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 | |
23 | ||
cb88320b | 24 | #include "psx_gpu_offsets.h" |
75e28f62 | 25 | |
cb88320b | 26 | #define psx_gpu_b_dx_offset (psx_gpu_b_block_span_offset + 4) |
75e28f62 | 27 | |
75e28f62 E |
28 | #define edge_data_left_x_offset 0 |
29 | #define edge_data_num_blocks_offset 2 | |
30 | #define edge_data_right_mask_offset 4 | |
31 | #define edge_data_y_offset 6 | |
32 | ||
33 | ||
34 | #define psx_gpu r0 | |
35 | #define v_a r1 | |
36 | #define v_b r2 | |
37 | #define v_c r3 | |
38 | ||
39 | #define x0 r4 | |
40 | #define x1 r5 | |
41 | #define x2 r6 | |
42 | #define x0_x1 r5 | |
43 | #define x1_x2 r6 | |
44 | #define y0 r7 | |
45 | #define y1 r8 | |
46 | #define y2 r9 | |
47 | #define y0_y1 r7 | |
48 | #define y1_y2 r8 | |
49 | #define b0 r9 | |
50 | #define b1 r10 | |
51 | #define b2 r11 | |
52 | #define b0_b1 r10 | |
53 | #define b1_b2 r11 | |
54 | ||
55 | ||
56 | #define area_r_s r5 | |
57 | ||
58 | #define g_bx0 r2 | |
59 | #define g_bx r3 | |
60 | #define g_bx2 r4 | |
61 | #define g_bx3 r5 | |
62 | #define b_base r6 | |
63 | #define g_by r8 | |
64 | ||
65 | #define gs_bx r7 | |
66 | #define gs_by r10 | |
67 | ||
68 | #define ga_bx g_bx | |
69 | #define ga_by g_by | |
70 | ||
71 | #define gw_bx_h g_bx | |
72 | #define gw_by_h g_by | |
73 | ||
74 | #define gw_bx_l r11 | |
75 | #define gw_by_l gw_bx_l | |
76 | ||
77 | #define store_a r0 | |
78 | #define store_b r1 | |
79 | #define store_inc r5 | |
80 | ||
81 | ||
82 | #define v0 q0 | |
83 | #define uvrgb0 d0 | |
84 | #define x0_y0 d1 | |
85 | ||
86 | #define v1 q1 | |
87 | #define uvrgb1 d2 | |
88 | #define x1_y1 d3 | |
89 | ||
90 | #define v2 q2 | |
91 | #define uvrgb2 d4 | |
92 | #define x2_y2 d5 | |
93 | ||
94 | #define x0_ab q3 | |
95 | #define uvrg_xxxx0 q3 | |
96 | #define uvrg0 d6 | |
97 | #define xxxx0 d7 | |
98 | ||
99 | #define x1_ab q4 | |
100 | #define uvrg_xxxx1 q4 | |
101 | #define uvrg1 d8 | |
102 | #define xxxx1 d9 | |
103 | ||
104 | #define x2_ab q5 | |
105 | #define uvrg_xxxx2 q5 | |
106 | #define uvrg2 d10 | |
107 | #define xxxx2 d11 | |
108 | ||
109 | #define y0_ab q6 | |
110 | #define yyyy_uvrg0 q6 | |
111 | #define yyyy0 d12 | |
112 | #define uvrg0b d13 | |
113 | ||
114 | #define y1_ab q7 | |
115 | #define yyyy_uvrg1 q7 | |
116 | #define yyyy1 d14 | |
117 | #define uvrg1b d15 | |
118 | ||
119 | #define y2_ab q8 | |
120 | #define yyyy_uvrg2 q8 | |
121 | #define yyyy2 d16 | |
122 | #define uvrg2b d17 | |
123 | ||
124 | #define d0_ab q9 | |
125 | #define d0_a d18 | |
126 | #define d0_b d19 | |
127 | ||
128 | #define d1_ab q10 | |
129 | #define d1_a d20 | |
130 | #define d1_b d21 | |
131 | ||
132 | #define d2_ab q11 | |
133 | #define d2_a d22 | |
134 | #define d2_b d23 | |
135 | ||
136 | #define d3_ab q12 | |
137 | #define d3_a d24 | |
138 | #define d3_b d25 | |
139 | ||
140 | #define ga_uvrg_x q1 | |
141 | #define ga_uvrg_y q4 | |
142 | ||
143 | #define dx x0_x1 | |
144 | #define dy y0_y1 | |
145 | #define db b0_b1 | |
146 | ||
147 | #define uvrg_base q11 | |
148 | ||
149 | #define gs_uvrg_x q5 | |
150 | #define gs_uvrg_y q6 | |
151 | ||
152 | #define g_uvrg_x q1 | |
153 | #define ga_uv_x d2 | |
154 | #define g_uv_x d2 | |
155 | #define ga_rg_x d3 | |
156 | #define g_rg_x d3 | |
157 | ||
158 | #define g_uvrg_y q4 | |
159 | #define ga_uv_y d8 | |
160 | #define g_uv_y d8 | |
161 | #define ga_rg_y d9 | |
162 | #define g_rg_y d9 | |
163 | ||
164 | #define gw_uv_x q1 | |
165 | #define gw_rg_x q2 | |
166 | #define gw_uv_y q4 | |
167 | #define gw_rg_y q3 | |
168 | ||
169 | #define w_mask q9 | |
170 | #define w_mask_l d18 | |
171 | ||
172 | #define r_shift q10 | |
173 | ||
174 | #define uvrg_dx0 q0 | |
175 | #define uvrg_dx0l d0 | |
176 | #define uvrg_dx0h d1 | |
177 | ||
178 | #define uvrg_dx1 q1 | |
179 | #define uvrg_dx1l d2 | |
180 | #define uvrg_dx1h d3 | |
181 | ||
182 | #define uvrg_dx2 q2 | |
183 | #define uvrg_dx2l d4 | |
184 | #define uvrg_dx2h d5 | |
185 | ||
186 | #define uvrg_dx3 q3 | |
187 | #define uvrg_dx3l d6 | |
188 | #define uvrg_dx3h d7 | |
189 | ||
c6063f89 | 190 | #define uvrgb_phase q13 |
75e28f62 E |
191 | |
192 | .align 4 | |
193 | ||
5d834c08 | 194 | /* FIXME: users of this should be in psx_gpu instead */ |
195 | #ifndef __PIC__ | |
196 | #define load_pointer(register, pointer) \ | |
197 | movw register, :lower16:pointer; \ | |
198 | movt register, :upper16:pointer; \ | |
199 | ||
200 | #else | |
201 | #define load_pointer(register, pointer) \ | |
202 | ldr register, =pointer \ | |
203 | ||
204 | #endif | |
205 | ||
75e28f62 E |
206 | #define function(name) \ |
207 | .global name; \ | |
208 | name: \ | |
209 | ||
210 | @ r0: psx_gpu | |
211 | @ r1: v_a | |
212 | @ r2: v_b | |
213 | @ r3: v_c | |
214 | ||
215 | function(compute_all_gradients) | |
216 | // First compute the triangle area reciprocal and shift. The division will | |
217 | // happen concurrently with much of the work which follows. | |
218 | @ r12 = psx_gpu->triangle_area | |
219 | ldr r12, [ psx_gpu, #psx_gpu_triangle_area_offset ] | |
220 | stmdb sp!, { r4 - r11, lr } | |
221 | ||
222 | @ load exponent of 62 into upper half of double | |
223 | movw r4, #0 | |
224 | clz r14, r12 @ r14 = shift | |
225 | ||
226 | movt r4, #((62 + 1023) << 4) | |
227 | mov r12, r12, lsl r14 @ r12 = triangle_area_normalized | |
228 | ||
229 | @ load area normalized into lower half of double | |
230 | mov r5, r12, lsr #10 | |
231 | vmov.f64 d30, r5, r4 @ d30 = (1 << 62) + ta_n | |
232 | ||
233 | movt r4, #((1022 + 31) << 4) | |
234 | mov r5, r12, lsl #20 | |
235 | ||
236 | add r4, r4, r12, lsr #11 | |
237 | vmov.f64 d31, r5, r4 | |
238 | ||
239 | vdiv.f64 d30, d30, d31 @ d30 = ((1 << 62) + ta_n) / ta_n | |
240 | ||
241 | // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) = | |
242 | // ( d0 * d1 ) - ( d2 * d3 ) = | |
243 | // ( m0 ) - ( m1 ) = gradient | |
244 | ||
245 | // This is split to do 12 elements at a time over three sets: a, b, and c. | |
246 | // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so | |
247 | // two of the slots are unused. | |
248 | ||
249 | // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as | |
250 | // is g. | |
251 | ||
252 | // First type is: uvrg bxxx xxxx | |
253 | // Second type is: yyyy ybyy uvrg | |
254 | // Since x_a and y_c are the same the same variable is used for both. | |
255 | ||
256 | vld1.u32 { v0 }, [ v_a, : 128 ] @ v0 = { uvrg0, b0, x0, y0 } | |
257 | ldrsh x0, [ v_a, #8 ] @ load x0 | |
258 | ||
259 | vld1.u32 { v1 }, [ v_b, : 128 ] @ v1 = { uvrg1, b1, x1, y1} | |
260 | ldrh x1, [ v_b, #8 ] @ load x1 | |
261 | ||
262 | vld1.u32 { v2 }, [ v_c, : 128 ] @ v2 = { uvrg2, b2, x2, y2 } | |
263 | ldrh x2, [ v_c, #8 ] @ load x2 | |
264 | ||
265 | vmovl.u8 uvrg_xxxx0, uvrgb0 @ uvrg_xxxx0 = { uv0, rg0, b0-, -- } | |
266 | ldrh y0, [ v_a, #10 ] @ load y0 | |
267 | ||
268 | vmovl.u8 uvrg_xxxx1, uvrgb1 @ uvrg_xxxx1 = { uv1, rg1, b1-, -- } | |
269 | ldrh y1, [ v_b, #10 ] @ load y1 | |
270 | ||
271 | vmovl.u8 uvrg_xxxx2, uvrgb2 @ uvrg_xxxx2 = { uv2, rg2, b2-, -- } | |
272 | ldrh y2, [ v_c, #10 ] @ load y2 | |
273 | ||
274 | vmov.u8 uvrg0b, uvrg0 @ uvrg0b = { uv0, rg0 } | |
275 | vdup.u16 xxxx0, x0_y0[0] @ xxxx0 = { xx0, xx0 } | |
276 | ||
277 | orr x1_x2, x1, x2, lsl #16 @ x1_x2 = { x1, x2 } | |
278 | pkhbt x0_x1, x0, x1, lsl #16 @ x0_x1 = { x0, x1 } | |
279 | ||
280 | vmov.u8 uvrg1b, uvrg1 @ uvrg1b = { uv1, rg1 } | |
281 | vdup.u16 xxxx1, x1_y1[0] @ xxxx1 = { xx1, xx1 } | |
282 | ||
283 | vmov.u8 uvrg2b, uvrg2 @ uvrg2b = { uv2, rg2 } | |
284 | vdup.u16 xxxx2, x2_y2[0] @ xxxx2 = { xx2, xx2 } | |
285 | ||
286 | ldrb b2, [ v_c, #4 ] @ load b2 | |
287 | orr y0_y1, y0, y1, lsl #16 @ y0_y1 = { y0, y1 } | |
288 | ||
289 | ldrb b1, [ v_b, #4 ] @ load b1 | |
290 | orr y1_y2, y1, y2, lsl #16 @ y1_y2 = { y1, y2 } | |
291 | ||
292 | vdup.u16 yyyy0, x0_y0[1] @ yyyy0 = { yy0, yy0 } | |
293 | vsub.s16 d0_ab, x1_ab, x0_ab | |
294 | ||
295 | ldrb b0, [ v_a, #4 ] @ load b0 | |
296 | orr b1_b2, b1, b2, lsl #16 @ b1_b2 = { b1, b2 } | |
297 | ||
298 | vdup.u16 yyyy1, x1_y1[1] @ yyyy1 = { yy1, yy1 } | |
299 | vsub.s16 d2_ab, x2_ab, x1_ab | |
300 | ||
301 | vdup.u16 yyyy2, x2_y2[1] @ yyyy2 = { yy2, yy2 } | |
302 | vsub.s16 d1_ab, y2_ab, y1_ab | |
303 | ||
304 | orr b0_b1, b0, b1, lsl #16 @ b1_b2 = { b1, b2 } | |
305 | ssub16 dx, x1_x2, x0_x1 @ dx = { x1 - x0, x2 - x1 } | |
306 | ||
307 | ssub16 dy, y1_y2, y0_y1 @ dy = { y1 - y0, y2 - y1 } | |
308 | ssub16 db, b1_b2, b0_b1 @ db = { b1 - b0, b2 - b1 } | |
309 | ||
310 | vsub.s16 d3_ab, y1_ab, y0_ab | |
311 | smusdx ga_by, dx, db @ ga_by = ((x1 - x0) * (b2 - b1)) - | |
312 | @ ((x2 - X1) * (b1 - b0)) | |
313 | vmull.s16 ga_uvrg_x, d0_a, d1_a | |
314 | smusdx ga_bx, db, dy @ ga_bx = ((b1 - b0) * (y2 - y1)) - | |
315 | @ ((b2 - b1) * (y1 - y0)) | |
316 | vmlsl.s16 ga_uvrg_x, d2_a, d3_a | |
317 | movs gs_bx, ga_bx, asr #31 | |
318 | ||
319 | vmull.s16 ga_uvrg_y, d0_b, d1_b | |
320 | rsbmi ga_bx, ga_bx, #0 | |
321 | ||
c6063f89 | 322 | @ r12 = psx_gpu->uvrgb_phase |
323 | ldr r12, [ psx_gpu, #psx_gpu_uvrgb_phase_offset ] | |
324 | ||
75e28f62 E |
325 | vmlsl.s16 ga_uvrg_y, d2_b, d3_b |
326 | movs gs_by, ga_by, asr #31 | |
327 | ||
328 | vshr.u64 d0, d30, #22 | |
c6063f89 | 329 | add b_base, r12, b0, lsl #16 |
330 | ||
331 | vdup.u32 uvrgb_phase, r12 | |
75e28f62 E |
332 | |
333 | rsbmi ga_by, ga_by, #0 | |
334 | vclt.s32 gs_uvrg_x, ga_uvrg_x, #0 @ gs_uvrg_x = ga_uvrg_x < 0 | |
335 | ||
336 | @ r12 = psx_gpu->triangle_winding_offset | |
337 | ldrb r12, [ psx_gpu, #psx_gpu_triangle_winding_offset ] | |
338 | vclt.s32 gs_uvrg_y, ga_uvrg_y, #0 @ gs_uvrg_y = ga_uvrg_y < 0 | |
339 | ||
75e28f62 E |
340 | rsb r12, r12, #0 @ r12 = -(triangle->winding) |
341 | ||
342 | vdup.u32 w_mask, r12 @ w_mask = { -w, -w, -w, -w } | |
343 | sub r14, r14, #(62 - 12) @ r14 = shift - (62 - FIXED_BITS) | |
344 | ||
345 | vshll.u16 uvrg_base, uvrg0, #16 @ uvrg_base = uvrg0 << 16 | |
346 | vdup.u32 r_shift, r14 @ r_shift = { shift, shift, shift, shift } | |
347 | ||
c6063f89 | 348 | vadd.u32 uvrg_base, uvrgb_phase |
75e28f62 E |
349 | vabs.s32 ga_uvrg_x, ga_uvrg_x @ ga_uvrg_x = abs(ga_uvrg_x) |
350 | ||
351 | vmov area_r_s, s0 @ area_r_s = triangle_reciprocal | |
352 | vabs.s32 ga_uvrg_y, ga_uvrg_y @ ga_uvrg_y = abs(ga_uvrg_y) | |
353 | ||
354 | vmull.u32 gw_rg_x, ga_rg_x, d0[0] | |
355 | vmull.u32 gw_uv_x, ga_uv_x, d0[0] | |
356 | vmull.u32 gw_rg_y, ga_rg_y, d0[0] | |
357 | vmull.u32 gw_uv_y, ga_uv_y, d0[0] | |
358 | ||
359 | vshl.u64 gw_rg_x, gw_rg_x, r_shift | |
360 | vshl.u64 gw_uv_x, gw_uv_x, r_shift | |
361 | vshl.u64 gw_rg_y, gw_rg_y, r_shift | |
362 | vshl.u64 gw_uv_y, gw_uv_y, r_shift | |
363 | ||
364 | veor.u32 gs_uvrg_x, gs_uvrg_x, w_mask | |
365 | vmovn.u64 g_uv_x, gw_uv_x | |
366 | ||
367 | veor.u32 gs_uvrg_y, gs_uvrg_y, w_mask | |
368 | vmovn.u64 g_rg_x, gw_rg_x | |
369 | ||
370 | veor.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
371 | vmovn.u64 g_uv_y, gw_uv_y | |
372 | ||
373 | vsub.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
374 | vmovn.u64 g_rg_y, gw_rg_y | |
375 | ||
376 | veor.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
377 | mov ga_bx, ga_bx, lsl #13 | |
378 | ||
379 | vsub.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
380 | mov ga_by, ga_by, lsl #13 | |
381 | ||
382 | vdup.u32 x0_y0, x0 | |
383 | umull gw_bx_l, gw_bx_h, ga_bx, area_r_s | |
384 | ||
385 | vshl.u32 g_uvrg_x, g_uvrg_x, #4 | |
386 | vshl.u32 g_uvrg_y, g_uvrg_y, #4 | |
387 | ||
388 | umull gw_by_l, gw_by_h, ga_by, area_r_s | |
389 | vmls.s32 uvrg_base, ga_uvrg_x, x0_y0[0] | |
390 | ||
391 | eor gs_bx, gs_bx, r12 | |
392 | vadd.u32 uvrg_dx2, uvrg_dx1, uvrg_dx1 | |
393 | ||
394 | veor.u32 uvrg_dx0, uvrg_dx0, uvrg_dx0 | |
395 | eor gs_by, gs_by, r12 | |
396 | ||
397 | rsb r11, r14, #0 @ r11 = negative shift for scalar lsr | |
398 | add store_a, psx_gpu, #psx_gpu_uvrg_offset | |
399 | ||
400 | sub r11, r11, #(32 - 13) | |
401 | ||
402 | add store_b, store_a, #16 | |
403 | mov store_inc, #32 | |
404 | ||
405 | vadd.u32 uvrg_dx3, uvrg_dx2, uvrg_dx1 | |
406 | vst1.u32 { uvrg_base }, [ store_a, : 128 ], store_inc | |
407 | ||
408 | vst1.u32 { uvrg_dx1 }, [ store_b, : 128 ], store_inc | |
409 | mov g_bx, gw_bx_h, lsr r11 | |
410 | ||
411 | vst1.u32 { g_uvrg_y }, [ store_a, : 128 ], store_inc | |
412 | mov g_by, gw_by_h, lsr r11 | |
413 | ||
414 | vst4.u32 { uvrg_dx0l, uvrg_dx1l, uvrg_dx2l, uvrg_dx3l }, \ | |
415 | [ store_b, : 128 ], store_inc | |
416 | eor g_bx, g_bx, gs_bx | |
417 | ||
418 | vst4.u32 { uvrg_dx0h, uvrg_dx1h, uvrg_dx2h, uvrg_dx3h }, \ | |
419 | [ store_b, : 128 ], store_inc | |
420 | sub g_bx, g_bx, gs_bx | |
421 | ||
422 | lsl g_bx, g_bx, #4 | |
423 | eor g_by, g_by, gs_by | |
424 | ||
425 | mls b_base, g_bx, x0, b_base | |
426 | sub g_by, g_by, gs_by | |
427 | ||
428 | lsl g_by, g_by, #4 | |
429 | mov g_bx0, #0 | |
430 | ||
431 | add g_bx2, g_bx, g_bx | |
432 | add g_bx3, g_bx, g_bx2 | |
433 | ||
434 | stmia store_b, { g_bx0, g_bx, g_bx2, g_bx3, b_base, g_by } | |
435 | ||
436 | ldmia sp!, { r4 - r11, pc } | |
437 | ||
438 | ||
439 | #define psx_gpu r0 | |
440 | #define v_a r1 | |
441 | #define v_b r2 | |
442 | #define v_c r3 | |
443 | ||
444 | #define temp r14 | |
445 | ||
446 | #define x_a r4 | |
447 | #define x_b r5 | |
448 | #define x_c r6 | |
449 | #define y_a r1 | |
450 | #define y_b r2 | |
451 | #define y_c r3 | |
452 | ||
453 | #define height_minor_a r7 | |
454 | #define height_minor_b r8 | |
455 | #define height_major r9 | |
456 | #define height r9 | |
457 | ||
458 | #define reciprocal_table_ptr r10 | |
459 | ||
460 | #define edge_alt_low r4 | |
461 | #define edge_alt_high r5 | |
462 | #define edge_dx_dy_alt r6 | |
463 | #define edge_shift_alt r10 | |
464 | ||
465 | #define edge_dx_dy_alt_low r4 | |
466 | #define edge_dx_dy_alt_high r5 | |
467 | ||
468 | #define span_edge_data r4 | |
469 | #define span_uvrg_offset r5 | |
470 | #define span_b_offset r6 | |
471 | ||
472 | #define clip r14 | |
473 | ||
474 | #define b r11 | |
475 | #define b_dy r12 | |
476 | ||
477 | ||
478 | #define alternate_x q0 | |
479 | #define alternate_dx_dy q1 | |
480 | #define alternate_x_32 q2 | |
481 | ||
482 | #define alternate_x_low d0 | |
483 | #define alternate_x_high d1 | |
484 | #define alternate_dx_dy_low d2 | |
485 | #define alternate_dx_dy_high d3 | |
486 | #define alternate_x_32_low d4 | |
487 | #define alternate_x_32_high d5 | |
488 | ||
489 | #define left_x q3 | |
490 | #define right_x q4 | |
491 | #define left_dx_dy q5 | |
492 | #define right_dx_dy q6 | |
493 | #define left_edge q7 | |
494 | #define right_edge q8 | |
495 | ||
496 | #define left_x_low d6 | |
497 | #define left_x_high d7 | |
498 | #define right_x_low d8 | |
499 | #define right_x_high d9 | |
500 | #define left_dx_dy_low d10 | |
501 | #define left_dx_dy_high d11 | |
502 | #define right_dx_dy_low d12 | |
503 | #define right_dx_dy_high d13 | |
504 | #define left_edge_low d14 | |
505 | #define left_edge_high d15 | |
506 | #define right_edge_low d16 | |
507 | #define right_edge_high d17 | |
508 | ||
509 | #define y_mid_point d18 | |
510 | #define c_0x0004 d19 | |
511 | ||
512 | #define left_right_x_16 q11 | |
513 | #define span_shifts_y q12 | |
514 | #define c_0x0001 q13 | |
515 | ||
516 | #define span_shifts d24 | |
517 | #define y_x4 d25 | |
518 | #define c_0xFFFE d26 | |
519 | #define c_0x0007 d27 | |
520 | ||
521 | #define left_right_x_16_low d22 | |
522 | #define left_right_x_16_high d23 | |
523 | ||
524 | #define uvrg q14 | |
525 | #define uvrg_dy q15 | |
526 | ||
527 | #define alternate_x_16 d4 | |
528 | ||
529 | #define v_clip q3 | |
530 | #define v_clip_low d6 | |
531 | ||
532 | #define right_x_32 q10 | |
533 | #define left_x_32 q11 | |
534 | #define alternate_select d24 | |
535 | ||
536 | #define right_x_32_low d20 | |
537 | #define right_x_32_high d21 | |
538 | #define left_x_32_low d22 | |
539 | #define left_x_32_high d23 | |
540 | ||
541 | #define edges_xy q0 | |
542 | #define edges_dx_dy d2 | |
543 | #define edge_shifts d3 | |
544 | #define edge_shifts_64 q2 | |
545 | ||
546 | #define edges_xy_left d0 | |
547 | #define edges_xy_right d1 | |
548 | ||
549 | #define height_reciprocals d6 | |
550 | #define heights d7 | |
551 | ||
552 | #define widths d8 | |
553 | #define c_0x01 d9 | |
554 | #define x_starts d10 | |
555 | #define x_ends d11 | |
556 | ||
557 | #define heights_b d12 | |
558 | #define edges_dx_dy_64 q10 | |
559 | ||
560 | #define edges_dx_dy_64_left d20 | |
561 | #define edges_dx_dy_64_right d21 | |
562 | ||
563 | ||
564 | #define setup_spans_prologue() \ | |
565 | stmdb sp!, { r4 - r11, lr }; \ | |
566 | \ | |
567 | ldrsh x_a, [ v_a, #8 ]; \ | |
568 | ldrsh x_b, [ v_b, #8 ]; \ | |
569 | ldrsh x_c, [ v_c, #8 ]; \ | |
570 | ldrsh y_a, [ v_a, #10 ]; \ | |
571 | ldrsh y_b, [ v_b, #10 ]; \ | |
572 | ldrsh y_c, [ v_c, #10 ]; \ | |
573 | \ | |
574 | add temp, psx_gpu, #psx_gpu_uvrg_offset; \ | |
575 | vld1.32 { uvrg }, [ temp ]; \ | |
576 | add temp, psx_gpu, #psx_gpu_uvrg_dy_offset; \ | |
577 | vld1.32 { uvrg_dy }, [ temp ]; \ | |
5d834c08 | 578 | load_pointer(reciprocal_table_ptr, reciprocal_table); \ |
75e28f62 E |
579 | \ |
580 | vmov.u32 c_0x01, #0x01 \ | |
581 | ||
582 | #define setup_spans_load_b() \ | |
583 | ldr b, [ psx_gpu, #psx_gpu_b_offset ]; \ | |
584 | ldr b_dy, [ psx_gpu, #psx_gpu_b_dy_offset ] \ | |
585 | ||
586 | #define setup_spans_prologue_b() \ | |
587 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
588 | add temp, psx_gpu, #psx_gpu_viewport_start_x_offset; \ | |
589 | \ | |
590 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
591 | vmov.u16 c_0x0004, #0x0004; \ | |
592 | \ | |
593 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
594 | vmov.u16 c_0x0001, #0x0001; \ | |
595 | \ | |
596 | vld1.u16 { left_edge_low[], left_edge_high[] }, [ temp ]; \ | |
597 | add temp, psx_gpu, #psx_gpu_viewport_end_x_offset; \ | |
598 | \ | |
599 | vld1.u16 { right_edge_low[], right_edge_high[] }, [ temp ]; \ | |
600 | vadd.u16 right_edge, right_edge, c_0x0001; \ | |
601 | \ | |
602 | vmov.u16 c_0x0007, #0x0007; \ | |
603 | vmvn.u16 c_0xFFFE, #0x0001 \ | |
604 | ||
605 | ||
606 | #define compute_edge_delta_x2() \ | |
607 | ldr temp, [ reciprocal_table_ptr, height, lsl #2 ]; \ | |
608 | \ | |
609 | vdup.u32 heights, height; \ | |
610 | vsub.u32 widths, x_ends, x_starts; \ | |
611 | \ | |
612 | vdup.u32 edge_shifts, temp; \ | |
613 | vsub.u32 heights_b, heights, c_0x01; \ | |
7d5140f5 | 614 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
75e28f62 E |
615 | \ |
616 | vmla.s32 heights_b, x_starts, heights; \ | |
617 | vbic.u16 edge_shifts, #0xE0; \ | |
618 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
619 | vmull.s32 edges_xy, heights_b, height_reciprocals \ | |
620 | ||
621 | #define width_alt r6 | |
622 | #define height_reciprocal_alt r11 | |
623 | #define height_b_alt r12 | |
624 | ||
625 | #define compute_edge_delta_x3(start_c, height_a, height_b) \ | |
626 | vmov.u32 heights, height_a, height_b; \ | |
627 | ldr temp, [ reciprocal_table_ptr, height_a, lsl #2 ]; \ | |
628 | vmov.u32 edge_shifts[0], temp; \ | |
629 | ldr temp, [ reciprocal_table_ptr, height_b, lsl #2 ]; \ | |
630 | vmov.u32 edge_shifts[1], temp; \ | |
631 | ldr edge_shift_alt, [ reciprocal_table_ptr, height_minor_b, lsl #2 ]; \ | |
632 | \ | |
633 | vsub.u32 widths, x_ends, x_starts; \ | |
634 | sub width_alt, x_c, start_c; \ | |
635 | \ | |
636 | vsub.u32 heights_b, heights, c_0x01; \ | |
637 | sub height_b_alt, height_minor_b, #1; \ | |
638 | \ | |
7d5140f5 E |
639 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
640 | lsr height_reciprocal_alt, edge_shift_alt, #10; \ | |
75e28f62 E |
641 | \ |
642 | vmla.s32 heights_b, x_starts, heights; \ | |
643 | mla height_b_alt, height_minor_b, start_c, height_b_alt; \ | |
644 | \ | |
645 | vbic.u16 edge_shifts, #0xE0; \ | |
646 | and edge_shift_alt, edge_shift_alt, #0x1F; \ | |
647 | \ | |
648 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
649 | mul edge_dx_dy_alt, width_alt, height_reciprocal_alt; \ | |
650 | \ | |
651 | vmull.s32 edges_xy, heights_b, height_reciprocals; \ | |
652 | smull edge_alt_low, edge_alt_high, height_b_alt, height_reciprocal_alt \ | |
653 | ||
654 | ||
655 | #define setup_spans_adjust_y_up() \ | |
656 | vsub.u32 y_x4, y_x4, c_0x0004 \ | |
657 | ||
658 | #define setup_spans_adjust_y_down() \ | |
659 | vadd.u32 y_x4, y_x4, c_0x0004 \ | |
660 | ||
661 | #define setup_spans_adjust_interpolants_up() \ | |
662 | vsub.u32 uvrg, uvrg, uvrg_dy; \ | |
663 | sub b, b, b_dy \ | |
664 | ||
665 | #define setup_spans_adjust_interpolants_down() \ | |
666 | vadd.u32 uvrg, uvrg, uvrg_dy; \ | |
667 | add b, b, b_dy \ | |
668 | ||
669 | ||
670 | #define setup_spans_clip_interpolants_increment() \ | |
671 | mla b, b_dy, clip, b; \ | |
672 | vmla.s32 uvrg, uvrg_dy, v_clip \ | |
673 | ||
674 | #define setup_spans_clip_interpolants_decrement() \ | |
675 | mls b, b_dy, clip, b; \ | |
676 | vmls.s32 uvrg, uvrg_dy, v_clip \ | |
677 | ||
678 | #define setup_spans_clip_alternate_yes() \ | |
679 | smlal edge_alt_low, edge_alt_high, edge_dx_dy_alt, clip \ | |
680 | ||
681 | #define setup_spans_clip_alternate_no() \ | |
682 | ||
683 | #define setup_spans_clip(direction, alternate_active) \ | |
684 | vdup.u32 v_clip, clip; \ | |
685 | setup_spans_clip_alternate_##alternate_active(); \ | |
686 | setup_spans_clip_interpolants_##direction(); \ | |
687 | vmlal.s32 edges_xy, edges_dx_dy, v_clip_low \ | |
688 | ||
689 | ||
690 | #define setup_spans_adjust_edges_alternate_no(left_index, right_index) \ | |
691 | vmovl.s32 edge_shifts_64, edge_shifts; \ | |
692 | vmovl.s32 edges_dx_dy_64, edges_dx_dy; \ | |
693 | \ | |
694 | vshl.s64 edges_xy, edges_xy, edge_shifts_64; \ | |
695 | vshl.s64 edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64; \ | |
696 | \ | |
697 | vmov left_x_low, edges_xy_##left_index; \ | |
698 | vmov right_x_low, edges_xy_##right_index; \ | |
699 | \ | |
700 | vmov left_dx_dy_low, edges_dx_dy_64_##left_index; \ | |
701 | vmov left_dx_dy_high, edges_dx_dy_64_##left_index; \ | |
702 | vmov right_dx_dy_low, edges_dx_dy_64_##right_index; \ | |
703 | vmov right_dx_dy_high, edges_dx_dy_64_##right_index; \ | |
704 | \ | |
705 | vadd.u64 left_x_high, left_x_low, left_dx_dy_low; \ | |
706 | vadd.u64 right_x_high, right_x_low, right_dx_dy_low; \ | |
707 | \ | |
708 | vadd.u64 left_dx_dy, left_dx_dy, left_dx_dy; \ | |
709 | vadd.u64 right_dx_dy, right_dx_dy, right_dx_dy \ | |
710 | ||
711 | ||
712 | #define setup_spans_adjust_edges_alternate_yes(left_index, right_index) \ | |
713 | setup_spans_adjust_edges_alternate_no(left_index, right_index); \ | |
714 | \ | |
715 | vdup.u16 y_mid_point, y_b; \ | |
716 | rsb temp, edge_shift_alt, #32; \ | |
717 | \ | |
718 | lsl edge_alt_high, edge_alt_high, edge_shift_alt; \ | |
719 | orr edge_alt_high, edge_alt_high, edge_alt_low, lsr temp; \ | |
720 | lsl edge_alt_low, edge_alt_low, edge_shift_alt; \ | |
721 | vmov alternate_x_low, edge_alt_low, edge_alt_high; \ | |
722 | \ | |
723 | asr edge_dx_dy_alt_high, edge_dx_dy_alt, temp; \ | |
724 | lsl edge_dx_dy_alt_low, edge_dx_dy_alt, edge_shift_alt; \ | |
725 | vmov alternate_dx_dy_low, edge_dx_dy_alt_low, edge_dx_dy_alt_high; \ | |
726 | vmov alternate_dx_dy_high, alternate_dx_dy_low; \ | |
727 | \ | |
728 | vadd.u64 alternate_x_high, alternate_x_low, alternate_dx_dy_low; \ | |
729 | vadd.u64 alternate_dx_dy, alternate_dx_dy, alternate_dx_dy \ | |
730 | ||
731 | ||
732 | #define setup_spans_y_select_up() \ | |
733 | vclt.s16 alternate_select, y_x4, y_mid_point \ | |
734 | ||
735 | #define setup_spans_y_select_down() \ | |
736 | vcgt.s16 alternate_select, y_x4, y_mid_point \ | |
737 | ||
738 | ||
739 | #define setup_spans_alternate_select_left() \ | |
740 | vbit.u16 left_right_x_16_low, alternate_x_16, alternate_select \ | |
741 | ||
742 | #define setup_spans_alternate_select_right() \ | |
743 | vbit.u16 left_right_x_16_high, alternate_x_16, alternate_select \ | |
744 | ||
745 | ||
746 | #define setup_spans_set_x4_alternate_yes(alternate, direction) \ | |
747 | vshrn.s64 alternate_x_32_low, alternate_x, #32; \ | |
748 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
749 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
750 | \ | |
751 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
752 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
753 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
754 | \ | |
755 | vshrn.s64 alternate_x_32_high, alternate_x, #32; \ | |
756 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
757 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
758 | \ | |
759 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
760 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
761 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
762 | \ | |
763 | vmovn.u32 alternate_x_16, alternate_x_32; \ | |
764 | setup_spans_y_select_##direction(); \ | |
765 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
766 | \ | |
767 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
768 | setup_spans_alternate_select_##alternate(); \ | |
769 | \ | |
770 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
771 | str b, [ span_b_offset ], #4; \ | |
772 | setup_spans_adjust_interpolants_##direction(); \ | |
773 | \ | |
774 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
775 | \ | |
776 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
777 | str b, [ span_b_offset ], #4; \ | |
778 | setup_spans_adjust_interpolants_##direction(); \ | |
779 | \ | |
780 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
781 | \ | |
782 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
783 | str b, [ span_b_offset ], #4; \ | |
784 | setup_spans_adjust_interpolants_##direction(); \ | |
785 | \ | |
786 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
787 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
