #psx_gpu_triangle_color_offset; \ |
e1f6de8f | 3108 | vld1.u32 { colors_r[] }, [color_ptr, :32]; \ |
75e28f62 E |
3109 | vdup.u8 colors_g, colors_r[1]; \ |
3110 | vdup.u8 colors_b, colors_r[2]; \ | |
3111 | vdup.u8 colors_r, colors_r[0] \ | |
3112 | ||
3113 | ||
3114 | #define shade_blocks_textured_modulated_load_dithered(target) \ | |
e1f6de8f | 3115 | vld1.u32 { target }, [block_ptr_load_b, :128] \ |
75e28f62 E |
3116 | |
3117 | #define shade_blocks_textured_modulated_load_last_dithered(target) \ | |
e1f6de8f | 3118 | vld1.u32 { target }, [block_ptr_load_b, :128], c_32 \ |
75e28f62 E |
3119 | |
3120 | #define shade_blocks_textured_modulated_load_undithered(target) \ | |
3121 | ||
3122 | #define shade_blocks_textured_modulated_load_last_undithered(target) \ | |
3123 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
3124 | ||
3125 | #define shade_blocks_textured_modulate_dithered(channel) \ | |
3126 | vmlal.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
3127 | ||
3128 | #define shade_blocks_textured_modulate_undithered(channel) \ | |
3129 | vmull.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
3130 | ||
3131 | ||
3132 | #define shade_blocks_textured_modulated_store_draw_mask_indirect(offset) \ | |
e1f6de8f | 3133 | vst1.u32 { draw_mask }, [block_ptr_store, :128]! \ |
75e28f62 E |
3134 | |
3135 | #define shade_blocks_textured_modulated_store_draw_mask_direct(offset) \ | |
e1f6de8f | 3136 | ldr fb_ptr, [block_ptr_load_b, #(offset - 64)]; \ |
3137 | vld1.u32 { fb_pixels }, [fb_ptr]; \ | |
75e28f62 E |
3138 | vbit.u16 pixels, fb_pixels, draw_mask \ |
3139 | ||
3140 | #define shade_blocks_textured_modulated_store_pixels_indirect() \ | |
e1f6de8f | 3141 | vst1.u32 { pixels }, [block_ptr_store, :128], c_48 \ |
75e28f62 E |
3142 | |
3143 | #define shade_blocks_textured_modulated_store_pixels_direct() \ | |
e1f6de8f | 3144 | vst1.u32 { pixels }, [fb_ptr] \ |
75e28f62 E |
3145 | |
3146 | ||
3147 | #define shade_blocks_textured_modulated_load_rg_shaded() \ | |
e1f6de8f | 3148 | vld1.u32 { colors_r, colors_g }, [block_ptr_load_b, :128], c_32 \ |
75e28f62 E |
3149 | |
3150 | #define shade_blocks_textured_modulated_load_rg_unshaded() \ | |
3151 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
3152 | ||
3153 | #define shade_blocks_textured_modulated_load_bdm_shaded() \ | |
e1f6de8f | 3154 | vld1.u32 { colors_b, draw_mask_bits }, [block_ptr_load_a, :128], c_32 \ |
75e28f62 E |
3155 | |
3156 | #define shade_blocks_textured_modulated_load_bdm_unshaded() \ | |
e1f6de8f | 3157 | ldr draw_mask_bits_scalar, [block_ptr_load_a, #8]; \ |
75e28f62 E |
3158 | add block_ptr_load_a, block_ptr_load_a, #32 \ |
3159 | ||
3160 | #define shade_blocks_textured_modulated_expand_draw_mask_shaded() \ | |
3161 | vdup.u16 draw_mask, draw_mask_bits[0] \ | |
3162 | ||
3163 | #define shade_blocks_textured_modulated_expand_draw_mask_unshaded() \ | |
3164 | vdup.u16 draw_mask, draw_mask_bits_scalar \ | |
3165 | ||
3166 | ||
3167 | #define shade_blocks_textured_modulated_apply_msb_mask_indirect() \ | |
3168 | ||
3169 | #define shade_blocks_textured_modulated_apply_msb_mask_direct() \ | |
3170 | vorr.u16 pixels, pixels, msb_mask \ | |
3171 | ||
3172 | ||
3173 | #define shade_blocks_textured_modulated_builder(shading, dithering, target) \ | |
3174 | .align 3; \ | |
3175 | \ | |
3176 | function(shade_blocks_##shading##_textured_modulated_##dithering##_##target) \ | |
3867c6ef | 3177 | shade_blocks_textured_modulated_prologue_##shading(dithering, target); \ |
75e28f62 | 3178 | stmdb sp!, { r4 - r5, lr }; \ |
3f0189c6 | 3179 | save_abi_regs(); \ |
e1f6de8f | 3180 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 | 3181 | \ |
e1f6de8f | 3182 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
3183 | \ |
3184 | shade_blocks_textured_modulated_prologue_##target(); \ | |
75e28f62 E |
3185 | \ |
3186 | add block_ptr_load_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
3187 | mov c_32, #32; \ | |
3188 | \ | |
3189 | add block_ptr_load_b, block_ptr_load_a, #16; \ | |
3190 | vmov.u8 d64_1, #1; \ | |
3191 | vmov.u8 d64_4, #4; \ | |
3192 | vmov.u8 d64_128, #128; \ | |
3193 | \ | |
e1f6de8f | 3194 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
75e28f62 E |
3195 | vmov.u8 d128_0x07, #0x07; \ |
3196 | \ | |
3197 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
3198 | vmov.u8 d128_0x1F, #0x1F; \ | |
3199 | \ | |
3200 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
3201 | vmov.u16 d128_0x8000, #0x8000; \ | |
3202 | \ | |
3203 | vmovn.u16 texels_r, texels; \ | |
3204 | vshrn.u16 texels_g, texels, #5; \ | |
3205 | \ | |
3206 | vshrn.u16 texels_b, texels, #7; \ | |
3207 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
3208 | \ | |
3209 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
3210 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
3211 | \ | |
3212 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
3213 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
3214 | \ | |
3215 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
3216 | vshr.u8 texels_b, texels_b, #3; \ | |
3217 | \ | |
3218 | shade_blocks_textured_modulate_##dithering(r); \ | |
3219 | shade_blocks_textured_modulate_##dithering(g); \ | |
3220 | shade_blocks_textured_modulate_##dithering(b); \ | |
3221 | \ | |
3222 | vand.u16 pixels, texels, d128_0x8000; \ | |
3223 | vceq.u16 zero_mask, texels, #0; \ | |
3224 | \ | |
3225 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
3226 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
3227 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
3228 | \ | |
3229 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
3230 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
3231 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
3232 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
3233 | \ | |
3234 | subs num_blocks, num_blocks, #1; \ | |
3235 | beq 1f; \ | |
3236 | \ | |
3237 | .align 3; \ | |
3238 | \ | |
3239 | 0: \ | |
e1f6de8f | 3240 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
75e28f62 E |
3241 | shade_blocks_textured_modulated_load_rg_##shading(); \ |
3242 | vshrn.u16 texels_g, texels, #5; \ | |
3243 | \ | |
3244 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
3245 | vshrn.u16 texels_b, texels, #7; \ | |
3246 | \ | |
e1f6de8f | 3247 | pld [block_ptr_load_a]; \ |
75e28f62 E |
3248 | vmovn.u16 texels_r, texels; \ |
3249 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
3250 | \ | |
3251 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
3252 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
3253 | shade_blocks_textured_modulated_store_draw_mask_##target(-4); \ | |
3254 | \ | |
3255 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
3256 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
3257 | \ | |
3258 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
3259 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
3260 | \ | |
3261 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
3262 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
3263 | \ | |
3264 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
3265 | vshr.u8 texels_b, texels_b, #3; \ | |
3266 | \ | |
3267 | shade_blocks_textured_modulate_##dithering(r); \ | |
3268 | shade_blocks_textured_modulate_##dithering(g); \ | |
3269 | shade_blocks_textured_modulate_##dithering(b); \ | |
3270 | \ | |
3271 | vand.u16 pixels, texels, d128_0x8000; \ | |
3272 | vceq.u16 zero_mask, texels, #0; \ | |
3273 | \ | |
3274 | subs num_blocks, num_blocks, #1; \ | |
3275 | \ | |
3276 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
3277 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
3278 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
3279 | \ | |
3280 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
3281 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
3282 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
3283 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
3284 | \ | |
3285 | bne 0b; \ | |
3286 | \ | |
3287 | 1: \ | |
3288 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
3289 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
3290 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
3291 | \ | |
3292 | shade_blocks_textured_modulated_store_draw_mask_##target(28); \ | |
3293 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
3294 | \ | |
3f0189c6 | 3295 | restore_abi_regs(); \ |
75e28f62 E |
3296 | ldmia sp!, { r4 - r5, pc } \ |
3297 | ||
3298 | ||
3299 | shade_blocks_textured_modulated_builder(shaded, dithered, direct); | |
3300 | shade_blocks_textured_modulated_builder(shaded, undithered, direct); | |
3301 | shade_blocks_textured_modulated_builder(unshaded, dithered, direct); | |
3302 | shade_blocks_textured_modulated_builder(unshaded, undithered, direct); | |
3303 | ||
3304 | shade_blocks_textured_modulated_builder(shaded, dithered, indirect); | |
3305 | shade_blocks_textured_modulated_builder(shaded, undithered, indirect); | |
3306 | shade_blocks_textured_modulated_builder(unshaded, dithered, indirect); | |
3307 | shade_blocks_textured_modulated_builder(unshaded, undithered, indirect); | |
3308 | ||
3309 | ||
3310 | #undef c_64 | |
3311 | #undef fb_ptr | |
3312 | #undef color_ptr | |
3313 | ||
3314 | #undef color_r | |
3315 | #undef color_g | |
3316 | #undef color_b | |
3317 | ||
3318 | #undef test_mask | |
3319 | #undef pixels | |
3320 | #undef draw_mask | |
3321 | #undef zero_mask | |
3322 | #undef fb_pixels | |
3323 | #undef msb_mask | |
3324 | #undef msb_mask_low | |
3325 | #undef msb_mask_high | |
3326 | ||
3327 | #define psx_gpu r0 | |
3328 | #define num_blocks r1 | |
3329 | #define mask_msb_ptr r2 | |
3330 | #define color_ptr r3 | |
3331 | ||
3332 | #define block_ptr_load r0 | |
3333 | #define draw_mask_store_ptr r3 | |
3334 | #define draw_mask_bits_ptr r12 | |
3335 | #define draw_mask_ptr r12 | |
3336 | #define pixel_store_ptr r14 | |
3337 | ||
3338 | #define fb_ptr_cmp r4 | |
3339 | ||
3340 | #define fb_ptr r3 | |
3341 | #define fb_ptr_next r14 | |
3342 | ||
3343 | #define c_64 r2 | |
3344 | ||
3345 | #define test_mask q0 | |
3346 | #define pixels q1 | |
3347 | #define draw_mask q2 | |
3348 | #define zero_mask q3 | |
3349 | #define draw_mask_combined q4 | |
3350 | #define fb_pixels q5 | |
3351 | #define fb_pixels_next q6 | |
3352 | #define msb_mask q7 | |
3353 | ||
3354 | #define draw_mask_low d4 | |
3355 | #define draw_mask_high d5 | |
3356 | #define msb_mask_low d14 | |
3357 | #define msb_mask_high d15 | |
3358 | ||
3359 | .align 3 | |
3360 | function(shade_blocks_textured_unmodulated_indirect) | |
3f0189c6 | 3361 | stmdb sp!, { r4, r14 } |
3362 | save_abi_regs() | |
75e28f62 E |
3363 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
3364 | ||
e1f6de8f | 3365 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
3366 | add pixel_store_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
3367 | ||
e1f6de8f | 3368 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 E |
3369 | add draw_mask_store_ptr, psx_gpu, #psx_gpu_blocks_offset |
3370 | ||
3371 | mov c_64, #64 | |
3372 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
3373 | ||
e1f6de8f | 3374 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
75e28f62 | 3375 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ |
e1f6de8f | 3376 | [draw_mask_bits_ptr, :16], c_64 |
75e28f62 E |
3377 | vceq.u16 zero_mask, pixels, #0 |
3378 | ||
3379 | vtst.u16 draw_mask, draw_mask, test_mask | |
e1f6de8f | 3380 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
75e28f62 E |
3381 | |
3382 | subs num_blocks, num_blocks, #1 | |
3383 | beq 1f | |
3384 | ||
3385 | 0: | |
e1f6de8f | 3386 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
75e28f62 E |
3387 | vorr.u16 draw_mask_combined, draw_mask, zero_mask |
3388 | ||
3389 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
e1f6de8f | 3390 | [draw_mask_bits_ptr, :16], c_64 |
75e28f62 E |
3391 | vceq.u16 zero_mask, pixels, #0 |
3392 | ||
3393 | vtst.u16 draw_mask, draw_mask, test_mask | |
e1f6de8f | 3394 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
75e28f62 | 3395 | |
e1f6de8f | 3396 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
75e28f62 E |
3397 | subs num_blocks, num_blocks, #1 |
3398 | ||
3399 | bne 0b | |
3400 | ||
3401 | 1: | |
3402 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
e1f6de8f | 3403 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
75e28f62 | 3404 | |
3f0189c6 | 3405 | restore_abi_regs() |
3406 | ldmia sp!, { r4, pc } | |
75e28f62 E |
3407 | |
3408 | ||
3409 | .align 3 | |
3410 | ||
3411 | function(shade_blocks_textured_unmodulated_direct) | |
3412 | stmdb sp!, { r4, r14 } | |
3f0189c6 | 3413 | save_abi_regs() |
75e28f62 E |
3414 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
3415 | ||
e1f6de8f | 3416 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
3417 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
3418 | ||
e1f6de8f | 3419 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
75e28f62 E |
3420 | mov c_64, #64 |
3421 | ||
e1f6de8f | 3422 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 E |
3423 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset |
3424 | ||
3425 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
e1f6de8f | 3426 | [draw_mask_bits_ptr, :16], c_64 |
3427 | ldr fb_ptr_next, [block_ptr_load, #44] | |
75e28f62 | 3428 | |
e1f6de8f | 3429 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
3430 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
75e28f62 E |
3431 | vceq.u16 zero_mask, pixels, #0 |
3432 | vtst.u16 draw_mask, draw_mask, test_mask | |
3433 | ||
3434 | subs num_blocks, num_blocks, #1 | |
3435 | beq 1f | |
3436 | ||
3437 | 0: | |
3438 | mov fb_ptr, fb_ptr_next | |
e1f6de8f | 3439 | ldr fb_ptr_next, [block_ptr_load, #44] |
75e28f62 E |
3440 | |
3441 | vorr.u16 pixels, pixels, msb_mask | |
3442 | ||
3443 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3444 | vmov fb_pixels, fb_pixels_next | |
3445 | ||
3446 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
e1f6de8f | 3447 | [draw_mask_bits_ptr, :16], c_64 |
75e28f62 E |
3448 | vbif.u16 fb_pixels, pixels, draw_mask_combined |
3449 | ||
75e28f62 | 3450 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr |
e1f6de8f | 3451 | pld [fb_ptr_next, #64] |
8438c3c7 | 3452 | |
75e28f62 | 3453 | add fb_ptr_cmp, fb_ptr_cmp, #14 |
e1f6de8f | 3454 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
8438c3c7 | 3455 | |
75e28f62 E |
3456 | cmp fb_ptr_cmp, #28 |
3457 | bls 4f | |
3458 | ||
e1f6de8f | 3459 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 E |
3460 | vceq.u16 zero_mask, pixels, #0 |
3461 | ||
e1f6de8f | 3462 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 E |
3463 | vtst.u16 draw_mask, draw_mask, test_mask |
3464 | ||
3465 | 3: | |
3466 | subs num_blocks, num_blocks, #1 | |
3467 | bne 0b | |
3468 | ||
3469 | 1: | |
3470 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3471 | vbif.u16 fb_pixels_next, pixels, draw_mask_combined | |
3472 | ||
e1f6de8f | 3473 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 | 3474 | |
3f0189c6 | 3475 | restore_abi_regs() |
75e28f62 E |
3476 | ldmia sp!, { r4, pc } |
3477 | ||
3478 | 4: | |