788 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
789 | \ | |
790 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
791 | str b, [ span_b_offset ], #4; \ | |
792 | setup_spans_adjust_interpolants_##direction(); \ | |
793 | \ | |
794 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
795 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
796 | \ | |
797 | vst4.u16 { left_right_x_16, span_shifts_y }, [ span_edge_data ]!; \ | |
798 | \ | |
799 | setup_spans_adjust_y_##direction() \ | |
800 | ||
801 | ||
802 | #define setup_spans_set_x4_alternate_no(alternate, direction) \ | |
803 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
804 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
805 | \ | |
806 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
807 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
808 | \ | |
809 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
810 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
811 | \ | |
812 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
813 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
814 | \ | |
815 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
816 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
817 | \ | |
818 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
819 | str b, [ span_b_offset ], #4; \ | |
820 | setup_spans_adjust_interpolants_##direction(); \ | |
821 | \ | |
822 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
823 | \ | |
824 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
825 | str b, [ span_b_offset ], #4; \ | |
826 | setup_spans_adjust_interpolants_##direction(); \ | |
827 | \ | |
828 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
829 | \ | |
830 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
831 | str b, [ span_b_offset ], #4; \ | |
832 | setup_spans_adjust_interpolants_##direction(); \ | |
833 | \ | |
834 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
835 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
836 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
837 | \ | |
838 | vst1.u32 { uvrg }, [ span_uvrg_offset, :128 ]!; \ | |
839 | str b, [ span_b_offset ], #4; \ | |
840 | setup_spans_adjust_interpolants_##direction(); \ | |
841 | \ | |
842 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
843 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
844 | \ | |
845 | vst4.u16 { left_right_x_16, span_shifts_y }, [ span_edge_data ]!; \ | |
846 | \ | |
847 | setup_spans_adjust_y_##direction() \ | |
848 | ||
849 | ||
850 | #define edge_adjust_low r11 | |
851 | #define edge_adjust_high r12 | |
852 | ||
853 | #define setup_spans_alternate_adjust_yes() \ | |
854 | smull edge_adjust_low, edge_adjust_high, edge_dx_dy_alt, height_minor_a; \ | |
855 | subs edge_alt_low, edge_alt_low, edge_adjust_low; \ | |
856 | sbc edge_alt_high, edge_alt_high, edge_adjust_high \ | |
857 | ||
858 | #define setup_spans_alternate_adjust_no() \ | |
859 | ||
860 | ||
861 | #define setup_spans_down(left_index, right_index, alternate, alternate_active) \ | |
862 | setup_spans_alternate_adjust_##alternate_active(); \ | |
863 | setup_spans_load_b(); \ | |
864 | \ | |
865 | ldrsh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ]; \ | |
866 | subs y_c, y_c, temp; \ | |
867 | subgt height, height, y_c; \ | |
868 | addgt height, height, #1; \ | |
869 | \ | |
870 | ldrsh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ]; \ | |
871 | subs clip, temp, y_a; \ | |
872 | ble 0f; \ | |
873 | \ | |
874 | sub height, height, clip; \ | |
875 | add y_a, y_a, clip; \ | |
876 | setup_spans_clip(increment, alternate_active); \ | |
877 | \ | |
878 | 0: \ | |
879 | cmp height, #0; \ | |
880 | ble 1f; \ | |
881 | \ | |
882 | orr temp, y_a, y_a, lsl #16; \ | |
883 | add temp, temp, #(1 << 16); \ | |
884 | add y_a, temp, #2; \ | |
885 | add y_a, y_a, #(2 << 16); \ | |
886 | vmov.u32 y_x4, temp, y_a; \ | |
887 | \ | |
888 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
889 | right_index); \ | |
890 | setup_spans_prologue_b(); \ | |
891 | \ | |
892 | strh height, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
893 | \ | |
894 | 2: \ | |
895 | setup_spans_set_x4_alternate_##alternate_active(alternate, down); \ | |
896 | subs height, height, #4; \ | |
897 | bhi 2b; \ | |
898 | \ | |
899 | 1: \ | |
900 | ||
901 | ||
902 | #define setup_spans_alternate_pre_increment_yes() \ | |
903 | adds edge_alt_low, edge_alt_low, edge_dx_dy_alt; \ | |
904 | adc edge_alt_high, edge_alt_high, edge_dx_dy_alt, asr #31 \ | |
905 | ||
906 | #define setup_spans_alternate_pre_increment_no() \ | |
907 | ||
908 | ||
909 | #define setup_spans_up_decrement_yes() \ | |
910 | suble height, height, #1 \ | |
911 | ||
912 | #define setup_spans_up_decrement_no() \ | |
913 | ||
914 | ||
915 | #define setup_spans_up(left_index, right_index, alternate, alternate_active) \ | |
916 | setup_spans_alternate_adjust_##alternate_active(); \ | |
917 | setup_spans_load_b(); \ | |
918 | sub y_a, y_a, #1; \ | |
919 | \ | |
920 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ]; \ | |
921 | subs temp, temp, y_c; \ | |
922 | subgt height, height, temp; \ | |
923 | setup_spans_up_decrement_##alternate_active(); \ | |
924 | \ | |
925 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ]; \ | |
926 | subs clip, y_a, temp; \ | |
927 | ble 0f; \ | |
928 | \ | |
929 | sub height, height, clip; \ | |
930 | sub y_a, y_a, clip; \ | |
931 | setup_spans_clip(decrement, alternate_active); \ | |
932 | \ | |
933 | 0: \ | |
934 | cmp height, #0; \ | |
935 | ble 1f; \ | |
936 | \ | |
937 | orr temp, y_a, y_a, lsl #16; \ | |
938 | sub temp, temp, #(1 << 16); \ | |
939 | sub y_a, temp, #2; \ | |
940 | sub y_a, y_a, #(2 << 16); \ | |
941 | vmov.u32 y_x4, temp, y_a; \ | |
942 | \ | |
943 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy; \ | |
944 | \ | |
945 | setup_spans_alternate_pre_increment_##alternate_active(); \ | |
946 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
947 | right_index); \ | |
948 | setup_spans_adjust_interpolants_up(); \ | |
949 | setup_spans_prologue_b(); \ | |
950 | \ | |
951 | strh height, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
952 | \ | |
953 | 2: \ | |
954 | setup_spans_set_x4_alternate_##alternate_active(alternate, up); \ | |
955 | subs height, height, #4; \ | |
956 | bhi 2b; \ | |
957 | \ | |
958 | 1: \ | |
959 | ||
960 | ||
961 | #define setup_spans_epilogue() \ | |
962 | ldmia sp!, { r4 - r11, pc } \ | |
963 | ||
964 | ||
965 | #define setup_spans_up_up(minor, major) \ | |
966 | setup_spans_prologue(); \ | |
967 | sub height_minor_a, y_a, y_b; \ | |
968 | sub height_minor_b, y_b, y_c; \ | |
969 | sub height, y_a, y_c; \ | |
970 | \ | |
971 | vdup.u32 x_starts, x_a; \ | |
972 | vmov.u32 x_ends, x_c, x_b; \ | |
973 | \ | |
974 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
975 | setup_spans_up(major, minor, minor, yes); \ | |
976 | setup_spans_epilogue() \ | |
977 | ||
978 | function(setup_spans_up_left) | |
979 | setup_spans_up_up(left, right) | |
980 | ||
981 | function(setup_spans_up_right) | |
982 | setup_spans_up_up(right, left) | |
983 | ||
5d834c08 | 984 | .pool |
75e28f62 E |
985 | |
986 | #define setup_spans_down_down(minor, major) \ | |
987 | setup_spans_prologue(); \ | |
988 | sub height_minor_a, y_b, y_a; \ | |
989 | sub height_minor_b, y_c, y_b; \ | |
990 | sub height, y_c, y_a; \ | |
991 | \ | |
992 | vdup.u32 x_starts, x_a; \ | |
993 | vmov.u32 x_ends, x_c, x_b; \ | |
994 | \ | |
995 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
996 | setup_spans_down(major, minor, minor, yes); \ | |
997 | setup_spans_epilogue() \ | |
998 | ||
999 | function(setup_spans_down_left) | |
1000 | setup_spans_down_down(left, right) | |
1001 | ||
1002 | function(setup_spans_down_right) | |
1003 | setup_spans_down_down(right, left) | |
1004 | ||
1005 | ||
1006 | #define setup_spans_up_flat() \ | |
1007 | sub height, y_a, y_c; \ | |
1008 | \ | |
1009 | compute_edge_delta_x2(); \ | |
1010 | setup_spans_up(left, right, none, no); \ | |
1011 | setup_spans_epilogue() \ | |
1012 | ||
1013 | function(setup_spans_up_a) | |
1014 | setup_spans_prologue() | |
1015 | ||
1016 | vmov.u32 x_starts, x_a, x_b | |
1017 | vdup.u32 x_ends, x_c | |
1018 | ||
1019 | setup_spans_up_flat() | |
1020 | ||
1021 | function(setup_spans_up_b) | |
1022 | setup_spans_prologue() | |
1023 | ||
1024 | vdup.u32 x_starts, x_a | |
1025 | vmov.u32 x_ends, x_b, x_c | |
1026 | ||
1027 | setup_spans_up_flat() | |
1028 | ||
1029 | #define setup_spans_down_flat() \ | |
1030 | sub height, y_c, y_a; \ | |
1031 | \ | |
1032 | compute_edge_delta_x2(); \ | |
1033 | setup_spans_down(left, right, none, no); \ | |
1034 | setup_spans_epilogue() \ | |
1035 | ||
1036 | function(setup_spans_down_a) | |
1037 | setup_spans_prologue() | |
1038 | ||
1039 | vmov.u32 x_starts, x_a, x_b | |
1040 | vdup.u32 x_ends, x_c | |
1041 | ||
1042 | setup_spans_down_flat() | |
1043 | ||
1044 | function(setup_spans_down_b) | |
1045 | setup_spans_prologue() | |
1046 | ||
1047 | vdup.u32 x_starts, x_a | |
1048 | vmov.u32 x_ends, x_b, x_c | |
1049 | ||
1050 | setup_spans_down_flat() | |
1051 | ||
1052 | ||
1053 | #define middle_y r9 | |
1054 | ||
1055 | #define edges_xy_b q11 | |
1056 | #define edges_dx_dy_b d26 | |
1057 | #define edge_shifts_b d27 | |
1058 | #define edges_dx_dy_and_shifts_b q13 | |
1059 | #define height_increment d20 | |
1060 | ||
1061 | #define edges_dx_dy_and_shifts q1 | |
1062 | ||
1063 | #define edges_xy_b_left d22 | |
1064 | #define edges_xy_b_right d23 | |
1065 | ||
1066 | #define setup_spans_up_down_load_edge_set_b() \ | |
1067 | vmov edges_xy, edges_xy_b; \ | |
1068 | vmov edges_dx_dy_and_shifts, edges_dx_dy_and_shifts_b \ | |
1069 | ||
1070 | ||
1071 | function(setup_spans_up_down) | |
1072 | setup_spans_prologue() | |
1073 | ||
1074 | // s32 middle_y = y_a; | |
1075 | sub height_minor_a, y_a, y_b | |
1076 | sub height_minor_b, y_c, y_a | |
1077 | sub height_major, y_c, y_b | |
1078 | ||
1079 | vmov.u32 x_starts, x_a, x_c | |
1080 | vdup.u32 x_ends, x_b | |
1081 | ||
1082 | compute_edge_delta_x3(x_a, height_minor_a, height_major) | |
1083 | ||
1084 | mov temp, #0 | |
1085 | vmov.u32 height_increment, temp, height_minor_b | |
1086 | vmlal.s32 edges_xy, edges_dx_dy, height_increment | |
1087 | ||
1088 | vmov edges_xy_b_left, edge_alt_low, edge_alt_high | |
1089 | vmov edges_xy_b_right, edges_xy_right | |
1090 | ||
1091 | vmov edge_shifts_b, edge_shifts | |
1092 | vmov.u32 edge_shifts_b[0], edge_shift_alt | |
1093 | ||
1094 | vneg.s32 edges_dx_dy_b, edges_dx_dy | |
1095 | vmov.u32 edges_dx_dy_b[0], edge_dx_dy_alt | |
1096 | ||
1097 | mov middle_y, y_a | |
1098 | ||
1099 | setup_spans_load_b() | |
1100 | sub y_a, y_a, #1 | |
1101 | ||
1102 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ] | |
1103 | subs temp, temp, y_b | |
1104 | subgt height_minor_a, height_minor_a, temp | |
1105 | ||
1106 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ] | |
1107 | subs clip, y_a, temp | |
1108 | ble 0f | |
1109 | ||
1110 | sub height_minor_a, height_minor_a, clip | |
1111 | sub y_a, y_a, clip | |
1112 | setup_spans_clip(decrement, no) | |
1113 | ||
1114 | 0: | |
1115 | cmp height_minor_a, #0 | |
1116 | ble 3f | |
1117 | ||
1118 | orr temp, y_a, y_a, lsl #16 | |
1119 | sub temp, temp, #(1 << 16) | |
1120 | sub y_a, temp, #2 | |
1121 | sub y_a, y_a, #(2 << 16) | |
1122 | vmov.u32 y_x4, temp, y_a | |
1123 | ||
1124 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy | |
1125 | ||
1126 | strh height_minor_a, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
1127 | ||
1128 | setup_spans_adjust_edges_alternate_no(left, right); | |
1129 | setup_spans_adjust_interpolants_up() | |
1130 | setup_spans_up_down_load_edge_set_b() | |
1131 | ||
1132 | setup_spans_prologue_b() | |
1133 | ||
1134 | ||
1135 | 2: | |
1136 | setup_spans_set_x4_alternate_no(none, up) | |
1137 | subs height_minor_a, height_minor_a, #4 | |
1138 | bhi 2b | |
1139 | ||
1140 | add span_edge_data, span_edge_data, height_minor_a, lsl #3 | |
1141 | add span_uvrg_offset, span_uvrg_offset, height_minor_a, lsl #4 | |
1142 | add span_b_offset, span_b_offset, height_minor_a, lsl #2 | |
1143 | ||
1144 | 4: | |
1145 | add temp, psx_gpu, #psx_gpu_uvrg_offset | |
1146 | vld1.32 { uvrg }, [ temp ] | |
1147 | mov y_a, middle_y | |
1148 | ||
1149 | setup_spans_load_b() | |
1150 | ||
1151 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_end_y_offset ] | |
1152 | subs y_c, y_c, temp | |
1153 | subgt height_minor_b, height_minor_b, y_c | |
1154 | addgt height_minor_b, height_minor_b, #1 | |
1155 | ||
1156 | ldrh temp, [ psx_gpu, #psx_gpu_viewport_start_y_offset ] | |
1157 | subs clip, temp, y_a | |
1158 | ble 0f | |
1159 | ||
1160 | sub height_minor_b, height_minor_b, clip | |
1161 | add y_a, y_a, clip | |
1162 | setup_spans_clip(increment, no) | |
1163 | ||
1164 | 0: | |
1165 | cmp height_minor_b, #0 | |
1166 | ble 1f | |
1167 | ||
1168 | orr temp, y_a, y_a, lsl #16 | |
1169 | add temp, temp, #(1 << 16) | |
1170 | add y_a, temp, #2 | |
1171 | add y_a, y_a, #(2 << 16) | |
1172 | vmov.u32 y_x4, temp, y_a | |
1173 | ||
1174 | setup_spans_adjust_edges_alternate_no(left, right) | |
1175 | ||
1176 | ldrh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
1177 | add temp, temp, height_minor_b | |
b7569147 | 1178 | |
1179 | cmp temp, #MAX_SPANS | |
1180 | beq 5f | |
1181 | ||
75e28f62 E |
1182 | strh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] |
1183 | ||
1184 | 2: | |
1185 | setup_spans_set_x4_alternate_no(none, down) | |
1186 | subs height_minor_b, height_minor_b, #4 | |
1187 | bhi 2b | |
1188 | ||
1189 | 1: | |
1190 | setup_spans_epilogue() | |
1191 | ||
1192 | 3: | |
1193 | setup_spans_up_down_load_edge_set_b() | |
1194 | setup_spans_prologue_b() | |
1195 | bal 4b | |
1196 | ||
b7569147 | 1197 | 5: |
1198 | // FIXME: overflow corner case | |
1199 | sub temp, temp, height_minor_b | |
1200 | bics height_minor_b, #3 | |
1201 | add temp, temp, height_minor_b | |
1202 | strh temp, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
1203 | bne 2b | |
1204 | bal 1b | |
1205 | ||
5d834c08 | 1206 | .pool |
75e28f62 E |
1207 | |
1208 | #undef span_uvrg_offset | |
1209 | #undef span_edge_data | |
1210 | #undef span_b_offset | |
1211 | #undef left_x | |
1212 | #undef b | |
1213 | ||
1214 | #define psx_gpu r0 | |
1215 | #define num_spans r1 | |
1216 | #define span_uvrg_offset r2 | |
1217 | #define span_edge_data r3 | |
1218 | #define span_b_offset r4 | |
1219 | #define b_dx r5 | |
1220 | #define span_num_blocks r6 | |
1221 | #define y r7 | |
1222 | #define left_x r8 | |
1223 | #define b r9 | |
1224 | #define dither_offset_ptr r10 | |
1225 | #define block_ptr_a r11 | |
1226 | #define fb_ptr r12 | |
1227 | #define num_blocks r14 | |
1228 | ||
1229 | #define uvrg_dx_ptr r2 | |
1230 | #define texture_mask_ptr r3 | |
1231 | #define dither_shift r8 | |
1232 | #define dither_row r10 | |
1233 | ||
1234 | #define c_32 r7 | |
1235 | #define b_dx4 r8 | |
1236 | #define b_dx8 r9 | |
1237 | #define block_ptr_b r10 | |
1238 | ||
1239 | #define block_span_ptr r10 | |
1240 | #define right_mask r8 | |
1241 | ||
1242 | #define color r2 | |
1243 | #define color_r r3 | |
1244 | #define color_g r4 | |
1245 | #define color_b r5 | |
1246 | ||
1247 | #undef uvrg | |
1248 | ||
1249 | #define u_block q0 | |
1250 | #define v_block q1 | |
1251 | #define r_block q2 | |
1252 | #define g_block q3 | |
1253 | #define b_block q4 | |
1254 | ||
1255 | #define uv_dx4 d10 | |
1256 | #define rg_dx4 d11 | |
1257 | #define uv_dx8 d12 | |
1258 | #define rg_dx8 d13 | |
1259 | #define b_whole_8 d14 | |
1260 | #define fb_mask_ptrs d15 | |
1261 | ||
1262 | #define uvrg_dx4 q5 | |
1263 | #define uvrg_dx8 q6 | |
1264 | #define uv_dx8 d12 | |
1265 | #define rg_dx8 d13 | |
1266 | ||
1267 | #define u_whole q8 | |
1268 | #define v_whole q9 | |
1269 | #define r_whole q10 | |
1270 | #define g_whole q11 | |
1271 | #define b_whole q12 | |
1272 | ||
1273 | #define u_whole_low d16 | |
1274 | #define u_whole_high d17 | |
1275 | #define v_whole_low d18 | |
1276 | #define v_whole_high d19 | |
1277 | #define r_whole_low d20 | |
1278 | #define r_whole_high d21 | |
1279 | #define g_whole_low d22 | |
1280 | #define g_whole_high d23 | |
1281 | #define b_whole_low d24 | |
1282 | #define b_whole_high d25 | |
1283 | ||
1284 | #define dx4 q13 | |
1285 | #define dx8 q13 | |
1286 | ||
1287 | #define u_whole_8 d26 | |
1288 | #define v_whole_8 d27 | |
1289 | #define u_whole_8b d24 | |
1290 | #define r_whole_8 d24 | |
1291 | #define g_whole_8 d25 | |
1292 | ||
1293 | #define uv_whole_8 q13 | |
1294 | #define uv_whole_8b q14 | |
1295 | ||
1296 | #define dither_offsets q14 | |
1297 | #define texture_mask q15 | |
1298 | #define texture_mask_u d30 | |
1299 | #define texture_mask_v d31 | |
1300 | ||
1301 | #define dither_offsets_short d28 | |
1302 | ||
1303 | #define v_left_x q8 | |
1304 | #define uvrg q9 | |
1305 | #define block_span q10 | |
1306 | ||
1307 | #define uv d18 | |
1308 | #define rg d19 | |
1309 | ||
1310 | #define draw_mask q1 | |
1311 | #define draw_mask_edge q13 | |
1312 | #define test_mask q0 | |
1313 | ||
1314 | #define uvrg_dx q3 | |
1315 | ||
1316 | #define colors q2 | |
1317 | ||
1318 | #define setup_blocks_texture_swizzled() \ | |
1319 | vand.u8 u_whole_8b, u_whole_8, texture_mask_u; \ | |
1320 | vsli.u8 u_whole_8, v_whole_8, #4; \ | |
1321 | vsri.u8 v_whole_8, u_whole_8b, #4 \ | |
1322 | ||
1323 | #define setup_blocks_texture_unswizzled() \ | |
1324 | ||
1325 | ||
1326 | #define setup_blocks_shaded_textured_builder(swizzling) \ | |
1327 | .align 3; \ | |
1328 | \ | |
1329 | function(setup_blocks_shaded_textured_dithered_##swizzling##_indirect) \ | |
1330 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
1331 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ | |
1332 | \ | |
1333 | vld1.u32 { uvrg_dx }, [ uvrg_dx_ptr, :128 ]; \ | |
1334 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ | |
1335 | \ | |
1336 | cmp num_spans, #0; \ | |
1337 | bxeq lr; \ | |
1338 | \ | |
1339 | stmdb sp!, { r4 - r11, r14 }; \ | |
1340 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ | |
1341 | \ | |
1342 | ldr b_dx, [ psx_gpu, #psx_gpu_b_dx_offset ]; \ | |
1343 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
1344 | \ | |
1345 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [ texture_mask_ptr, :16 ]; \ | |
1346 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
1347 | \ | |
1348 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
1349 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
1350 | \ | |
1351 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
1352 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1353 | \ | |
1354 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
1355 | \ | |
1356 | 0: \ | |
1357 | vmov.u8 fb_mask_ptrs, #0; \ | |
1358 | \ | |
1359 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
1360 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
1361 | \ | |
1362 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
c1817bd9 | 1363 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ |
75e28f62 E |
1364 | \ |
1365 | cmp span_num_blocks, #0; \ | |
1366 | beq 1f; \ | |
1367 | \ | |
1368 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
1369 | add num_blocks, span_num_blocks, num_blocks; \ | |
1370 | \ | |
1371 | cmp num_blocks, #MAX_BLOCKS; \ | |
1372 | bgt 2f; \ | |
1373 | \ | |
1374 | 3: \ | |
1375 | ldr b, [ span_b_offset ]; \ | |
1376 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
1377 | \ | |
1378 | vdup.u32 v_left_x, left_x; \ | |
1379 | and y, y, #0x3; \ | |
1380 | \ | |
1381 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
1382 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
1383 | \ | |
1384 | mla b, b_dx, left_x, b; \ | |
1385 | and dither_shift, left_x, #0x03; \ | |
1386 | \ | |
1387 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
1388 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ | |
1389 | \ | |
1390 | mov dither_shift, dither_shift, lsl #3; \ | |
1391 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
1392 | \ | |
1393 | mov c_32, #32; \ | |
1394 | subs span_num_blocks, span_num_blocks, #1; \ | |
1395 | \ | |
1396 | mov dither_row, dither_row, ror dither_shift; \ | |
1397 | mov b_dx4, b_dx, lsl #2; \ | |
1398 | \ | |
1399 | vdup.u32 dither_offsets_short, dither_row; \ | |
1400 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
1401 | \ | |
1402 | vdup.u32 b_block, b; \ | |
1403 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
1404 | \ | |
1405 | vdup.u32 u_block, uv[0]; \ | |
1406 | mov b_dx8, b_dx, lsl #3; \ | |
1407 | \ | |
1408 | vdup.u32 v_block, uv[1]; \ | |
1409 | vdup.u32 r_block, rg[0]; \ | |
1410 | vdup.u32 g_block, rg[1]; \ | |
1411 | \ | |
1412 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
1413 | \ | |
1414 | vadd.u32 u_block, u_block, block_span; \ | |
1415 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
1416 | \ | |
1417 | vadd.u32 v_block, v_block, block_span; \ | |
1418 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
1419 | \ | |
1420 | vadd.u32 r_block, r_block, block_span; \ | |
1421 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
1422 | \ | |
1423 | vadd.u32 g_block, g_block, block_span; \ | |
1424 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]; \ | |
1425 | \ | |
1426 | vadd.u32 b_block, b_block, block_span; \ | |
1427 | add block_ptr_b, block_ptr_a, #16; \ | |
1428 | \ | |
1429 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1430 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1431 | vshrn.u32 r_whole_low, r_block, #16; \ | |
1432 | vshrn.u32 g_whole_low, g_block, #16; \ | |
1433 | \ | |
1434 | vdup.u32 dx4, uv_dx4[0]; \ | |
1435 | vshrn.u32 b_whole_low, b_block, #16; \ | |
1436 | \ | |
1437 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1438 | vdup.u32 dx4, uv_dx4[1]; \ | |
1439 | \ | |
1440 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1441 | vdup.u32 dx4, rg_dx4[0]; \ | |
1442 | \ | |
1443 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
1444 | vdup.u32 dx4, rg_dx4[1]; \ | |
1445 | \ | |
1446 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
1447 | vdup.u32 dx4, b_dx4; \ | |
1448 | \ | |
1449 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
1450 | vdup.u32 dx8, uv_dx8[0]; \ | |
1451 | \ | |
1452 | vadd.u32 u_block, u_block, dx8; \ | |
1453 | vdup.u32 dx8, uv_dx8[1]; \ | |
1454 | \ | |
1455 | vadd.u32 v_block, v_block, dx8; \ | |
1456 | vdup.u32 dx8, rg_dx8[0]; \ | |
1457 | \ | |
1458 | vadd.u32 r_block, r_block, dx8; \ | |
1459 | vdup.u32 dx8, rg_dx8[1]; \ | |
1460 | \ | |
1461 | vadd.u32 g_block, g_block, dx8; \ | |
1462 | vdup.u32 dx8, b_dx8; \ | |
1463 | \ | |
1464 | vadd.u32 b_block, b_block, dx8; \ | |
1465 | vmovn.u16 u_whole_8, u_whole; \ | |
1466 | \ | |
1467 | vmovn.u16 v_whole_8, v_whole; \ | |
1468 | \ | |
1469 | vmovn.u16 b_whole_8, b_whole; \ | |
1470 | pld [ fb_ptr ]; \ | |
1471 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1472 | \ | |
1473 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1474 | setup_blocks_texture_##swizzling(); \ | |
1475 | \ | |
1476 | vmovn.u16 r_whole_8, r_whole; \ | |
1477 | beq 5f; \ | |
1478 | \ | |
1479 | 4: \ | |
1480 | vmovn.u16 g_whole_8, g_whole; \ | |
1481 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1482 | \ | |
1483 | vst2.u8 { u_whole_8, v_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
1484 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1485 | \ | |
1486 | vst1.u32 { r_whole_8, g_whole_8 }, [ block_ptr_b, :128 ], c_32; \ | |
1487 | vshrn.u32 r_whole_low, r_block, #16; \ | |
1488 | \ | |
1489 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
1490 | vshrn.u32 g_whole_low, g_block, #16; \ | |
1491 | \ | |
1492 | vdup.u32 dx4, uv_dx4[0]; \ | |
1493 | vshrn.u32 b_whole_low, b_block, #16; \ | |
1494 | \ | |
1495 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1496 | vdup.u32 dx4, uv_dx4[1]; \ | |
1497 | \ | |
1498 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1499 | vdup.u32 dx4, rg_dx4[0]; \ | |
1500 | \ | |
1501 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
1502 | vdup.u32 dx4, rg_dx4[1]; \ | |
1503 | \ | |
1504 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
1505 | vdup.u32 dx4, b_dx4; \ | |
1506 | \ | |
1507 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
1508 | vdup.u32 dx8, uv_dx8[0]; \ | |
1509 | \ | |
1510 | vadd.u32 u_block, u_block, dx8; \ | |
1511 | vdup.u32 dx8, uv_dx8[1]; \ | |
1512 | \ | |
1513 | vadd.u32 v_block, v_block, dx8; \ | |
1514 | vdup.u32 dx8, rg_dx8[0]; \ | |
1515 | \ | |
1516 | vadd.u32 r_block, r_block, dx8; \ | |
1517 | vdup.u32 dx8, rg_dx8[1]; \ | |
1518 | \ | |
1519 | vadd.u32 g_block, g_block, dx8; \ | |
1520 | vdup.u32 dx8, b_dx8; \ | |
1521 | \ | |
1522 | vadd.u32 b_block, b_block, dx8; \ | |
1523 | vmovn.u16 u_whole_8, u_whole; \ | |
1524 | \ | |
1525 | add fb_ptr, fb_ptr, #16; \ | |
1526 | vmovn.u16 v_whole_8, v_whole; \ | |
1527 | \ | |
1528 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
1529 | vmovn.u16 b_whole_8, b_whole; \ | |
1530 | \ | |
1531 | pld [ fb_ptr ]; \ | |
1532 | \ | |
1533 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1534 | subs span_num_blocks, span_num_blocks, #1; \ | |
1535 | \ | |
1536 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1537 | setup_blocks_texture_##swizzling(); \ | |
1538 | \ | |
1539 | vmovn.u16 r_whole_8, r_whole; \ | |
1540 | bne 4b; \ | |
1541 | \ | |
1542 | 5: \ | |
1543 | vmovn.u16 g_whole_8, g_whole; \ | |
1544 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ | |
1545 | \ | |
1546 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
1547 | vdup.u8 draw_mask, right_mask; \ | |
1548 | \ | |
1549 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
1550 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
1551 | vzip.u8 u_whole_8, v_whole_8; \ | |