e1f6de8f | 3479 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 E |
3480 | vceq.u16 zero_mask, pixels, #0 |
3481 | ||
e1f6de8f | 3482 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 E |
3483 | vtst.u16 draw_mask, draw_mask, test_mask |
3484 | ||
3485 | bal 3b | |
3486 | ||
3487 | ||
3488 | function(shade_blocks_unshaded_untextured_indirect) | |
3489 | bx lr | |
3490 | ||
3491 | .align 3 | |
3492 | ||
3493 | function(shade_blocks_unshaded_untextured_direct) | |
3494 | stmdb sp!, { r4, r14 } | |
3f0189c6 | 3495 | save_abi_regs() |
75e28f62 E |
3496 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset |
3497 | ||
e1f6de8f | 3498 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
3499 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
3500 | ||
e1f6de8f | 3501 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
75e28f62 E |
3502 | add color_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
3503 | ||
3504 | add block_ptr_load, psx_gpu, #(psx_gpu_blocks_offset + 44) | |
e1f6de8f | 3505 | vld1.u16 { pixels }, [color_ptr, :128] |
75e28f62 E |
3506 | |
3507 | mov c_64, #64 | |
e1f6de8f | 3508 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
75e28f62 E |
3509 | |
3510 | vorr.u16 pixels, pixels, msb_mask | |
3511 | subs num_blocks, num_blocks, #1 | |
3512 | ||
e1f6de8f | 3513 | ldr fb_ptr_next, [block_ptr_load], #64 |
75e28f62 | 3514 | |
e1f6de8f | 3515 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 E |
3516 | beq 1f |
3517 | ||
3518 | 0: | |
3519 | vmov fb_pixels, fb_pixels_next | |
3520 | mov fb_ptr, fb_ptr_next | |
e1f6de8f | 3521 | ldr fb_ptr_next, [block_ptr_load], #64 |
75e28f62 E |
3522 | |
3523 | vbif.u16 fb_pixels, pixels, draw_mask | |
e1f6de8f | 3524 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
75e28f62 E |
3525 | |
3526 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr | |
3527 | add fb_ptr_cmp, fb_ptr_cmp, #14 | |
3528 | cmp fb_ptr_cmp, #28 | |
3529 | bls 4f | |
3530 | ||
e1f6de8f | 3531 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
3532 | vst1.u16 { fb_pixels }, [fb_ptr] | |
75e28f62 E |
3533 | |
3534 | 3: | |
3535 | subs num_blocks, num_blocks, #1 | |
3536 | bne 0b | |
3537 | ||
3538 | 1: | |
3539 | vbif.u16 fb_pixels_next, pixels, draw_mask | |
e1f6de8f | 3540 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 | 3541 | |
3f0189c6 | 3542 | restore_abi_regs() |
75e28f62 E |
3543 | ldmia sp!, { r4, pc } |
3544 | ||
3545 | 4: | |
e1f6de8f | 3546 | vst1.u16 { fb_pixels }, [fb_ptr] |
3547 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
75e28f62 E |
3548 | bal 3b |
3549 | ||
3550 | ||
3551 | #undef draw_mask_ptr | |
3552 | #undef c_64 | |
3553 | #undef fb_ptr | |
3554 | #undef fb_ptr_next | |
3555 | #undef fb_ptr_cmp | |
3556 | ||
3557 | #define psx_gpu r0 | |
3558 | #define num_blocks r1 | |
3559 | #define msb_mask_ptr r2 | |
3560 | #define pixel_ptr r3 | |
3561 | #define draw_mask_ptr r0 | |
3562 | #define c_64 r2 | |
3563 | #define fb_ptr r12 | |
3564 | #define fb_ptr_next r14 | |
3565 | #define fb_ptr_cmp r4 | |
3566 | ||
3567 | #undef msb_mask | |
3568 | #undef draw_mask | |
3569 | #undef pixels | |
3570 | #undef fb_pixels | |
3571 | #undef d128_0x8000 | |
3572 | #undef msb_mask_low | |
3573 | #undef msb_mask_high | |
3574 | #undef draw_mask_next | |
3575 | #undef pixels_g | |
3576 | #undef blend_pixels | |
3577 | #undef fb_pixels_next | |
3578 | ||
3579 | #define msb_mask q0 | |
3580 | #define draw_mask q1 | |
3581 | #define pixels q2 | |
3582 | #define fb_pixels q3 | |
3583 | #define blend_pixels q4 | |
3584 | #define pixels_no_msb q5 | |
3585 | #define blend_mask q6 | |
3586 | #define fb_pixels_no_msb q7 | |
3587 | #define d128_0x8000 q8 | |
3588 | #define d128_0x0421 q9 | |
3589 | #define fb_pixels_next q10 | |
3590 | #define blend_pixels_next q11 | |
3591 | #define pixels_next q12 | |
3592 | #define draw_mask_next q13 | |
3593 | #define write_mask q14 | |
3594 | ||
3595 | #define pixels_rb q5 | |
3596 | #define pixels_mg q7 | |
3597 | #define pixels_g q7 | |
3598 | #define d128_0x7C1F q8 | |
3599 | #define d128_0x03E0 q9 | |
3600 | #define fb_pixels_rb q10 | |
3601 | #define fb_pixels_g q11 | |
3602 | #define fb_pixels_masked q11 | |
3603 | #define d128_0x83E0 q15 | |
3604 | #define pixels_fourth q7 | |
3605 | #define d128_0x1C07 q12 | |
3606 | #define d128_0x00E0 q13 | |
3607 | #define d128_0x80E0 q13 | |
3608 | ||
3609 | #define msb_mask_low d0 | |
3610 | #define msb_mask_high d1 | |
3611 | ||
3612 | #define blend_blocks_average_set_blend_mask_textured(source) \ | |
3613 | vclt.s16 blend_mask, source, #0 \ | |
3614 | ||
3615 | #define blend_blocks_average_set_stp_bit_textured() \ | |
3616 | vorr.u16 blend_pixels, #0x8000 \ | |
3617 | ||
3618 | #define blend_blocks_average_combine_textured(source) \ | |
3619 | vbif.u16 blend_pixels, source, blend_mask \ | |
3620 | ||
3621 | #define blend_blocks_average_set_blend_mask_untextured(source) \ | |
3622 | ||
3623 | #define blend_blocks_average_set_stp_bit_untextured() \ | |
3624 | ||
3625 | #define blend_blocks_average_combine_untextured(source) \ | |
3626 | ||
3627 | #define blend_blocks_average_mask_set_on() \ | |
3628 | vclt.s16 write_mask, fb_pixels_next, #0 \ | |
3629 | ||
3630 | #define blend_blocks_average_mask_copy_on() \ | |
3631 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
3632 | ||
3633 | #define blend_blocks_average_mask_copy_b_on() \ | |
3634 | vorr.u16 draw_mask_next, draw_mask_next, write_mask \ | |
3635 | ||
3636 | #define blend_blocks_average_mask_set_off() \ | |
3637 | ||
3638 | #define blend_blocks_average_mask_copy_off() \ | |
3639 | vmov draw_mask, draw_mask_next \ | |
3640 | ||
3641 | #define blend_blocks_average_mask_copy_b_off() \ | |
3642 | ||
3643 | #define blend_blocks_average_builder(texturing, mask_evaluate) \ | |
3644 | .align 3; \ | |
3645 | \ | |
3646 | function(blend_blocks_##texturing##_average_##mask_evaluate) \ | |
3647 | stmdb sp!, { r4, r14 }; \ | |
3f0189c6 | 3648 | save_abi_regs(); \ |
75e28f62 | 3649 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 3650 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
3651 | \ |
3652 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 3653 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
3654 | \ |
3655 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3656 | mov c_64, #64; \ | |
3657 | \ | |
3658 | vmov.u16 d128_0x8000, #0x8000; \ | |
e1f6de8f | 3659 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
3660 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
75e28f62 E |
3661 | \ |
3662 | vmov.u16 d128_0x0421, #0x0400; \ | |
e1f6de8f | 3663 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3664 | \ |
3665 | vorr.u16 d128_0x0421, #0x0021; \ | |
e1f6de8f | 3666 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 E |
3667 | \ |
3668 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
3669 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3670 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3671 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3672 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
3673 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3674 | \ | |
3675 | subs num_blocks, num_blocks, #1; \ | |
3676 | beq 1f; \ | |
3677 | \ | |
3678 | 0: \ | |
3679 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 3680 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 E |
3681 | \ |
3682 | vmov pixels, pixels_next; \ | |
e1f6de8f | 3683 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3684 | \ |
3685 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
3686 | \ | |
3687 | blend_blocks_average_mask_copy_##mask_evaluate(); \ | |
e1f6de8f | 3688 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
3689 | \ |
3690 | blend_blocks_average_set_blend_mask_##texturing(pixels); \ | |
3691 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
3692 | vmov fb_pixels, fb_pixels_next; \ | |
3693 | blend_blocks_average_combine_##texturing(pixels); \ | |
3694 | \ | |
3695 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
3696 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
3697 | cmp fb_ptr_cmp, #28; \ | |
3698 | bls 2f; \ | |
3699 | \ | |
e1f6de8f | 3700 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 E |
3701 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
3702 | \ | |
3703 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3704 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3705 | \ | |
3706 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3707 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
3708 | \ | |
3709 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3710 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3711 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
e1f6de8f | 3712 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
75e28f62 E |
3713 | \ |
3714 | 3: \ | |
3715 | subs num_blocks, num_blocks, #1; \ | |
3716 | bne 0b; \ | |
3717 | \ | |
3718 | 1: \ | |
3719 | blend_blocks_average_mask_copy_b_##mask_evaluate(); \ | |
3720 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
3721 | \ | |
3722 | blend_blocks_average_set_blend_mask_##texturing(pixels_next); \ | |
3723 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
3724 | blend_blocks_average_combine_##texturing(pixels_next); \ | |
3725 | \ | |
3726 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3727 | vbif.u16 fb_pixels_next, blend_pixels, draw_mask_next; \ | |
e1f6de8f | 3728 | vst1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 | 3729 | \ |
3f0189c6 | 3730 | restore_abi_regs(); \ |
75e28f62 E |
3731 | ldmia sp!, { r4, pc }; \ |
3732 | \ | |
3733 | 2: \ | |
3734 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3735 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
e1f6de8f | 3736 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
75e28f62 | 3737 | \ |
e1f6de8f | 3738 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 E |
3739 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
3740 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3741 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3742 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3743 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3744 | \ | |
3745 | bal 3b \ | |
3746 | ||
3747 | blend_blocks_average_builder(textured, off) | |
3748 | blend_blocks_average_builder(untextured, off) | |
3749 | blend_blocks_average_builder(textured, on) | |
3750 | blend_blocks_average_builder(untextured, on) | |
3751 | ||
3752 | ||
3753 | #define blend_blocks_add_mask_set_on() \ | |
3754 | vclt.s16 write_mask, fb_pixels, #0 \ | |
3755 | ||
3756 | #define blend_blocks_add_mask_copy_on() \ | |
3757 | vorr.u16 draw_mask, draw_mask, write_mask \ | |
3758 | ||
3759 | #define blend_blocks_add_mask_set_off() \ | |
3760 | ||
3761 | #define blend_blocks_add_mask_copy_off() \ | |
3762 | ||
3763 | ||
3764 | #define blend_blocks_add_textured_builder(mask_evaluate) \ | |
3765 | .align 3; \ | |
3766 | \ | |
3767 | function(blend_blocks_textured_add_##mask_evaluate) \ | |
3768 | stmdb sp!, { r4, r14 }; \ | |
3f0189c6 | 3769 | save_abi_regs(); \ |
75e28f62 | 3770 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 3771 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
3772 | \ |
3773 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 3774 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
3775 | \ |
3776 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3777 | mov c_64, #64; \ | |
3778 | \ | |
3779 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
3780 | vmov.u16 d128_0x03E0, #0x0300; \ | |
3781 | vmov.u16 d128_0x83E0, #0x8000; \ | |
3782 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
3783 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
3784 | vorr.u16 d128_0x83E0, d128_0x83E0, d128_0x03E0; \ | |
3785 | \ | |
e1f6de8f | 3786 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
3787 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
3788 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
75e28f62 | 3789 | vclt.s16 blend_mask, pixels, #0; \ |
e1f6de8f | 3790 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3791 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3792 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3793 | \ | |
3794 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3795 | vorr.u16 pixels, pixels, msb_mask; \ | |
3796 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3797 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
3798 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
3799 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
3800 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3801 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
3802 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3803 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
3804 | \ | |
3805 | subs num_blocks, num_blocks, #1; \ | |
3806 | beq 1f; \ | |
3807 | \ | |
3808 | 0: \ | |
3809 | mov fb_ptr, fb_ptr_next; \ | |
3810 | \ | |
e1f6de8f | 3811 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 3812 | \ |
e1f6de8f | 3813 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3814 | vclt.s16 blend_mask, pixels, #0; \ |
3815 | \ | |
3816 | vorr.u16 pixels, pixels, msb_mask; \ | |
3817 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3818 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
3819 | \ | |
8438c3c7 | 3820 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ |
e1f6de8f | 3821 | pld [fb_ptr_next, #64]; \ |
75e28f62 E |
3822 | \ |
3823 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
8438c3c7 | 3824 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
3825 | \ | |
75e28f62 | 3826 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ |
e1f6de8f | 3827 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
8438c3c7 | 3828 | \ |
75e28f62 E |
3829 | cmp fb_ptr_cmp, #28; \ |
3830 | bls 2f; \ | |
3831 | \ | |
e1f6de8f | 3832 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3833 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3834 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3835 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3836 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3837 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
e1f6de8f | 3838 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3839 | \ |
3840 | 3: \ | |
3841 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
3842 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3843 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
3844 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3845 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
3846 | \ | |
3847 | subs num_blocks, num_blocks, #1; \ | |
3848 | bne 0b; \ | |
3849 | \ | |
3850 | 1: \ | |
3851 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3852 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 3853 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 3854 | \ |
3f0189c6 | 3855 | restore_abi_regs(); \ |
75e28f62 E |
3856 | ldmia sp!, { r4, pc }; \ |
3857 | \ | |
3858 | 2: \ | |
e1f6de8f | 3859 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3860 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
3861 | \ | |
e1f6de8f | 3862 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3863 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3864 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3865 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3866 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
3867 | bal 3b \ | |
3868 | ||
3869 | ||
3870 | #define blend_blocks_add_untextured_builder(mask_evaluate) \ | |
3871 | .align 3; \ | |
3872 | \ | |
3873 | function(blend_blocks_untextured_add_##mask_evaluate) \ | |
3874 | stmdb sp!, { r4, r14 }; \ | |
3f0189c6 | 3875 | save_abi_regs(); \ |
75e28f62 | 3876 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 3877 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
3878 | \ |
3879 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 3880 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
3881 | \ |
3882 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3883 | mov c_64, #64; \ | |
3884 | \ | |
3885 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
3886 | vmov.u16 d128_0x03E0, #0x0300; \ | |
3887 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
3888 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
3889 | \ | |
e1f6de8f | 3890 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
3891 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
3892 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
3893 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
75e28f62 E |
3894 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3895 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3896 | \ | |
3897 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3898 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
3899 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3900 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
3901 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3902 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
3903 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3904 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
3905 | \ | |
3906 | subs num_blocks, num_blocks, #1; \ | |
3907 | beq 1f; \ | |
3908 | \ | |
3909 | 0: \ | |
3910 | mov fb_ptr, fb_ptr_next; \ | |
3911 | \ | |
e1f6de8f | 3912 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 3913 | \ |
e1f6de8f | 3914 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3915 | \ |
3916 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3917 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3918 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
3919 | \ | |
3920 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 3921 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
3922 | \ |
3923 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
3924 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
3925 | cmp fb_ptr_cmp, #28; \ | |
3926 | bls 2f; \ | |
3927 | \ | |
e1f6de8f | 3928 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3929 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3930 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3931 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3932 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
e1f6de8f | 3933 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3934 | \ |
3935 | 3: \ | |
3936 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
3937 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3938 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
3939 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3940 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
3941 | \ | |
3942 | subs num_blocks, num_blocks, #1; \ | |
3943 | bne 0b; \ | |
3944 | \ | |
3945 | 1: \ | |
3946 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3947 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3948 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 3949 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 3950 | \ |
3f0189c6 | 3951 | restore_abi_regs(); \ |
75e28f62 E |
3952 | ldmia sp!, { r4, pc }; \ |
3953 | \ | |
3954 | 2: \ | |
e1f6de8f | 3955 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3956 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
3957 | \ | |
e1f6de8f | 3958 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3959 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3960 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3961 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3962 | bal 3b \ | |
3963 | ||
3964 | ||
3965 | blend_blocks_add_textured_builder(off) | |
3966 | blend_blocks_add_textured_builder(on) | |
3967 | blend_blocks_add_untextured_builder(off) | |
3968 | blend_blocks_add_untextured_builder(on) | |
3969 | ||
3970 | #define blend_blocks_subtract_set_blend_mask_textured() \ | |
3971 | vclt.s16 blend_mask, pixels_next, #0 \ | |
3972 | ||
3973 | #define blend_blocks_subtract_combine_textured() \ | |
3974 | vbif.u16 blend_pixels, pixels, blend_mask \ | |
3975 | ||
718a9e58 | 3976 | #define blend_blocks_subtract_set_stp_textured() \ |
75e28f62 E |
3977 | vorr.u16 blend_pixels, #0x8000 \ |
3978 | ||
3979 | #define blend_blocks_subtract_msb_mask_textured() \ | |
3980 | vorr.u16 pixels, pixels_next, msb_mask \ | |
3981 | ||
3982 | #define blend_blocks_subtract_set_blend_mask_untextured() \ | |
3983 | ||
3984 | #define blend_blocks_subtract_combine_untextured() \ | |
3985 | ||
718a9e58 | 3986 | #define blend_blocks_subtract_set_stp_untextured() \ |
75e28f62 E |
3987 | vorr.u16 blend_pixels, blend_pixels, msb_mask \ |
3988 | ||
3989 | #define blend_blocks_subtract_msb_mask_untextured() \ | |
3990 | ||
3991 | ||
3992 | #define blend_blocks_subtract_mask_set_on() \ | |
3993 | vclt.s16 write_mask, fb_pixels, #0 \ | |
3994 | ||
3995 | #define blend_blocks_subtract_mask_copy_on() \ | |
3996 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
3997 | ||
3998 | #define blend_blocks_subtract_mask_set_off() \ | |
3999 | ||
4000 | #define blend_blocks_subtract_mask_copy_off() \ | |
4001 | vmov draw_mask, draw_mask_next \ | |
4002 | ||
4003 | ||
4004 | #define blend_blocks_subtract_builder(texturing, mask_evaluate) \ | |
4005 | .align 3; \ | |
4006 | \ | |
4007 | function(blend_blocks_##texturing##_subtract_##mask_evaluate) \ | |
4008 | stmdb sp!, { r4, r14 }; \ | |
3f0189c6 | 4009 | save_abi_regs(); \ |
75e28f62 | 4010 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 4011 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
4012 | \ |
4013 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 4014 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
4015 | \ |
4016 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4017 | mov c_64, #64; \ | |
4018 | \ | |
4019 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4020 | vmov.u16 d128_0x03E0, #0x0300; \ | |
4021 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
4022 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
4023 | \ | |
e1f6de8f | 4024 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
4025 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
4026 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ | |
75e28f62 | 4027 | blend_blocks_subtract_set_blend_mask_##texturing(); \ |
e1f6de8f | 4028 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4029 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
4030 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
4031 | \ | |
4032 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
4033 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4034 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4035 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4036 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4037 | \ | |
4038 | subs num_blocks, num_blocks, #1; \ | |
4039 | beq 1f; \ | |
4040 | \ | |
4041 | 0: \ | |
4042 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
4043 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 4044 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 4045 | \ |
e1f6de8f | 4046 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
4047 | blend_blocks_subtract_msb_mask_##texturing(); \ |
4048 | \ | |
e1f6de8f | 4049 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
4050 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
4051 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
718a9e58 | 4052 | blend_blocks_subtract_set_stp_##texturing(); \ |
75e28f62 E |
4053 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ |
4054 | blend_blocks_subtract_combine_##texturing(); \ | |
4055 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
4056 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
4057 | \ | |
4058 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4059 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4060 | cmp fb_ptr_cmp, #28; \ | |
4061 | bls 2f; \ | |
4062 | \ | |
e1f6de8f | 4063 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4064 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
4065 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4066 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4067 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
e1f6de8f | 4068 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4069 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
4070 | \ | |
4071 | 3: \ | |
4072 | subs num_blocks, num_blocks, #1; \ | |
4073 | bne 0b; \ | |
4074 | \ | |
4075 | 1: \ | |
4076 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
4077 | \ | |
4078 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
4079 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
718a9e58 | 4080 | blend_blocks_subtract_set_stp_##texturing(); \ |
75e28f62 E |
4081 | blend_blocks_subtract_combine_##texturing(); \ |
4082 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4083 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4084 | \ |
3f0189c6 | 4085 | restore_abi_regs(); \ |
75e28f62 E |
4086 | ldmia sp!, { r4, pc }; \ |
4087 | \ | |
4088 | 2: \ | |
e1f6de8f | 4089 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
4090 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
75e28f62 E |
4091 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
4092 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4093 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4094 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4095 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4096 | bal 3b \ | |
4097 | ||
4098 | ||
4099 | blend_blocks_subtract_builder(textured, off) | |
4100 | blend_blocks_subtract_builder(textured, on) | |
4101 | blend_blocks_subtract_builder(untextured, off) | |
4102 | blend_blocks_subtract_builder(untextured, on) | |
4103 | ||
4104 | ||
4105 | #define blend_blocks_add_fourth_textured_builder(mask_evaluate) \ | |
4106 | .align 3; \ | |
4107 | \ | |
4108 | function(blend_blocks_textured_add_fourth_##mask_evaluate) \ | |
4109 | stmdb sp!, { r4, r14 }; \ | |
3f0189c6 | 4110 | save_abi_regs(); \ |
75e28f62 | 4111 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 4112 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
4113 | \ |
4114 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 4115 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
4116 | \ |
4117 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4118 | mov c_64, #64; \ | |
4119 | \ | |
4120 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4121 | vmov.u16 d128_0x03E0, #0x0300; \ | |
75e28f62 | 4122 | vmov.u16 d128_0x1C07, #0x1C00; \ |
d1c75d1e | 4123 | vmov.u16 d128_0x00E0, #0x00E0; \ |
75e28f62 E |
4124 | vorr.u16 d128_0x7C1F, #0x001F; \ |
4125 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
75e28f62 | 4126 | vorr.u16 d128_0x1C07, #0x0007; \ |
75e28f62 | 4127 | \ |
e1f6de8f | 4128 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
4129 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
4130 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
75e28f62 | 4131 | vclt.s16 blend_mask, pixels, #0; \ |
e1f6de8f | 4132 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4133 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4134 | vshr.s16 pixels_fourth, pixels, #2; \ | |
d1c75d1e | 4135 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
75e28f62 E |
4136 | \ |
4137 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
d1c75d1e E |
4138 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4139 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4140 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
75e28f62 | 4141 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
d1c75d1e | 4142 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
75e28f62 | 4143 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
d1c75d1e | 4144 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
75e28f62 E |
4145 | \ |
4146 | subs num_blocks, num_blocks, #1; \ | |
4147 | beq 1f; \ | |
4148 | \ | |
4149 | 0: \ | |
4150 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 4151 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 4152 | \ |
d1c75d1e | 4153 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
718a9e58 | 4154 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
d1c75d1e E |
4155 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
4156 | \ | |
e1f6de8f | 4157 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 | 4158 | vclt.s16 blend_mask, pixels, #0; \ |
75e28f62 | 4159 | vshr.s16 pixels_fourth, pixels, #2; \ |
d1c75d1e | 4160 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
75e28f62 E |
4161 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
4162 | \ | |
4163 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4164 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
4165 | \ |
4166 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4167 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4168 | cmp fb_ptr_cmp, #28; \ | |
4169 | bls 2f; \ | |
4170 | \ | |
e1f6de8f | 4171 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4172 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
75e28f62 | 4173 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
d1c75d1e E |
4174 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4175 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
e1f6de8f | 4176 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4177 | \ |
4178 | 3: \ | |
d1c75d1e | 4179 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ |
75e28f62 | 4180 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
d1c75d1e | 4181 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
75e28f62 | 4182 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
d1c75d1e | 4183 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
75e28f62 E |
4184 | \ |
4185 | subs num_blocks, num_blocks, #1; \ | |
4186 | bne 0b; \ | |
4187 | \ | |
4188 | 1: \ | |
4189 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
718a9e58 | 4190 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
d1c75d1e | 4191 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
718a9e58 | 4192 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
75e28f62 | 4193 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
e1f6de8f | 4194 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4195 | \ |
3f0189c6 | 4196 | restore_abi_regs(); \ |
75e28f62 E |