1552 | \ | |
1553 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
1554 | vst1.u32 { r_whole_8, g_whole_8 }, [ block_ptr_b, :128 ], c_32; \ | |
1555 | vst1.u32 { uv_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
1556 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
1557 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
1558 | \ | |
1559 | 1: \ | |
1560 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
1561 | add span_b_offset, span_b_offset, #4; \ | |
1562 | \ | |
1563 | add span_edge_data, span_edge_data, #8; \ | |
1564 | subs num_spans, num_spans, #1; \ | |
1565 | \ | |
1566 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
1567 | bne 0b; \ | |
1568 | \ | |
1569 | ldmia sp!, { r4 - r11, pc }; \ | |
1570 | \ | |
1571 | 2: \ | |
1572 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
1573 | vpush { texture_mask }; \ | |
1574 | vpush { uvrg_dx4 }; \ | |
1575 | \ | |
1576 | stmdb sp!, { r0 - r3, r12, r14 }; \ | |
1577 | bl flush_render_block_buffer; \ | |
1578 | ldmia sp!, { r0 - r3, r12, r14 }; \ | |
1579 | \ | |
1580 | vpop { uvrg_dx4 }; \ | |
1581 | vpop { texture_mask }; \ | |
1582 | \ | |
1583 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
1584 | vmov.u8 fb_mask_ptrs, #0; \ | |
1585 | \ | |
1586 | mov num_blocks, span_num_blocks; \ | |
1587 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1588 | bal 3b \ | |
1589 | ||
1590 | ||
1591 | setup_blocks_shaded_textured_builder(swizzled) | |
1592 | setup_blocks_shaded_textured_builder(unswizzled) | |
1593 | ||
1594 | ||
1595 | #define setup_blocks_unshaded_textured_builder(swizzling) \ | |
1596 | .align 3; \ | |
1597 | \ | |
1598 | function(setup_blocks_unshaded_textured_dithered_##swizzling##_indirect) \ | |
1599 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
1600 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ | |
1601 | \ | |
1602 | vld1.u32 { uvrg_dx }, [ uvrg_dx_ptr, :128 ]; \ | |
1603 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ | |
1604 | \ | |
1605 | cmp num_spans, #0; \ | |
1606 | bxeq lr; \ | |
1607 | \ | |
1608 | stmdb sp!, { r4 - r11, r14 }; \ | |
1609 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ | |
1610 | \ | |
1611 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
1612 | \ | |
1613 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [ texture_mask_ptr, :16 ]; \ | |
1614 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
1615 | \ | |
1616 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
1617 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
1618 | \ | |
1619 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1620 | \ | |
1621 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
1622 | \ | |
1623 | 0: \ | |
1624 | vmov.u8 fb_mask_ptrs, #0; \ | |
1625 | \ | |
1626 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
1627 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
1628 | \ | |
1629 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
c1817bd9 | 1630 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ |
75e28f62 E |
1631 | \ |
1632 | cmp span_num_blocks, #0; \ | |
1633 | beq 1f; \ | |
1634 | \ | |
1635 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
1636 | add num_blocks, span_num_blocks, num_blocks; \ | |
1637 | \ | |
1638 | cmp num_blocks, #MAX_BLOCKS; \ | |
1639 | bgt 2f; \ | |
1640 | \ | |
1641 | 3: \ | |
1642 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
1643 | \ | |
1644 | vdup.u32 v_left_x, left_x; \ | |
1645 | and y, y, #0x3; \ | |
1646 | \ | |
1647 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
1648 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
1649 | \ | |
1650 | and dither_shift, left_x, #0x03; \ | |
1651 | \ | |
1652 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
1653 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ | |
1654 | \ | |
1655 | mov dither_shift, dither_shift, lsl #3; \ | |
1656 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
1657 | \ | |
1658 | mov c_32, #32; \ | |
1659 | subs span_num_blocks, span_num_blocks, #1; \ | |
1660 | \ | |
1661 | mov dither_row, dither_row, ror dither_shift; \ | |
1662 | \ | |
1663 | vdup.u32 dither_offsets_short, dither_row; \ | |
1664 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
1665 | \ | |
1666 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
1667 | \ | |
1668 | vdup.u32 u_block, uv[0]; \ | |
1669 | \ | |
1670 | vdup.u32 v_block, uv[1]; \ | |
1671 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
1672 | \ | |
1673 | vadd.u32 u_block, u_block, block_span; \ | |
1674 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
1675 | \ | |
1676 | vadd.u32 v_block, v_block, block_span; \ | |
1677 | add block_ptr_b, block_ptr_a, #16; \ | |
1678 | \ | |
1679 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1680 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1681 | \ | |
1682 | vdup.u32 dx4, uv_dx4[0]; \ | |
1683 | \ | |
1684 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1685 | vdup.u32 dx4, uv_dx4[1]; \ | |
1686 | \ | |
1687 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1688 | vdup.u32 dx8, uv_dx8[0]; \ | |
1689 | \ | |
1690 | vadd.u32 u_block, u_block, dx8; \ | |
1691 | vdup.u32 dx8, uv_dx8[1]; \ | |
1692 | \ | |
1693 | vadd.u32 v_block, v_block, dx8; \ | |
1694 | vmovn.u16 u_whole_8, u_whole; \ | |
1695 | \ | |
1696 | vmovn.u16 v_whole_8, v_whole; \ | |
1697 | \ | |
1698 | pld [ fb_ptr ]; \ | |
1699 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1700 | \ | |
1701 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1702 | setup_blocks_texture_##swizzling(); \ | |
1703 | \ | |
1704 | beq 5f; \ | |
1705 | \ | |
1706 | 4: \ | |
1707 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1708 | \ | |
1709 | vst2.u8 { u_whole_8, v_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
1710 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1711 | \ | |
1712 | add block_ptr_b, block_ptr_b, #32; \ | |
1713 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
1714 | \ | |
1715 | vdup.u32 dx4, uv_dx4[0]; \ | |
1716 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1717 | vdup.u32 dx4, uv_dx4[1]; \ | |
1718 | \ | |
1719 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1720 | vdup.u32 dx8, uv_dx8[0]; \ | |
1721 | \ | |
1722 | vadd.u32 u_block, u_block, dx8; \ | |
1723 | vdup.u32 dx8, uv_dx8[1]; \ | |
1724 | \ | |
1725 | vadd.u32 v_block, v_block, dx8; \ | |
1726 | vmovn.u16 u_whole_8, u_whole; \ | |
1727 | \ | |
1728 | add fb_ptr, fb_ptr, #16; \ | |
1729 | vmovn.u16 v_whole_8, v_whole; \ | |
1730 | \ | |
1731 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
1732 | pld [ fb_ptr ]; \ | |
1733 | \ | |
1734 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1735 | subs span_num_blocks, span_num_blocks, #1; \ | |
1736 | \ | |
1737 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1738 | setup_blocks_texture_##swizzling(); \ | |
1739 | \ | |
1740 | bne 4b; \ | |
1741 | \ | |
1742 | 5: \ | |
1743 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ | |
1744 | \ | |
1745 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
1746 | vdup.u8 draw_mask, right_mask; \ | |
1747 | \ | |
1748 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
1749 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
1750 | vzip.u8 u_whole_8, v_whole_8; \ | |
1751 | \ | |
1752 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
1753 | add block_ptr_b, block_ptr_b, #32; \ | |
1754 | vst1.u32 { uv_whole_8 }, [ block_ptr_a, :128 ], c_32; \ | |
1755 | vst1.u32 { dither_offsets }, [ block_ptr_b, :128 ], c_32; \ | |
1756 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32; \ | |
1757 | \ | |
1758 | 1: \ | |
1759 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
1760 | add span_edge_data, span_edge_data, #8; \ | |
1761 | subs num_spans, num_spans, #1; \ | |
1762 | \ | |
1763 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
1764 | bne 0b; \ | |
1765 | \ | |
1766 | ldmia sp!, { r4 - r11, pc }; \ | |
1767 | \ | |
1768 | 2: \ | |
1769 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
1770 | vpush { texture_mask }; \ | |
1771 | vpush { uvrg_dx4 }; \ | |
1772 | \ | |
1773 | stmdb sp!, { r0 - r3, r12, r14 }; \ | |
1774 | bl flush_render_block_buffer; \ | |
1775 | ldmia sp!, { r0 - r3, r12, r14 }; \ | |
1776 | \ | |
1777 | vpop { uvrg_dx4 }; \ | |
1778 | vpop { texture_mask }; \ | |
1779 | \ | |
1780 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
1781 | vmov.u8 fb_mask_ptrs, #0; \ | |
1782 | \ | |
1783 | mov num_blocks, span_num_blocks; \ | |
1784 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1785 | bal 3b \ | |
1786 | ||
1787 | ||
1788 | setup_blocks_unshaded_textured_builder(swizzled) | |
1789 | setup_blocks_unshaded_textured_builder(unswizzled) | |
1790 | ||
1791 | ||
1792 | .align 3 | |
1793 | ||
1794 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_indirect) | |
1795 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
1796 | veor.u32 draw_mask, draw_mask, draw_mask | |
1797 | ||
1798 | cmp num_spans, #0 | |
1799 | bxeq lr | |
1800 | ||
1801 | stmdb sp!, { r4 - r11, r14 } | |
1802 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
1803 | ||
1804 | ldr color, [ psx_gpu, #psx_gpu_triangle_color_offset ] | |
1805 | ||
1806 | ubfx color_r, color, #3, #5 | |
1807 | ubfx color_g, color, #11, #5 | |
1808 | ubfx color_b, color, #19, #5 | |
1809 | ||
1810 | orr color, color_r, color_b, lsl #10 | |
1811 | orr color, color, color_g, lsl #5 | |
1812 | ||
1813 | vdup.u16 colors, color | |
1814 | ||
1815 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
1816 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
1817 | ||
1818 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
1819 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6 | |
1820 | ||
1821 | 0: | |
1822 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ] | |
1823 | ldrh y, [ span_edge_data, #edge_data_y_offset ] | |
1824 | ||
c1817bd9 | 1825 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] |
75e28f62 E |
1826 | |
1827 | cmp span_num_blocks, #0 | |
1828 | beq 1f | |
1829 | ||
1830 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ] | |
1831 | add num_blocks, span_num_blocks, num_blocks | |
1832 | ||
1833 | cmp num_blocks, #MAX_BLOCKS | |
1834 | bgt 2f | |
1835 | ||
1836 | 3: | |
1837 | add fb_ptr, fb_ptr, y, lsl #11 | |
1838 | and y, y, #0x3 | |
1839 | ||
1840 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
1841 | mov c_32, #32 | |
1842 | ||
1843 | subs span_num_blocks, span_num_blocks, #1 | |
1844 | ||
1845 | add block_ptr_b, block_ptr_a, #16 | |
1846 | pld [ fb_ptr ] | |
1847 | ||
1848 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
1849 | beq 5f | |
1850 | ||
1851 | 4: | |
1852 | vst1.u32 { draw_mask }, [ block_ptr_a, :128 ], c_32 | |
1853 | vst1.u32 { colors }, [ block_ptr_b, :128 ], c_32 | |
1854 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32 | |
1855 | ||
1856 | add fb_ptr, fb_ptr, #16 | |
1857 | add block_ptr_b, block_ptr_b, #32 | |
1858 | ||
1859 | pld [ fb_ptr ] | |
1860 | ||
1861 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
1862 | subs span_num_blocks, span_num_blocks, #1 | |
1863 | ||
1864 | bne 4b | |
1865 | ||
1866 | 5: | |
1867 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ] | |
1868 | ||
1869 | vdup.u8 draw_mask_edge, right_mask | |
1870 | vtst.u16 draw_mask_edge, draw_mask_edge, test_mask | |
1871 | ||
1872 | vst1.u32 { colors }, [ block_ptr_b, :128 ], c_32 | |
1873 | vst1.u32 { draw_mask_edge }, [ block_ptr_a, :128 ], c_32 | |
1874 | add block_ptr_b, block_ptr_b, #32 | |
1875 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [ block_ptr_a, :128 ], c_32 | |
1876 | ||
1877 | 1: | |
1878 | add span_edge_data, span_edge_data, #8 | |
1879 | subs num_spans, num_spans, #1 | |
1880 | ||
1881 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
1882 | bne 0b | |
1883 | ||
1884 | ldmia sp!, { r4 - r11, pc } | |
1885 | ||
1886 | 2: | |
1887 | vpush { colors } | |
1888 | ||
1889 | stmdb sp!, { r0 - r3, r12, r14 } | |
1890 | bl flush_render_block_buffer | |
1891 | ldmia sp!, { r0 - r3, r12, r14 } | |
1892 | ||
1893 | vpop { colors } | |
1894 | ||
1895 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
1896 | veor.u32 draw_mask, draw_mask, draw_mask | |
1897 | ||
1898 | mov num_blocks, span_num_blocks | |
1899 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
1900 | bal 3b | |
1901 | ||
1902 | ||
1903 | #define mask_msb_scalar r14 | |
1904 | ||
1905 | #define msb_mask q15 | |
1906 | ||
1907 | #define pixels_low d16 | |
1908 | ||
1909 | #define msb_mask_low d30 | |
1910 | #define msb_mask_high d31 | |
1911 | ||
1912 | ||
1913 | .align 3 | |
1914 | ||
1915 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_direct) | |
1916 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ] | |
1917 | ||
1918 | cmp num_spans, #0 | |
1919 | bxeq lr | |
1920 | ||
1921 | stmdb sp!, { r4 - r11, r14 } | |
1922 | ||
1923 | ldr color, [ psx_gpu, #psx_gpu_triangle_color_offset ] | |
1924 | ||
1925 | ubfx color_r, color, #3, #5 | |
1926 | ubfx color_g, color, #11, #5 | |
1927 | ||
1928 | ldrh mask_msb_scalar, [ psx_gpu, #psx_gpu_mask_msb_offset ] | |
1929 | ubfx color_b, color, #19, #5 | |
1930 | ||
1931 | orr color, color_r, color_b, lsl #10 | |
1932 | orr color, color, color_g, lsl #5 | |
1933 | orr color, color, mask_msb_scalar | |
1934 | ||
1935 | vdup.u16 colors, color | |
1936 | ||
1937 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
3867c6ef E |
1938 | orr color, color, lsl #16 |
1939 | ||
75e28f62 E |
1940 | |
1941 | 0: | |
1942 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ] | |
1943 | ldrh y, [ span_edge_data, #edge_data_y_offset ] | |
1944 | ||
c1817bd9 | 1945 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] |
75e28f62 E |
1946 | |
1947 | cmp span_num_blocks, #0 | |
1948 | beq 1f | |
1949 | ||
1950 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ] | |
1951 | ||
1952 | add fb_ptr, fb_ptr, y, lsl #11 | |
1953 | subs span_num_blocks, span_num_blocks, #1 | |
1954 | ||
1955 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
1956 | beq 3f | |
1957 | ||
1958 | 2: | |
1959 | vst1.u32 { colors }, [ fb_ptr ]! | |
1960 | subs span_num_blocks, span_num_blocks, #1 | |
1961 | ||
1962 | bne 2b | |
1963 | ||
1964 | 3: | |
1965 | ldrb right_mask, [ span_edge_data, #edge_data_right_mask_offset ] | |
75e28f62 | 1966 | |
3867c6ef E |
1967 | cmp right_mask, #0x0 |
1968 | beq 5f | |
1969 | ||
1970 | tst right_mask, #0xF | |
1971 | streq color, [ fb_ptr ], #4 | |
1972 | moveq right_mask, right_mask, lsr #4 | |
1973 | streq color, [ fb_ptr ], #4 | |
1974 | ||
1975 | tst right_mask, #0x3 | |
1976 | streq color, [ fb_ptr ], #4 | |
1977 | moveq right_mask, right_mask, lsr #2 | |
1978 | ||
1979 | tst right_mask, #0x1 | |
1980 | streqh color, [ fb_ptr ] | |
75e28f62 E |
1981 | |
1982 | 1: | |
1983 | add span_edge_data, span_edge_data, #8 | |
1984 | subs num_spans, num_spans, #1 | |
75e28f62 E |
1985 | bne 0b |
1986 | ||
1987 | ldmia sp!, { r4 - r11, pc } | |
1988 | ||
3867c6ef E |
1989 | 5: |
1990 | vst1.u32 { colors }, [ fb_ptr ] | |
1991 | bal 1b | |
75e28f62 E |
1992 | |
1993 | ||
1994 | #undef c_64 | |
1995 | ||
1996 | #define c_64 r7 | |
1997 | #define rg_dx_ptr r2 | |
1998 | ||
1999 | ||
2000 | #undef r_block | |
2001 | #undef g_block | |
2002 | #undef b_block | |
2003 | #undef r_whole | |
2004 | #undef g_whole | |
2005 | #undef b_whole | |
2006 | #undef r_whole_low | |
2007 | #undef r_whole_high | |
2008 | #undef g_whole_low | |
2009 | #undef g_whole_high | |
2010 | #undef b_whole_low | |
2011 | #undef b_whole_high | |
2012 | #undef r_whole_8 | |
2013 | #undef g_whole_8 | |
2014 | #undef b_whole_8 | |
2015 | #undef dither_offsets | |
2016 | #undef rg_dx4 | |
2017 | #undef rg_dx8 | |
2018 | #undef dx4 | |
2019 | #undef dx8 | |
2020 | #undef v_left_x | |
2021 | #undef uvrg | |
2022 | #undef block_span | |
2023 | #undef rg | |
2024 | #undef draw_mask | |
2025 | #undef test_mask | |
2026 | ||
2027 | #define r_block q0 | |
2028 | #define g_block q1 | |
2029 | #define b_block q2 | |
2030 | ||
2031 | #define r_whole q3 | |
2032 | #define g_whole q4 | |
2033 | #define b_whole q5 | |
2034 | ||
2035 | #define r_whole_low d6 | |
2036 | #define r_whole_high d7 | |
2037 | #define g_whole_low d8 | |
2038 | #define g_whole_high d9 | |
2039 | #define b_whole_low d10 | |
2040 | #define b_whole_high d11 | |
2041 | ||
2042 | #define gb_whole_8 q6 | |
2043 | ||
2044 | #define g_whole_8 d12 | |
2045 | #define b_whole_8 d13 | |
2046 | ||
2047 | #define r_whole_8 d14 | |
2048 | ||
2049 | #define pixels q8 | |
2050 | ||
2051 | #define rg_dx4 d18 | |
2052 | #define rg_dx8 d19 | |
2053 | ||
2054 | #define dx4 q10 | |
2055 | #define dx8 q10 | |
2056 | ||
2057 | #define v_left_x d6 | |
2058 | #define uvrg q4 | |
2059 | #define block_span q5 | |
2060 | ||
2061 | #define rg d9 | |
2062 | ||
2063 | #define d64_1 d22 | |
2064 | #define d64_128 d23 | |
2065 | ||
2066 | #define d128_4 q12 | |
2067 | #define d128_0x7 q13 | |
2068 | ||
2069 | #define d64_4 d24 | |
2070 | ||
2071 | #define dither_offsets q14 | |
2072 | #define draw_mask q15 | |
2073 | ||
2074 | #define dither_offsets_low d28 | |
2075 | ||
2076 | #define rg_dx d0 | |
2077 | #define test_mask q10 | |
2078 | ||
2079 | ||
2080 | #define setup_blocks_shaded_untextured_dither_a_dithered() \ | |
2081 | vqadd.u8 r_whole_8, r_whole_8, dither_offsets_low; \ | |
2082 | vqadd.u8 gb_whole_8, gb_whole_8, dither_offsets; \ | |
2083 | ||
2084 | #define setup_blocks_shaded_untextured_dither_b_dithered() \ | |
2085 | vqsub.u8 r_whole_8, r_whole_8, d64_4; \ | |
2086 | vqsub.u8 gb_whole_8, gb_whole_8, d128_4 \ | |
2087 | ||
2088 | #define setup_blocks_shaded_untextured_dither_a_undithered() \ | |
2089 | ||
2090 | #define setup_blocks_shaded_untextured_dither_b_undithered() \ | |
2091 | ||
2092 | ||
2093 | #define setup_blocks_shaded_untextured_indirect_builder(dithering) \ | |
2094 | .align 3; \ | |
2095 | \ | |
2096 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_indirect) \ | |
2097 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
2098 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ | |
2099 | \ | |
2100 | vld1.u32 { rg_dx }, [ rg_dx_ptr, :64 ]; \ | |
2101 | \ | |
2102 | cmp num_spans, #0; \ | |
2103 | bxeq lr; \ | |
2104 | \ | |
2105 | stmdb sp!, { r4 - r11, r14 }; \ | |
2106 | vshl.u32 rg_dx4, rg_dx, #2; \ | |
2107 | \ | |
2108 | ldr b_dx, [ psx_gpu, #psx_gpu_b_dx_offset ]; \ | |
2109 | vshl.u32 rg_dx8, rg_dx, #3; \ | |
2110 | \ | |
2111 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
2112 | \ | |
2113 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
2114 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
2115 | \ | |
2116 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
2117 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
2118 | \ | |
2119 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
2120 | vmov.u8 d64_1, #1; \ | |
2121 | \ | |
2122 | vmov.u8 d128_4, #4; \ | |
2123 | vmov.u8 d64_128, #128; \ | |
2124 | \ | |
2125 | vmov.u8 d128_0x7, #0x7; \ | |
2126 | \ | |
2127 | 0: \ | |
2128 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
2129 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
2130 | \ | |
2131 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
c1817bd9 | 2132 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ |
75e28f62 E |
2133 | \ |
2134 | cmp span_num_blocks, #0; \ | |
2135 | beq 1f; \ | |
2136 | \ | |
2137 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
2138 | add num_blocks, span_num_blocks, num_blocks; \ | |
2139 | \ | |
2140 | cmp num_blocks, #MAX_BLOCKS; \ | |
2141 | bgt 2f; \ | |
2142 | \ | |
2143 | 3: \ | |
2144 | ldr b, [ span_b_offset ]; \ | |
2145 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
2146 | \ | |
2147 | vdup.u32 v_left_x, left_x; \ | |
2148 | and y, y, #0x3; \ | |
2149 | \ | |
2150 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
2151 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
2152 | \ | |
2153 | mla b, b_dx, left_x, b; \ | |
2154 | and dither_shift, left_x, #0x03; \ | |
2155 | \ | |
2156 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
2157 | vshr.u32 rg_dx, rg_dx4, #2; \ | |
2158 | \ | |
2159 | mov dither_shift, dither_shift, lsl #3; \ | |
2160 | vmla.u32 rg, rg_dx, v_left_x; \ | |
2161 | \ | |
2162 | mov c_64, #64; \ | |
2163 | subs span_num_blocks, span_num_blocks, #1; \ | |
2164 | \ | |
2165 | mov dither_row, dither_row, ror dither_shift; \ | |
2166 | mov b_dx4, b_dx, lsl #2; \ | |
2167 | \ | |
2168 | vdup.u32 dither_offsets, dither_row; \ | |
2169 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
2170 | \ | |
2171 | vdup.u32 b_block, b; \ | |
2172 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
2173 | \ | |
2174 | mov b_dx8, b_dx, lsl #3; \ | |
2175 | vdup.u32 r_block, rg[0]; \ | |
2176 | vdup.u32 g_block, rg[1]; \ | |
2177 | \ | |
2178 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
2179 | \ | |
2180 | vadd.u32 r_block, r_block, block_span; \ | |
2181 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
2182 | \ | |
2183 | vadd.u32 g_block, g_block, block_span; \ | |
2184 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]; \ | |
2185 | \ | |
2186 | vadd.u32 b_block, b_block, block_span; \ | |
2187 | add block_ptr_b, block_ptr_a, #16; \ | |
2188 | \ | |
2189 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2190 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2191 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2192 | vdup.u32 dx4, rg_dx4[0]; \ | |
2193 | \ | |
2194 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2195 | vdup.u32 dx4, rg_dx4[1]; \ | |
2196 | \ | |
2197 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2198 | vdup.u32 dx4, b_dx4; \ | |
2199 | \ | |
2200 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2201 | vdup.u32 dx8, rg_dx8[0]; \ | |
2202 | \ | |
2203 | vadd.u32 r_block, r_block, dx8; \ | |
2204 | vdup.u32 dx8, rg_dx8[1]; \ | |
2205 | \ | |
2206 | vadd.u32 g_block, g_block, dx8; \ | |
2207 | vdup.u32 dx8, b_dx8; \ | |
2208 | \ | |
2209 | vadd.u32 b_block, b_block, dx8; \ | |
2210 | \ | |
2211 | vmovn.u16 r_whole_8, r_whole; \ | |
2212 | vmovn.u16 g_whole_8, g_whole; \ | |
2213 | vmovn.u16 b_whole_8, b_whole; \ | |
2214 | \ | |
2215 | beq 5f; \ | |
2216 | veor.u32 draw_mask, draw_mask, draw_mask; \ | |
2217 | \ | |
2218 | 4: \ | |
2219 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2220 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2221 | \ | |
2222 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
2223 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2224 | \ | |
2225 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2226 | str fb_ptr, [ block_ptr_a, #44 ]; \ | |
2227 | \ | |
2228 | vdup.u32 dx4, rg_dx4[0]; \ | |
2229 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2230 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
2231 | \ | |
2232 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2233 | vdup.u32 dx4, rg_dx4[1]; \ | |
2234 | \ | |
2235 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2236 | vdup.u32 dx4, b_dx4; \ | |
2237 | \ | |
2238 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2239 | vdup.u32 dx8, rg_dx8[0]; \ | |
2240 | \ | |
2241 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
2242 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2243 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2244 | \ | |
2245 | vadd.u32 r_block, r_block, dx8; \ | |
2246 | vdup.u32 dx8, rg_dx8[1]; \ | |
2247 | \ | |
2248 | vadd.u32 g_block, g_block, dx8; \ | |
2249 | vdup.u32 dx8, b_dx8; \ | |
2250 | \ | |
2251 | vadd.u32 b_block, b_block, dx8; \ | |
2252 | add fb_ptr, fb_ptr, #16; \ | |
2253 | \ | |
2254 | vmovn.u16 r_whole_8, r_whole; \ | |
2255 | vmovn.u16 g_whole_8, g_whole; \ | |
2256 | vmovn.u16 b_whole_8, b_whole; \ | |
2257 | \ | |
2258 | vst1.u32 { draw_mask }, [ block_ptr_a, :128 ], c_64; \ | |
2259 | vst1.u32 { pixels }, [ block_ptr_b, :128 ], c_64; \ | |
2260 | \ | |
2261 | pld [ fb_ptr ]; \ | |
2262 | \ | |
2263 | subs span_num_blocks, span_num_blocks, #1; \ | |
2264 | bne 4b; \ | |
2265 | \ | |
2266 | 5: \ | |
2267 | str fb_ptr, [ block_ptr_a, #44 ]; \ | |
2268 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2269 | \ | |
2270 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ | |
2271 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
2272 | \ | |
2273 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2274 | vdup.u8 draw_mask, right_mask; \ | |
2275 | \ | |
2276 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
2277 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
2278 | \ | |
2279 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
2280 | \ | |
2281 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
2282 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2283 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2284 | \ | |
2285 | vst1.u32 { draw_mask }, [ block_ptr_a, :128 ], c_64; \ | |
2286 | vst1.u32 { pixels }, [ block_ptr_b, :128 ], c_64; \ | |
2287 | \ | |
2288 | 1: \ | |
2289 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
2290 | add span_b_offset, span_b_offset, #4; \ | |
2291 | \ | |
2292 | add span_edge_data, span_edge_data, #8; \ | |
2293 | subs num_spans, num_spans, #1; \ | |
2294 | \ | |
2295 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
2296 | bne 0b; \ | |
2297 | \ | |
2298 | ldmia sp!, { r4 - r11, pc }; \ | |
2299 | \ | |
2300 | 2: \ | |
2301 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
2302 | vpush { rg_dx4 }; \ | |
2303 | \ | |
2304 | stmdb sp!, { r0 - r3, r12, r14 }; \ | |
2305 | bl flush_render_block_buffer; \ | |
2306 | ldmia sp!, { r0 - r3, r12, r14 }; \ | |
2307 | \ | |
2308 | vpop { rg_dx4 }; \ | |
2309 | \ | |
2310 | vmov.u8 d64_1, #1; \ | |
2311 | vmov.u8 d128_4, #4; \ | |
2312 | vmov.u8 d64_128, #128; \ | |
2313 | vmov.u8 d128_0x7, #0x7; \ | |