4197 | ldmia sp!, { r4, pc }; \ |
4198 | \ | |
4199 | 2: \ | |
e1f6de8f | 4200 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
d1c75d1e | 4201 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
75e28f62 | 4202 | \ |
e1f6de8f | 4203 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4204 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
75e28f62 | 4205 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
d1c75d1e | 4206 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ |
75e28f62 E |
4207 | bal 3b \ |
4208 | ||
4209 | ||
d1c75d1e | 4210 | |
75e28f62 E |
4211 | #define blend_blocks_add_fourth_untextured_builder(mask_evaluate) \ |
4212 | .align 3; \ | |
4213 | \ | |
4214 | function(blend_blocks_untextured_add_fourth_##mask_evaluate) \ | |
4215 | stmdb sp!, { r4, r14 }; \ | |
3f0189c6 | 4216 | save_abi_regs(); \ |
75e28f62 | 4217 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 4218 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
4219 | \ |
4220 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 4221 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
4222 | \ |
4223 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4224 | mov c_64, #64; \ | |
4225 | \ | |
4226 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4227 | vmov.u16 d128_0x03E0, #0x0300; \ | |
75e28f62 E |
4228 | vmov.u16 d128_0x1C07, #0x1C00; \ |
4229 | vmov.u16 d128_0x00E0, #0x00E0; \ | |
4230 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
4231 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
75e28f62 E |
4232 | vorr.u16 d128_0x1C07, #0x0007; \ |
4233 | \ | |
e1f6de8f | 4234 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
4235 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
4236 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
4237 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
75e28f62 E |
4238 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4239 | vshr.s16 pixels_fourth, pixels, #2; \ | |
4240 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
4241 | \ | |
4242 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4243 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
4244 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4245 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4246 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4247 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4248 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
4249 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
4250 | \ | |
4251 | subs num_blocks, num_blocks, #1; \ | |
4252 | beq 1f; \ | |
4253 | \ | |
4254 | 0: \ | |
4255 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 4256 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 4257 | \ |
e1f6de8f | 4258 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
4259 | \ |
4260 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
4261 | vshr.s16 pixels_fourth, pixels, #2; \ | |
4262 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
4263 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
4264 | \ | |
4265 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4266 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
4267 | \ |
4268 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4269 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4270 | cmp fb_ptr_cmp, #28; \ | |
4271 | bls 2f; \ | |
4272 | \ | |
e1f6de8f | 4273 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4274 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4275 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4276 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
4277 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
e1f6de8f | 4278 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4279 | \ |
4280 | 3: \ | |
4281 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4282 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4283 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4284 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
4285 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
4286 | \ | |
4287 | subs num_blocks, num_blocks, #1; \ | |
4288 | bne 0b; \ | |
4289 | \ | |
4290 | 1: \ | |
4291 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
4292 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
4293 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4294 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4295 | \ |
3f0189c6 | 4296 | restore_abi_regs(); \ |
75e28f62 E |
4297 | ldmia sp!, { r4, pc }; \ |
4298 | \ | |
4299 | 2: \ | |
e1f6de8f | 4300 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4301 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4302 | \ | |
e1f6de8f | 4303 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4304 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4305 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4306 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4307 | bal 3b \ | |
4308 | ||
4309 | ||
4310 | blend_blocks_add_fourth_textured_builder(off) | |
4311 | blend_blocks_add_fourth_textured_builder(on) | |
4312 | blend_blocks_add_fourth_untextured_builder(off) | |
4313 | blend_blocks_add_fourth_untextured_builder(on) | |
4314 | ||
4315 | // TODO: Optimize this more. Need a scene that actually uses it for | |
4316 | // confirmation.. | |
4317 | ||
4318 | .align 3 | |
4319 | ||
4320 | function(blend_blocks_textured_unblended_on) | |
4321 | stmdb sp!, { r4, r14 } | |
3f0189c6 | 4322 | save_abi_regs() |
75e28f62 | 4323 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
e1f6de8f | 4324 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
4325 | |
4326 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
e1f6de8f | 4327 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
75e28f62 E |
4328 | |
4329 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
4330 | mov c_64, #64 | |
4331 | ||
e1f6de8f | 4332 | ldr fb_ptr, [pixel_ptr, #28] |
4333 | vld1.u16 { fb_pixels }, [fb_ptr] | |
4334 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
75e28f62 | 4335 | vclt.s16 write_mask, fb_pixels, #0 |
e1f6de8f | 4336 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
75e28f62 E |
4337 | |
4338 | subs num_blocks, num_blocks, #1 | |
4339 | beq 1f | |
4340 | ||
4341 | 0: | |
134f81ec | 4342 | vorr.u16 pixels, pixels, msb_mask |
75e28f62 E |
4343 | vorr.u16 draw_mask, draw_mask, write_mask |
4344 | vbif.u16 fb_pixels, pixels, draw_mask | |
e1f6de8f | 4345 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 | 4346 | |
e1f6de8f | 4347 | ldr fb_ptr, [pixel_ptr, #28] |
4348 | vld1.u16 { fb_pixels }, [fb_ptr] | |
4349 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
75e28f62 | 4350 | vclt.s16 write_mask, fb_pixels, #0 |
e1f6de8f | 4351 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
75e28f62 E |
4352 | |
4353 | subs num_blocks, num_blocks, #1 | |
4354 | bne 0b | |
4355 | ||
4356 | 1: | |
134f81ec | 4357 | vorr.u16 pixels, pixels, msb_mask |
75e28f62 E |
4358 | vorr.u16 draw_mask, draw_mask, write_mask |
4359 | vbif.u16 fb_pixels, pixels, draw_mask | |
e1f6de8f | 4360 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 | 4361 | |
3f0189c6 | 4362 | restore_abi_regs() |
75e28f62 E |
4363 | ldmia sp!, { r4, pc } |
4364 | ||
4365 | ||
4366 | function(blend_blocks_textured_unblended_off) | |
4367 | bx lr | |
4368 | ||
4369 | ||
4370 | function(warmup) | |
4371 | mov r3, #64 | |
4372 | cmp r0, #0 | |
4373 | bxeq lr | |
4374 | ||
4375 | 0: | |
e1f6de8f | 4376 | vld1.u32 { u_whole_8, v_whole_8 }, [r1, :128], r3 |
75e28f62 E |
4377 | |
4378 | subs r0, r0, #1 | |
4379 | bne 0b | |
4380 | ||
4381 | bx lr | |
4382 | ||
6c4a10c4 | 4383 | #undef vram_ptr |
75e28f62 | 4384 | #undef color |
6c4a10c4 | 4385 | #undef width |
75e28f62 | 4386 | #undef height |
6c4a10c4 | 4387 | #undef pitch |
75e28f62 E |
4388 | |
4389 | #define vram_ptr r0 | |
6c4a10c4 E |
4390 | #define color r1 |
4391 | #define width r2 | |
4392 | #define height r3 | |
75e28f62 | 4393 | |
6c4a10c4 | 4394 | #define pitch r1 |
75e28f62 | 4395 | |
6c4a10c4 | 4396 | #define num_width r12 |
75e28f62 | 4397 | |
87c45ad1 E |
4398 | #undef colors_a |
4399 | #undef colors_b | |
75e28f62 | 4400 | |
87c45ad1 E |
4401 | #define colors_a q0 |
4402 | #define colors_b q1 | |
75e28f62 E |
4403 | |
4404 | .align 3 | |
4405 | ||
4406 | function(render_block_fill_body) | |
87c45ad1 | 4407 | vdup.u16 colors_a, color |
6c4a10c4 | 4408 | mov pitch, #2048 |
75e28f62 | 4409 | |
87c45ad1 | 4410 | vmov colors_b, colors_a |
75e28f62 | 4411 | sub pitch, pitch, width, lsl #1 |
75e28f62 | 4412 | |
6c4a10c4 | 4413 | mov num_width, width |
75e28f62 | 4414 | |
6c4a10c4 | 4415 | 0: |
e1f6de8f | 4416 | vst1.u32 { colors_a, colors_b }, [vram_ptr, :256]! |
75e28f62 | 4417 | |
d1c75d1e | 4418 | subs num_width, num_width, #16 |
6c4a10c4 | 4419 | bne 0b |
75e28f62 | 4420 | |
75e28f62 | 4421 | add vram_ptr, vram_ptr, pitch |
6c4a10c4 E |
4422 | mov num_width, width |
4423 | ||
75e28f62 E |
4424 | subs height, height, #1 |
4425 | bne 0b | |
75e28f62 | 4426 | |
6c4a10c4 E |
4427 | bx lr |
4428 | ||
75e28f62 E |
4429 | |
4430 | #undef x | |
4431 | #undef y | |
4432 | #undef width | |
4433 | #undef height | |
4434 | #undef fb_ptr | |
4435 | #undef texture_mask | |
4436 | #undef num_blocks | |
4437 | #undef temp | |
4438 | #undef dirty_textures_mask | |
4439 | #undef clut_ptr | |
4440 | #undef current_texture_mask | |
4441 | ||
4442 | #define psx_gpu r0 | |
4443 | #define x r1 | |
4444 | #define y r2 | |
4445 | #define u r3 | |
4446 | #define v r4 | |
4447 | #define width r5 | |
4448 | #define height r6 | |
4449 | #define offset_u r8 | |
4450 | #define offset_v r9 | |
4451 | #define offset_u_right r10 | |
4452 | #define width_rounded r11 | |
4453 | #define height_rounded r12 | |
4454 | ||
4455 | #define texture_offset_base r1 | |
4456 | #define tile_width r2 | |
4457 | #define tile_height r3 | |
4458 | #define num_blocks r4 | |
4459 | #define block r5 | |
4460 | #define sub_tile_height r6 | |
4461 | #define fb_ptr r7 | |
4462 | #define texture_mask r8 | |
4463 | #define column_data r9 | |
4464 | #define texture_offset r10 | |
4465 | #define tiles_remaining r11 | |
4466 | #define fb_ptr_advance_column r12 | |
4467 | #define texture_block_ptr r14 | |
4468 | ||
8184d7c5 | 4469 | #define temp r14 |
4470 | ||
75e28f62 E |
4471 | #define texture_page_ptr r3 |
4472 | #define left_block_mask r4 | |
4473 | #define right_block_mask r5 | |
4474 | #define texture_mask_rev r10 | |
4475 | #define control_mask r11 | |
4476 | ||
4477 | #define dirty_textures_mask r4 | |
4478 | #define clut_ptr r5 | |
4479 | #define current_texture_mask r6 | |
4480 | ||
4481 | ||
4482 | #undef texels | |
4483 | #undef clut_low_a | |
4484 | #undef clut_low_b | |
4485 | #undef clut_high_a | |
4486 | #undef clut_high_b | |
4487 | #undef clut_a | |
4488 | #undef clut_b | |
4489 | #undef texels_low | |
4490 | #undef texels_high | |
4491 | ||
4492 | #define texels d0 | |
4493 | #define draw_masks_fb_ptrs q1 | |
4494 | ||
4495 | #define draw_mask_fb_ptr_left d2 | |
4496 | #define draw_mask_fb_ptr_right d3 | |
4497 | ||
59d15d23 | 4498 | #define draw_mask_fb_ptr_left_a d2 |
4499 | #define draw_mask_fb_ptr_left_b d3 | |
4500 | #define draw_mask_fb_ptr_right_a d10 | |
4501 | #define draw_mask_fb_ptr_right_b d11 | |
4502 | #define draw_masks_fb_ptrs2 q5 | |
4503 | ||
75e28f62 E |
4504 | #define clut_low_a d4 |
4505 | #define clut_low_b d5 | |
4506 | #define clut_high_a d6 | |
4507 | #define clut_high_b d7 | |
4508 | ||
4509 | #define block_masks d8 | |
4510 | #define block_masks_shifted d9 | |
4511 | ||
4512 | #define clut_a q2 | |
4513 | #define clut_b q3 | |
4514 | ||
59d15d23 | 4515 | #define texels_low d12 |
4516 | #define texels_high d13 | |
75e28f62 | 4517 | |
59d15d23 | 4518 | #define texels_wide_low d14 |
4519 | #define texels_wide_high d15 | |
4520 | #define texels_wide q7 | |
75e28f62 E |
4521 | |
4522 | ||
59d15d23 | 4523 | setup_sprite_flush_blocks: |
4524 | vpush { q1 - q5 } | |
75e28f62 | 4525 | |
4d646738 | 4526 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 4527 | bl flush_render_block_buffer |
4d646738 | 4528 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 4529 | |
59d15d23 | 4530 | vpop { q1 - q5 } |
75e28f62 E |
4531 | |
4532 | add block, psx_gpu, #psx_gpu_blocks_offset | |
75e28f62 E |
4533 | bx lr |
4534 | ||
4535 | ||
4536 | setup_sprite_update_texture_4bpp_cache: | |
4537 | stmdb sp!, { r0 - r3, r14 } | |
4538 | bl update_texture_4bpp_cache | |
4539 | ldmia sp!, { r0 - r3, pc } | |
4540 | ||
4541 | ||
4542 | setup_sprite_update_texture_8bpp_cache: | |
4d646738 | 4543 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r14 } |
75e28f62 | 4544 | bl update_texture_8bpp_cache |
4d646738 | 4545 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS pc } |
75e28f62 E |
4546 | |
4547 | ||
4548 | #define setup_sprite_tiled_initialize_4bpp() \ | |
4549 | ldr dirty_textures_mask, \ | |
e1f6de8f | 4550 | [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset]; \ |
4551 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset]; \ | |
75e28f62 | 4552 | \ |
e1f6de8f | 4553 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset]; \ |
4554 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128]; \ | |
75e28f62 E |
4555 | \ |
4556 | tst current_texture_mask, dirty_textures_mask; \ | |
4557 | vuzp.u8 clut_a, clut_b; \ | |
4558 | \ | |
4559 | blne setup_sprite_update_texture_4bpp_cache \ | |
4560 | ||
4561 | #define setup_sprite_tiled_initialize_8bpp() \ | |
4562 | ldr dirty_textures_mask, \ | |
e1f6de8f | 4563 | [psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset]; \ |
4564 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset]; \ | |
75e28f62 E |
4565 | \ |
4566 | tst current_texture_mask, dirty_textures_mask; \ | |
4567 | blne setup_sprite_update_texture_8bpp_cache \ | |
4568 | ||
4569 | ||
75e28f62 E |
4570 | #define setup_sprite_block_count_single() \ |
4571 | sub_tile_height \ | |
4572 | ||
4573 | #define setup_sprite_block_count_double() \ | |
4574 | sub_tile_height, lsl #1 \ | |
4575 | ||
4576 | #define setup_sprite_tile_add_blocks(type) \ | |
4577 | add num_blocks, num_blocks, setup_sprite_block_count_##type(); \ | |
4578 | cmp num_blocks, #MAX_BLOCKS; \ | |
4579 | \ | |
59d15d23 | 4580 | movgt num_blocks, setup_sprite_block_count_##type(); \ |
4581 | blgt setup_sprite_flush_blocks \ | |
75e28f62 E |
4582 | |
4583 | ||
4584 | #define setup_sprite_tile_full_4bpp(edge) \ | |
4585 | setup_sprite_tile_add_blocks(double); \ | |
4586 | \ | |
4587 | 4: \ | |
4588 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4589 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
4590 | \ | |
e1f6de8f | 4591 | pld [fb_ptr]; \ |
75e28f62 | 4592 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
e1f6de8f | 4593 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4594 | \ |
4595 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4596 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4597 | \ | |
e1f6de8f | 4598 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
75e28f62 E |
4599 | add texture_block_ptr, texture_offset, #8; \ |
4600 | \ | |
4601 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4602 | add block, block, #40; \ | |
4603 | \ | |
4604 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4605 | add fb_ptr, fb_ptr, #16; \ | |
4606 | \ | |
e1f6de8f | 4607 | vst1.u32 { draw_mask_fb_ptr_left }, [block, :64]; \ |
75e28f62 E |
4608 | add block, block, #24; \ |
4609 | \ | |
e1f6de8f | 4610 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4611 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ |
4612 | \ | |
e1f6de8f | 4613 | pld [fb_ptr]; \ |
75e28f62 E |
4614 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ |
4615 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4616 | \ | |
e1f6de8f | 4617 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
75e28f62 E |
4618 | add block, block, #40; \ |
4619 | \ | |
4620 | add texture_offset, texture_offset, #0x10; \ | |
4621 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
4622 | \ | |
e1f6de8f | 4623 | vst1.u32 { draw_mask_fb_ptr_right }, [block, :64]; \ |
75e28f62 E |
4624 | add block, block, #24; \ |
4625 | \ | |
4626 | subs sub_tile_height, sub_tile_height, #1; \ | |
4627 | bne 4b; \ | |
4628 | \ | |
4629 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4630 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4631 | |
4632 | ||
4633 | #define setup_sprite_tile_half_4bpp(edge) \ | |
4634 | setup_sprite_tile_add_blocks(single); \ | |
4635 | \ | |
4636 | 4: \ | |
4637 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4638 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
4639 | \ | |
e1f6de8f | 4640 | pld [fb_ptr]; \ |
75e28f62 | 4641 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
e1f6de8f | 4642 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4643 | \ |
4644 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4645 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4646 | \ | |
e1f6de8f | 4647 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
75e28f62 E |
4648 | add block, block, #40; \ |
4649 | \ | |
4650 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 4651 | vst1.u32 { draw_mask_fb_ptr_##edge }, [block, :64]; \ |
75e28f62 E |
4652 | \ |
4653 | add block, block, #24; \ | |
4654 | add texture_offset, texture_offset, #0x10; \ | |
4655 | \ | |
4656 | add fb_ptr, fb_ptr, #2048; \ | |
4657 | subs sub_tile_height, sub_tile_height, #1; \ | |
4658 | \ | |
4659 | bne 4b; \ | |
4660 | \ | |
4661 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4662 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4663 | |
4664 | ||
4665 | #define setup_sprite_tile_full_8bpp(edge) \ | |
4666 | setup_sprite_tile_add_blocks(double); \ | |
4667 | add block, block, #16; \ | |
4668 | \ | |
4669 | 4: \ | |
4670 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4671 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
4672 | \ | |
e1f6de8f | 4673 | pld [fb_ptr]; \ |
75e28f62 | 4674 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
e1f6de8f | 4675 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4676 | \ |
4677 | add texture_block_ptr, texture_offset, #8; \ | |
e1f6de8f | 4678 | vst1.u32 { texels }, [block, :64]; \ |
75e28f62 E |
4679 | \ |
4680 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4681 | add block, block, #24; \ | |
4682 | \ | |
4683 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4684 | \ | |
4685 | add fb_ptr, fb_ptr, #16; \ | |
e1f6de8f | 4686 | vst1.u32 { draw_mask_fb_ptr_left }, [block, :64]; \ |
75e28f62 E |
4687 | \ |
4688 | add block, block, #40; \ | |
e1f6de8f | 4689 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
4690 | pld [fb_ptr]; \ | |
75e28f62 E |
4691 | \ |
4692 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ | |
e1f6de8f | 4693 | vst1.u32 { texels }, [block, :64]; \ |
75e28f62 E |
4694 | add block, block, #24; \ |
4695 | \ | |
4696 | add texture_offset, texture_offset, #0x10; \ | |
4697 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
4698 | \ | |
e1f6de8f | 4699 | vst1.u32 { draw_mask_fb_ptr_right }, [block, :64]; \ |
75e28f62 E |
4700 | add block, block, #40; \ |
4701 | \ | |
4702 | subs sub_tile_height, sub_tile_height, #1; \ | |
4703 | bne 4b; \ | |
4704 | \ | |
4705 | sub block, block, #16; \ | |
4706 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4707 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4708 | |
4709 | ||
4710 | #define setup_sprite_tile_half_8bpp(edge) \ | |
4711 | setup_sprite_tile_add_blocks(single); \ | |
4712 | add block, block, #16; \ | |
4713 | \ | |
4714 | 4: \ | |
4715 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4716 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
e1f6de8f | 4717 | pld [fb_ptr]; \ |
75e28f62 E |
4718 | \ |
4719 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 4720 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 | 4721 | \ |
e1f6de8f | 4722 | vst1.u32 { texels }, [block, :64]; \ |
75e28f62 E |
4723 | add block, block, #24; \ |
4724 | \ | |
e1f6de8f | 4725 | vst1.u32 { draw_mask_fb_ptr_##edge }, [block, :64]; \ |
75e28f62 E |
4726 | add block, block, #40; \ |
4727 | \ | |
4728 | add texture_offset, texture_offset, #0x10; \ | |
4729 | add fb_ptr, fb_ptr, #2048; \ | |
4730 | \ | |
4731 | subs sub_tile_height, sub_tile_height, #1; \ | |
4732 | bne 4b; \ | |
4733 | \ | |
4734 | sub block, block, #16; \ | |
4735 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4736 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4737 | |
4738 | ||
4739 | #define setup_sprite_tile_column_edge_pre_adjust_half_right() \ | |
4740 | add texture_offset, texture_offset_base, #8; \ | |
4741 | add fb_ptr, fb_ptr, #16 \ | |
4742 | ||
4743 | #define setup_sprite_tile_column_edge_pre_adjust_half_left() \ | |
4744 | mov texture_offset, texture_offset_base \ | |
4745 | ||
4746 | #define setup_sprite_tile_column_edge_pre_adjust_half(edge) \ | |
4747 | setup_sprite_tile_column_edge_pre_adjust_half_##edge() \ | |
4748 | ||
4749 | #define setup_sprite_tile_column_edge_pre_adjust_full(edge) \ | |
4750 | mov texture_offset, texture_offset_base \ | |
4751 | ||
4752 | #define setup_sprite_tile_column_edge_post_adjust_half_right() \ | |
4753 | sub fb_ptr, fb_ptr, #16 \ | |
4754 | ||
4755 | #define setup_sprite_tile_column_edge_post_adjust_half_left() \ | |
4756 | ||
4757 | #define setup_sprite_tile_column_edge_post_adjust_half(edge) \ | |
4758 | setup_sprite_tile_column_edge_post_adjust_half_##edge() \ | |
4759 | ||
4760 | #define setup_sprite_tile_column_edge_post_adjust_full(edge) \ | |
4761 | ||
4762 | ||
59d15d23 | 4763 | #define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode, \ |
4764 | x4mode) \ | |
75e28f62 | 4765 | mov sub_tile_height, column_data; \ |
59d15d23 | 4766 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
4767 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
4768 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
75e28f62 | 4769 | |
59d15d23 | 4770 | #define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode, \ |
4771 | x4mode) \ | |
75e28f62 E |
4772 | and sub_tile_height, column_data, #0xFF; \ |
4773 | mov tiles_remaining, column_data, lsr #16; \ | |
59d15d23 | 4774 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
4775 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
75e28f62 E |
4776 | \ |
4777 | subs tiles_remaining, tiles_remaining, #1; \ | |
4778 | beq 2f; \ | |
4779 | \ | |
4780 | 3: \ | |
4781 | mov sub_tile_height, #16; \ | |
59d15d23 | 4782 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
75e28f62 E |
4783 | subs tiles_remaining, tiles_remaining, #1; \ |
4784 | bne 3b; \ | |
4785 | \ | |
4786 | 2: \ | |
4787 | uxtb sub_tile_height, column_data, ror #8; \ | |
59d15d23 | 4788 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
4789 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
75e28f62 E |
4790 | |
4791 | ||
4792 | #define setup_sprite_column_data_single() \ | |
4793 | mov column_data, height; \ | |
e1f6de8f | 4794 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] \ |
75e28f62 E |
4795 | |
4796 | #define setup_sprite_column_data_multi() \ | |
4797 | and height_rounded, height_rounded, #0xF; \ | |
4798 | rsb column_data, offset_v, #16; \ | |
4799 | \ | |
4800 | add height_rounded, height_rounded, #1; \ | |
4801 | sub tile_height, tile_height, #1; \ | |
4802 | \ | |
4803 | orr column_data, column_data, tile_height, lsl #16; \ | |
e1f6de8f | 4804 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset]; \ |
75e28f62 E |
4805 | \ |
4806 | orr column_data, column_data, height_rounded, lsl #8 \ | |
4807 | ||
59d15d23 | 4808 | #define setup_sprite_setup_left_draw_mask_fb_ptr() \ |
4809 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
4810 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ | |
4811 | ||
4812 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column() \ | |
4813 | mov fb_ptr_advance_column, #32; \ | |
4814 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
4815 | \ | |
ed0fd81d | 4816 | sub fb_ptr_advance_column, fb_ptr_advance_column, height, lsl #11; \ |
59d15d23 | 4817 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ |
4818 | ||
4819 | #define setup_sprite_setup_right_draw_mask_fb_ptr() \ | |
4820 | vdup.u8 draw_mask_fb_ptr_left, block_masks[4]; \ | |
4821 | vdup.u8 draw_mask_fb_ptr_right, block_masks[5] \ | |
4822 | ||
4823 | #define setup_sprite_tile_column_width_single(tm, multi_height, edge_mode, \ | |
4824 | edge, x4mode) \ | |
4825 | setup_sprite_##tm##_single_##multi_height##_##edge_mode##_##edge##x4mode: \ | |
75e28f62 E |
4826 | setup_sprite_column_data_##multi_height(); \ |
4827 | vext.32 block_masks_shifted, block_masks, block_masks, #1; \ | |
4828 | vorr.u32 block_masks, block_masks, block_masks_shifted; \ | |
59d15d23 | 4829 | setup_sprite_setup_left_draw_mask_fb_ptr##x4mode(); \ |
75e28f62 | 4830 | \ |
59d15d23 | 4831 | setup_sprite_tile_column_height_##multi_height(edge_mode, edge, tm, x4mode); \ |
3f0189c6 | 4832 | restore_abi_regs(); \ |
75e28f62 E |
4833 | ldmia sp!, { r4 - r11, pc } \ |
4834 | ||
4835 | #define setup_sprite_tiled_advance_column() \ | |
4836 | add texture_offset_base, texture_offset_base, #0x100; \ | |
4837 | tst texture_offset_base, #0xF00; \ | |
4838 | subeq texture_offset_base, texture_offset_base, #(0x100 + 0xF00) \ | |
4839 | ||
4840 | #define setup_sprite_tile_column_width_multi(tm, multi_height, left_mode, \ | |
59d15d23 | 4841 | right_mode, x4mode) \ |
4842 | setup_sprite_##tm##_multi_##multi_height##_##left_mode##_##right_mode##x4mode:\ | |
75e28f62 | 4843 | setup_sprite_column_data_##multi_height(); \ |
75e28f62 | 4844 | \ |
59d15d23 | 4845 | setup_sprite_setup_left_draw_mask_fb_ptr_advance_column##x4mode(); \ |
75e28f62 | 4846 | \ |
59d15d23 | 4847 | setup_sprite_tile_column_height_##multi_height(left_mode, right, tm, x4mode);\ |
75e28f62 E |
4848 | \ |
4849 | subs tile_width, tile_width, #2; \ | |
4850 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ | |
4851 | \ | |
75e28f62 E |
4852 | beq 1f; \ |
4853 | \ | |
59d15d23 | 4854 | vmov.u8 draw_masks_fb_ptrs, #0; \ |
4855 | vmov.u8 draw_masks_fb_ptrs2, #0; \ | |
4856 | \ | |
75e28f62 E |
4857 | 0: \ |
4858 | setup_sprite_tiled_advance_column(); \ | |
59d15d23 | 4859 | setup_sprite_tile_column_height_##multi_height(full, none, tm, x4mode); \ |
75e28f62 E |
4860 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ |
4861 | subs tile_width, tile_width, #1; \ | |
4862 | bne 0b; \ | |
4863 | \ | |
4864 | 1: \ | |
59d15d23 | 4865 | setup_sprite_setup_right_draw_mask_fb_ptr##x4mode(); \ |
75e28f62 E |
4866 | \ |
4867 | setup_sprite_tiled_advance_column(); \ | |
59d15d23 | 4868 | setup_sprite_tile_column_height_##multi_height(right_mode, left, tm, x4mode);\ |
3f0189c6 | 4869 | restore_abi_regs(); \ |
75e28f62 E |
4870 | ldmia sp!, { r4 - r11, pc } \ |
4871 | ||
4872 | ||
59d15d23 | 4873 | #define setup_sprite_offset_u_adjust() \ |
4874 | ||
4875 | #define setup_sprite_get_left_block_mask() \ | |
4876 | and left_block_mask, left_block_mask, #0xFF \ | |
4877 | ||
4878 | #define setup_sprite_compare_left_block_mask() \ | |
4879 | cmp left_block_mask, #0xFF \ | |
4880 | ||
4881 | #define setup_sprite_get_right_block_mask() \ | |
4882 | uxtb right_block_mask, right_block_mask, ror #8 \ | |
4883 | ||
4884 | #define setup_sprite_compare_right_block_mask() \ | |
4885 | cmp right_block_mask, #0xFF \ | |
4886 | ||
4887 | ||
4888 | ||
4889 | /* 4x stuff */ | |
4890 | #define fb_ptr2 column_data | |
4891 | ||
4892 | #define setup_sprite_offset_u_adjust_4x() \ | |
4893 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
4894 | lsl offset_u_right, #1; \ | |
4895 | lsl offset_u, #1; \ | |
4896 | add offset_u_right, #1 \ | |
4897 | ||
4898 | #define setup_sprite_get_left_block_mask_4x() \ | |
4899 | sxth left_block_mask, left_block_mask \ | |
4900 | ||
4901 | #define setup_sprite_compare_left_block_mask_4x() \ | |
4902 | cmp left_block_mask, #0xFFFFFFFF \ | |
4903 | ||
4904 | #define setup_sprite_get_right_block_mask_4x() \ | |
4905 | sxth right_block_mask, right_block_mask, ror #16 \ | |
4906 | ||
4907 | #define setup_sprite_compare_right_block_mask_4x() \ | |
4908 | cmp right_block_mask, #0xFFFFFFFF \ | |
4909 | ||
4910 | ||
4911 | #define widen_texels_16bpp(texels_) \ | |
4912 | vmov texels_wide_low, texels_; \ | |
4913 | vmov texels_wide_high, texels_; \ | |
4914 | vzip.16 texels_wide_low, texels_wide_high \ | |
4915 | ||
4916 | #define widen_texels_8bpp(texels_) \ | |
4917 | vmov texels_wide_low, texels_; \ | |
4918 | vmov texels_wide_high, texels_; \ | |
4919 | vzip.8 texels_wide_low, texels_wide_high \ | |
4920 | ||
4921 | #define write_block_16bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
e1f6de8f | 4922 | vst1.u32 { texels_ }, [block_, :128]; \ |
59d15d23 | 4923 | add block_, block_, #40; \ |
4924 | \ | |
4925 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
e1f6de8f | 4926 | vst1.u32 { draw_mask_fb_ptr_ }, [block_, :64]; \ |
59d15d23 | 4927 | add block_, block_, #24 \ |
4928 | ||
4929 | /* assumes 16-byte offset already added to block_ */ | |
4930 | #define write_block_8bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
e1f6de8f | 4931 | vst1.u32 { texels_ }, [block_, :64]; \ |
59d15d23 | 4932 | add block_, block_, #24; \ |
4933 | \ | |
4934 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
e1f6de8f | 4935 | vst1.u32 { draw_mask_fb_ptr_ }, [block_, :64]; \ |
59d15d23 | 4936 | add block_, block_, #40 \ |
4937 | ||
4938 | #define do_texture_block_16bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
4939 | draw_mask_fb_ptr_b_) \ | |
4940 | widen_texels_16bpp(texels_low); \ | |
4941 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
4942 | \ | |
4943 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
4944 | \ | |
4945 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
4946 | widen_texels_16bpp(texels_high); \ | |
4947 | \ | |
4948 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
4949 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
4950 | \ | |
4951 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
4952 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
4953 | ||
4954 | #define do_texture_block_8bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
4955 | draw_mask_fb_ptr_b_) \ | |
4956 | widen_texels_8bpp(texels); \ | |
4957 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