2314 | \ | |
2315 | vadd.u32 rg_dx8, rg_dx4, rg_dx4; \ | |
2316 | \ | |
2317 | mov num_blocks, span_num_blocks; \ | |
2318 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
2319 | bal 3b \ | |
2320 | ||
2321 | ||
2322 | setup_blocks_shaded_untextured_indirect_builder(undithered) | |
2323 | setup_blocks_shaded_untextured_indirect_builder(dithered) | |
2324 | ||
2325 | ||
2326 | #undef draw_mask | |
2327 | ||
2328 | #define mask_msb_ptr r14 | |
2329 | ||
2330 | #define draw_mask q0 | |
2331 | #define pixels_low d16 | |
3867c6ef | 2332 | #define pixels_high d17 |
75e28f62 E |
2333 | |
2334 | ||
2335 | ||
2336 | #define setup_blocks_shaded_untextured_direct_builder(dithering) \ | |
2337 | .align 3; \ | |
2338 | \ | |
2339 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_direct) \ | |
2340 | ldrh num_spans, [ psx_gpu, #psx_gpu_num_spans_offset ]; \ | |
2341 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ | |
2342 | \ | |
2343 | vld1.u32 { rg_dx }, [ rg_dx_ptr, :64 ]; \ | |
2344 | \ | |
2345 | cmp num_spans, #0; \ | |
2346 | bxeq lr; \ | |
2347 | \ | |
2348 | stmdb sp!, { r4 - r11, r14 }; \ | |
2349 | vshl.u32 rg_dx4, rg_dx, #2; \ | |
2350 | \ | |
2351 | ldr b_dx, [ psx_gpu, #psx_gpu_b_dx_offset ]; \ | |
2352 | vshl.u32 rg_dx8, rg_dx, #3; \ | |
2353 | \ | |
2354 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
2355 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
2356 | \ | |
2357 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
2358 | vmov.u8 d64_1, #1; \ | |
2359 | \ | |
2360 | vmov.u8 d128_4, #4; \ | |
2361 | vmov.u8 d64_128, #128; \ | |
2362 | \ | |
2363 | vmov.u8 d128_0x7, #0x7; \ | |
2364 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
2365 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
2366 | \ | |
2367 | 0: \ | |
2368 | ldrh span_num_blocks, [ span_edge_data, #edge_data_num_blocks_offset ]; \ | |
2369 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ | |
2370 | \ | |
2371 | ldrh y, [ span_edge_data, #edge_data_y_offset ]; \ | |
c1817bd9 | 2372 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ |
75e28f62 E |
2373 | \ |
2374 | cmp span_num_blocks, #0; \ | |
2375 | beq 1f; \ | |
2376 | \ | |
2377 | ldrh left_x, [ span_edge_data, #edge_data_left_x_offset ]; \ | |
2378 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
2379 | \ | |
2380 | ldr b, [ span_b_offset ]; \ | |
2381 | vdup.u32 v_left_x, left_x; \ | |
2382 | and y, y, #0x3; \ | |
2383 | \ | |
2384 | ldr dither_row, [ dither_offset_ptr, y, lsl #2 ]; \ | |
2385 | add fb_ptr, fb_ptr, left_x, lsl #1; \ | |
2386 | \ | |
2387 | mla b, b_dx, left_x, b; \ | |
2388 | and dither_shift, left_x, #0x03; \ | |
2389 | \ | |
2390 | vld1.u32 { uvrg }, [ span_uvrg_offset, :128 ]; \ | |
2391 | vshr.u32 rg_dx, rg_dx4, #2; \ | |
2392 | \ | |
2393 | mov dither_shift, dither_shift, lsl #3; \ | |
2394 | vmla.u32 rg, rg_dx, v_left_x; \ | |
2395 | \ | |
2396 | subs span_num_blocks, span_num_blocks, #1; \ | |
2397 | \ | |
2398 | mov dither_row, dither_row, ror dither_shift; \ | |
2399 | mov b_dx4, b_dx, lsl #2; \ | |
2400 | \ | |
2401 | vdup.u32 dither_offsets, dither_row; \ | |
2402 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
2403 | \ | |
2404 | vdup.u32 b_block, b; \ | |
2405 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
2406 | \ | |
2407 | mov b_dx8, b_dx, lsl #3; \ | |
2408 | vdup.u32 r_block, rg[0]; \ | |
2409 | vdup.u32 g_block, rg[1]; \ | |
2410 | \ | |
2411 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
2412 | \ | |
2413 | vadd.u32 r_block, r_block, block_span; \ | |
2414 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]!; \ | |
2415 | \ | |
2416 | vadd.u32 g_block, g_block, block_span; \ | |
2417 | vld1.u32 { block_span }, [ block_span_ptr, :128 ]; \ | |
2418 | \ | |
2419 | vadd.u32 b_block, b_block, block_span; \ | |
2420 | add block_ptr_b, block_ptr_a, #16; \ | |
2421 | \ | |
2422 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2423 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2424 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2425 | vdup.u32 dx4, rg_dx4[0]; \ | |
2426 | \ | |
2427 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2428 | vdup.u32 dx4, rg_dx4[1]; \ | |
2429 | \ | |
2430 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2431 | vdup.u32 dx4, b_dx4; \ | |
2432 | \ | |
2433 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2434 | vdup.u32 dx8, rg_dx8[0]; \ | |
2435 | \ | |
2436 | vadd.u32 r_block, r_block, dx8; \ | |
2437 | vdup.u32 dx8, rg_dx8[1]; \ | |
2438 | \ | |
2439 | vadd.u32 g_block, g_block, dx8; \ | |
2440 | vdup.u32 dx8, b_dx8; \ | |
2441 | \ | |
2442 | vadd.u32 b_block, b_block, dx8; \ | |
2443 | \ | |
2444 | vmovn.u16 r_whole_8, r_whole; \ | |
2445 | vmovn.u16 g_whole_8, g_whole; \ | |
2446 | vmovn.u16 b_whole_8, b_whole; \ | |
2447 | \ | |
2448 | beq 3f; \ | |
2449 | \ | |
2450 | 2: \ | |
2451 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2452 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2453 | \ | |
2454 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
2455 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2456 | \ | |
2457 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2458 | \ | |
2459 | vdup.u32 dx4, rg_dx4[0]; \ | |
2460 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2461 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
2462 | \ | |
2463 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2464 | vdup.u32 dx4, rg_dx4[1]; \ | |
2465 | \ | |
2466 | vmov pixels, msb_mask; \ | |
2467 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2468 | vdup.u32 dx4, b_dx4; \ | |
2469 | \ | |
2470 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2471 | vdup.u32 dx8, rg_dx8[0]; \ | |
2472 | \ | |
2473 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
2474 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2475 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2476 | \ | |
2477 | vadd.u32 r_block, r_block, dx8; \ | |
2478 | vdup.u32 dx8, rg_dx8[1]; \ | |
2479 | \ | |
2480 | vadd.u32 g_block, g_block, dx8; \ | |
2481 | vdup.u32 dx8, b_dx8; \ | |
2482 | \ | |
2483 | vadd.u32 b_block, b_block, dx8; \ | |
2484 | \ | |
2485 | vmovn.u16 r_whole_8, r_whole; \ | |
2486 | vmovn.u16 g_whole_8, g_whole; \ | |
2487 | vmovn.u16 b_whole_8, b_whole; \ | |
2488 | \ | |
2489 | vst1.u32 { pixels }, [ fb_ptr ]!; \ | |
2490 | subs span_num_blocks, span_num_blocks, #1; \ | |
2491 | bne 2b; \ | |
2492 | \ | |
2493 | 3: \ | |
2494 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2495 | \ | |
3867c6ef | 2496 | ldrh right_mask, [ span_edge_data, #edge_data_right_mask_offset ]; \ |
75e28f62 E |
2497 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
2498 | \ | |
2499 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
3867c6ef | 2500 | rbit right_mask, right_mask; \ |
75e28f62 E |
2501 | vmov pixels, msb_mask; \ |
2502 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
3867c6ef | 2503 | clz right_mask, right_mask; \ |
75e28f62 E |
2504 | \ |
2505 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
2506 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2507 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2508 | \ | |
3867c6ef E |
2509 | ldr pc, [ pc, right_mask, lsl #2 ]; \ |
2510 | nop; \ | |
2511 | nop; \ | |
2512 | .word 4f; \ | |
2513 | .word 5f; \ | |
2514 | .word 6f; \ | |
2515 | .word 7f; \ | |
2516 | .word 8f; \ | |
2517 | .word 9f; \ | |
2518 | .word 10f; \ | |
2519 | .word 11f; \ | |
2520 | \ | |
75e28f62 | 2521 | 4: \ |
3867c6ef E |
2522 | vst1.u16 { pixels_low[0] }, [ fb_ptr ]; \ |
2523 | bal 1f; \ | |
2524 | \ | |
2525 | 5: \ | |
2526 | vst1.u32 { pixels_low[0] }, [ fb_ptr ]; \ | |
2527 | bal 1f; \ | |
2528 | \ | |
2529 | 6: \ | |
2530 | vst1.u32 { pixels_low[0] }, [ fb_ptr ]!; \ | |
2531 | vst1.u16 { pixels_low[2] }, [ fb_ptr ]; \ | |
2532 | bal 1f; \ | |
2533 | \ | |
2534 | 7: \ | |
2535 | vst1.u32 { pixels_low }, [ fb_ptr ]; \ | |
2536 | bal 1f; \ | |
2537 | \ | |
2538 | 8: \ | |
2539 | vst1.u32 { pixels_low }, [ fb_ptr ]!; \ | |
2540 | vst1.u16 { pixels_high[0] }, [ fb_ptr ]; \ | |
2541 | bal 1f; \ | |
2542 | \ | |
2543 | 9: \ | |
2544 | vst1.u32 { pixels_low }, [ fb_ptr ]!; \ | |
2545 | vst1.u32 { pixels_high[0] }, [ fb_ptr ]!; \ | |
2546 | bal 1f; \ | |
2547 | \ | |
2548 | 10: \ | |
2549 | vst1.u32 { pixels_low }, [ fb_ptr ]!; \ | |
2550 | vst1.u32 { pixels_high[0] }, [ fb_ptr ]!; \ | |
2551 | vst1.u16 { pixels_high[2] }, [ fb_ptr ]; \ | |
2552 | bal 1f; \ | |
2553 | \ | |
2554 | 11: \ | |
2555 | vst1.u32 { pixels }, [ fb_ptr ]; \ | |
2556 | bal 1f; \ | |
75e28f62 E |
2557 | \ |
2558 | 1: \ | |
2559 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
2560 | add span_b_offset, span_b_offset, #4; \ | |
2561 | \ | |
2562 | add span_edge_data, span_edge_data, #8; \ | |
2563 | subs num_spans, num_spans, #1; \ | |
2564 | \ | |
2565 | bne 0b; \ | |
2566 | \ | |
2567 | ldmia sp!, { r4 - r11, pc } \ | |
2568 | ||
2569 | setup_blocks_shaded_untextured_direct_builder(undithered) | |
2570 | setup_blocks_shaded_untextured_direct_builder(dithered) | |
2571 | ||
2572 | ||
2573 | #undef psx_gpu | |
2574 | #undef num_blocks | |
2575 | #undef triangle | |
2576 | #undef c_64 | |
2577 | ||
2578 | #define psx_gpu r0 | |
2579 | #define block_ptr r1 | |
2580 | #define num_blocks r2 | |
2581 | #define uv_01 r3 | |
2582 | #define uv_23 r4 | |
2583 | #define uv_45 r5 | |
2584 | #define uv_67 r6 | |
2585 | #define uv_0 r7 | |
2586 | #define uv_1 r3 | |
2587 | #define uv_2 r8 | |
2588 | #define uv_3 r4 | |
2589 | #define uv_4 r9 | |
2590 | #define uv_5 r5 | |
2591 | #define uv_6 r10 | |
2592 | #define uv_7 r6 | |
2593 | #define texture_ptr r11 | |
2594 | ||
2595 | #define pixel_0 r7 | |
2596 | #define pixel_1 r3 | |
2597 | #define pixel_2 r8 | |
2598 | #define pixel_3 r4 | |
2599 | #define pixel_4 r9 | |
2600 | #define pixel_5 r5 | |
2601 | #define pixel_6 r10 | |
2602 | #define pixel_7 r6 | |
2603 | ||
2604 | #define pixels_a r7 | |
2605 | #define pixels_b r9 | |
2606 | #define pixels_c r8 | |
2607 | #define pixels_d r10 | |
2608 | ||
2609 | #define c_64 r0 | |
2610 | ||
2611 | #define clut_ptr r12 | |
2612 | #define current_texture_mask r5 | |
2613 | #define dirty_textures_mask r6 | |
2614 | ||
2615 | #define texels d0 | |
2616 | ||
2617 | #define clut_low_a d2 | |
2618 | #define clut_low_b d3 | |
2619 | #define clut_high_a d4 | |
2620 | #define clut_high_b d5 | |
2621 | ||
2622 | #define clut_a q1 | |
2623 | #define clut_b q2 | |
2624 | ||
2625 | #define texels_low d6 | |
2626 | #define texels_high d7 | |
2627 | ||
2628 | .align 3 | |
2629 | ||
2630 | function(texture_blocks_untextured) | |
2631 | bx lr | |
2632 | ||
2633 | ||
2634 | .align 3 | |
2635 | ||
2636 | function(texture_blocks_4bpp) | |
2637 | stmdb sp!, { r3 - r11, r14 } | |
2638 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2639 | ||
2640 | ldr texture_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
2641 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
2642 | ||
2643 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ] | |
2644 | vld1.u32 { clut_a, clut_b }, [ clut_ptr, :128 ] | |
2645 | ||
2646 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ] | |
2647 | vuzp.u8 clut_a, clut_b | |
2648 | ||
2649 | ldr dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ] | |
2650 | tst dirty_textures_mask, current_texture_mask | |
2651 | ||
2652 | bne 1f | |
2653 | mov c_64, #64 | |
2654 | ||
2655 | 0: | |
2656 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
2657 | ||
2658 | uxtah uv_0, texture_ptr, uv_01 | |
2659 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
2660 | ||
2661 | uxtah uv_2, texture_ptr, uv_23 | |
2662 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
2663 | ||
2664 | uxtah uv_4, texture_ptr, uv_45 | |
2665 | ldrb pixel_0, [ uv_0 ] | |
2666 | ||
2667 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
2668 | ldrb pixel_1, [ uv_1 ] | |
2669 | ||
2670 | uxtah uv_6, texture_ptr, uv_67 | |
2671 | ldrb pixel_2, [ uv_2 ] | |
2672 | ||
2673 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
2674 | ldrb pixel_3, [ uv_3 ] | |
2675 | ||
2676 | ldrb pixel_4, [ uv_4 ] | |
2677 | subs num_blocks, num_blocks, #1 | |
2678 | ||
2679 | ldrb pixel_5, [ uv_5 ] | |
2680 | orr pixels_a, pixel_0, pixel_1, lsl #8 | |
2681 | ||
2682 | ldrb pixel_6, [ uv_6 ] | |
2683 | orr pixels_b, pixel_4, pixel_5, lsl #8 | |
2684 | ||
2685 | ldrb pixel_7, [ uv_7 ] | |
2686 | orr pixels_a, pixels_a, pixel_2, lsl #16 | |
2687 | ||
2688 | orr pixels_b, pixels_b, pixel_6, lsl #16 | |
2689 | orr pixels_a, pixels_a, pixel_3, lsl #24 | |
2690 | ||
2691 | orr pixels_b, pixels_b, pixel_7, lsl #24 | |
2692 | vmov.u32 texels, pixels_a, pixels_b | |
2693 | ||
2694 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels | |
2695 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels | |
2696 | ||
2697 | vst2.u8 { texels_low, texels_high }, [ block_ptr, :128 ], c_64 | |
2698 | bne 0b | |
2699 | ||
2700 | ldmia sp!, { r3 - r11, pc } | |
2701 | ||
2702 | 1: | |
2703 | stmdb sp!, { r1 - r2 } | |
2704 | bl update_texture_4bpp_cache | |
2705 | ||
2706 | mov c_64, #64 | |
2707 | ldmia sp!, { r1 - r2 } | |
2708 | bal 0b | |
2709 | ||
2710 | ||
2711 | .align 3 | |
2712 | ||
2713 | function(texture_blocks_8bpp) | |
2714 | stmdb sp!, { r3 - r11, r14 } | |
2715 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2716 | ||
2717 | ldr texture_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
2718 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
2719 | ||
2720 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ] | |
2721 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ] | |
2722 | ||
2723 | ldr dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset ] | |
2724 | tst dirty_textures_mask, current_texture_mask | |
2725 | ||
2726 | bne 1f | |
2727 | nop | |
2728 | ||
2729 | 0: | |
2730 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
2731 | ||
2732 | uxtah uv_0, texture_ptr, uv_01 | |
2733 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
2734 | ||
2735 | uxtah uv_2, texture_ptr, uv_23 | |
2736 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
2737 | ||
2738 | uxtah uv_4, texture_ptr, uv_45 | |
2739 | ldrb pixel_0, [ uv_0 ] | |
2740 | ||
2741 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
2742 | ldrb pixel_1, [ uv_1 ] | |
2743 | ||
2744 | uxtah uv_6, texture_ptr, uv_67 | |
2745 | ldrb pixel_2, [ uv_2 ] | |
2746 | ||
2747 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
2748 | ldrb pixel_3, [ uv_3 ] | |
2749 | ||
2750 | ldrb pixel_4, [ uv_4 ] | |
2751 | add pixel_0, pixel_0, pixel_0 | |
2752 | ||
2753 | ldrb pixel_5, [ uv_5 ] | |
2754 | add pixel_1, pixel_1, pixel_1 | |
2755 | ||
2756 | ldrb pixel_6, [ uv_6 ] | |
2757 | add pixel_2, pixel_2, pixel_2 | |
2758 | ||
2759 | ldrb pixel_7, [ uv_7 ] | |
2760 | add pixel_3, pixel_3, pixel_3 | |
2761 | ||
2762 | ldrh pixel_0, [ clut_ptr, pixel_0 ] | |
2763 | add pixel_4, pixel_4, pixel_4 | |
2764 | ||
2765 | ldrh pixel_1, [ clut_ptr, pixel_1 ] | |
2766 | add pixel_5, pixel_5, pixel_5 | |
2767 | ||
2768 | ldrh pixel_2, [ clut_ptr, pixel_2 ] | |
2769 | add pixel_6, pixel_6, pixel_6 | |
2770 | ||
2771 | ldrh pixel_3, [ clut_ptr, pixel_3 ] | |
2772 | add pixel_7, pixel_7, pixel_7 | |
2773 | ||
2774 | ldrh pixel_4, [ clut_ptr, pixel_4 ] | |
2775 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
2776 | ||
2777 | ldrh pixel_5, [ clut_ptr, pixel_5 ] | |
2778 | orr pixels_c, pixel_2, pixel_3, lsl #16 | |
2779 | ||
2780 | ldrh pixel_6, [ clut_ptr, pixel_6 ] | |
2781 | subs num_blocks, num_blocks, #1 | |
2782 | ||
2783 | ldrh pixel_7, [ clut_ptr, pixel_7 ] | |
2784 | orr pixels_b, pixel_4, pixel_5, lsl #16 | |
2785 | ||
2786 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
2787 | stm block_ptr, { pixels_a, pixels_c, pixels_b, pixels_d } | |
2788 | ||
2789 | add block_ptr, block_ptr, #64 | |
2790 | bne 0b | |
2791 | ||
2792 | ldmia sp!, { r3 - r11, pc } | |
2793 | ||
2794 | 1: | |
2795 | stmdb sp!, { r1 - r2, r12 } | |
2796 | ||
2797 | bl update_texture_8bpp_cache | |
2798 | ||
2799 | ldmia sp!, { r1 - r2, r12 } | |
2800 | bal 0b | |
2801 | ||
2802 | ||
2803 | #undef uv_0 | |
2804 | #undef uv_1 | |
2805 | #undef uv_2 | |
2806 | #undef uv_3 | |
2807 | #undef uv_4 | |
2808 | #undef uv_5 | |
2809 | #undef uv_6 | |
2810 | #undef uv_7 | |
2811 | ||
2812 | #undef pixel_0 | |
2813 | #undef pixel_1 | |
2814 | #undef pixel_2 | |
2815 | #undef pixel_3 | |
2816 | #undef pixel_4 | |
2817 | #undef pixel_5 | |
2818 | #undef pixel_6 | |
2819 | #undef pixel_7 | |
2820 | ||
2821 | #undef texture_ptr | |
2822 | ||
2823 | #undef pixels_a | |
2824 | #undef pixels_b | |
2825 | #undef pixels_c | |
2826 | #undef pixels_d | |
2827 | ||
2828 | #define psx_gpu r0 | |
2829 | #define block_ptr r1 | |
2830 | #define num_blocks r2 | |
2831 | ||
2832 | #define uv_0 r3 | |
2833 | #define uv_1 r4 | |
2834 | #define u_0 r3 | |
2835 | #define u_1 r4 | |
2836 | #define v_0 r5 | |
2837 | #define v_1 r6 | |
2838 | ||
2839 | #define uv_2 r5 | |
2840 | #define uv_3 r6 | |
2841 | #define u_2 r5 | |
2842 | #define u_3 r6 | |
2843 | #define v_2 r7 | |
2844 | #define v_3 r8 | |
2845 | ||
2846 | #define uv_4 r7 | |
2847 | #define uv_5 r8 | |
2848 | #define u_4 r7 | |
2849 | #define u_5 r8 | |
2850 | #define v_4 r9 | |
2851 | #define v_5 r10 | |
2852 | ||
2853 | #define uv_6 r9 | |
2854 | #define uv_7 r10 | |
2855 | #define u_6 r9 | |
2856 | #define u_7 r10 | |
2857 | #define v_6 r11 | |
2858 | #define v_7 r0 | |
2859 | ||
2860 | #define pixel_0 r3 | |
2861 | #define pixel_1 r4 | |
2862 | #define pixel_2 r5 | |
2863 | #define pixel_3 r6 | |
2864 | #define pixel_4 r7 | |
2865 | #define pixel_5 r8 | |
2866 | #define pixel_6 r9 | |
2867 | #define pixel_7 r10 | |
2868 | ||
2869 | #define pixels_a r3 | |
2870 | #define pixels_b r5 | |
2871 | #define pixels_c r7 | |
2872 | #define pixels_d r9 | |
2873 | ||
2874 | #define texture_ptr r12 | |
2875 | ||
2876 | ||
2877 | .align 3 | |
2878 | ||
2879 | function(texture_blocks_16bpp) | |
2880 | stmdb sp!, { r3 - r11, r14 } | |
2881 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2882 | ||
2883 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
2884 | ldr texture_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
2885 | ||
2886 | 0: | |
2887 | ldrh uv_0, [ block_ptr ] | |
2888 | subs num_blocks, num_blocks, #1 | |
2889 | ||
2890 | ldrh uv_1, [ block_ptr, #2 ] | |
2891 | ||
2892 | and v_0, uv_0, #0xFF00 | |
2893 | and v_1, uv_1, #0xFF00 | |
2894 | ||
2895 | and u_0, uv_0, #0xFF | |
2896 | and u_1, uv_1, #0xFF | |
2897 | ||
2898 | add uv_0, u_0, v_0, lsl #2 | |
2899 | ldrh uv_2, [ block_ptr, #4 ] | |
2900 | ||
2901 | add uv_1, u_1, v_1, lsl #2 | |
2902 | ldrh uv_3, [ block_ptr, #6 ] | |
2903 | ||
2904 | add uv_0, uv_0, uv_0 | |
2905 | add uv_1, uv_1, uv_1 | |
2906 | ||
2907 | and v_2, uv_2, #0xFF00 | |
2908 | and v_3, uv_3, #0xFF00 | |
2909 | ||
2910 | and u_2, uv_2, #0xFF | |
2911 | and u_3, uv_3, #0xFF | |
2912 | ||
2913 | add uv_2, u_2, v_2, lsl #2 | |
2914 | ldrh uv_4, [ block_ptr, #8 ] | |
2915 | ||
2916 | add uv_3, u_3, v_3, lsl #2 | |
2917 | ldrh uv_5, [ block_ptr, #10 ] | |
2918 | ||
2919 | add uv_2, uv_2, uv_2 | |
2920 | add uv_3, uv_3, uv_3 | |
2921 | ||
2922 | and v_4, uv_4, #0xFF00 | |
2923 | and v_5, uv_5, #0xFF00 | |
2924 | ||
2925 | and u_4, uv_4, #0xFF | |
2926 | and u_5, uv_5, #0xFF | |
2927 | ||
2928 | add uv_4, u_4, v_4, lsl #2 | |
2929 | ldrh uv_6, [ block_ptr, #12 ] | |
2930 | ||
2931 | add uv_5, u_5, v_5, lsl #2 | |
2932 | ldrh uv_7, [ block_ptr, #14 ] | |
2933 | ||
2934 | add uv_4, uv_4, uv_4 | |
2935 | ldrh pixel_0, [ texture_ptr, uv_0 ] | |
2936 | ||
2937 | add uv_5, uv_5, uv_5 | |
2938 | ldrh pixel_1, [ texture_ptr, uv_1 ] | |
2939 | ||
2940 | and v_6, uv_6, #0xFF00 | |
2941 | ldrh pixel_2, [ texture_ptr, uv_2 ] | |
2942 | ||
2943 | and v_7, uv_7, #0xFF00 | |
2944 | ldrh pixel_3, [ texture_ptr, uv_3 ] | |
2945 | ||
2946 | and u_6, uv_6, #0xFF | |
2947 | ldrh pixel_4, [ texture_ptr, uv_4 ] | |
2948 | ||
2949 | and u_7, uv_7, #0xFF | |
2950 | ldrh pixel_5, [ texture_ptr, uv_5 ] | |
2951 | ||
2952 | add uv_6, u_6, v_6, lsl #2 | |
2953 | add uv_7, u_7, v_7, lsl #2 | |
2954 | ||
2955 | add uv_6, uv_6, uv_6 | |
2956 | add uv_7, uv_7, uv_7 | |
2957 | ||
2958 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
2959 | orr pixels_b, pixel_2, pixel_3, lsl #16 | |
2960 | ||
2961 | ldrh pixel_6, [ texture_ptr, uv_6 ] | |
2962 | orr pixels_c, pixel_4, pixel_5, lsl #16 | |
2963 | ||
2964 | ldrh pixel_7, [ texture_ptr, uv_7 ] | |
2965 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
2966 | ||
2967 | stm block_ptr, { pixels_a, pixels_b, pixels_c, pixels_d } | |
2968 | add block_ptr, block_ptr, #64 | |
2969 | ||
2970 | bne 0b | |
2971 | ||
2972 | ldmia sp!, { r3 - r11, pc } | |
2973 | ||
2974 | ||
2975 | #undef num_blocks | |
2976 | ||
2977 | #undef test_mask | |
2978 | #undef texels | |
2979 | #undef pixels_b | |
2980 | #undef pixels | |
2981 | #undef d64_1 | |
2982 | #undef d64_4 | |
2983 | #undef d64_128 | |
2984 | #undef draw_mask | |
2985 | #undef msb_mask | |
2986 | #undef msb_mask_low | |
2987 | #undef msb_mask_high | |
2988 | #undef fb_pixels | |
2989 | ||
2990 | #undef c_32 | |
2991 | #undef fb_ptr | |
2992 | #undef mask_msb_ptr | |
2993 | ||
2994 | #define psx_gpu r0 | |
2995 | #define num_blocks r1 | |
2996 | #define color_ptr r2 | |
3867c6ef E |
2997 | #define colors_scalar r2 |
2998 | #define colors_scalar_compare r3 | |
75e28f62 E |
2999 | #define mask_msb_ptr r2 |
3000 | ||
3001 | #define block_ptr_load_a r0 | |
3002 | #define block_ptr_store r3 | |
3003 | #define block_ptr_load_b r12 | |
3004 | #define c_32 r2 | |
3005 | ||
3006 | #define c_48 r4 | |
3007 | #define fb_ptr r14 | |
3008 | #define draw_mask_bits_scalar r5 | |
3009 | ||
3010 | #define d128_0x07 q0 | |
3011 | #define d128_0x1F q1 | |
3012 | #define d128_0x8000 q2 | |
3013 | #define test_mask q3 | |
3014 | #define texels q4 | |
3015 | #define colors_rg q5 | |
3016 | #define colors_b_dm_bits q6 | |
3017 | #define texels_rg q7 | |
3018 | #define pixels_r q8 | |
3019 | #define pixels_g q9 | |
3020 | #define pixels_b q10 | |
3021 | #define pixels q11 | |
3022 | #define zero_mask q4 | |
3023 | #define draw_mask q12 | |
3024 | #define msb_mask q13 | |
3025 | ||
3026 | #define fb_pixels q8 | |
3027 | ||
3028 | #define pixels_gb_low q9 | |
3029 | ||
3030 | #define colors_r d10 | |
3031 | #define colors_g d11 | |
3032 | #define colors_b d12 | |
3033 | #define draw_mask_bits d13 | |
3034 | #define texels_r d14 | |
3035 | #define texels_g d15 | |
3036 | #define pixels_r_low d16 | |
3037 | #define pixels_g_low d18 | |
3038 | #define pixels_b_low d19 | |
3039 | #define msb_mask_low d26 | |
3040 | #define msb_mask_high d27 | |
3041 | ||
3042 | #define d64_1 d28 | |
3043 | #define d64_4 d29 | |
3044 | #define d64_128 d30 | |
3045 | #define texels_b d31 | |
3046 | ||
3047 | #define shade_blocks_textured_modulated_prologue_indirect() \ | |
3048 | mov c_48, #48; \ | |
3049 | add block_ptr_store, psx_gpu, #psx_gpu_blocks_offset \ | |
3050 | ||
3051 | #define shade_blocks_textured_modulated_prologue_direct() \ | |
3052 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
3053 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] \ | |
3054 | ||
75e28f62 | 3055 | |
3867c6ef E |
3056 | #define shade_blocks_textured_modulated_prologue_shaded(dithering, target) \ |
3057 | ||
3058 | #define shade_blocks_textured_false_modulation_check_undithered(target) \ | |
3059 | ldr colors_scalar, [ psx_gpu, #psx_gpu_triangle_color_offset ]; \ | |
3060 | movw colors_scalar_compare, #0x8080; \ | |
3061 | \ | |
3062 | movt colors_scalar_compare, #0x80; \ | |
3063 | cmp colors_scalar, colors_scalar_compare; \ | |
3064 | beq shade_blocks_textured_unmodulated_##target \ | |
3065 | ||
3066 | #define shade_blocks_textured_false_modulation_check_dithered(target) \ | |
3067 | ||
3068 | #define shade_blocks_textured_modulated_prologue_unshaded(dithering, target) \ | |
3069 | shade_blocks_textured_false_modulation_check_##dithering(target); \ | |
75e28f62 E |
3070 | add color_ptr, psx_gpu, #psx_gpu_triangle_color_offset; \ |
3071 | vld1.u32 { colors_r[] }, [ color_ptr, :32 ]; \ | |
3072 | vdup.u8 colors_g, colors_r[1]; \ | |
3073 | vdup.u8 colors_b, colors_r[2]; \ | |
3074 | vdup.u8 colors_r, colors_r[0] \ | |
3075 | ||
3076 | ||
3077 | #define shade_blocks_textured_modulated_load_dithered(target) \ | |
3078 | vld1.u32 { target }, [ block_ptr_load_b, :128 ] \ | |
3079 | ||
3080 | #define shade_blocks_textured_modulated_load_last_dithered(target) \ | |
3081 | vld1.u32 { target }, [ block_ptr_load_b, :128 ], c_32 \ | |
3082 | ||
3083 | #define shade_blocks_textured_modulated_load_undithered(target) \ | |
3084 | ||
3085 | #define shade_blocks_textured_modulated_load_last_undithered(target) \ | |
3086 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
3087 | ||
3088 | #define shade_blocks_textured_modulate_dithered(channel) \ | |
3089 | vmlal.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
3090 | ||
3091 | #define shade_blocks_textured_modulate_undithered(channel) \ | |
3092 | vmull.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
3093 | ||
3094 | ||
3095 | #define shade_blocks_textured_modulated_store_draw_mask_indirect(offset) \ | |
3096 | vst1.u32 { draw_mask }, [ block_ptr_store, :128 ]! \ | |
3097 | ||
3098 | #define shade_blocks_textured_modulated_store_draw_mask_direct(offset) \ | |
3099 | ldr fb_ptr, [ block_ptr_load_b, #(offset - 64) ]; \ | |
3100 | vld1.u32 { fb_pixels }, [ fb_ptr ]; \ | |
3101 | vbit.u16 pixels, fb_pixels, draw_mask \ | |
3102 | ||
3103 | #define shade_blocks_textured_modulated_store_pixels_indirect() \ | |
3104 | vst1.u32 { pixels }, [ block_ptr_store, :128 ], c_48 \ | |
3105 | ||
3106 | #define shade_blocks_textured_modulated_store_pixels_direct() \ | |
3107 | vst1.u32 { pixels }, [ fb_ptr ] \ | |
3108 | ||
3109 | ||
3110 | #define shade_blocks_textured_modulated_load_rg_shaded() \ | |
3111 | vld1.u32 { colors_r, colors_g }, [ block_ptr_load_b, :128 ], c_32 \ | |
3112 | ||
3113 | #define shade_blocks_textured_modulated_load_rg_unshaded() \ | |
3114 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
3115 | ||
3116 | #define shade_blocks_textured_modulated_load_bdm_shaded() \ | |
3117 | vld1.u32 { colors_b, draw_mask_bits }, [ block_ptr_load_a, :128 ], c_32 \ | |
3118 | ||
3119 | #define shade_blocks_textured_modulated_load_bdm_unshaded() \ | |
3120 | ldr draw_mask_bits_scalar, [ block_ptr_load_a, #8 ]; \ | |
3121 | add block_ptr_load_a, block_ptr_load_a, #32 \ | |
3122 | ||
3123 | #define shade_blocks_textured_modulated_expand_draw_mask_shaded() \ | |
3124 | vdup.u16 draw_mask, draw_mask_bits[0] \ | |
3125 | ||
3126 | #define shade_blocks_textured_modulated_expand_draw_mask_unshaded() \ | |
3127 | vdup.u16 draw_mask, draw_mask_bits_scalar \ | |
3128 | ||
3129 | ||
3130 | #define shade_blocks_textured_modulated_apply_msb_mask_indirect() \ | |
3131 | ||
3132 | #define shade_blocks_textured_modulated_apply_msb_mask_direct() \ | |
3133 | vorr.u16 pixels, pixels, msb_mask \ | |
3134 | ||
3135 | ||
3136 | #define shade_blocks_textured_modulated_builder(shading, dithering, target) \ | |
3137 | .align 3; \ | |
3138 | \ | |
3139 | function(shade_blocks_##shading##_textured_modulated_##dithering##_##target) \ | |
3867c6ef | 3140 | shade_blocks_textured_modulated_prologue_##shading(dithering, target); \ |
75e28f62 E |
3141 | stmdb sp!, { r4 - r5, lr }; \ |
3142 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
3143 | \ | |
3144 | vld1.u32 { test_mask }, [ psx_gpu, :128 ]; \ | |
3145 | \ | |
3146 | shade_blocks_textured_modulated_prologue_##target(); \ | |
75e28f62 E |
3147 | \ |
3148 | add block_ptr_load_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
3149 | mov c_32, #32; \ | |
3150 | \ | |
3151 | add block_ptr_load_b, block_ptr_load_a, #16; \ | |
3152 | vmov.u8 d64_1, #1; \ | |
3153 | vmov.u8 d64_4, #4; \ | |
3154 | vmov.u8 d64_128, #128; \ | |
3155 | \ | |
3156 | vld1.u32 { texels }, [ block_ptr_load_a, :128 ], c_32; \ | |
3157 | vmov.u8 d128_0x07, #0x07; \ | |
3158 | \ | |
3159 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
3160 | vmov.u8 d128_0x1F, #0x1F; \ | |
3161 | \ | |
3162 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
3163 | vmov.u16 d128_0x8000, #0x8000; \ | |
3164 | \ | |
3165 | vmovn.u16 texels_r, texels; \ | |
3166 | vshrn.u16 texels_g, texels, #5; \ | |
3167 | \ | |
3168 | vshrn.u16 texels_b, texels, #7; \ | |
3169 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
3170 | \ | |
3171 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
3172 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
3173 | \ | |
3174 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
3175 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
3176 | \ | |
3177 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
3178 | vshr.u8 texels_b, texels_b, #3; \ | |
3179 | \ | |
3180 | shade_blocks_textured_modulate_##dithering(r); \ | |
3181 | shade_blocks_textured_modulate_##dithering(g); \ | |
3182 | shade_blocks_textured_modulate_##dithering(b); \ | |
3183 | \ | |
3184 | vand.u16 pixels, texels, d128_0x8000; \ | |
3185 | vceq.u16 zero_mask, texels, #0; \ | |
3186 | \ | |
3187 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
3188 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
3189 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
3190 | \ | |
3191 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
3192 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
3193 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
3194 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
3195 | \ | |
3196 | subs num_blocks, num_blocks, #1; \ | |
3197 | beq 1f; \ | |
3198 | \ | |
3199 | .align 3; \ | |
3200 | \ | |
3201 | 0: \ | |
3202 | vld1.u32 { texels }, [ block_ptr_load_a, :128 ], c_32; \ | |
3203 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
3204 | vshrn.u16 texels_g, texels, #5; \ | |
3205 | \ | |
3206 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
3207 | vshrn.u16 texels_b, texels, #7; \ | |
3208 | \ | |
59d15d23 | 3209 | pld [ block_ptr_load_a ]; \ |
75e28f62 E |
3210 | vmovn.u16 texels_r, texels; \ |
3211 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
3212 | \ | |
3213 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
3214 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
3215 | shade_blocks_textured_modulated_store_draw_mask_##target(-4); \ | |
3216 | \ | |
3217 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
3218 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
3219 | \ | |
3220 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
3221 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
3222 | \ | |
3223 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
3224 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
3225 | \ | |
3226 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
3227 | vshr.u8 texels_b, texels_b, #3; \ | |
3228 | \ | |
3229 | shade_blocks_textured_modulate_##dithering(r); \ | |
3230 | shade_blocks_textured_modulate_##dithering(g); \ | |
3231 | shade_blocks_textured_modulate_##dithering(b); \ | |
3232 | \ | |
3233 | vand.u16 pixels, texels, d128_0x8000; \ | |
3234 | vceq.u16 zero_mask, texels, #0; \ | |
3235 | \ | |
3236 | subs num_blocks, num_blocks, #1; \ | |
3237 | \ | |
3238 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
3239 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
3240 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
3241 | \ | |
3242 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
3243 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
3244 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
3245 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
3246 | \ | |
3247 | bne 0b; \ | |
3248 | \ | |
3249 | 1: \ | |
3250 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
3251 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
3252 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
3253 | \ | |
3254 | shade_blocks_textured_modulated_store_draw_mask_##target(28); \ | |
3255 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
3256 | \ | |
3257 | ldmia sp!, { r4 - r5, pc } \ | |
3258 | ||
3259 | ||
3260 | shade_blocks_textured_modulated_builder(shaded, dithered, direct); | |
3261 | shade_blocks_textured_modulated_builder(shaded, undithered, direct); | |
3262 | shade_blocks_textured_modulated_builder(unshaded, dithered, direct); | |
3263 | shade_blocks_textured_modulated_builder(unshaded, undithered, direct); | |
3264 | ||
3265 | shade_blocks_textured_modulated_builder(shaded, dithered, indirect); | |
3266 | shade_blocks_textured_modulated_builder(shaded, undithered, indirect); | |
3267 | shade_blocks_textured_modulated_builder(unshaded, dithered, indirect); | |
3268 | shade_blocks_textured_modulated_builder(unshaded, undithered, indirect); | |
3269 | ||
3270 | ||
3271 | #undef c_64 | |
3272 | #undef fb_ptr | |
3273 | #undef color_ptr | |
3274 | ||
3275 | #undef color_r | |
3276 | #undef color_g | |
3277 | #undef color_b | |
3278 | ||
3279 | #undef test_mask | |
3280 | #undef pixels | |
3281 | #undef draw_mask | |
3282 | #undef zero_mask | |
3283 | #undef fb_pixels | |
3284 | #undef msb_mask | |
3285 | #undef msb_mask_low | |
3286 | #undef msb_mask_high | |
3287 | ||
3288 | #define psx_gpu r0 | |
3289 | #define num_blocks r1 | |
3290 | #define mask_msb_ptr r2 | |
3291 | #define color_ptr r3 | |
3292 | ||
3293 | #define block_ptr_load r0 | |
3294 | #define draw_mask_store_ptr r3 | |
3295 | #define draw_mask_bits_ptr r12 | |
3296 | #define draw_mask_ptr r12 | |
3297 | #define pixel_store_ptr r14 | |
3298 | ||
3299 | #define fb_ptr_cmp r4 | |
3300 | ||
3301 | #define fb_ptr r3 | |
3302 | #define fb_ptr_next r14 | |
3303 | ||
3304 | #define c_64 r2 | |
3305 | ||
3306 | #define test_mask q0 | |
3307 | #define pixels q1 | |
3308 | #define draw_mask q2 | |
3309 | #define zero_mask q3 | |
3310 | #define draw_mask_combined q4 | |
3311 | #define fb_pixels q5 | |
3312 | #define fb_pixels_next q6 | |
3313 | #define msb_mask q7 | |
3314 | ||
3315 | #define draw_mask_low d4 | |
3316 | #define draw_mask_high d5 | |
3317 | #define msb_mask_low d14 | |
3318 | #define msb_mask_high d15 | |
3319 | ||
3320 | .align 3 | |
3321 | function(shade_blocks_textured_unmodulated_indirect) | |
3322 | str r14, [ sp, #-4 ] | |
3323 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) | |
3324 | ||
3325 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
3326 | add pixel_store_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
3327 | ||
3328 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
3329 | add draw_mask_store_ptr, psx_gpu, #psx_gpu_blocks_offset | |
3330 | ||
3331 | mov c_64, #64 | |
3332 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
3333 | ||
3334 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 | |
3335 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
3336 | [ draw_mask_bits_ptr, :16 ], c_64 | |
3337 | vceq.u16 zero_mask, pixels, #0 | |
3338 | ||
3339 | vtst.u16 draw_mask, draw_mask, test_mask | |
3340 | vst1.u32 { pixels }, [ pixel_store_ptr, :128 ], c_64 | |
3341 | ||
3342 | subs num_blocks, num_blocks, #1 | |
3343 | beq 1f | |
3344 | ||
3345 | 0: | |
3346 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 | |
3347 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3348 | ||
3349 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
3350 | [ draw_mask_bits_ptr, :16 ], c_64 | |
3351 | vceq.u16 zero_mask, pixels, #0 | |
3352 | ||
3353 | vtst.u16 draw_mask, draw_mask, test_mask | |
3354 | vst1.u32 { pixels }, [ pixel_store_ptr, :128 ], c_64 | |
3355 | ||
3356 | vst1.u32 { draw_mask_combined }, [ draw_mask_store_ptr, :128 ], c_64 | |
3357 | subs num_blocks, num_blocks, #1 | |
3358 | ||
3359 | bne 0b | |
3360 | ||
3361 | 1: | |
3362 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3363 | vst1.u32 { draw_mask_combined }, [ draw_mask_store_ptr, :128 ], c_64 | |
3364 | ||
3365 | ldr pc, [ sp, #-4 ] | |
3366 | ||
3367 | ||
3368 | .align 3 | |
3369 | ||
3370 | function(shade_blocks_textured_unmodulated_direct) | |
3371 | stmdb sp!, { r4, r14 } | |
3372 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) | |
3373 | ||
3374 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
3375 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset | |
3376 | ||
3377 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] | |
3378 | mov c_64, #64 | |
3379 | ||
3380 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
3381 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
3382 | ||
3383 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
3384 | [ draw_mask_bits_ptr, :16 ], c_64 | |
3385 | ldr fb_ptr_next, [ block_ptr_load, #44 ] | |
3386 | ||
3387 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 | |
3388 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3389 | vceq.u16 zero_mask, pixels, #0 | |
3390 | vtst.u16 draw_mask, draw_mask, test_mask | |
3391 | ||
3392 | subs num_blocks, num_blocks, #1 | |
3393 | beq 1f | |
3394 | ||
3395 | 0: | |
3396 | mov fb_ptr, fb_ptr_next | |
3397 | ldr fb_ptr_next, [ block_ptr_load, #44 ] | |
3398 | ||
3399 | vorr.u16 pixels, pixels, msb_mask | |
3400 | ||
3401 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3402 | vmov fb_pixels, fb_pixels_next | |
3403 | ||
3404 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
3405 | [ draw_mask_bits_ptr, :16 ], c_64 | |
3406 | vbif.u16 fb_pixels, pixels, draw_mask_combined | |
3407 | ||
75e28f62 | 3408 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr |
8438c3c7 | 3409 | pld [ fb_ptr_next, #64 ] |
3410 | ||
75e28f62 | 3411 | add fb_ptr_cmp, fb_ptr_cmp, #14 |
8438c3c7 | 3412 | vld1.u32 { pixels }, [ block_ptr_load, :128 ], c_64 |
3413 | ||
75e28f62 E |
3414 | cmp fb_ptr_cmp, #28 |
3415 | bls 4f | |
3416 | ||
3417 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3418 | vceq.u16 zero_mask, pixels, #0 | |
3419 | ||
3420 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
3421 | vtst.u16 draw_mask, draw_mask, test_mask | |
3422 | ||
3423 | 3: | |
3424 | subs num_blocks, num_blocks, #1 | |
3425 | bne 0b | |
3426 | ||
3427 | 1: | |
3428 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3429 | vbif.u16 fb_pixels_next, pixels, draw_mask_combined | |
3430 | ||
3431 | vst1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3432 | ||
3433 | ldmia sp!, { r4, pc } | |
3434 | ||
3435 | 4: | |
3436 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
3437 | vceq.u16 zero_mask, pixels, #0 | |
3438 | ||
3439 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3440 | vtst.u16 draw_mask, draw_mask, test_mask | |
3441 | ||
3442 | bal 3b | |
3443 | ||
3444 | ||
3445 | function(shade_blocks_unshaded_untextured_indirect) | |
3446 | bx lr | |
3447 | ||
3448 | .align 3 | |
3449 | ||
3450 | function(shade_blocks_unshaded_untextured_direct) | |
3451 | stmdb sp!, { r4, r14 } | |
3452 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
3453 | ||
3454 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
3455 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset | |
3456 | ||
3457 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] | |
3458 | add color_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
3459 | ||
3460 | add block_ptr_load, psx_gpu, #(psx_gpu_blocks_offset + 44) | |
3461 | vld1.u16 { pixels }, [ color_ptr, :128 ] | |
3462 | ||
3463 | mov c_64, #64 | |
3464 | vld1.u16 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
3465 | ||
3466 | vorr.u16 pixels, pixels, msb_mask | |
3467 | subs num_blocks, num_blocks, #1 | |
3468 | ||
3469 | ldr fb_ptr_next, [ block_ptr_load ], #64 | |
3470 | ||
3471 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3472 | beq 1f | |
3473 | ||
3474 | 0: | |
3475 | vmov fb_pixels, fb_pixels_next | |
3476 | mov fb_ptr, fb_ptr_next | |
3477 | ldr fb_ptr_next, [ block_ptr_load ], #64 | |
3478 | ||
3479 | vbif.u16 fb_pixels, pixels, draw_mask | |
3480 | vld1.u16 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
3481 | ||
3482 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr | |
3483 | add fb_ptr_cmp, fb_ptr_cmp, #14 | |
3484 | cmp fb_ptr_cmp, #28 | |
3485 | bls 4f | |
3486 | ||
3487 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3488 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
3489 | ||
3490 | 3: | |
3491 | subs num_blocks, num_blocks, #1 | |
3492 | bne 0b | |
3493 | ||
3494 | 1: | |
3495 | vbif.u16 fb_pixels_next, pixels, draw_mask | |
3496 | vst1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3497 | ||
3498 | ldmia sp!, { r4, pc } | |
3499 | ||
3500 | 4: | |
3501 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
3502 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ] | |
3503 | bal 3b | |
3504 | ||
3505 | ||
3506 | #undef draw_mask_ptr | |
3507 | #undef c_64 | |
3508 | #undef fb_ptr | |
3509 | #undef fb_ptr_next | |
3510 | #undef fb_ptr_cmp | |
3511 | ||
3512 | #define psx_gpu r0 | |
3513 | #define num_blocks r1 | |
3514 | #define msb_mask_ptr r2 | |
3515 | #define pixel_ptr r3 | |
3516 | #define draw_mask_ptr r0 | |
3517 | #define c_64 r2 | |
3518 | #define fb_ptr r12 | |
3519 | #define fb_ptr_next r14 | |
3520 | #define fb_ptr_cmp r4 | |
3521 | ||
3522 | #undef msb_mask | |
3523 | #undef draw_mask | |
3524 | #undef pixels | |
3525 | #undef fb_pixels | |
3526 | #undef d128_0x8000 | |
3527 | #undef msb_mask_low | |
3528 | #undef msb_mask_high | |
3529 | #undef draw_mask_next | |
3530 | #undef pixels_g | |
3531 | #undef blend_pixels | |
3532 | #undef fb_pixels_next | |
3533 | ||
3534 | #define msb_mask q0 | |
3535 | #define draw_mask q1 | |
3536 | #define pixels q2 | |
3537 | #define fb_pixels q3 | |
3538 | #define blend_pixels q4 | |
3539 | #define pixels_no_msb q5 | |
3540 | #define blend_mask q6 | |
3541 | #define fb_pixels_no_msb q7 | |
3542 | #define d128_0x8000 q8 | |
3543 | #define d128_0x0421 q9 | |
3544 | #define fb_pixels_next q10 | |
3545 | #define blend_pixels_next q11 | |
3546 | #define pixels_next q12 | |
3547 | #define draw_mask_next q13 | |
3548 | #define write_mask q14 | |
3549 | ||
3550 | #define pixels_rb q5 | |
3551 | #define pixels_mg q7 | |
3552 | #define pixels_g q7 | |
3553 | #define d128_0x7C1F q8 | |
3554 | #define d128_0x03E0 q9 | |
3555 | #define fb_pixels_rb q10 | |
3556 | #define fb_pixels_g q11 | |
3557 | #define fb_pixels_masked q11 | |
3558 | #define d128_0x83E0 q15 | |
3559 | #define pixels_fourth q7 | |
3560 | #define d128_0x1C07 q12 | |
3561 | #define d128_0x00E0 q13 | |
3562 | #define d128_0x80E0 q13 | |
3563 | ||
3564 | #define msb_mask_low d0 | |
3565 | #define msb_mask_high d1 | |
3566 | ||
3567 | #define blend_blocks_average_set_blend_mask_textured(source) \ | |
3568 | vclt.s16 blend_mask, source, #0 \ | |
3569 | ||
3570 | #define blend_blocks_average_set_stp_bit_textured() \ | |
3571 | vorr.u16 blend_pixels, #0x8000 \ | |
3572 | ||
3573 | #define blend_blocks_average_combine_textured(source) \ | |
3574 | vbif.u16 blend_pixels, source, blend_mask \ | |
3575 | ||
3576 | #define blend_blocks_average_set_blend_mask_untextured(source) \ | |
3577 | ||
3578 | #define blend_blocks_average_set_stp_bit_untextured() \ | |
3579 | ||
3580 | #define blend_blocks_average_combine_untextured(source) \ | |
3581 | ||
3582 | #define blend_blocks_average_mask_set_on() \ | |
3583 | vclt.s16 write_mask, fb_pixels_next, #0 \ | |
3584 | ||
3585 | #define blend_blocks_average_mask_copy_on() \ | |
3586 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
3587 | ||
3588 | #define blend_blocks_average_mask_copy_b_on() \ | |
3589 | vorr.u16 draw_mask_next, draw_mask_next, write_mask \ | |
3590 | ||
3591 | #define blend_blocks_average_mask_set_off() \ | |
3592 | ||
3593 | #define blend_blocks_average_mask_copy_off() \ | |
3594 | vmov draw_mask, draw_mask_next \ | |
3595 | ||
3596 | #define blend_blocks_average_mask_copy_b_off() \ | |
3597 | ||
3598 | #define blend_blocks_average_builder(texturing, mask_evaluate) \ | |
3599 | .align 3; \ | |
3600 | \ | |
3601 | function(blend_blocks_##texturing##_average_##mask_evaluate) \ | |
3602 | stmdb sp!, { r4, r14 }; \ | |
3603 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
3604 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
3605 | \ | |
3606 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
3607 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
3608 | \ | |
3609 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3610 | mov c_64, #64; \ | |
3611 | \ | |
3612 | vmov.u16 d128_0x8000, #0x8000; \ | |
3613 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
3614 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3615 | \ | |
3616 | vmov.u16 d128_0x0421, #0x0400; \ | |
3617 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
3618 | \ | |
3619 | vorr.u16 d128_0x0421, #0x0021; \ | |
3620 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
3621 | \ | |
3622 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
3623 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3624 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3625 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3626 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
3627 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3628 | \ | |
3629 | subs num_blocks, num_blocks, #1; \ | |
3630 | beq 1f; \ | |
3631 | \ | |
3632 | 0: \ | |
3633 | mov fb_ptr, fb_ptr_next; \ | |
3634 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3635 | \ | |
3636 | vmov pixels, pixels_next; \ | |
3637 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
3638 | \ | |
3639 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
3640 | \ | |
3641 | blend_blocks_average_mask_copy_##mask_evaluate(); \ | |
3642 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
3643 | \ | |
3644 | blend_blocks_average_set_blend_mask_##texturing(pixels); \ | |
3645 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
3646 | vmov fb_pixels, fb_pixels_next; \ | |
3647 | blend_blocks_average_combine_##texturing(pixels); \ | |
3648 | \ | |
3649 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
3650 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
3651 | cmp fb_ptr_cmp, #28; \ | |
3652 | bls 2f; \ | |
3653 | \ | |
3654 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
3655 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
3656 | \ | |
3657 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3658 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3659 | \ | |
3660 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3661 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
3662 | \ | |
3663 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3664 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3665 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
3666 | vst1.u16 { fb_pixels }, [ fb_ptr ]; \ | |
3667 | \ | |
3668 | 3: \ | |
3669 | subs num_blocks, num_blocks, #1; \ | |
3670 | bne 0b; \ | |
3671 | \ | |
3672 | 1: \ | |
3673 | blend_blocks_average_mask_copy_b_##mask_evaluate(); \ | |
3674 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
3675 | \ | |
3676 | blend_blocks_average_set_blend_mask_##texturing(pixels_next); \ | |
3677 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
3678 | blend_blocks_average_combine_##texturing(pixels_next); \ | |
3679 | \ | |
3680 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3681 | vbif.u16 fb_pixels_next, blend_pixels, draw_mask_next; \ | |
3682 | vst1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
3683 | \ | |
3684 | ldmia sp!, { r4, pc }; \ | |
3685 | \ | |
3686 | 2: \ | |
3687 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3688 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
3689 | vst1.u16 { fb_pixels }, [ fb_ptr ]; \ | |
3690 | \ | |
3691 | vld1.u16 { fb_pixels_next }, [ fb_ptr_next ]; \ | |
3692 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
3693 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3694 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3695 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3696 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3697 | \ | |
3698 | bal 3b \ | |
3699 | ||
3700 | blend_blocks_average_builder(textured, off) | |
3701 | blend_blocks_average_builder(untextured, off) | |
3702 | blend_blocks_average_builder(textured, on) | |
3703 | blend_blocks_average_builder(untextured, on) | |
3704 | ||
3705 | ||
3706 | #define blend_blocks_add_mask_set_on() \ | |
3707 | vclt.s16 write_mask, fb_pixels, #0 \ | |
3708 | ||
3709 | #define blend_blocks_add_mask_copy_on() \ | |
3710 | vorr.u16 draw_mask, draw_mask, write_mask \ | |
3711 | ||
3712 | #define blend_blocks_add_mask_set_off() \ | |
3713 | ||
3714 | #define blend_blocks_add_mask_copy_off() \ | |
3715 | ||
3716 | ||
3717 | #define blend_blocks_add_textured_builder(mask_evaluate) \ | |
3718 | .align 3; \ | |
3719 | \ | |
3720 | function(blend_blocks_textured_add_##mask_evaluate) \ | |
3721 | stmdb sp!, { r4, r14 }; \ | |
3722 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
3723 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
3724 | \ | |
3725 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
3726 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
3727 | \ | |
3728 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3729 | mov c_64, #64; \ | |
3730 | \ | |
3731 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
3732 | vmov.u16 d128_0x03E0, #0x0300; \ | |
3733 | vmov.u16 d128_0x83E0, #0x8000; \ | |
3734 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
3735 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
3736 | vorr.u16 d128_0x83E0, d128_0x83E0, d128_0x03E0; \ | |
3737 | \ | |
3738 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
3739 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3740 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
3741 | vclt.s16 blend_mask, pixels, #0; \ | |
3742 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
3743 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
3744 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3745 | \ | |
3746 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3747 | vorr.u16 pixels, pixels, msb_mask; \ | |
3748 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3749 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
3750 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
3751 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
3752 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3753 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
3754 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3755 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
3756 | \ | |
3757 | subs num_blocks, num_blocks, #1; \ | |
3758 | beq 1f; \ | |
3759 | \ | |
3760 | 0: \ | |
3761 | mov fb_ptr, fb_ptr_next; \ | |
3762 | \ | |
3763 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3764 | \ | |
3765 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
3766 | vclt.s16 blend_mask, pixels, #0; \ | |
3767 | \ | |
3768 | vorr.u16 pixels, pixels, msb_mask; \ | |
3769 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3770 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
3771 | \ | |
8438c3c7 | 3772 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ |
3773 | pld [ fb_ptr_next, #64 ]; \ | |
75e28f62 E |
3774 | \ |
3775 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
8438c3c7 | 3776 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
3777 | \ | |
75e28f62 | 3778 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ |
8438c3c7 | 3779 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ |
3780 | \ | |
75e28f62 E |
3781 | cmp fb_ptr_cmp, #28; \ |
3782 | bls 2f; \ | |
3783 | \ | |