4958 | \ | |
4959 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
4960 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
4961 | \ | |
4962 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
4963 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
4964 | \ | |
4965 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
4966 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
4967 | ||
4968 | ||
4969 | #define setup_sprite_tiled_initialize_4bpp_4x() \ | |
e1f6de8f | 4970 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset]; \ |
4971 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128]; \ | |
59d15d23 | 4972 | \ |
4973 | vuzp.u8 clut_a, clut_b \ | |
4974 | ||
4975 | #define setup_sprite_tiled_initialize_8bpp_4x() \ | |
4976 | ||
4977 | ||
4978 | #define setup_sprite_block_count_single_4x() \ | |
4979 | sub_tile_height, lsl #2 \ | |
4980 | ||
4981 | #define setup_sprite_block_count_double_4x() \ | |
4982 | sub_tile_height, lsl #(1+2) \ | |
4983 | ||
4984 | #define setup_sprite_tile_full_4bpp_4x(edge) \ | |
4985 | setup_sprite_tile_add_blocks(double_4x); \ | |
4986 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
4987 | \ | |
4988 | 4: \ | |
4989 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 4990 | pld [fb_ptr]; \ |
59d15d23 | 4991 | \ |
4992 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 4993 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 4994 | \ |
4995 | add texture_block_ptr, texture_offset, #8; \ | |
4996 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4997 | \ | |
4998 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4999 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
5000 | \ | |
5001 | vzip.8 texels_low, texels_high; \ | |
5002 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
5003 | draw_mask_fb_ptr_left_b); \ | |
5004 | \ | |
5005 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5006 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5007 | \ |
e1f6de8f | 5008 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
8438c3c7 | 5009 | add fb_ptr, fb_ptr, #16*2; \ |
59d15d23 | 5010 | \ |
8438c3c7 | 5011 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ |
59d15d23 | 5012 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ |
5013 | \ | |
5014 | vzip.8 texels_low, texels_high; \ | |
5015 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
5016 | draw_mask_fb_ptr_right_b); \ | |
5017 | \ | |
5018 | add texture_offset, texture_offset, #0x10; \ | |
5019 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
5020 | \ | |
5021 | subs sub_tile_height, sub_tile_height, #1; \ | |
5022 | bne 4b; \ | |
5023 | \ | |
5024 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5025 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5026 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5027 | |
5028 | ||
5029 | #define setup_sprite_tile_half_4bpp_4x(edge) \ | |
5030 | setup_sprite_tile_add_blocks(single_4x); \ | |
5031 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
5032 | \ | |
5033 | 4: \ | |
5034 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 5035 | pld [fb_ptr]; \ |
59d15d23 | 5036 | \ |
5037 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5038 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5039 | \ |
5040 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
5041 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
5042 | \ | |
5043 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
5044 | add texture_offset, texture_offset, #0x10; \ | |
5045 | \ | |
5046 | vzip.8 texels_low, texels_high; \ | |
5047 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ | |
5048 | draw_mask_fb_ptr_##edge##_b); \ | |
5049 | \ | |
e1f6de8f | 5050 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5051 | add fb_ptr, fb_ptr, #2048 * 2; \ |
59d15d23 | 5052 | \ |
8438c3c7 | 5053 | subs sub_tile_height, sub_tile_height, #1; \ |
59d15d23 | 5054 | bne 4b; \ |
5055 | \ | |
5056 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5057 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5058 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5059 | |
5060 | ||
5061 | #define setup_sprite_tile_full_8bpp_4x(edge) \ | |
5062 | setup_sprite_tile_add_blocks(double_4x); \ | |
5063 | add block, block, #16; \ | |
5064 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
5065 | \ | |
5066 | 4: \ | |
5067 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 5068 | pld [fb_ptr]; \ |
59d15d23 | 5069 | \ |
5070 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5071 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5072 | \ |
5073 | add texture_block_ptr, texture_offset, #8; \ | |
5074 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
5075 | draw_mask_fb_ptr_left_b); \ | |
5076 | \ | |
e1f6de8f | 5077 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5078 | and texture_block_ptr, texture_block_ptr, texture_mask; \ |
5079 | \ | |
5080 | add fb_ptr, fb_ptr, #16*2; \ | |
5081 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
5082 | \ | |
e1f6de8f | 5083 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5084 | \ |
5085 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
5086 | draw_mask_fb_ptr_right_b); \ | |
5087 | \ | |
5088 | add texture_offset, texture_offset, #0x10; \ | |
5089 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
5090 | \ | |
5091 | subs sub_tile_height, sub_tile_height, #1; \ | |
5092 | bne 4b; \ | |
5093 | \ | |
5094 | sub block, block, #16; \ | |
5095 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5096 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5097 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5098 | |
5099 | ||
5100 | #define setup_sprite_tile_half_8bpp_4x(edge) \ | |
5101 | setup_sprite_tile_add_blocks(single_4x); \ | |
5102 | add block, block, #16; \ | |
5103 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
5104 | \ | |
5105 | 4: \ | |
5106 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 5107 | pld [fb_ptr]; \ |
59d15d23 | 5108 | \ |
5109 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5110 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5111 | \ |
e1f6de8f | 5112 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5113 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ |
5114 | draw_mask_fb_ptr_##edge##_b); \ | |
5115 | \ | |
5116 | add texture_offset, texture_offset, #0x10; \ | |
5117 | add fb_ptr, fb_ptr, #2048 * 2; \ | |
5118 | \ | |
5119 | subs sub_tile_height, sub_tile_height, #1; \ | |
5120 | bne 4b; \ | |
5121 | \ | |
5122 | sub block, block, #16; \ | |
5123 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5124 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5125 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5126 | |
5127 | ||
5128 | #define setup_sprite_tile_column_edge_pre_adjust_half_right_4x() \ | |
5129 | add texture_offset, texture_offset_base, #8; \ | |
5130 | add fb_ptr, fb_ptr, #16 * 2 \ | |
5131 | ||
5132 | #define setup_sprite_tile_column_edge_pre_adjust_half_left_4x() \ | |
5133 | mov texture_offset, texture_offset_base \ | |
5134 | ||
5135 | #define setup_sprite_tile_column_edge_pre_adjust_half_4x(edge) \ | |
5136 | setup_sprite_tile_column_edge_pre_adjust_half_##edge##_4x() \ | |
5137 | ||
5138 | #define setup_sprite_tile_column_edge_pre_adjust_full_4x(edge) \ | |
5139 | mov texture_offset, texture_offset_base \ | |
5140 | ||
5141 | #define setup_sprite_tile_column_edge_post_adjust_half_right_4x() \ | |
5142 | sub fb_ptr, fb_ptr, #16 * 2 \ | |
5143 | ||
5144 | #define setup_sprite_tile_column_edge_post_adjust_half_left_4x() \ | |
5145 | ||
5146 | #define setup_sprite_tile_column_edge_post_adjust_half_4x(edge) \ | |
5147 | setup_sprite_tile_column_edge_post_adjust_half_##edge##_4x() \ | |
5148 | ||
5149 | #define setup_sprite_tile_column_edge_post_adjust_full_4x(edge) \ | |
5150 | ||
5151 | ||
5152 | #define setup_sprite_setup_left_draw_mask_fb_ptr_4x() \ | |
5153 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
5154 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
5155 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ | |
5156 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
5157 | ||
5158 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column_4x() \ | |
5159 | mov fb_ptr_advance_column, #32 * 2; \ | |
5160 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
5161 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
ed0fd81d | 5162 | sub fb_ptr_advance_column, fb_ptr_advance_column, height, lsl #11 + 1; \ |
59d15d23 | 5163 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ |
5164 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
5165 | ||
5166 | #define setup_sprite_setup_right_draw_mask_fb_ptr_4x() \ | |
5167 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[4]; \ | |
5168 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[5]; \ | |
5169 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[6]; \ | |
5170 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[7] \ | |
5171 | ||
5172 | ||
75e28f62 E |
5173 | // r0: psx_gpu |
5174 | // r1: x | |
5175 | // r2: y | |
5176 | // r3: u | |
e1f6de8f | 5177 | // [sp]: v |
5178 | // [sp + 4]: width | |
5179 | // [sp + 8]: height | |
5180 | // [sp + 12]: color (unused) | |
75e28f62 | 5181 | |
59d15d23 | 5182 | #define setup_sprite_tiled_builder(texture_mode, x4mode) \ |
5183 | \ | |
5184 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, full, \ | |
5185 | x4mode); \ | |
5186 | setup_sprite_tile_column_width_single(texture_mode, multi, full, none, \ | |
5187 | x4mode); \ | |
5188 | setup_sprite_tile_column_width_multi(texture_mode, single, full, full, \ | |
5189 | x4mode); \ | |
5190 | setup_sprite_tile_column_width_single(texture_mode, single, full, none, \ | |
5191 | x4mode); \ | |
5192 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, full, \ | |
5193 | x4mode); \ | |
5194 | setup_sprite_tile_column_width_single(texture_mode, multi, half, right, \ | |
5195 | x4mode); \ | |
5196 | setup_sprite_tile_column_width_multi(texture_mode, single, half, full, \ | |
5197 | x4mode); \ | |
5198 | setup_sprite_tile_column_width_single(texture_mode, single, half, right, \ | |
5199 | x4mode); \ | |
5200 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, half, \ | |
5201 | x4mode); \ | |
5202 | setup_sprite_tile_column_width_single(texture_mode, multi, half, left, \ | |
5203 | x4mode); \ | |
5204 | setup_sprite_tile_column_width_multi(texture_mode, single, full, half, \ | |
5205 | x4mode); \ | |
5206 | setup_sprite_tile_column_width_single(texture_mode, single, half, left, \ | |
5207 | x4mode); \ | |
5208 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, half, \ | |
5209 | x4mode); \ | |
5210 | setup_sprite_tile_column_width_multi(texture_mode, single, half, half, \ | |
5211 | x4mode); \ | |
75e28f62 E |
5212 | \ |
5213 | .align 4; \ | |
5214 | \ | |
59d15d23 | 5215 | function(setup_sprite_##texture_mode##x4mode) \ |
75e28f62 | 5216 | stmdb sp!, { r4 - r11, r14 }; \ |
59d15d23 | 5217 | setup_sprite_tiled_initialize_##texture_mode##x4mode(); \ |
75e28f62 | 5218 | \ |
e1f6de8f | 5219 | ldr v, [sp, #36]; \ |
75e28f62 E |
5220 | and offset_u, u, #0xF; \ |
5221 | \ | |
e1f6de8f | 5222 | ldr width, [sp, #40]; \ |
5223 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 | 5224 | \ |
e1f6de8f | 5225 | ldr height, [sp, #44]; \ |
75e28f62 E |
5226 | add fb_ptr, fb_ptr, y, lsl #11; \ |
5227 | \ | |
3f0189c6 | 5228 | save_abi_regs(); \ |
5229 | \ | |
75e28f62 E |
5230 | add fb_ptr, fb_ptr, x, lsl #1; \ |
5231 | and offset_v, v, #0xF; \ | |
5232 | \ | |
5233 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
5234 | add width_rounded, offset_u, width; \ | |
5235 | \ | |
5236 | add height_rounded, offset_v, height; \ | |
5237 | add width_rounded, width_rounded, #15; \ | |
5238 | \ | |
5239 | add height_rounded, height_rounded, #15; \ | |
5240 | mov tile_width, width_rounded, lsr #4; \ | |
5241 | \ | |
5242 | /* texture_offset_base = VH-VL-00-00 */\ | |
5243 | mov texture_offset_base, v, lsl #8; \ | |
5244 | and offset_u_right, width_rounded, #0xF; \ | |
5245 | \ | |
5246 | /* texture_offset_base = VH-UH-UL-00 */\ | |
5247 | bfi texture_offset_base, u, #4, #8; \ | |
59d15d23 | 5248 | mov right_block_mask, #0xFFFFFFFE; \ |
5249 | \ | |
5250 | setup_sprite_offset_u_adjust##x4mode(); \ | |
75e28f62 E |
5251 | \ |
5252 | /* texture_offset_base = VH-UH-VL-00 */\ | |
5253 | bfi texture_offset_base, v, #4, #4; \ | |
59d15d23 | 5254 | mov left_block_mask, #0xFFFFFFFF; \ |
75e28f62 E |
5255 | \ |
5256 | mov tile_height, height_rounded, lsr #4; \ | |
5257 | mvn left_block_mask, left_block_mask, lsl offset_u; \ | |
5258 | \ | |
5259 | /* texture_mask = HH-HL-WH-WL */\ | |
e1f6de8f | 5260 | ldrh texture_mask, [psx_gpu, #psx_gpu_texture_mask_width_offset]; \ |
75e28f62 E |
5261 | mov right_block_mask, right_block_mask, lsl offset_u_right; \ |
5262 | \ | |
5263 | /* texture_mask_rev = WH-WL-HH-HL */\ | |
5264 | rev16 texture_mask_rev, texture_mask; \ | |
5265 | vmov block_masks, left_block_mask, right_block_mask; \ | |
5266 | \ | |
5267 | /* texture_mask = HH-HL-HL-WL */\ | |
5268 | bfi texture_mask, texture_mask_rev, #4, #4; \ | |
5269 | /* texture_mask_rev = 00-00-00-WH */\ | |
5270 | mov texture_mask_rev, texture_mask_rev, lsr #12; \ | |
5271 | \ | |
5272 | /* texture_mask = HH-WH-HL-WL */\ | |
5273 | bfi texture_mask, texture_mask_rev, #8, #4; \ | |
59d15d23 | 5274 | setup_sprite_get_left_block_mask##x4mode(); \ |
75e28f62 E |
5275 | \ |
5276 | mov control_mask, #0; \ | |
59d15d23 | 5277 | setup_sprite_compare_left_block_mask##x4mode(); \ |
75e28f62 | 5278 | \ |
59d15d23 | 5279 | setup_sprite_get_right_block_mask##x4mode(); \ |
75e28f62 E |
5280 | orreq control_mask, control_mask, #0x4; \ |
5281 | \ | |
e1f6de8f | 5282 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
59d15d23 | 5283 | setup_sprite_compare_right_block_mask##x4mode(); \ |
75e28f62 E |
5284 | \ |
5285 | orreq control_mask, control_mask, #0x8; \ | |
5286 | cmp tile_width, #1; \ | |
5287 | \ | |
5288 | add block, psx_gpu, #psx_gpu_blocks_offset; \ | |
5289 | orreq control_mask, control_mask, #0x1; \ | |
5290 | \ | |
5291 | cmp tile_height, #1; \ | |
5292 | add block, block, num_blocks, lsl #6; \ | |
5293 | \ | |
5294 | orreq control_mask, control_mask, #0x2; \ | |
8184d7c5 | 5295 | JT_OP_REL(9f, control_mask, temp); \ |
e1f6de8f | 5296 | JT_OP(ldr pc, [pc, control_mask, lsl #2]); \ |
75e28f62 E |
5297 | nop; \ |
5298 | \ | |
8184d7c5 | 5299 | 9: \ |
5300 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_full_full##x4mode); \ | |
5301 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_full_none##x4mode); \ | |
5302 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_full_full##x4mode); \ | |
5303 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_full_none##x4mode); \ | |
5304 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_half_full##x4mode); \ | |
5305 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_half_right##x4mode); \ | |
5306 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_half_full##x4mode); \ | |
5307 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_half_right##x4mode);\ | |