3784 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
3785 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
3786 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3787 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3788 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3789 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
3790 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
3791 | \ | |
3792 | 3: \ | |
3793 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
3794 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3795 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
3796 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3797 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
3798 | \ | |
3799 | subs num_blocks, num_blocks, #1; \ | |
3800 | bne 0b; \ | |
3801 | \ | |
3802 | 1: \ | |
3803 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3804 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
3805 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
3806 | \ | |
3807 | ldmia sp!, { r4, pc }; \ | |
3808 | \ | |
3809 | 2: \ | |
3810 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
3811 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3812 | \ | |
3813 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
3814 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
3815 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3816 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3817 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
3818 | bal 3b \ | |
3819 | ||
3820 | ||
3821 | #define blend_blocks_add_untextured_builder(mask_evaluate) \ | |
3822 | .align 3; \ | |
3823 | \ | |
3824 | function(blend_blocks_untextured_add_##mask_evaluate) \ | |
3825 | stmdb sp!, { r4, r14 }; \ | |
3826 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
3827 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
3828 | \ | |
3829 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
3830 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
3831 | \ | |
3832 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3833 | mov c_64, #64; \ | |
3834 | \ | |
3835 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
3836 | vmov.u16 d128_0x03E0, #0x0300; \ | |
3837 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
3838 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
3839 | \ | |
3840 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
3841 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3842 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
3843 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
3844 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
3845 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3846 | \ | |
3847 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3848 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
3849 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3850 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
3851 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3852 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
3853 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3854 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
3855 | \ | |
3856 | subs num_blocks, num_blocks, #1; \ | |
3857 | beq 1f; \ | |
3858 | \ | |
3859 | 0: \ | |
3860 | mov fb_ptr, fb_ptr_next; \ | |
3861 | \ | |
3862 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3863 | \ | |
3864 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
3865 | \ | |
3866 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3867 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3868 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
3869 | \ | |
3870 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
3871 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
3872 | \ | |
3873 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
3874 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
3875 | cmp fb_ptr_cmp, #28; \ | |
3876 | bls 2f; \ | |
3877 | \ | |
3878 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
3879 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
3880 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3881 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3882 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3883 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
3884 | \ | |
3885 | 3: \ | |
3886 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
3887 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3888 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
3889 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3890 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
3891 | \ | |
3892 | subs num_blocks, num_blocks, #1; \ | |
3893 | bne 0b; \ | |
3894 | \ | |
3895 | 1: \ | |
3896 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3897 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3898 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
3899 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
3900 | \ | |
3901 | ldmia sp!, { r4, pc }; \ | |
3902 | \ | |
3903 | 2: \ | |
3904 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
3905 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3906 | \ | |
3907 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
3908 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
3909 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3910 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3911 | bal 3b \ | |
3912 | ||
3913 | ||
3914 | blend_blocks_add_textured_builder(off) | |
3915 | blend_blocks_add_textured_builder(on) | |
3916 | blend_blocks_add_untextured_builder(off) | |
3917 | blend_blocks_add_untextured_builder(on) | |
3918 | ||
3919 | #define blend_blocks_subtract_set_blend_mask_textured() \ | |
3920 | vclt.s16 blend_mask, pixels_next, #0 \ | |
3921 | ||
3922 | #define blend_blocks_subtract_combine_textured() \ | |
3923 | vbif.u16 blend_pixels, pixels, blend_mask \ | |
3924 | ||
3925 | #define blend_blocks_subtract_set_stb_textured() \ | |
3926 | vorr.u16 blend_pixels, #0x8000 \ | |
3927 | ||
3928 | #define blend_blocks_subtract_msb_mask_textured() \ | |
3929 | vorr.u16 pixels, pixels_next, msb_mask \ | |
3930 | ||
3931 | #define blend_blocks_subtract_set_blend_mask_untextured() \ | |
3932 | ||
3933 | #define blend_blocks_subtract_combine_untextured() \ | |
3934 | ||
3935 | #define blend_blocks_subtract_set_stb_untextured() \ | |
3936 | vorr.u16 blend_pixels, blend_pixels, msb_mask \ | |
3937 | ||
3938 | #define blend_blocks_subtract_msb_mask_untextured() \ | |
3939 | ||
3940 | ||
3941 | #define blend_blocks_subtract_mask_set_on() \ | |
3942 | vclt.s16 write_mask, fb_pixels, #0 \ | |
3943 | ||
3944 | #define blend_blocks_subtract_mask_copy_on() \ | |
3945 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
3946 | ||
3947 | #define blend_blocks_subtract_mask_set_off() \ | |
3948 | ||
3949 | #define blend_blocks_subtract_mask_copy_off() \ | |
3950 | vmov draw_mask, draw_mask_next \ | |
3951 | ||
3952 | ||
3953 | #define blend_blocks_subtract_builder(texturing, mask_evaluate) \ | |
3954 | .align 3; \ | |
3955 | \ | |
3956 | function(blend_blocks_##texturing##_subtract_##mask_evaluate) \ | |
3957 | stmdb sp!, { r4, r14 }; \ | |
3958 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
3959 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
3960 | \ | |
3961 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
3962 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
3963 | \ | |
3964 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3965 | mov c_64, #64; \ | |
3966 | \ | |
3967 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
3968 | vmov.u16 d128_0x03E0, #0x0300; \ | |
3969 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
3970 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
3971 | \ | |
3972 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
3973 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3974 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
3975 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
3976 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
3977 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ | |
3978 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
3979 | \ | |
3980 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
3981 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3982 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
3983 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3984 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
3985 | \ | |
3986 | subs num_blocks, num_blocks, #1; \ | |
3987 | beq 1f; \ | |
3988 | \ | |
3989 | 0: \ | |
3990 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
3991 | mov fb_ptr, fb_ptr_next; \ | |
3992 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
3993 | \ | |
3994 | vld1.u32 { draw_mask_next }, [ draw_mask_ptr, :128 ], c_64; \ | |
3995 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
3996 | \ | |
3997 | vld1.u32 { pixels_next }, [ pixel_ptr, :128 ], c_64; \ | |
3998 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3999 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
4000 | blend_blocks_subtract_set_stb_##texturing(); \ | |
4001 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
4002 | blend_blocks_subtract_combine_##texturing(); \ | |
4003 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
4004 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
4005 | \ | |
4006 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4007 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4008 | cmp fb_ptr_cmp, #28; \ | |
4009 | bls 2f; \ | |
4010 | \ | |
4011 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4012 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ | |
4013 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4014 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4015 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4016 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
4017 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4018 | \ | |
4019 | 3: \ | |
4020 | subs num_blocks, num_blocks, #1; \ | |
4021 | bne 0b; \ | |
4022 | \ | |
4023 | 1: \ | |
4024 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
4025 | \ | |
4026 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
4027 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
4028 | blend_blocks_subtract_set_stb_##texturing(); \ | |
4029 | blend_blocks_subtract_combine_##texturing(); \ | |
4030 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
4031 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
4032 | \ | |
4033 | ldmia sp!, { r4, pc }; \ | |
4034 | \ | |
4035 | 2: \ | |
4036 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
4037 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4038 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ | |
4039 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4040 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4041 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4042 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4043 | bal 3b \ | |
4044 | ||
4045 | ||
4046 | blend_blocks_subtract_builder(textured, off) | |
4047 | blend_blocks_subtract_builder(textured, on) | |
4048 | blend_blocks_subtract_builder(untextured, off) | |
4049 | blend_blocks_subtract_builder(untextured, on) | |
4050 | ||
4051 | ||
4052 | #define blend_blocks_add_fourth_textured_builder(mask_evaluate) \ | |
4053 | .align 3; \ | |
4054 | \ | |
4055 | function(blend_blocks_textured_add_fourth_##mask_evaluate) \ | |
4056 | stmdb sp!, { r4, r14 }; \ | |
4057 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
4058 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
4059 | \ | |
4060 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
4061 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
4062 | \ | |
4063 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4064 | mov c_64, #64; \ | |
4065 | \ | |
4066 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4067 | vmov.u16 d128_0x03E0, #0x0300; \ | |
75e28f62 | 4068 | vmov.u16 d128_0x1C07, #0x1C00; \ |
d1c75d1e | 4069 | vmov.u16 d128_0x00E0, #0x00E0; \ |
75e28f62 E |
4070 | vorr.u16 d128_0x7C1F, #0x001F; \ |
4071 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
75e28f62 | 4072 | vorr.u16 d128_0x1C07, #0x0007; \ |
75e28f62 E |
4073 | \ |
4074 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
4075 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
4076 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
4077 | vclt.s16 blend_mask, pixels, #0; \ | |
4078 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4079 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
4080 | vshr.s16 pixels_fourth, pixels, #2; \ | |
d1c75d1e | 4081 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
75e28f62 E |
4082 | \ |
4083 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
d1c75d1e E |
4084 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4085 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4086 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
75e28f62 | 4087 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
d1c75d1e | 4088 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
75e28f62 | 4089 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
d1c75d1e | 4090 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
75e28f62 E |
4091 | \ |
4092 | subs num_blocks, num_blocks, #1; \ | |
4093 | beq 1f; \ | |
4094 | \ | |
4095 | 0: \ | |
4096 | mov fb_ptr, fb_ptr_next; \ | |
75e28f62 E |
4097 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ |
4098 | \ | |
d1c75d1e E |
4099 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
4100 | vbif.u16 blend_pixels, pixels, blend_mask; \ | |
4101 | \ | |
75e28f62 E |
4102 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ |
4103 | vclt.s16 blend_mask, pixels, #0; \ | |
75e28f62 | 4104 | vshr.s16 pixels_fourth, pixels, #2; \ |
d1c75d1e | 4105 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
75e28f62 E |
4106 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
4107 | \ | |
4108 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
4109 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
4110 | \ | |
4111 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4112 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4113 | cmp fb_ptr_cmp, #28; \ | |
4114 | bls 2f; \ | |
4115 | \ | |
4116 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4117 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
75e28f62 | 4118 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
d1c75d1e E |
4119 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4120 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
75e28f62 E |
4121 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ |
4122 | \ | |
4123 | 3: \ | |
d1c75d1e | 4124 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ |
75e28f62 | 4125 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
d1c75d1e | 4126 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
75e28f62 | 4127 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
d1c75d1e | 4128 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
75e28f62 E |
4129 | \ |
4130 | subs num_blocks, num_blocks, #1; \ | |
4131 | bne 0b; \ | |
4132 | \ | |
4133 | 1: \ | |
4134 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
d1c75d1e E |
4135 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
4136 | vbif.u16 blend_pixels, pixels, blend_mask; \ | |
75e28f62 E |
4137 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
4138 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
4139 | \ | |
4140 | ldmia sp!, { r4, pc }; \ | |
4141 | \ | |
4142 | 2: \ | |
4143 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
d1c75d1e | 4144 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
75e28f62 E |
4145 | \ |
4146 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4147 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
75e28f62 | 4148 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
d1c75d1e | 4149 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ |
75e28f62 E |
4150 | bal 3b \ |
4151 | ||
4152 | ||
d1c75d1e | 4153 | |
75e28f62 E |
4154 | #define blend_blocks_add_fourth_untextured_builder(mask_evaluate) \ |
4155 | .align 3; \ | |
4156 | \ | |
4157 | function(blend_blocks_untextured_add_fourth_##mask_evaluate) \ | |
4158 | stmdb sp!, { r4, r14 }; \ | |
4159 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
4160 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
4161 | \ | |
4162 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
4163 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ]; \ | |
4164 | \ | |
4165 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4166 | mov c_64, #64; \ | |
4167 | \ | |
4168 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4169 | vmov.u16 d128_0x03E0, #0x0300; \ | |
75e28f62 E |
4170 | vmov.u16 d128_0x1C07, #0x1C00; \ |
4171 | vmov.u16 d128_0x00E0, #0x00E0; \ | |
4172 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
4173 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
75e28f62 E |
4174 | vorr.u16 d128_0x1C07, #0x0007; \ |
4175 | \ | |
4176 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
4177 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ | |
4178 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
4179 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4180 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
4181 | vshr.s16 pixels_fourth, pixels, #2; \ | |
4182 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
4183 | \ | |
4184 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4185 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
4186 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4187 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4188 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4189 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4190 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
4191 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
4192 | \ | |
4193 | subs num_blocks, num_blocks, #1; \ | |
4194 | beq 1f; \ | |
4195 | \ | |
4196 | 0: \ | |
4197 | mov fb_ptr, fb_ptr_next; \ | |
75e28f62 E |
4198 | ldr fb_ptr_next, [ pixel_ptr, #28 ]; \ |
4199 | \ | |
4200 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64; \ | |
4201 | \ | |
4202 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
4203 | vshr.s16 pixels_fourth, pixels, #2; \ | |
4204 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
4205 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
4206 | \ | |
4207 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
4208 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64; \ | |
4209 | \ | |
4210 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4211 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4212 | cmp fb_ptr_cmp, #28; \ | |
4213 | bls 2f; \ | |
4214 | \ | |
4215 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4216 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
4217 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4218 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
4219 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4220 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
4221 | \ | |
4222 | 3: \ | |
4223 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4224 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4225 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4226 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
4227 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
4228 | \ | |
4229 | subs num_blocks, num_blocks, #1; \ | |
4230 | bne 0b; \ | |
4231 | \ | |
4232 | 1: \ | |
4233 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
4234 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
4235 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
4236 | vst1.u16 { blend_pixels }, [ fb_ptr_next ]; \ | |
4237 | \ | |
4238 | ldmia sp!, { r4, pc }; \ | |
4239 | \ | |
4240 | 2: \ | |
4241 | vst1.u16 { blend_pixels }, [ fb_ptr ]; \ | |
4242 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
4243 | \ | |
4244 | vld1.u16 { fb_pixels }, [ fb_ptr_next ]; \ | |
4245 | blend_blocks_add_mask_set_##mask_evaluate(); \ | |
4246 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4247 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4248 | bal 3b \ | |
4249 | ||
4250 | ||
4251 | blend_blocks_add_fourth_textured_builder(off) | |
4252 | blend_blocks_add_fourth_textured_builder(on) | |
4253 | blend_blocks_add_fourth_untextured_builder(off) | |
4254 | blend_blocks_add_fourth_untextured_builder(on) | |
4255 | ||
4256 | // TODO: Optimize this more. Need a scene that actually uses it for | |
4257 | // confirmation.. | |
4258 | ||
4259 | .align 3 | |
4260 | ||
4261 | function(blend_blocks_textured_unblended_on) | |
4262 | stmdb sp!, { r4, r14 } | |
4263 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset | |
4264 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
4265 | ||
4266 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
4267 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [ mask_msb_ptr, :16 ] | |
4268 | ||
4269 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
4270 | mov c_64, #64 | |
4271 | ||
4272 | ldr fb_ptr, [ pixel_ptr, #28 ] | |
4273 | vld1.u16 { fb_pixels }, [ fb_ptr ] | |
4274 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
4275 | vclt.s16 write_mask, fb_pixels, #0 | |
4276 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64 | |
4277 | ||
4278 | subs num_blocks, num_blocks, #1 | |
4279 | beq 1f | |
4280 | ||
4281 | 0: | |
4282 | vorr.u16 draw_mask, draw_mask, write_mask | |
4283 | vbif.u16 fb_pixels, pixels, draw_mask | |
4284 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
4285 | ||
4286 | ldr fb_ptr, [ pixel_ptr, #28 ] | |
4287 | vld1.u16 { fb_pixels }, [ fb_ptr ] | |
4288 | vld1.u32 { draw_mask }, [ draw_mask_ptr, :128 ], c_64 | |
4289 | vclt.s16 write_mask, fb_pixels, #0 | |
4290 | vld1.u32 { pixels }, [ pixel_ptr, :128 ], c_64 | |
4291 | ||
4292 | subs num_blocks, num_blocks, #1 | |
4293 | bne 0b | |
4294 | ||
4295 | 1: | |
4296 | vorr.u16 draw_mask, draw_mask, write_mask | |
4297 | vbif.u16 fb_pixels, pixels, draw_mask | |
4298 | vst1.u16 { fb_pixels }, [ fb_ptr ] | |
4299 | ||
4300 | ldmia sp!, { r4, pc } | |
4301 | ||
4302 | ||
4303 | function(blend_blocks_textured_unblended_off) | |
4304 | bx lr | |
4305 | ||
4306 | ||
4307 | function(warmup) | |
4308 | mov r3, #64 | |
4309 | cmp r0, #0 | |
4310 | bxeq lr | |
4311 | ||
4312 | 0: | |
4313 | vld1.u32 { u_whole_8, v_whole_8 }, [ r1, :128 ], r3 | |
4314 | ||
4315 | subs r0, r0, #1 | |
4316 | bne 0b | |
4317 | ||
4318 | bx lr | |
4319 | ||
6c4a10c4 | 4320 | #undef vram_ptr |
75e28f62 | 4321 | #undef color |
6c4a10c4 | 4322 | #undef width |
75e28f62 | 4323 | #undef height |
6c4a10c4 | 4324 | #undef pitch |
75e28f62 E |
4325 | |
4326 | #define vram_ptr r0 | |
6c4a10c4 E |
4327 | #define color r1 |
4328 | #define width r2 | |
4329 | #define height r3 | |
75e28f62 | 4330 | |
6c4a10c4 | 4331 | #define pitch r1 |
75e28f62 | 4332 | |
6c4a10c4 | 4333 | #define num_width r12 |
75e28f62 | 4334 | |
87c45ad1 E |
4335 | #undef colors_a |
4336 | #undef colors_b | |
75e28f62 | 4337 | |
87c45ad1 E |
4338 | #define colors_a q0 |
4339 | #define colors_b q1 | |
75e28f62 E |
4340 | |
4341 | .align 3 | |
4342 | ||
4343 | function(render_block_fill_body) | |
87c45ad1 | 4344 | vdup.u16 colors_a, color |
6c4a10c4 | 4345 | mov pitch, #2048 |
75e28f62 | 4346 | |
87c45ad1 | 4347 | vmov colors_b, colors_a |
75e28f62 | 4348 | sub pitch, pitch, width, lsl #1 |
75e28f62 | 4349 | |
6c4a10c4 | 4350 | mov num_width, width |
75e28f62 | 4351 | |
6c4a10c4 E |
4352 | 0: |
4353 | vst1.u32 { colors_a, colors_b }, [ vram_ptr, :256 ]! | |
75e28f62 | 4354 | |
d1c75d1e | 4355 | subs num_width, num_width, #16 |
6c4a10c4 | 4356 | bne 0b |
75e28f62 | 4357 | |
75e28f62 | 4358 | add vram_ptr, vram_ptr, pitch |
6c4a10c4 E |
4359 | mov num_width, width |
4360 | ||
75e28f62 E |
4361 | subs height, height, #1 |
4362 | bne 0b | |
75e28f62 | 4363 | |
6c4a10c4 E |
4364 | bx lr |
4365 | ||
75e28f62 E |
4366 | |
4367 | #undef x | |
4368 | #undef y | |
4369 | #undef width | |
4370 | #undef height | |
4371 | #undef fb_ptr | |
4372 | #undef texture_mask | |
4373 | #undef num_blocks | |
4374 | #undef temp | |
4375 | #undef dirty_textures_mask | |
4376 | #undef clut_ptr | |
4377 | #undef current_texture_mask | |
4378 | ||
4379 | #define psx_gpu r0 | |
4380 | #define x r1 | |
4381 | #define y r2 | |
4382 | #define u r3 | |
4383 | #define v r4 | |
4384 | #define width r5 | |
4385 | #define height r6 | |
4386 | #define offset_u r8 | |
4387 | #define offset_v r9 | |
4388 | #define offset_u_right r10 | |
4389 | #define width_rounded r11 | |
4390 | #define height_rounded r12 | |
4391 | ||
4392 | #define texture_offset_base r1 | |
4393 | #define tile_width r2 | |
4394 | #define tile_height r3 | |
4395 | #define num_blocks r4 | |
4396 | #define block r5 | |
4397 | #define sub_tile_height r6 | |
4398 | #define fb_ptr r7 | |
4399 | #define texture_mask r8 | |
4400 | #define column_data r9 | |
4401 | #define texture_offset r10 | |
4402 | #define tiles_remaining r11 | |
4403 | #define fb_ptr_advance_column r12 | |
4404 | #define texture_block_ptr r14 | |
4405 | ||
4406 | #define texture_page_ptr r3 | |
4407 | #define left_block_mask r4 | |
4408 | #define right_block_mask r5 | |
4409 | #define texture_mask_rev r10 | |
4410 | #define control_mask r11 | |
4411 | ||
4412 | #define dirty_textures_mask r4 | |
4413 | #define clut_ptr r5 | |
4414 | #define current_texture_mask r6 | |
4415 | ||
4416 | ||
4417 | #undef texels | |
4418 | #undef clut_low_a | |
4419 | #undef clut_low_b | |
4420 | #undef clut_high_a | |
4421 | #undef clut_high_b | |