5308 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_full_half##x4mode); \ | |
5309 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_half_left##x4mode); \ | |
5310 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_full_half##x4mode); \ | |
5311 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_half_left##x4mode); \ | |
5312 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_half_half##x4mode); \ | |
75e28f62 | 5313 | .word 0x00000000; \ |
8184d7c5 | 5314 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_half_half##x4mode); \ |
59d15d23 | 5315 | |
5316 | ||
5317 | setup_sprite_tiled_builder(4bpp,); | |
5318 | setup_sprite_tiled_builder(8bpp,); | |
75e28f62 | 5319 | |
59d15d23 | 5320 | #undef draw_mask_fb_ptr_left |
5321 | #undef draw_mask_fb_ptr_right | |
75e28f62 | 5322 | |
59d15d23 | 5323 | setup_sprite_tiled_builder(4bpp, _4x); |
5324 | setup_sprite_tiled_builder(8bpp, _4x); | |
75e28f62 E |
5325 | |
5326 | ||
5327 | #undef block_ptr | |
5328 | #undef num_blocks | |
5329 | #undef clut_ptr | |
5330 | ||
5331 | #define psx_gpu r0 | |
5332 | #define block_ptr r0 | |
5333 | #define num_blocks r1 | |
5334 | #define clut_ptr r2 | |
5335 | #define texel_shift_mask r3 | |
5336 | #define block_pixels_a r4 | |
5337 | #define block_pixels_b r5 | |
5338 | #define texel_0 r6 | |
5339 | #define texel_2 r7 | |
5340 | #define texel_4 r8 | |
5341 | #define texel_6 r9 | |
5342 | #define texel_1 r10 | |
5343 | #define texel_3 r11 | |
5344 | #define texel_5 r12 | |
5345 | #define texel_7 r14 | |
5346 | #define texels_01 r6 | |
5347 | #define texels_23 r7 | |
5348 | #define texels_45 r8 | |
5349 | #define texels_67 r9 | |
5350 | ||
5351 | function(texture_sprite_blocks_8bpp) | |
5352 | stmdb sp!, { r4 - r11, r14 } | |
5353 | movw texel_shift_mask, #(0xFF << 1) | |
5354 | ||
e1f6de8f | 5355 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
5356 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] | |
75e28f62 E |
5357 | |
5358 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
e1f6de8f | 5359 | ldr block_pixels_a, [block_ptr, #16] |
75e28f62 E |
5360 | |
5361 | 0: | |
5362 | and texel_0, texel_shift_mask, block_pixels_a, lsl #1 | |
e1f6de8f | 5363 | ldr block_pixels_b, [block_ptr, #20] |
75e28f62 E |
5364 | |
5365 | and texel_1, texel_shift_mask, block_pixels_a, lsr #7 | |
e1f6de8f | 5366 | ldrh texel_0, [clut_ptr, texel_0] |
75e28f62 E |
5367 | |
5368 | and texel_2, texel_shift_mask, block_pixels_a, lsr #15 | |
e1f6de8f | 5369 | ldrh texel_1, [clut_ptr, texel_1] |
75e28f62 E |
5370 | |
5371 | and texel_3, texel_shift_mask, block_pixels_a, lsr #23 | |
e1f6de8f | 5372 | ldr block_pixels_a, [block_ptr, #(64 + 16)] |
75e28f62 | 5373 | |
e1f6de8f | 5374 | ldrh texel_2, [clut_ptr, texel_2] |
75e28f62 E |
5375 | and texel_4, texel_shift_mask, block_pixels_b, lsl #1 |
5376 | ||
e1f6de8f | 5377 | ldrh texel_3, [clut_ptr, texel_3] |
75e28f62 E |
5378 | and texel_5, texel_shift_mask, block_pixels_b, lsr #7 |
5379 | ||
e1f6de8f | 5380 | ldrh texel_4, [clut_ptr, texel_4] |
75e28f62 E |
5381 | and texel_6, texel_shift_mask, block_pixels_b, lsr #15 |
5382 | ||
e1f6de8f | 5383 | ldrh texel_5, [clut_ptr, texel_5] |
75e28f62 E |
5384 | and texel_7, texel_shift_mask, block_pixels_b, lsr #23 |
5385 | ||
e1f6de8f | 5386 | ldrh texel_6, [clut_ptr, texel_6] |
75e28f62 E |
5387 | orr texels_01, texel_0, texel_1, lsl #16 |
5388 | ||
e1f6de8f | 5389 | ldrh texel_7, [clut_ptr, texel_7] |
75e28f62 E |
5390 | orr texels_23, texel_2, texel_3, lsl #16 |
5391 | ||
5392 | orr texels_45, texel_4, texel_5, lsl #16 | |
e1f6de8f | 5393 | str texels_01, [block_ptr, #0] |
75e28f62 E |
5394 | |
5395 | orr texels_67, texel_6, texel_7, lsl #16 | |
e1f6de8f | 5396 | str texels_23, [block_ptr, #4] |
75e28f62 E |
5397 | |
5398 | subs num_blocks, num_blocks, #1 | |
e1f6de8f | 5399 | str texels_45, [block_ptr, #8] |
75e28f62 | 5400 | |
e1f6de8f | 5401 | str texels_67, [block_ptr, #12] |
75e28f62 E |
5402 | add block_ptr, block_ptr, #64 |
5403 | ||
5404 | bne 0b | |
5405 | ||
5406 | ldmia sp!, { r4 - r11, pc } | |
5407 | ||
5408 | ||
5409 | #undef width_rounded | |
5410 | #undef texture_mask | |
5411 | #undef num_blocks | |
5412 | #undef texture_offset | |
59d15d23 | 5413 | #undef texels_low |
5414 | #undef texels_high | |
5415 | #undef texels_wide_low | |
5416 | #undef texels_wide_high | |
5417 | #undef texels_wide | |
5418 | #undef fb_ptr2 | |
8184d7c5 | 5419 | #undef temp |
75e28f62 E |
5420 | |
5421 | #define psx_gpu r0 | |
5422 | #define x r1 | |
5423 | #define y r2 | |
5424 | #define u r3 | |
5425 | #define v r4 | |
5426 | #define width r5 | |
5427 | #define height r6 | |
5428 | #define left_offset r8 | |
5429 | #define width_rounded r9 | |
5430 | #define right_width r10 | |
59d15d23 | 5431 | |
75e28f62 E |
5432 | #define block_width r11 |
5433 | ||
5434 | #define texture_offset_base r1 | |
5435 | #define texture_mask r2 | |
5436 | #define texture_page_ptr r3 | |
5437 | #define num_blocks r4 | |
5438 | #define block r5 | |
5439 | #define fb_ptr r7 | |
5440 | #define texture_offset r8 | |
5441 | #define blocks_remaining r9 | |
59d15d23 | 5442 | #define fb_ptr2 r10 |
75e28f62 E |
5443 | #define fb_ptr_pitch r12 |
5444 | #define texture_block_ptr r14 | |
5445 | ||
5446 | #define texture_mask_width r2 | |
5447 | #define texture_mask_height r3 | |
5448 | #define left_mask_bits r4 | |
5449 | #define right_mask_bits r5 | |
5450 | ||
5451 | ||
5452 | #undef block_masks | |
5453 | #undef block_masks_shifted | |
5454 | #undef texels | |
5455 | ||
5456 | #define block_masks d0 | |
5457 | #define block_masks_shifted d1 | |
5458 | #define draw_mask_fb_ptr d2 | |
5459 | #define texels q2 | |
5460 | ||
59d15d23 | 5461 | #define draw_mask_fb_ptr_a d2 |
5462 | #define draw_mask_fb_ptr_b d3 | |
5463 | #define texels_low d4 | |
5464 | #define texels_high d5 | |
5465 | #define texels_wide_low d6 | |
5466 | #define texels_wide_high d7 | |
5467 | #define texels_wide q3 | |
75e28f62 | 5468 | |
75e28f62 | 5469 | |
59d15d23 | 5470 | setup_sprites_16bpp_flush: |
5471 | vpush { d0 - d3 } | |
75e28f62 | 5472 | |
4d646738 | 5473 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 5474 | bl flush_render_block_buffer |
4d646738 | 5475 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 5476 | |
59d15d23 | 5477 | vpop { d0 - d3 } |
75e28f62 E |
5478 | |
5479 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5480 | mov num_blocks, block_width | |
5481 | ||
5482 | bx lr | |
5483 | ||
5484 | function(setup_sprite_16bpp) | |
5485 | stmdb sp!, { r4 - r11, r14 } | |
e1f6de8f | 5486 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
75e28f62 | 5487 | |
e1f6de8f | 5488 | ldr v, [sp, #36] |
75e28f62 E |
5489 | add fb_ptr, fb_ptr, y, lsl #11 |
5490 | ||
e1f6de8f | 5491 | ldr width, [sp, #40] |
75e28f62 E |
5492 | add fb_ptr, fb_ptr, x, lsl #1 |
5493 | ||
e1f6de8f | 5494 | ldr height, [sp, #44] |
75e28f62 E |
5495 | and left_offset, u, #0x7 |
5496 | ||
5497 | add texture_offset_base, u, u | |
5498 | add width_rounded, width, #7 | |
5499 | ||
ed0fd81d | 5500 | add texture_offset_base, texture_offset_base, v, lsl #11 |
75e28f62 E |
5501 | mov left_mask_bits, #0xFF |
5502 | ||
e1f6de8f | 5503 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
75e28f62 E |
5504 | add width_rounded, width_rounded, left_offset |
5505 | ||
e1f6de8f | 5506 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
75e28f62 E |
5507 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
5508 | ||
5509 | add texture_mask, texture_mask_width, texture_mask_width | |
5510 | mov right_mask_bits, #0xFE | |
5511 | ||
5512 | and right_width, width_rounded, #0x7 | |
5513 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
5514 | ||
ed0fd81d | 5515 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
75e28f62 E |
5516 | mov block_width, width_rounded, lsr #3 |
5517 | ||
5518 | mov right_mask_bits, right_mask_bits, lsl right_width | |
5519 | movw fb_ptr_pitch, #(2048 + 16) | |
5520 | ||
5521 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4 | |
5522 | vmov block_masks, left_mask_bits, right_mask_bits | |
5523 | ||
e1f6de8f | 5524 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
5525 | add block, psx_gpu, #psx_gpu_blocks_offset |
5526 | ||
6ea0f7bf | 5527 | bic texture_offset_base, texture_offset_base, #0xF |
75e28f62 E |
5528 | cmp block_width, #1 |
5529 | ||
e1f6de8f | 5530 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
75e28f62 E |
5531 | add block, block, num_blocks, lsl #6 |
5532 | ||
5533 | bne 0f | |
5534 | ||
5535 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
5536 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
5537 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
5538 | ||
5539 | 1: | |
5540 | add num_blocks, num_blocks, #1 | |
5541 | cmp num_blocks, #MAX_BLOCKS | |
59d15d23 | 5542 | blgt setup_sprites_16bpp_flush |
75e28f62 E |
5543 | |
5544 | and texture_block_ptr, texture_offset_base, texture_mask | |
5545 | subs height, height, #1 | |
5546 | ||
5547 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5548 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 | 5549 | |
e1f6de8f | 5550 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5551 | add block, block, #40 |
5552 | ||
5553 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5554 | pld [fb_ptr] |
75e28f62 | 5555 | |
e1f6de8f | 5556 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5557 | |
5558 | add block, block, #24 | |
5559 | add texture_offset_base, texture_offset_base, #2048 | |
5560 | add fb_ptr, fb_ptr, #2048 | |
e1f6de8f | 5561 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
5562 | bne 1b |
5563 | ||
5564 | ldmia sp!, { r4 - r11, pc } | |
5565 | ||
5566 | 0: | |
5567 | add num_blocks, num_blocks, block_width | |
5568 | mov texture_offset, texture_offset_base | |
5569 | ||
5570 | cmp num_blocks, #MAX_BLOCKS | |
59d15d23 | 5571 | blgt setup_sprites_16bpp_flush |
75e28f62 E |
5572 | |
5573 | add texture_offset_base, texture_offset_base, #2048 | |
5574 | and texture_block_ptr, texture_offset, texture_mask | |
5575 | ||
5576 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5577 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 | 5578 | |
e1f6de8f | 5579 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5580 | add block, block, #40 |
5581 | ||
5582 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
5583 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5584 | pld [fb_ptr] |
75e28f62 | 5585 | |
e1f6de8f | 5586 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5587 | subs blocks_remaining, block_width, #2 |
5588 | ||
5589 | add texture_offset, texture_offset, #16 | |
5590 | add fb_ptr, fb_ptr, #16 | |
5591 | ||
5592 | vmov.u8 draw_mask_fb_ptr, #0 | |
5593 | ||
5594 | add block, block, #24 | |
5595 | beq 2f | |
5596 | ||
5597 | 1: | |
5598 | and texture_block_ptr, texture_offset, texture_mask | |
5599 | subs blocks_remaining, blocks_remaining, #1 | |
5600 | ||
5601 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5602 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 | 5603 | |
e1f6de8f | 5604 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5605 | add block, block, #40 |
5606 | ||
5607 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5608 | pld [fb_ptr] |
75e28f62 | 5609 | |
e1f6de8f | 5610 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5611 | |
5612 | add texture_offset, texture_offset, #16 | |
5613 | add fb_ptr, fb_ptr, #16 | |
5614 | ||
5615 | add block, block, #24 | |
5616 | bne 1b | |
5617 | ||
5618 | 2: | |
5619 | and texture_block_ptr, texture_offset, texture_mask | |
5620 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5621 | ||
e1f6de8f | 5622 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 E |
5623 | vdup.u8 draw_mask_fb_ptr, block_masks[4] |
5624 | ||
e1f6de8f | 5625 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5626 | add block, block, #40 |
5627 | ||
5628 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5629 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5630 | |
5631 | add block, block, #24 | |
5632 | subs height, height, #1 | |
5633 | ||
5634 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
e1f6de8f | 5635 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
5636 | |
5637 | bne 0b | |
5638 | ||
5639 | ldmia sp!, { r4 - r11, pc } | |
5640 | ||
5641 | ||
59d15d23 | 5642 | // 4x version |
5643 | // FIXME: duplicate code with normal version :( | |
5644 | #undef draw_mask_fb_ptr | |
5645 | ||
5646 | function(setup_sprite_16bpp_4x) | |
5647 | stmdb sp!, { r4 - r11, r14 } | |
e1f6de8f | 5648 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
59d15d23 | 5649 | |
e1f6de8f | 5650 | ldr v, [sp, #36] |
59d15d23 | 5651 | add fb_ptr, fb_ptr, y, lsl #11 |
5652 | ||
e1f6de8f | 5653 | ldr width, [sp, #40] |
59d15d23 | 5654 | add fb_ptr, fb_ptr, x, lsl #1 |
5655 | ||
e1f6de8f | 5656 | ldr height, [sp, #44] |
59d15d23 | 5657 | and left_offset, u, #0x7 |
5658 | ||
5659 | add texture_offset_base, u, u | |
5660 | add width_rounded, width, #7 | |
5661 | ||
ed0fd81d | 5662 | add texture_offset_base, texture_offset_base, v, lsl #11 |
59d15d23 | 5663 | movw left_mask_bits, #0xFFFF |
5664 | ||
e1f6de8f | 5665 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
59d15d23 | 5666 | add width_rounded, width_rounded, left_offset |
5667 | ||
5668 | lsl left_offset, #1 | |
5669 | ||
e1f6de8f | 5670 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
59d15d23 | 5671 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
5672 | ||
5673 | add texture_mask, texture_mask_width, texture_mask_width | |
5674 | movw right_mask_bits, #0xFFFC | |
5675 | ||
5676 | and right_width, width_rounded, #0x7 | |
5677 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
5678 | ||
5679 | lsl right_width, #1 | |
5680 | ||
ed0fd81d | 5681 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
59d15d23 | 5682 | mov block_width, width_rounded, lsr #3 |
5683 | ||
5684 | mov right_mask_bits, right_mask_bits, lsl right_width | |
5685 | movw fb_ptr_pitch, #(2048 + 16) * 2 | |
5686 | ||
5687 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4+1 | |
5688 | vmov block_masks, left_mask_bits, right_mask_bits | |
5689 | ||
e1f6de8f | 5690 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
59d15d23 | 5691 | add block, psx_gpu, #psx_gpu_blocks_offset |
5692 | ||
5693 | bic texture_offset_base, texture_offset_base, #0xF | |
5694 | cmp block_width, #1 | |
5695 | ||
e1f6de8f | 5696 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
59d15d23 | 5697 | add block, block, num_blocks, lsl #6 |
5698 | ||
5699 | lsl block_width, #2 | |
5700 | bne 0f | |
5701 | ||
5702 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
5703 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
5704 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] | |
5705 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
5706 | ||
5707 | 1: | |
5708 | add num_blocks, num_blocks, block_width | |
5709 | cmp num_blocks, #MAX_BLOCKS | |
5710 | blgt setup_sprites_16bpp_flush | |
5711 | ||
5712 | and texture_block_ptr, texture_offset_base, texture_mask | |
5713 | subs height, height, #1 | |
5714 | ||
5715 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5716 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5717 | |
5718 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5719 | ||
5720 | add texture_offset_base, texture_offset_base, #2048 | |
5721 | add fb_ptr, fb_ptr, #2048*2 | |
e1f6de8f | 5722 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
59d15d23 | 5723 | bne 1b |
5724 | ||
5725 | ldmia sp!, { r4 - r11, pc } | |
5726 | ||
5727 | 0: | |
5728 | add num_blocks, num_blocks, block_width | |
5729 | mov texture_offset, texture_offset_base | |
5730 | ||
5731 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] // left_mask_bits | |
5732 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
5733 | ||
5734 | cmp num_blocks, #MAX_BLOCKS | |
5735 | blgt setup_sprites_16bpp_flush | |
5736 | ||
5737 | add texture_offset_base, texture_offset_base, #2048 | |
5738 | and texture_block_ptr, texture_offset, texture_mask | |
5739 | ||
5740 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5741 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5742 | |
5743 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5744 | ||
5745 | subs blocks_remaining, block_width, #2*4 | |
5746 | add texture_offset, texture_offset, #16 | |
5747 | ||
5748 | vmov.u8 draw_mask_fb_ptr_a, #0 | |
5749 | vmov.u8 draw_mask_fb_ptr_b, #0 | |
5750 | ||
5751 | add fb_ptr, fb_ptr, #16*2 | |
5752 | beq 2f | |
5753 | ||
5754 | 1: | |
5755 | and texture_block_ptr, texture_offset, texture_mask | |
5756 | subs blocks_remaining, blocks_remaining, #4 | |
5757 | ||
5758 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5759 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5760 | |
5761 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5762 | add texture_offset, texture_offset, #16 | |
5763 | ||
5764 | add fb_ptr, fb_ptr, #16*2 | |
5765 | bgt 1b | |
5766 | ||
5767 | 2: | |
5768 | vdup.u8 draw_mask_fb_ptr_a, block_masks[4] // right_mask_bits | |
5769 | vdup.u8 draw_mask_fb_ptr_b, block_masks[5] | |
5770 | ||
5771 | and texture_block_ptr, texture_offset, texture_mask | |
5772 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5773 | ||
e1f6de8f | 5774 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5775 | |
5776 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5777 | subs height, height, #1 | |
5778 | ||
5779 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
e1f6de8f | 5780 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
59d15d23 | 5781 | |
5782 | bne 0b | |
5783 | ||
5784 | ldmia sp!, { r4 - r11, pc } | |
5785 | ||
5786 | ||
f0931e56 | 5787 | #undef width |
5788 | #undef right_width | |
5789 | #undef right_mask_bits | |
5790 | #undef color | |
5791 | #undef height | |
5792 | #undef blocks_remaining | |
5793 | #undef colors | |
5794 | #undef right_mask | |
5795 | #undef test_mask | |
5796 | #undef draw_mask | |
5797 | ||
5798 | #define psx_gpu r0 | |
5799 | #define x r1 | |
5800 | #define y r2 | |
5801 | #define width r3 | |
5802 | #define right_width r5 | |
5803 | #define right_mask_bits r6 | |
5804 | #define fb_ptr r7 | |
5805 | #define color r8 | |
5806 | #define height r9 | |
5807 | #define fb_ptr_pitch r12 | |
5808 | ||
5809 | // referenced by setup_sprites_16bpp_flush | |
5810 | #define num_blocks r4 | |
5811 | #define block r5 | |
5812 | #define block_width r11 | |
5813 | ||
5814 | #define color_r r1 | |
5815 | #define color_g r2 | |
5816 | #define color_b r8 | |
5817 | #define blocks_remaining r6 | |
5818 | ||
5819 | #define colors q0 | |
5820 | #define right_mask q1 | |
5821 | #define test_mask q2 | |
5822 | #define draw_mask q2 | |
5823 | #define draw_mask_bits_fb_ptr d6 | |
5824 | ||
5825 | ||
5826 | .align 3 | |
5827 | ||
5828 | function(setup_sprite_untextured) | |
e1f6de8f | 5829 | ldrh r12, [psx_gpu, #psx_gpu_render_state_offset] |
f0931e56 | 5830 | tst r12, #(RENDER_STATE_MASK_EVALUATE | RENDER_FLAGS_MODULATE_TEXELS \ |
5831 | | RENDER_FLAGS_BLEND) | |
e1f6de8f | 5832 | ldrbeq r12, [psx_gpu, #psx_gpu_render_mode_offset] |
d5c08ed3 | 5833 | tsteq r12, #RENDER_INTERLACE_ENABLED |
f0931e56 | 5834 | beq setup_sprite_untextured_simple |
5835 | ||
5836 | stmdb sp!, { r4 - r11, r14 } | |
5837 | ||
e1f6de8f | 5838 | ldr width, [sp, #40] |
5839 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] | |
f0931e56 | 5840 | |
e1f6de8f | 5841 | ldr height, [sp, #44] |
f0931e56 | 5842 | add fb_ptr, fb_ptr, y, lsl #11 |
5843 | ||
5844 | add fb_ptr, fb_ptr, x, lsl #1 | |
5845 | sub right_width, width, #1 | |
5846 | ||
e1f6de8f | 5847 | ldr color, [sp, #48] |
f0931e56 | 5848 | and right_width, #7 |
5849 | ||
5850 | add block_width, width, #7 | |
5851 | add right_width, #1 | |
5852 | ||
5853 | lsr block_width, #3 | |
5854 | mov right_mask_bits, #0xff | |
5855 | ||
5856 | sub fb_ptr_pitch, block_width, #1 | |
5857 | lsl right_mask_bits, right_width | |
5858 | ||
5859 | lsl fb_ptr_pitch, #3+1 | |
5860 | ubfx color_r, color, #3, #5 | |
5861 | ||
5862 | rsb fb_ptr_pitch, #1024*2 | |
5863 | ubfx color_g, color, #11, #5 | |
5864 | ||
e1f6de8f | 5865 | vld1.u32 { test_mask }, [psx_gpu, :128] |
f0931e56 | 5866 | ubfx color_b, color, #19, #5 |
5867 | ||
5868 | vdup.u16 right_mask, right_mask_bits | |
5869 | orr color, color_r, color_b, lsl #10 | |
5870 | ||
e1f6de8f | 5871 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
f0931e56 | 5872 | orr color, color, color_g, lsl #5 |
5873 | ||
5874 | vtst.u16 right_mask, right_mask, test_mask | |
5875 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5876 | ||
5877 | vdup.u16 colors, color | |
5878 | add block, block, num_blocks, lsl #6 | |
5879 | ||
5880 | ||
5881 | setup_sprite_untextured_height_loop: | |
5882 | add num_blocks, block_width | |
5883 | sub blocks_remaining, block_width, #1 | |
5884 | ||
5885 | cmp num_blocks, #MAX_BLOCKS | |
5886 | blgt setup_sprites_16bpp_flush | |
5887 | ||
5888 | cmp blocks_remaining, #0 | |
5889 | ble 1f | |
5890 | ||
5891 | vmov.u8 draw_mask, #0 /* zero_mask */ | |
5892 | vmov.u8 draw_mask_bits_fb_ptr, #0 | |
5893 | ||
5894 | 0: | |
e1f6de8f | 5895 | vst1.u32 { draw_mask }, [block, :128]! |
f0931e56 | 5896 | subs blocks_remaining, #1 |
5897 | ||
e1f6de8f | 5898 | vst1.u32 { colors }, [block, :128] |
f0931e56 | 5899 | add block, block, #24 |
5900 | ||
5901 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
e1f6de8f | 5902 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
f0931e56 | 5903 | |
5904 | add block, block, #24 | |
5905 | add fb_ptr, #8*2 | |
5906 | bgt 0b | |
5907 | ||
5908 | 1: | |
e1f6de8f | 5909 | vst1.u32 { right_mask }, [block, :128]! |
f0931e56 | 5910 | subs height, #1 |
5911 | ||
e1f6de8f | 5912 | vst1.u32 { colors }, [block, :128] |
f0931e56 | 5913 | add block, block, #24 |
5914 | ||
5915 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
e1f6de8f | 5916 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
f0931e56 | 5917 | |
5918 | add block, block, #24 | |
5919 | add fb_ptr, fb_ptr_pitch | |
5920 | ||
e1f6de8f | 5921 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
f0931e56 | 5922 | bgt setup_sprite_untextured_height_loop |
5923 | ||
5924 | ldmia sp!, { r4 - r11, pc } | |
5925 | ||
5926 | ||
5927 | ||
75e28f62 E |
5928 | #undef texture_page_ptr |
5929 | #undef vram_ptr | |
5930 | #undef dirty_textures_mask | |
5931 | #undef current_texture_mask | |
5932 | ||
5933 | #define psx_gpu r0 | |
5934 | #define current_texture_page r1 | |
5935 | #define texture_page_ptr r2 | |
5936 | #define vram_ptr_a r3 | |
5937 | #define current_texture_page_x r12 | |
5938 | #define current_texture_page_y r4 | |
5939 | #define dirty_textures_mask r5 | |
5940 | #define tile_y r6 | |
5941 | #define tile_x r7 | |
5942 | #define sub_y r8 | |
5943 | #define current_texture_mask r9 | |
5944 | #define c_4096 r10 | |
5945 | #define vram_ptr_b r11 | |
5946 | ||
5947 | #define texel_block_a d0 | |
5948 | #define texel_block_b d1 | |
5949 | #define texel_block_expanded_a q1 | |
5950 | #define texel_block_expanded_b q2 | |
5951 | #define texel_block_expanded_ab q2 | |
5952 | #define texel_block_expanded_c q3 | |
5953 | #define texel_block_expanded_d q4 | |
5954 | #define texel_block_expanded_cd q3 | |
5955 | ||
5956 | function(update_texture_4bpp_cache) | |
5957 | stmdb sp!, { r4 - r11, r14 } | |
5958 | vpush { q0 - q3 } | |
5959 | ||
e1f6de8f | 5960 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
75e28f62 | 5961 | |
e1f6de8f | 5962 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
5963 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
75e28f62 E |
5964 | |
5965 | and current_texture_page_x, current_texture_page, #0xF | |
e1f6de8f | 5966 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
75e28f62 E |
5967 | |
5968 | mov current_texture_page_y, current_texture_page, lsr #4 | |
e1f6de8f | 5969 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
75e28f62 E |
5970 | |
5971 | add vram_ptr_a, vram_ptr_a, current_texture_page_y, lsl #19 | |
5972 | mov tile_y, #16 | |
5973 | ||
5974 | add vram_ptr_a, vram_ptr_a, current_texture_page_x, lsl #7 | |
5975 | bic dirty_textures_mask, current_texture_mask | |
5976 | ||
5977 | mov tile_x, #16 | |
e1f6de8f | 5978 | str dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
75e28f62 E |
5979 | |
5980 | mov sub_y, #8 | |
5981 | movw c_4096, #4096 | |
5982 | ||
5983 | add vram_ptr_b, vram_ptr_a, #2048 | |
5984 | ||
5985 | 0: | |
e1f6de8f | 5986 | vld1.u32 { texel_block_a }, [vram_ptr_a, :64], c_4096 |
5987 | vld1.u32 { texel_block_b }, [vram_ptr_b, :64], c_4096 | |
75e28f62 E |
5988 | |
5989 | vmovl.u8 texel_block_expanded_a, texel_block_a | |
5990 | vshll.u8 texel_block_expanded_b, texel_block_a, #4 | |
5991 | vmovl.u8 texel_block_expanded_c, texel_block_b | |
5992 | vshll.u8 texel_block_expanded_d, texel_block_b, #4 | |
5993 | ||
5994 | vbic.u16 texel_block_expanded_a, #0x00F0 | |
5995 | vbic.u16 texel_block_expanded_b, #0x00F0 | |
5996 | vbic.u16 texel_block_expanded_c, #0x00F0 | |
5997 | vbic.u16 texel_block_expanded_d, #0x00F0 | |
5998 | ||
5999 | vorr.u16 texel_block_expanded_ab, texel_block_expanded_a, \ | |
6000 | texel_block_expanded_b | |
6001 | vorr.u16 texel_block_expanded_cd, texel_block_expanded_c, \ | |
6002 | texel_block_expanded_d | |
6003 | ||
6004 | vst1.u32 { texel_block_expanded_ab, texel_block_expanded_cd }, \ | |
e1f6de8f | 6005 | [texture_page_ptr, :256]! |
75e28f62 E |
6006 | |
6007 | subs sub_y, sub_y, #1 | |
6008 | bne 0b | |
6009 | ||
6010 | mov sub_y, #8 | |
6011 | add vram_ptr_a, vram_ptr_a, #8 | |
6012 | add vram_ptr_b, vram_ptr_b, #8 | |
6013 | ||
6014 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6015 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6016 | ||
6017 | subs tile_x, tile_x, #1 | |
6018 | bne 0b | |
6019 | ||
6020 | mov tile_x, #16 | |
6021 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6022 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6023 | ||
6024 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
6025 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
6026 | ||
6027 | subs tile_y, tile_y, #1 | |
6028 | bne 0b | |
6029 | ||
6030 | vpop { q0 - q3 } | |
6031 | ldmia sp!, { r4 - r11, pc } | |
6032 | ||
6033 | ||
6034 | #undef current_texture_page | |
6035 | ||
6036 | #define psx_gpu r0 | |
6037 | #define texture_page r1 | |
6038 | #define texture_page_ptr r2 | |
6039 | #define vram_ptr_a r3 | |
6040 | #define texture_page_x r12 | |
6041 | #define texture_page_y r4 | |
6042 | #define current_texture_page r5 | |
6043 | #define tile_y r6 | |
6044 | #define tile_x r7 | |
6045 | #define sub_y r8 | |
6046 | #define c_4096 r10 | |
6047 | #define vram_ptr_b r11 | |
6048 | ||
6049 | ||
6050 | #undef texels_a | |
6051 | #undef texels_b | |
6052 | ||
6053 | #define texels_a q0 | |
6054 | #define texels_b q1 | |
6055 | #define texels_c q2 | |
6056 | #define texels_d q3 | |
6057 | ||
6058 | ||
6059 | function(update_texture_8bpp_cache_slice) | |
6060 | stmdb sp!, { r4 - r11, r14 } | |
6061 | vpush { q0 - q3 } | |
6062 | ||
e1f6de8f | 6063 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
6064 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
75e28f62 | 6065 | |
e1f6de8f | 6066 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
75e28f62 E |
6067 | mov tile_y, #16 |
6068 | ||
6069 | and texture_page_x, texture_page, #0xF | |
6070 | mov texture_page_y, texture_page, lsr #4 | |
6071 | ||
6072 | add vram_ptr_a, vram_ptr_a, texture_page_x, lsl #7 | |
6073 | mov tile_x, #8 | |
6074 | ||
6075 | add vram_ptr_a, vram_ptr_a, texture_page_y, lsl #19 | |
6076 | eor current_texture_page, current_texture_page, texture_page | |
6077 | ||
6078 | ands current_texture_page, current_texture_page, #0x1 | |
6079 | mov sub_y, #4 | |
6080 | ||
6081 | addne texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
6082 | movw c_4096, #4096 | |
6083 | ||
6084 | add vram_ptr_b, vram_ptr_a, #2048 | |
6085 | ||
6086 | 0: | |
e1f6de8f | 6087 | vld1.u32 { texels_a }, [vram_ptr_a, :128], c_4096 |
6088 | vld1.u32 { texels_b }, [vram_ptr_b, :128], c_4096 | |
6089 | vld1.u32 { texels_c }, [vram_ptr_a, :128], c_4096 | |
6090 | vld1.u32 { texels_d }, [vram_ptr_b, :128], c_4096 | |
75e28f62 | 6091 | |
e1f6de8f | 6092 | vst1.u32 { texels_a, texels_b }, [texture_page_ptr, :256]! |
6093 | vst1.u32 { texels_c, texels_d }, [texture_page_ptr, :256]! | |
75e28f62 E |
6094 | |
6095 | subs sub_y, sub_y, #1 | |
6096 | bne 0b | |
6097 | ||
6098 | mov sub_y, #4 | |
6099 | ||
6100 | add vram_ptr_a, vram_ptr_a, #16 | |
6101 | add vram_ptr_b, vram_ptr_b, #16 | |
6102 | ||
6103 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6104 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6105 | ||
6106 | subs tile_x, tile_x, #1 | |
6107 | bne 0b | |
6108 | ||
6109 | mov tile_x, #8 | |
6110 | ||
6111 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6112 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6113 | ||
6114 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
6115 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
6116 | ||
6117 | subs tile_y, tile_y, #1 | |
6118 | add texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
6119 | ||
6120 | bne 0b | |
6121 | ||
6122 | vpop { q0 - q3 } | |
6123 | ldmia sp!, { r4 - r11, pc } | |
6124 | ||
50f9355a | 6125 | |
6126 | /* void scale2x_tiles8(void *dst, const void *src, int w8, int h) */ | |
6127 | function(scale2x_tiles8) | |
6128 | push { r4, r14 } | |
6129 | ||
6130 | mov r4, r1 | |
6131 | add r12, r0, #1024*2 | |
6132 | mov r14, r2 | |
6133 | ||
6134 | 0: | |
e1f6de8f | 6135 | vld1.u16 { q0 }, [r1, :128]! |
6136 | vld1.u16 { q2 }, [r1, :128]! | |
50f9355a | 6137 | vmov q1, q0 |
6138 | vmov q3, q2 | |
6139 | vzip.16 q0, q1 | |
6140 | vzip.16 q2, q3 | |
6141 | subs r14, #2 | |
e1f6de8f | 6142 | vst1.u16 { q0, q1 }, [r0, :128]! |
6143 | vst1.u16 { q0, q1 }, [r12, :128]! | |
50f9355a | 6144 | blt 1f |
e1f6de8f | 6145 | vst1.u16 { q2, q3 }, [r0, :128]! |
6146 | vst1.u16 { q2, q3 }, [r12, :128]! | |
50f9355a | 6147 | bgt 0b |
6148 | 1: | |
6149 | subs r3, #1 | |
6150 | mov r14, r2 | |
6151 | add r0, #1024*2*2 | |
6152 | add r4, #1024*2 | |
ed0fd81d | 6153 | sub r0, r0, r2, lsl #4+1 |
50f9355a | 6154 | mov r1, r4 |
6155 | add r12, r0, #1024*2 | |
6156 | bgt 0b | |
6157 | nop | |
6158 | ||
6159 | pop { r4, pc } | |
59d15d23 | 6160 | |
6161 | // vim:filetype=armasm |