4422 | #undef clut_a | |
4423 | #undef clut_b | |
4424 | #undef texels_low | |
4425 | #undef texels_high | |
4426 | ||
4427 | #define texels d0 | |
4428 | #define draw_masks_fb_ptrs q1 | |
4429 | ||
4430 | #define draw_mask_fb_ptr_left d2 | |
4431 | #define draw_mask_fb_ptr_right d3 | |
4432 | ||
59d15d23 | 4433 | #define draw_mask_fb_ptr_left_a d2 |
4434 | #define draw_mask_fb_ptr_left_b d3 | |
4435 | #define draw_mask_fb_ptr_right_a d10 | |
4436 | #define draw_mask_fb_ptr_right_b d11 | |
4437 | #define draw_masks_fb_ptrs2 q5 | |
4438 | ||
75e28f62 E |
4439 | #define clut_low_a d4 |
4440 | #define clut_low_b d5 | |
4441 | #define clut_high_a d6 | |
4442 | #define clut_high_b d7 | |
4443 | ||
4444 | #define block_masks d8 | |
4445 | #define block_masks_shifted d9 | |
4446 | ||
4447 | #define clut_a q2 | |
4448 | #define clut_b q3 | |
4449 | ||
59d15d23 | 4450 | #define texels_low d12 |
4451 | #define texels_high d13 | |
75e28f62 | 4452 | |
59d15d23 | 4453 | #define texels_wide_low d14 |
4454 | #define texels_wide_high d15 | |
4455 | #define texels_wide q7 | |
75e28f62 E |
4456 | |
4457 | ||
59d15d23 | 4458 | setup_sprite_flush_blocks: |
4459 | vpush { q1 - q5 } | |
75e28f62 E |
4460 | |
4461 | stmdb sp!, { r0 - r3, r12, r14 } | |
4462 | bl flush_render_block_buffer | |
4463 | ldmia sp!, { r0 - r3, r12, r14 } | |
4464 | ||
59d15d23 | 4465 | vpop { q1 - q5 } |
75e28f62 E |
4466 | |
4467 | add block, psx_gpu, #psx_gpu_blocks_offset | |
75e28f62 E |
4468 | bx lr |
4469 | ||
4470 | ||
4471 | setup_sprite_update_texture_4bpp_cache: | |
4472 | stmdb sp!, { r0 - r3, r14 } | |
4473 | bl update_texture_4bpp_cache | |
4474 | ldmia sp!, { r0 - r3, pc } | |
4475 | ||
4476 | ||
4477 | setup_sprite_update_texture_8bpp_cache: | |
4478 | stmdb sp!, { r0 - r3, r14 } | |
4479 | bl update_texture_8bpp_cache | |
4480 | ldmia sp!, { r0 - r3, pc } | |
4481 | ||
4482 | ||
4483 | #define setup_sprite_tiled_initialize_4bpp() \ | |
4484 | ldr dirty_textures_mask, \ | |
4485 | [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ]; \ | |
4486 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ]; \ | |
4487 | \ | |
4488 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ]; \ | |
4489 | vld1.u32 { clut_a, clut_b }, [ clut_ptr, :128 ]; \ | |
4490 | \ | |
4491 | tst current_texture_mask, dirty_textures_mask; \ | |
4492 | vuzp.u8 clut_a, clut_b; \ | |
4493 | \ | |
4494 | blne setup_sprite_update_texture_4bpp_cache \ | |
4495 | ||
4496 | #define setup_sprite_tiled_initialize_8bpp() \ | |
4497 | ldr dirty_textures_mask, \ | |
4498 | [ psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset ]; \ | |
4499 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ]; \ | |
4500 | \ | |
4501 | tst current_texture_mask, dirty_textures_mask; \ | |
4502 | blne setup_sprite_update_texture_8bpp_cache \ | |
4503 | ||
4504 | ||
75e28f62 E |
4505 | #define setup_sprite_block_count_single() \ |
4506 | sub_tile_height \ | |
4507 | ||
4508 | #define setup_sprite_block_count_double() \ | |
4509 | sub_tile_height, lsl #1 \ | |
4510 | ||
4511 | #define setup_sprite_tile_add_blocks(type) \ | |
4512 | add num_blocks, num_blocks, setup_sprite_block_count_##type(); \ | |
4513 | cmp num_blocks, #MAX_BLOCKS; \ | |
4514 | \ | |
59d15d23 | 4515 | movgt num_blocks, setup_sprite_block_count_##type(); \ |
4516 | blgt setup_sprite_flush_blocks \ | |
75e28f62 E |
4517 | |
4518 | ||
4519 | #define setup_sprite_tile_full_4bpp(edge) \ | |
4520 | setup_sprite_tile_add_blocks(double); \ | |
4521 | \ | |
4522 | 4: \ | |
4523 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4524 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
4525 | \ | |
4526 | pld [ fb_ptr ]; \ | |
4527 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4528 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4529 | \ | |
4530 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4531 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4532 | \ | |
4533 | vst2.u8 { texels_low, texels_high }, [ block, :128 ]; \ | |
4534 | add texture_block_ptr, texture_offset, #8; \ | |
4535 | \ | |
4536 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4537 | add block, block, #40; \ | |
4538 | \ | |
4539 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4540 | add fb_ptr, fb_ptr, #16; \ | |
4541 | \ | |
4542 | vst1.u32 { draw_mask_fb_ptr_left }, [ block, :64 ]; \ | |
4543 | add block, block, #24; \ | |
4544 | \ | |
4545 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4546 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4547 | \ | |
4548 | pld [ fb_ptr ]; \ | |
4549 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ | |
4550 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4551 | \ | |
4552 | vst2.u8 { texels_low, texels_high }, [ block, :128 ]; \ | |
4553 | add block, block, #40; \ | |
4554 | \ | |
4555 | add texture_offset, texture_offset, #0x10; \ | |
4556 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
4557 | \ | |
4558 | vst1.u32 { draw_mask_fb_ptr_right }, [ block, :64 ]; \ | |
4559 | add block, block, #24; \ | |
4560 | \ | |
4561 | subs sub_tile_height, sub_tile_height, #1; \ | |
4562 | bne 4b; \ | |
4563 | \ | |
4564 | add texture_offset, texture_offset, #0xF00; \ | |
4565 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
4566 | ||
4567 | ||
4568 | #define setup_sprite_tile_half_4bpp(edge) \ | |
4569 | setup_sprite_tile_add_blocks(single); \ | |
4570 | \ | |
4571 | 4: \ | |
4572 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4573 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
4574 | \ | |
4575 | pld [ fb_ptr ]; \ | |
4576 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4577 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4578 | \ | |
4579 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4580 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4581 | \ | |
4582 | vst2.u8 { texels_low, texels_high }, [ block, :128 ]; \ | |
4583 | add block, block, #40; \ | |
4584 | \ | |
4585 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4586 | vst1.u32 { draw_mask_fb_ptr_##edge }, [ block, :64 ]; \ | |
4587 | \ | |
4588 | add block, block, #24; \ | |
4589 | add texture_offset, texture_offset, #0x10; \ | |
4590 | \ | |
4591 | add fb_ptr, fb_ptr, #2048; \ | |
4592 | subs sub_tile_height, sub_tile_height, #1; \ | |
4593 | \ | |
4594 | bne 4b; \ | |
4595 | \ | |
4596 | add texture_offset, texture_offset, #0xF00; \ | |
4597 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
4598 | ||
4599 | ||
4600 | #define setup_sprite_tile_full_8bpp(edge) \ | |
4601 | setup_sprite_tile_add_blocks(double); \ | |
4602 | add block, block, #16; \ | |
4603 | \ | |
4604 | 4: \ | |
4605 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4606 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
4607 | \ | |
4608 | pld [ fb_ptr ]; \ | |
4609 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4610 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4611 | \ | |
4612 | add texture_block_ptr, texture_offset, #8; \ | |
4613 | vst1.u32 { texels }, [ block, :64 ]; \ | |
4614 | \ | |
4615 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4616 | add block, block, #24; \ | |
4617 | \ | |
4618 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4619 | \ | |
4620 | add fb_ptr, fb_ptr, #16; \ | |
4621 | vst1.u32 { draw_mask_fb_ptr_left }, [ block, :64 ]; \ | |
4622 | \ | |
4623 | add block, block, #40; \ | |
4624 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4625 | pld [ fb_ptr ]; \ | |
4626 | \ | |
4627 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ | |
4628 | vst1.u32 { texels }, [ block, :64 ]; \ | |
4629 | add block, block, #24; \ | |
4630 | \ | |
4631 | add texture_offset, texture_offset, #0x10; \ | |
4632 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
4633 | \ | |
4634 | vst1.u32 { draw_mask_fb_ptr_right }, [ block, :64 ]; \ | |
4635 | add block, block, #40; \ | |
4636 | \ | |
4637 | subs sub_tile_height, sub_tile_height, #1; \ | |
4638 | bne 4b; \ | |
4639 | \ | |
4640 | sub block, block, #16; \ | |
4641 | add texture_offset, texture_offset, #0xF00; \ | |
4642 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
4643 | ||
4644 | ||
4645 | #define setup_sprite_tile_half_8bpp(edge) \ | |
4646 | setup_sprite_tile_add_blocks(single); \ | |
4647 | add block, block, #16; \ | |
4648 | \ | |
4649 | 4: \ | |
4650 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4651 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
4652 | pld [ fb_ptr ]; \ | |
4653 | \ | |
4654 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4655 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4656 | \ | |
4657 | vst1.u32 { texels }, [ block, :64 ]; \ | |
4658 | add block, block, #24; \ | |
4659 | \ | |
4660 | vst1.u32 { draw_mask_fb_ptr_##edge }, [ block, :64 ]; \ | |
4661 | add block, block, #40; \ | |
4662 | \ | |
4663 | add texture_offset, texture_offset, #0x10; \ | |
4664 | add fb_ptr, fb_ptr, #2048; \ | |
4665 | \ | |
4666 | subs sub_tile_height, sub_tile_height, #1; \ | |
4667 | bne 4b; \ | |
4668 | \ | |
4669 | sub block, block, #16; \ | |
4670 | add texture_offset, texture_offset, #0xF00; \ | |
4671 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
4672 | ||
4673 | ||
4674 | #define setup_sprite_tile_column_edge_pre_adjust_half_right() \ | |
4675 | add texture_offset, texture_offset_base, #8; \ | |
4676 | add fb_ptr, fb_ptr, #16 \ | |
4677 | ||
4678 | #define setup_sprite_tile_column_edge_pre_adjust_half_left() \ | |
4679 | mov texture_offset, texture_offset_base \ | |
4680 | ||
4681 | #define setup_sprite_tile_column_edge_pre_adjust_half(edge) \ | |
4682 | setup_sprite_tile_column_edge_pre_adjust_half_##edge() \ | |
4683 | ||
4684 | #define setup_sprite_tile_column_edge_pre_adjust_full(edge) \ | |
4685 | mov texture_offset, texture_offset_base \ | |
4686 | ||
4687 | #define setup_sprite_tile_column_edge_post_adjust_half_right() \ | |
4688 | sub fb_ptr, fb_ptr, #16 \ | |
4689 | ||
4690 | #define setup_sprite_tile_column_edge_post_adjust_half_left() \ | |
4691 | ||
4692 | #define setup_sprite_tile_column_edge_post_adjust_half(edge) \ | |
4693 | setup_sprite_tile_column_edge_post_adjust_half_##edge() \ | |
4694 | ||
4695 | #define setup_sprite_tile_column_edge_post_adjust_full(edge) \ | |
4696 | ||
4697 | ||
59d15d23 | 4698 | #define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode, \ |
4699 | x4mode) \ | |
75e28f62 | 4700 | mov sub_tile_height, column_data; \ |
59d15d23 | 4701 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
4702 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
4703 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
75e28f62 | 4704 | |
59d15d23 | 4705 | #define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode, \ |
4706 | x4mode) \ | |
75e28f62 E |
4707 | and sub_tile_height, column_data, #0xFF; \ |
4708 | mov tiles_remaining, column_data, lsr #16; \ | |
59d15d23 | 4709 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
4710 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
75e28f62 E |
4711 | \ |
4712 | subs tiles_remaining, tiles_remaining, #1; \ | |
4713 | beq 2f; \ | |
4714 | \ | |
4715 | 3: \ | |
4716 | mov sub_tile_height, #16; \ | |
59d15d23 | 4717 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
75e28f62 E |
4718 | subs tiles_remaining, tiles_remaining, #1; \ |
4719 | bne 3b; \ | |
4720 | \ | |
4721 | 2: \ | |
4722 | uxtb sub_tile_height, column_data, ror #8; \ | |
59d15d23 | 4723 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
4724 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
75e28f62 E |
4725 | |
4726 | ||
4727 | #define setup_sprite_column_data_single() \ | |
4728 | mov column_data, height; \ | |
4729 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] \ | |
4730 | ||
4731 | #define setup_sprite_column_data_multi() \ | |
4732 | and height_rounded, height_rounded, #0xF; \ | |
4733 | rsb column_data, offset_v, #16; \ | |
4734 | \ | |
4735 | add height_rounded, height_rounded, #1; \ | |
4736 | sub tile_height, tile_height, #1; \ | |
4737 | \ | |
4738 | orr column_data, column_data, tile_height, lsl #16; \ | |
4739 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ]; \ | |
4740 | \ | |
4741 | orr column_data, column_data, height_rounded, lsl #8 \ | |
4742 | ||
59d15d23 | 4743 | #define setup_sprite_setup_left_draw_mask_fb_ptr() \ |
4744 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
4745 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ | |
4746 | ||
4747 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column() \ | |
4748 | mov fb_ptr_advance_column, #32; \ | |
4749 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
4750 | \ | |
4751 | sub fb_ptr_advance_column, height, lsl #11; \ | |
4752 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ | |
4753 | ||
4754 | #define setup_sprite_setup_right_draw_mask_fb_ptr() \ | |
4755 | vdup.u8 draw_mask_fb_ptr_left, block_masks[4]; \ | |
4756 | vdup.u8 draw_mask_fb_ptr_right, block_masks[5] \ | |
4757 | ||
4758 | #define setup_sprite_tile_column_width_single(tm, multi_height, edge_mode, \ | |
4759 | edge, x4mode) \ | |
4760 | setup_sprite_##tm##_single_##multi_height##_##edge_mode##_##edge##x4mode: \ | |
75e28f62 E |
4761 | setup_sprite_column_data_##multi_height(); \ |
4762 | vext.32 block_masks_shifted, block_masks, block_masks, #1; \ | |
4763 | vorr.u32 block_masks, block_masks, block_masks_shifted; \ | |
59d15d23 | 4764 | setup_sprite_setup_left_draw_mask_fb_ptr##x4mode(); \ |
75e28f62 | 4765 | \ |
59d15d23 | 4766 | setup_sprite_tile_column_height_##multi_height(edge_mode, edge, tm, x4mode); \ |
75e28f62 E |
4767 | ldmia sp!, { r4 - r11, pc } \ |
4768 | ||
4769 | #define setup_sprite_tiled_advance_column() \ | |
4770 | add texture_offset_base, texture_offset_base, #0x100; \ | |
4771 | tst texture_offset_base, #0xF00; \ | |
4772 | subeq texture_offset_base, texture_offset_base, #(0x100 + 0xF00) \ | |
4773 | ||
4774 | #define setup_sprite_tile_column_width_multi(tm, multi_height, left_mode, \ | |
59d15d23 | 4775 | right_mode, x4mode) \ |
4776 | setup_sprite_##tm##_multi_##multi_height##_##left_mode##_##right_mode##x4mode:\ | |
75e28f62 | 4777 | setup_sprite_column_data_##multi_height(); \ |
75e28f62 | 4778 | \ |
59d15d23 | 4779 | setup_sprite_setup_left_draw_mask_fb_ptr_advance_column##x4mode(); \ |
75e28f62 | 4780 | \ |
59d15d23 | 4781 | setup_sprite_tile_column_height_##multi_height(left_mode, right, tm, x4mode);\ |
75e28f62 E |
4782 | \ |
4783 | subs tile_width, tile_width, #2; \ | |
4784 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ | |
4785 | \ | |
75e28f62 E |
4786 | beq 1f; \ |
4787 | \ | |
59d15d23 | 4788 | vmov.u8 draw_masks_fb_ptrs, #0; \ |
4789 | vmov.u8 draw_masks_fb_ptrs2, #0; \ | |
4790 | \ | |
75e28f62 E |
4791 | 0: \ |
4792 | setup_sprite_tiled_advance_column(); \ | |
59d15d23 | 4793 | setup_sprite_tile_column_height_##multi_height(full, none, tm, x4mode); \ |
75e28f62 E |
4794 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ |
4795 | subs tile_width, tile_width, #1; \ | |
4796 | bne 0b; \ | |
4797 | \ | |
4798 | 1: \ | |
59d15d23 | 4799 | setup_sprite_setup_right_draw_mask_fb_ptr##x4mode(); \ |
75e28f62 E |
4800 | \ |
4801 | setup_sprite_tiled_advance_column(); \ | |
59d15d23 | 4802 | setup_sprite_tile_column_height_##multi_height(right_mode, left, tm, x4mode);\ |
75e28f62 E |
4803 | ldmia sp!, { r4 - r11, pc } \ |
4804 | ||
4805 | ||
59d15d23 | 4806 | #define setup_sprite_offset_u_adjust() \ |
4807 | ||
4808 | #define setup_sprite_get_left_block_mask() \ | |
4809 | and left_block_mask, left_block_mask, #0xFF \ | |
4810 | ||
4811 | #define setup_sprite_compare_left_block_mask() \ | |
4812 | cmp left_block_mask, #0xFF \ | |
4813 | ||
4814 | #define setup_sprite_get_right_block_mask() \ | |
4815 | uxtb right_block_mask, right_block_mask, ror #8 \ | |
4816 | ||
4817 | #define setup_sprite_compare_right_block_mask() \ | |
4818 | cmp right_block_mask, #0xFF \ | |
4819 | ||
4820 | ||
4821 | ||
4822 | /* 4x stuff */ | |
4823 | #define fb_ptr2 column_data | |
4824 | ||
4825 | #define setup_sprite_offset_u_adjust_4x() \ | |
4826 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
4827 | lsl offset_u_right, #1; \ | |
4828 | lsl offset_u, #1; \ | |
4829 | add offset_u_right, #1 \ | |
4830 | ||
4831 | #define setup_sprite_get_left_block_mask_4x() \ | |
4832 | sxth left_block_mask, left_block_mask \ | |
4833 | ||
4834 | #define setup_sprite_compare_left_block_mask_4x() \ | |
4835 | cmp left_block_mask, #0xFFFFFFFF \ | |
4836 | ||
4837 | #define setup_sprite_get_right_block_mask_4x() \ | |
4838 | sxth right_block_mask, right_block_mask, ror #16 \ | |
4839 | ||
4840 | #define setup_sprite_compare_right_block_mask_4x() \ | |
4841 | cmp right_block_mask, #0xFFFFFFFF \ | |
4842 | ||
4843 | ||
4844 | #define widen_texels_16bpp(texels_) \ | |
4845 | vmov texels_wide_low, texels_; \ | |
4846 | vmov texels_wide_high, texels_; \ | |
4847 | vzip.16 texels_wide_low, texels_wide_high \ | |
4848 | ||
4849 | #define widen_texels_8bpp(texels_) \ | |
4850 | vmov texels_wide_low, texels_; \ | |
4851 | vmov texels_wide_high, texels_; \ | |
4852 | vzip.8 texels_wide_low, texels_wide_high \ | |
4853 | ||
4854 | #define write_block_16bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
4855 | vst1.u32 { texels_ }, [ block_, :128 ]; \ | |
4856 | add block_, block_, #40; \ | |
4857 | \ | |
4858 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
4859 | vst1.u32 { draw_mask_fb_ptr_ }, [ block_, :64 ]; \ | |
4860 | add block_, block_, #24 \ | |
4861 | ||
4862 | /* assumes 16-byte offset already added to block_ */ | |
4863 | #define write_block_8bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
4864 | vst1.u32 { texels_ }, [ block_, :64 ]; \ | |
4865 | add block_, block_, #24; \ | |
4866 | \ | |
4867 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
4868 | vst1.u32 { draw_mask_fb_ptr_ }, [ block_, :64 ]; \ | |
4869 | add block_, block_, #40 \ | |
4870 | ||
4871 | #define do_texture_block_16bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
4872 | draw_mask_fb_ptr_b_) \ | |
4873 | widen_texels_16bpp(texels_low); \ | |
4874 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
4875 | \ | |
4876 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
4877 | \ | |
4878 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
4879 | widen_texels_16bpp(texels_high); \ | |
4880 | \ | |
4881 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
4882 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
4883 | \ | |
4884 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
4885 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
4886 | ||
4887 | #define do_texture_block_8bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
4888 | draw_mask_fb_ptr_b_) \ | |
4889 | widen_texels_8bpp(texels); \ | |
4890 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
4891 | \ | |
4892 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
4893 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
4894 | \ | |
4895 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
4896 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
4897 | \ | |
4898 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
4899 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
4900 | ||
4901 | ||
4902 | #define setup_sprite_tiled_initialize_4bpp_4x() \ | |
4903 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ]; \ | |
4904 | vld1.u32 { clut_a, clut_b }, [ clut_ptr, :128 ]; \ | |
4905 | \ | |
4906 | vuzp.u8 clut_a, clut_b \ | |
4907 | ||
4908 | #define setup_sprite_tiled_initialize_8bpp_4x() \ | |
4909 | ||
4910 | ||
4911 | #define setup_sprite_block_count_single_4x() \ | |
4912 | sub_tile_height, lsl #2 \ | |
4913 | ||
4914 | #define setup_sprite_block_count_double_4x() \ | |
4915 | sub_tile_height, lsl #(1+2) \ | |
4916 | ||
4917 | #define setup_sprite_tile_full_4bpp_4x(edge) \ | |
4918 | setup_sprite_tile_add_blocks(double_4x); \ | |
4919 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
4920 | \ | |
4921 | 4: \ | |
4922 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4923 | pld [ fb_ptr ]; \ | |
4924 | \ | |
4925 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4926 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4927 | \ | |
4928 | add texture_block_ptr, texture_offset, #8; \ | |
4929 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4930 | \ | |
4931 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4932 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4933 | \ | |
4934 | vzip.8 texels_low, texels_high; \ | |
4935 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
4936 | draw_mask_fb_ptr_left_b); \ | |
4937 | \ | |
4938 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
8438c3c7 | 4939 | pld [ fb_ptr, #2048 ]; \ |
59d15d23 | 4940 | \ |
4941 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
8438c3c7 | 4942 | add fb_ptr, fb_ptr, #16*2; \ |
59d15d23 | 4943 | \ |
8438c3c7 | 4944 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ |
59d15d23 | 4945 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ |
4946 | \ | |
4947 | vzip.8 texels_low, texels_high; \ | |
4948 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
4949 | draw_mask_fb_ptr_right_b); \ | |
4950 | \ | |
4951 | add texture_offset, texture_offset, #0x10; \ | |
4952 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
4953 | \ | |
4954 | subs sub_tile_height, sub_tile_height, #1; \ | |
4955 | bne 4b; \ | |
4956 | \ | |
4957 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
4958 | add texture_offset, texture_offset, #0xF00; \ | |
4959 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
4960 | ||
4961 | ||
4962 | #define setup_sprite_tile_half_4bpp_4x(edge) \ | |
4963 | setup_sprite_tile_add_blocks(single_4x); \ | |
4964 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
4965 | \ | |
4966 | 4: \ | |
4967 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4968 | pld [ fb_ptr ]; \ | |
4969 | \ | |
4970 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4971 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
4972 | \ | |
4973 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4974 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4975 | \ | |
4976 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4977 | add texture_offset, texture_offset, #0x10; \ | |
4978 | \ | |
4979 | vzip.8 texels_low, texels_high; \ | |
4980 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ | |
4981 | draw_mask_fb_ptr_##edge##_b); \ | |
4982 | \ | |
8438c3c7 | 4983 | pld [ fb_ptr, #2048 ]; \ |
59d15d23 | 4984 | add fb_ptr, fb_ptr, #2048 * 2; \ |
59d15d23 | 4985 | \ |
8438c3c7 | 4986 | subs sub_tile_height, sub_tile_height, #1; \ |
59d15d23 | 4987 | bne 4b; \ |
4988 | \ | |
4989 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
4990 | add texture_offset, texture_offset, #0xF00; \ | |
4991 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
4992 | ||
4993 | ||
4994 | #define setup_sprite_tile_full_8bpp_4x(edge) \ | |
4995 | setup_sprite_tile_add_blocks(double_4x); \ | |
4996 | add block, block, #16; \ | |
4997 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
4998 | \ | |
4999 | 4: \ | |
5000 | and texture_block_ptr, texture_offset, texture_mask; \ | |
5001 | pld [ fb_ptr ]; \ | |
5002 | \ | |
5003 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
5004 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
5005 | \ | |
5006 | add texture_block_ptr, texture_offset, #8; \ | |
5007 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
5008 | draw_mask_fb_ptr_left_b); \ | |
5009 | \ | |
8438c3c7 | 5010 | pld [ fb_ptr, #2048 ]; \ |
59d15d23 | 5011 | and texture_block_ptr, texture_block_ptr, texture_mask; \ |
5012 | \ | |
5013 | add fb_ptr, fb_ptr, #16*2; \ | |
5014 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
5015 | \ | |
5016 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
59d15d23 | 5017 | \ |
5018 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
5019 | draw_mask_fb_ptr_right_b); \ | |
5020 | \ | |
5021 | add texture_offset, texture_offset, #0x10; \ | |
5022 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
5023 | \ | |
5024 | subs sub_tile_height, sub_tile_height, #1; \ | |
5025 | bne 4b; \ | |
5026 | \ | |
5027 | sub block, block, #16; \ | |
5028 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5029 | add texture_offset, texture_offset, #0xF00; \ | |
5030 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
5031 | ||
5032 | ||
5033 | #define setup_sprite_tile_half_8bpp_4x(edge) \ | |
5034 | setup_sprite_tile_add_blocks(single_4x); \ | |
5035 | add block, block, #16; \ | |
5036 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
5037 | \ | |
5038 | 4: \ | |
5039 | and texture_block_ptr, texture_offset, texture_mask; \ | |
5040 | pld [ fb_ptr ]; \ | |
5041 | \ | |
5042 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
5043 | vld1.u32 { texels }, [ texture_block_ptr, :64 ]; \ | |
5044 | \ | |
8438c3c7 | 5045 | pld [ fb_ptr, #2048 ]; \ |
59d15d23 | 5046 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ |
5047 | draw_mask_fb_ptr_##edge##_b); \ | |
5048 | \ | |
5049 | add texture_offset, texture_offset, #0x10; \ | |
5050 | add fb_ptr, fb_ptr, #2048 * 2; \ | |
5051 | \ | |
5052 | subs sub_tile_height, sub_tile_height, #1; \ | |
5053 | bne 4b; \ | |
5054 | \ | |
5055 | sub block, block, #16; \ | |
5056 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5057 | add texture_offset, texture_offset, #0xF00; \ | |
5058 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] \ | |
5059 | ||
5060 | ||
5061 | #define setup_sprite_tile_column_edge_pre_adjust_half_right_4x() \ | |
5062 | add texture_offset, texture_offset_base, #8; \ | |
5063 | add fb_ptr, fb_ptr, #16 * 2 \ | |
5064 | ||
5065 | #define setup_sprite_tile_column_edge_pre_adjust_half_left_4x() \ | |
5066 | mov texture_offset, texture_offset_base \ | |
5067 | ||
5068 | #define setup_sprite_tile_column_edge_pre_adjust_half_4x(edge) \ | |
5069 | setup_sprite_tile_column_edge_pre_adjust_half_##edge##_4x() \ | |
5070 | ||
5071 | #define setup_sprite_tile_column_edge_pre_adjust_full_4x(edge) \ | |
5072 | mov texture_offset, texture_offset_base \ | |
5073 | ||
5074 | #define setup_sprite_tile_column_edge_post_adjust_half_right_4x() \ | |
5075 | sub fb_ptr, fb_ptr, #16 * 2 \ | |
5076 | ||
5077 | #define setup_sprite_tile_column_edge_post_adjust_half_left_4x() \ | |
5078 | ||
5079 | #define setup_sprite_tile_column_edge_post_adjust_half_4x(edge) \ | |
5080 | setup_sprite_tile_column_edge_post_adjust_half_##edge##_4x() \ | |
5081 | ||
5082 | #define setup_sprite_tile_column_edge_post_adjust_full_4x(edge) \ | |
5083 | ||
5084 | ||
5085 | #define setup_sprite_setup_left_draw_mask_fb_ptr_4x() \ | |
5086 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
5087 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
5088 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ | |
5089 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
5090 | ||
5091 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column_4x() \ | |
5092 | mov fb_ptr_advance_column, #32 * 2; \ | |
5093 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
5094 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
5095 | sub fb_ptr_advance_column, height, lsl #11 + 1; \ | |
5096 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ | |
5097 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
5098 | ||
5099 | #define setup_sprite_setup_right_draw_mask_fb_ptr_4x() \ | |
5100 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[4]; \ | |
5101 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[5]; \ | |
5102 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[6]; \ | |
5103 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[7] \ | |
5104 | ||
5105 | ||
75e28f62 E |
5106 | // r0: psx_gpu |
5107 | // r1: x | |
5108 | // r2: y | |
5109 | // r3: u | |
5110 | // [ sp ]: v | |
5111 | // [ sp + 4 ]: width | |
5112 | // [ sp + 8 ]: height | |
5113 | // [ sp + 12 ]: color (unused) | |
5114 | ||
59d15d23 | 5115 | #define setup_sprite_tiled_builder(texture_mode, x4mode) \ |
5116 | \ | |
5117 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, full, \ | |
5118 | x4mode); \ | |
5119 | setup_sprite_tile_column_width_single(texture_mode, multi, full, none, \ | |
5120 | x4mode); \ | |
5121 | setup_sprite_tile_column_width_multi(texture_mode, single, full, full, \ | |
5122 | x4mode); \ | |
5123 | setup_sprite_tile_column_width_single(texture_mode, single, full, none, \ | |
5124 | x4mode); \ | |
5125 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, full, \ | |
5126 | x4mode); \ | |
5127 | setup_sprite_tile_column_width_single(texture_mode, multi, half, right, \ | |
5128 | x4mode); \ | |
5129 | setup_sprite_tile_column_width_multi(texture_mode, single, half, full, \ | |
5130 | x4mode); \ | |
5131 | setup_sprite_tile_column_width_single(texture_mode, single, half, right, \ | |
5132 | x4mode); \ | |
5133 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, half, \ | |
5134 | x4mode); \ | |
5135 | setup_sprite_tile_column_width_single(texture_mode, multi, half, left, \ | |
5136 | x4mode); \ | |
5137 | setup_sprite_tile_column_width_multi(texture_mode, single, full, half, \ | |
5138 | x4mode); \ | |
5139 | setup_sprite_tile_column_width_single(texture_mode, single, half, left, \ | |
5140 | x4mode); \ | |
5141 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, half, \ | |
5142 | x4mode); \ | |
5143 | setup_sprite_tile_column_width_multi(texture_mode, single, half, half, \ | |
5144 | x4mode); \ | |
75e28f62 E |
5145 | \ |
5146 | .align 4; \ | |
5147 | \ | |
59d15d23 | 5148 | function(setup_sprite_##texture_mode##x4mode) \ |
75e28f62 | 5149 | stmdb sp!, { r4 - r11, r14 }; \ |
59d15d23 | 5150 | setup_sprite_tiled_initialize_##texture_mode##x4mode(); \ |
75e28f62 E |
5151 | \ |
5152 | ldr v, [ sp, #36 ]; \ | |
5153 | and offset_u, u, #0xF; \ | |
5154 | \ | |
5155 | ldr width, [ sp, #40 ]; \ | |
c1817bd9 | 5156 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ]; \ |
75e28f62 E |
5157 | \ |
5158 | ldr height, [ sp, #44 ]; \ | |
5159 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
5160 | \ | |
5161 | add fb_ptr, fb_ptr, x, lsl #1; \ | |
5162 | and offset_v, v, #0xF; \ | |
5163 | \ | |
5164 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
5165 | add width_rounded, offset_u, width; \ | |
5166 | \ | |
5167 | add height_rounded, offset_v, height; \ | |
5168 | add width_rounded, width_rounded, #15; \ | |
5169 | \ | |
5170 | add height_rounded, height_rounded, #15; \ | |
5171 | mov tile_width, width_rounded, lsr #4; \ | |
5172 | \ | |
5173 | /* texture_offset_base = VH-VL-00-00 */\ | |
5174 | mov texture_offset_base, v, lsl #8; \ | |
5175 | and offset_u_right, width_rounded, #0xF; \ | |
5176 | \ | |
5177 | /* texture_offset_base = VH-UH-UL-00 */\ | |
5178 | bfi texture_offset_base, u, #4, #8; \ | |
59d15d23 | 5179 | mov right_block_mask, #0xFFFFFFFE; \ |
5180 | \ | |
5181 | setup_sprite_offset_u_adjust##x4mode(); \ | |
75e28f62 E |
5182 | \ |
5183 | /* texture_offset_base = VH-UH-VL-00 */\ | |
5184 | bfi texture_offset_base, v, #4, #4; \ | |
59d15d23 | 5185 | mov left_block_mask, #0xFFFFFFFF; \ |
75e28f62 E |
5186 | \ |
5187 | mov tile_height, height_rounded, lsr #4; \ | |
5188 | mvn left_block_mask, left_block_mask, lsl offset_u; \ | |
5189 | \ | |
5190 | /* texture_mask = HH-HL-WH-WL */\ | |
5191 | ldrh texture_mask, [ psx_gpu, #psx_gpu_texture_mask_width_offset ]; \ | |
5192 | mov right_block_mask, right_block_mask, lsl offset_u_right; \ | |
5193 | \ | |
5194 | /* texture_mask_rev = WH-WL-HH-HL */\ | |
5195 | rev16 texture_mask_rev, texture_mask; \ | |
5196 | vmov block_masks, left_block_mask, right_block_mask; \ | |
5197 | \ | |
5198 | /* texture_mask = HH-HL-HL-WL */\ | |
5199 | bfi texture_mask, texture_mask_rev, #4, #4; \ | |
5200 | /* texture_mask_rev = 00-00-00-WH */\ | |
5201 | mov texture_mask_rev, texture_mask_rev, lsr #12; \ | |
5202 | \ | |
5203 | /* texture_mask = HH-WH-HL-WL */\ | |
5204 | bfi texture_mask, texture_mask_rev, #8, #4; \ | |
59d15d23 | 5205 | setup_sprite_get_left_block_mask##x4mode(); \ |
75e28f62 E |
5206 | \ |
5207 | mov control_mask, #0; \ | |
59d15d23 | 5208 | setup_sprite_compare_left_block_mask##x4mode(); \ |
75e28f62 | 5209 | \ |
59d15d23 | 5210 | setup_sprite_get_right_block_mask##x4mode(); \ |
75e28f62 E |
5211 | orreq control_mask, control_mask, #0x4; \ |
5212 | \ | |
5213 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ]; \ | |
59d15d23 | 5214 | setup_sprite_compare_right_block_mask##x4mode(); \ |
75e28f62 E |
5215 | \ |
5216 | orreq control_mask, control_mask, #0x8; \ | |
5217 | cmp tile_width, #1; \ | |
5218 | \ | |
5219 | add block, psx_gpu, #psx_gpu_blocks_offset; \ | |
5220 | orreq control_mask, control_mask, #0x1; \ | |
5221 | \ | |
5222 | cmp tile_height, #1; \ | |
5223 | add block, block, num_blocks, lsl #6; \ | |
5224 | \ | |
5225 | orreq control_mask, control_mask, #0x2; \ | |
5226 | ldr pc, [ pc, control_mask, lsl #2 ]; \ | |
5227 | nop; \ | |
5228 | \ | |
59d15d23 | 5229 | .word setup_sprite_##texture_mode##_multi_multi_full_full##x4mode; \ |
5230 | .word setup_sprite_##texture_mode##_single_multi_full_none##x4mode; \ | |
5231 | .word setup_sprite_##texture_mode##_multi_single_full_full##x4mode; \ | |
5232 | .word setup_sprite_##texture_mode##_single_single_full_none##x4mode; \ | |
5233 | .word setup_sprite_##texture_mode##_multi_multi_half_full##x4mode; \ | |
5234 | .word setup_sprite_##texture_mode##_single_multi_half_right##x4mode; \ | |
5235 | .word setup_sprite_##texture_mode##_multi_single_half_full##x4mode; \ | |
5236 | .word setup_sprite_##texture_mode##_single_single_half_right##x4mode; \ | |
5237 | .word setup_sprite_##texture_mode##_multi_multi_full_half##x4mode; \ | |
5238 | .word setup_sprite_##texture_mode##_single_multi_half_left##x4mode; \ | |
5239 | .word setup_sprite_##texture_mode##_multi_single_full_half##x4mode; \ | |
5240 | .word setup_sprite_##texture_mode##_single_single_half_left##x4mode; \ | |
5241 | .word setup_sprite_##texture_mode##_multi_multi_half_half##x4mode; \ | |
75e28f62 | 5242 | .word 0x00000000; \ |
59d15d23 | 5243 | .word setup_sprite_##texture_mode##_multi_single_half_half##x4mode; \ |
5244 | ||
5245 | ||
5246 | setup_sprite_tiled_builder(4bpp,); | |
5247 | setup_sprite_tiled_builder(8bpp,); | |
75e28f62 | 5248 | |
59d15d23 | 5249 | #undef draw_mask_fb_ptr_left |
5250 | #undef draw_mask_fb_ptr_right | |
75e28f62 | 5251 | |
59d15d23 | 5252 | setup_sprite_tiled_builder(4bpp, _4x); |
5253 | setup_sprite_tiled_builder(8bpp, _4x); | |
75e28f62 E |
5254 | |
5255 | ||
5256 | #undef block_ptr | |
5257 | #undef num_blocks | |
5258 | #undef clut_ptr | |
5259 | ||
5260 | #define psx_gpu r0 | |
5261 | #define block_ptr r0 | |
5262 | #define num_blocks r1 | |
5263 | #define clut_ptr r2 | |
5264 | #define texel_shift_mask r3 | |
5265 | #define block_pixels_a r4 | |
5266 | #define block_pixels_b r5 | |
5267 | #define texel_0 r6 | |
5268 | #define texel_2 r7 | |
5269 | #define texel_4 r8 | |
5270 | #define texel_6 r9 | |
5271 | #define texel_1 r10 | |
5272 | #define texel_3 r11 | |
5273 | #define texel_5 r12 | |
5274 | #define texel_7 r14 | |
5275 | #define texels_01 r6 | |
5276 | #define texels_23 r7 | |
5277 | #define texels_45 r8 | |
5278 | #define texels_67 r9 | |
5279 | ||
5280 | function(texture_sprite_blocks_8bpp) | |
5281 | stmdb sp!, { r4 - r11, r14 } | |
5282 | movw texel_shift_mask, #(0xFF << 1) | |
5283 | ||
5284 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5285 | ldr clut_ptr, [ psx_gpu, #psx_gpu_clut_ptr_offset ] | |
5286 | ||
5287 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
5288 | ldr block_pixels_a, [ block_ptr, #16 ] | |
5289 | ||
5290 | 0: | |
5291 | and texel_0, texel_shift_mask, block_pixels_a, lsl #1 | |
5292 | ldr block_pixels_b, [ block_ptr, #20 ] | |
5293 | ||
5294 | and texel_1, texel_shift_mask, block_pixels_a, lsr #7 | |
5295 | ldrh texel_0, [ clut_ptr, texel_0 ] | |
5296 | ||
5297 | and texel_2, texel_shift_mask, block_pixels_a, lsr #15 | |
5298 | ldrh texel_1, [ clut_ptr, texel_1 ] | |
5299 | ||
5300 | and texel_3, texel_shift_mask, block_pixels_a, lsr #23 | |
5301 | ldr block_pixels_a, [ block_ptr, #(64 + 16) ] | |
5302 | ||
5303 | ldrh texel_2, [ clut_ptr, texel_2 ] | |
5304 | and texel_4, texel_shift_mask, block_pixels_b, lsl #1 | |
5305 | ||
5306 | ldrh texel_3, [ clut_ptr, texel_3 ] | |
5307 | and texel_5, texel_shift_mask, block_pixels_b, lsr #7 | |
5308 | ||
5309 | ldrh texel_4, [ clut_ptr, texel_4 ] | |
5310 | and texel_6, texel_shift_mask, block_pixels_b, lsr #15 | |
5311 | ||
5312 | ldrh texel_5, [ clut_ptr, texel_5 ] | |
5313 | and texel_7, texel_shift_mask, block_pixels_b, lsr #23 | |
5314 | ||
5315 | ldrh texel_6, [ clut_ptr, texel_6 ] | |
5316 | orr texels_01, texel_0, texel_1, lsl #16 | |
5317 | ||
5318 | ldrh texel_7, [ clut_ptr, texel_7 ] | |
5319 | orr texels_23, texel_2, texel_3, lsl #16 | |
5320 | ||
5321 | orr texels_45, texel_4, texel_5, lsl #16 | |
5322 | str texels_01, [ block_ptr, #0 ] | |
5323 | ||
5324 | orr texels_67, texel_6, texel_7, lsl #16 | |
5325 | str texels_23, [ block_ptr, #4 ] | |
5326 | ||
5327 | subs num_blocks, num_blocks, #1 | |
5328 | str texels_45, [ block_ptr, #8 ] | |
5329 | ||
5330 | str texels_67, [ block_ptr, #12 ] | |
5331 | add block_ptr, block_ptr, #64 | |
5332 | ||
5333 | bne 0b | |
5334 | ||
5335 | ldmia sp!, { r4 - r11, pc } | |
5336 | ||
5337 | ||
5338 | #undef width_rounded | |
5339 | #undef texture_mask | |
5340 | #undef num_blocks | |
5341 | #undef texture_offset | |
59d15d23 | 5342 | #undef texels_low |
5343 | #undef texels_high | |
5344 | #undef texels_wide_low | |
5345 | #undef texels_wide_high | |
5346 | #undef texels_wide | |
5347 | #undef fb_ptr2 | |
75e28f62 E |
5348 | |
5349 | #define psx_gpu r0 | |
5350 | #define x r1 | |
5351 | #define y r2 | |
5352 | #define u r3 | |
5353 | #define v r4 | |
5354 | #define width r5 | |
5355 | #define height r6 | |
5356 | #define left_offset r8 | |
5357 | #define width_rounded r9 | |
5358 | #define right_width r10 | |
59d15d23 | 5359 | |
75e28f62 E |
5360 | #define block_width r11 |
5361 | ||
5362 | #define texture_offset_base r1 | |
5363 | #define texture_mask r2 | |
5364 | #define texture_page_ptr r3 | |
5365 | #define num_blocks r4 | |
5366 | #define block r5 | |
5367 | #define fb_ptr r7 | |
5368 | #define texture_offset r8 | |
5369 | #define blocks_remaining r9 | |
59d15d23 | 5370 | #define fb_ptr2 r10 |
75e28f62 E |
5371 | #define fb_ptr_pitch r12 |
5372 | #define texture_block_ptr r14 | |
5373 | ||
5374 | #define texture_mask_width r2 | |
5375 | #define texture_mask_height r3 | |
5376 | #define left_mask_bits r4 | |
5377 | #define right_mask_bits r5 | |
5378 | ||
5379 | ||
5380 | #undef block_masks | |
5381 | #undef block_masks_shifted | |
5382 | #undef texels | |
5383 | ||
5384 | #define block_masks d0 | |
5385 | #define block_masks_shifted d1 | |
5386 | #define draw_mask_fb_ptr d2 | |
5387 | #define texels q2 | |
5388 | ||
59d15d23 | 5389 | #define draw_mask_fb_ptr_a d2 |
5390 | #define draw_mask_fb_ptr_b d3 | |
5391 | #define texels_low d4 | |
5392 | #define texels_high d5 | |
5393 | #define texels_wide_low d6 | |
5394 | #define texels_wide_high d7 | |
5395 | #define texels_wide q3 | |
75e28f62 | 5396 | |
75e28f62 | 5397 | |
59d15d23 | 5398 | setup_sprites_16bpp_flush: |
5399 | vpush { d0 - d3 } | |
75e28f62 E |
5400 | |
5401 | stmdb sp!, { r0 - r3, r12, r14 } | |
5402 | bl flush_render_block_buffer | |
5403 | ldmia sp!, { r0 - r3, r12, r14 } | |
5404 | ||
59d15d23 | 5405 | vpop { d0 - d3 } |
75e28f62 E |
5406 | |
5407 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5408 | mov num_blocks, block_width | |
5409 | ||
5410 | bx lr | |
5411 | ||
5412 | function(setup_sprite_16bpp) | |
5413 | stmdb sp!, { r4 - r11, r14 } | |
c1817bd9 | 5414 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] |
75e28f62 E |
5415 | |
5416 | ldr v, [ sp, #36 ] | |
5417 | add fb_ptr, fb_ptr, y, lsl #11 | |
5418 | ||
5419 | ldr width, [ sp, #40 ] | |
5420 | add fb_ptr, fb_ptr, x, lsl #1 | |
5421 | ||
5422 | ldr height, [ sp, #44 ] | |
5423 | and left_offset, u, #0x7 | |
5424 | ||
5425 | add texture_offset_base, u, u | |
5426 | add width_rounded, width, #7 | |
5427 | ||
5428 | add texture_offset_base, v, lsl #11 | |
5429 | mov left_mask_bits, #0xFF | |
5430 | ||
5431 | ldrb texture_mask_width, [ psx_gpu, #psx_gpu_texture_mask_width_offset ] | |
5432 | add width_rounded, width_rounded, left_offset | |
5433 | ||
5434 | ldrb texture_mask_height, [ psx_gpu, #psx_gpu_texture_mask_height_offset ] | |
5435 | sub fb_ptr, fb_ptr, left_offset, lsl #1 | |
5436 | ||
5437 | add texture_mask, texture_mask_width, texture_mask_width | |
5438 | mov right_mask_bits, #0xFE | |
5439 | ||
5440 | and right_width, width_rounded, #0x7 | |
5441 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
5442 | ||
5443 | add texture_mask, texture_mask_height, lsl #11 | |
5444 | mov block_width, width_rounded, lsr #3 | |
5445 | ||
5446 | mov right_mask_bits, right_mask_bits, lsl right_width | |
5447 | movw fb_ptr_pitch, #(2048 + 16) | |
5448 | ||
5449 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4 | |
5450 | vmov block_masks, left_mask_bits, right_mask_bits | |
5451 | ||
5452 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5453 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5454 | ||
6ea0f7bf | 5455 | bic texture_offset_base, texture_offset_base, #0xF |
75e28f62 E |
5456 | cmp block_width, #1 |
5457 | ||
5458 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
5459 | add block, block, num_blocks, lsl #6 | |
5460 | ||
5461 | bne 0f | |
5462 | ||
5463 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
5464 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
5465 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
5466 | ||
5467 | 1: | |
5468 | add num_blocks, num_blocks, #1 | |
5469 | cmp num_blocks, #MAX_BLOCKS | |
59d15d23 | 5470 | blgt setup_sprites_16bpp_flush |
75e28f62 E |
5471 | |
5472 | and texture_block_ptr, texture_offset_base, texture_mask | |
5473 | subs height, height, #1 | |
5474 | ||
5475 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5476 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5477 | ||
5478 | vst1.u32 { texels }, [ block, :128 ] | |
5479 | add block, block, #40 | |
5480 | ||
5481 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
5482 | pld [ fb_ptr ] | |
5483 | ||
5484 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
5485 | ||
5486 | add block, block, #24 | |
5487 | add texture_offset_base, texture_offset_base, #2048 | |
5488 | add fb_ptr, fb_ptr, #2048 | |
5489 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5490 | bne 1b | |
5491 | ||
5492 | ldmia sp!, { r4 - r11, pc } | |
5493 | ||
5494 | 0: | |
5495 | add num_blocks, num_blocks, block_width | |
5496 | mov texture_offset, texture_offset_base | |
5497 | ||
5498 | cmp num_blocks, #MAX_BLOCKS | |
59d15d23 | 5499 | blgt setup_sprites_16bpp_flush |
75e28f62 E |
5500 | |
5501 | add texture_offset_base, texture_offset_base, #2048 | |
5502 | and texture_block_ptr, texture_offset, texture_mask | |
5503 | ||
5504 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5505 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5506 | ||
5507 | vst1.u32 { texels }, [ block, :128 ] | |
5508 | add block, block, #40 | |
5509 | ||
5510 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
5511 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
5512 | pld [ fb_ptr ] | |
5513 | ||
5514 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
5515 | subs blocks_remaining, block_width, #2 | |
5516 | ||
5517 | add texture_offset, texture_offset, #16 | |
5518 | add fb_ptr, fb_ptr, #16 | |
5519 | ||
5520 | vmov.u8 draw_mask_fb_ptr, #0 | |
5521 | ||
5522 | add block, block, #24 | |
5523 | beq 2f | |
5524 | ||
5525 | 1: | |
5526 | and texture_block_ptr, texture_offset, texture_mask | |
5527 | subs blocks_remaining, blocks_remaining, #1 | |
5528 | ||
5529 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5530 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5531 | ||
5532 | vst1.u32 { texels }, [ block, :128 ] | |
5533 | add block, block, #40 | |
5534 | ||
5535 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
5536 | pld [ fb_ptr ] | |
5537 | ||
5538 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
5539 | ||
5540 | add texture_offset, texture_offset, #16 | |
5541 | add fb_ptr, fb_ptr, #16 | |
5542 | ||
5543 | add block, block, #24 | |
5544 | bne 1b | |
5545 | ||
5546 | 2: | |
5547 | and texture_block_ptr, texture_offset, texture_mask | |
5548 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5549 | ||
5550 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5551 | vdup.u8 draw_mask_fb_ptr, block_masks[4] | |
5552 | ||
5553 | vst1.u32 { texels }, [ block, :128 ] | |
5554 | add block, block, #40 | |
5555 | ||
5556 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
5557 | vst1.u32 { draw_mask_fb_ptr }, [ block, :64 ] | |
5558 | ||
5559 | add block, block, #24 | |
5560 | subs height, height, #1 | |
5561 | ||
5562 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
5563 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5564 | ||
5565 | bne 0b | |
5566 | ||
5567 | ldmia sp!, { r4 - r11, pc } | |
5568 | ||
5569 | ||
59d15d23 | 5570 | // 4x version |
5571 | // FIXME: duplicate code with normal version :( | |
5572 | #undef draw_mask_fb_ptr | |
5573 | ||
5574 | function(setup_sprite_16bpp_4x) | |
5575 | stmdb sp!, { r4 - r11, r14 } | |
5576 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] | |
5577 | ||
5578 | ldr v, [ sp, #36 ] | |
5579 | add fb_ptr, fb_ptr, y, lsl #11 | |
5580 | ||
5581 | ldr width, [ sp, #40 ] | |
5582 | add fb_ptr, fb_ptr, x, lsl #1 | |
5583 | ||
5584 | ldr height, [ sp, #44 ] | |
5585 | and left_offset, u, #0x7 | |
5586 | ||
5587 | add texture_offset_base, u, u | |
5588 | add width_rounded, width, #7 | |
5589 | ||
5590 | add texture_offset_base, v, lsl #11 | |
5591 | movw left_mask_bits, #0xFFFF | |
5592 | ||
5593 | ldrb texture_mask_width, [ psx_gpu, #psx_gpu_texture_mask_width_offset ] | |
5594 | add width_rounded, width_rounded, left_offset | |
5595 | ||
5596 | lsl left_offset, #1 | |
5597 | ||
5598 | ldrb texture_mask_height, [ psx_gpu, #psx_gpu_texture_mask_height_offset ] | |
5599 | sub fb_ptr, fb_ptr, left_offset, lsl #1 | |
5600 | ||
5601 | add texture_mask, texture_mask_width, texture_mask_width | |
5602 | movw right_mask_bits, #0xFFFC | |
5603 | ||
5604 | and right_width, width_rounded, #0x7 | |
5605 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
5606 | ||
5607 | lsl right_width, #1 | |
5608 | ||
5609 | add texture_mask, texture_mask_height, lsl #11 | |
5610 | mov block_width, width_rounded, lsr #3 | |
5611 | ||
5612 | mov right_mask_bits, right_mask_bits, lsl right_width | |
5613 | movw fb_ptr_pitch, #(2048 + 16) * 2 | |
5614 | ||
5615 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4+1 | |
5616 | vmov block_masks, left_mask_bits, right_mask_bits | |
5617 | ||
5618 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5619 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5620 | ||
5621 | bic texture_offset_base, texture_offset_base, #0xF | |
5622 | cmp block_width, #1 | |
5623 | ||
5624 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_ptr_offset ] | |
5625 | add block, block, num_blocks, lsl #6 | |
5626 | ||
5627 | lsl block_width, #2 | |
5628 | bne 0f | |
5629 | ||
5630 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
5631 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
5632 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] | |
5633 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
5634 | ||
5635 | 1: | |
5636 | add num_blocks, num_blocks, block_width | |
5637 | cmp num_blocks, #MAX_BLOCKS | |
5638 | blgt setup_sprites_16bpp_flush | |
5639 | ||
5640 | and texture_block_ptr, texture_offset_base, texture_mask | |
5641 | subs height, height, #1 | |
5642 | ||
5643 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5644 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5645 | ||
5646 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5647 | ||
5648 | add texture_offset_base, texture_offset_base, #2048 | |
5649 | add fb_ptr, fb_ptr, #2048*2 | |
5650 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5651 | bne 1b | |
5652 | ||
5653 | ldmia sp!, { r4 - r11, pc } | |
5654 | ||
5655 | 0: | |
5656 | add num_blocks, num_blocks, block_width | |
5657 | mov texture_offset, texture_offset_base | |
5658 | ||
5659 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] // left_mask_bits | |
5660 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
5661 | ||
5662 | cmp num_blocks, #MAX_BLOCKS | |
5663 | blgt setup_sprites_16bpp_flush | |
5664 | ||
5665 | add texture_offset_base, texture_offset_base, #2048 | |
5666 | and texture_block_ptr, texture_offset, texture_mask | |
5667 | ||
5668 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5669 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5670 | ||
5671 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5672 | ||
5673 | subs blocks_remaining, block_width, #2*4 | |
5674 | add texture_offset, texture_offset, #16 | |
5675 | ||
5676 | vmov.u8 draw_mask_fb_ptr_a, #0 | |
5677 | vmov.u8 draw_mask_fb_ptr_b, #0 | |
5678 | ||
5679 | add fb_ptr, fb_ptr, #16*2 | |
5680 | beq 2f | |
5681 | ||
5682 | 1: | |
5683 | and texture_block_ptr, texture_offset, texture_mask | |
5684 | subs blocks_remaining, blocks_remaining, #4 | |
5685 | ||
5686 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5687 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5688 | ||
5689 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5690 | add texture_offset, texture_offset, #16 | |
5691 | ||
5692 | add fb_ptr, fb_ptr, #16*2 | |
5693 | bgt 1b | |
5694 | ||
5695 | 2: | |
5696 | vdup.u8 draw_mask_fb_ptr_a, block_masks[4] // right_mask_bits | |
5697 | vdup.u8 draw_mask_fb_ptr_b, block_masks[5] | |
5698 | ||
5699 | and texture_block_ptr, texture_offset, texture_mask | |
5700 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5701 | ||
5702 | vld1.u32 { texels }, [ texture_block_ptr, :128 ] | |
5703 | ||
5704 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5705 | subs height, height, #1 | |
5706 | ||
5707 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
5708 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5709 | ||
5710 | bne 0b | |
5711 | ||
5712 | ldmia sp!, { r4 - r11, pc } | |
5713 | ||
5714 | ||
f0931e56 | 5715 | #undef width |
5716 | #undef right_width | |
5717 | #undef right_mask_bits | |
5718 | #undef color | |
5719 | #undef height | |
5720 | #undef blocks_remaining | |
5721 | #undef colors | |
5722 | #undef right_mask | |
5723 | #undef test_mask | |
5724 | #undef draw_mask | |
5725 | ||
5726 | #define psx_gpu r0 | |
5727 | #define x r1 | |
5728 | #define y r2 | |
5729 | #define width r3 | |
5730 | #define right_width r5 | |
5731 | #define right_mask_bits r6 | |
5732 | #define fb_ptr r7 | |
5733 | #define color r8 | |
5734 | #define height r9 | |
5735 | #define fb_ptr_pitch r12 | |
5736 | ||
5737 | // referenced by setup_sprites_16bpp_flush | |
5738 | #define num_blocks r4 | |
5739 | #define block r5 | |
5740 | #define block_width r11 | |
5741 | ||
5742 | #define color_r r1 | |
5743 | #define color_g r2 | |
5744 | #define color_b r8 | |
5745 | #define blocks_remaining r6 | |
5746 | ||
5747 | #define colors q0 | |
5748 | #define right_mask q1 | |
5749 | #define test_mask q2 | |
5750 | #define draw_mask q2 | |
5751 | #define draw_mask_bits_fb_ptr d6 | |
5752 | ||
5753 | ||
5754 | .align 3 | |
5755 | ||
5756 | function(setup_sprite_untextured) | |
5757 | ldrh r12, [ psx_gpu, #psx_gpu_render_state_offset ] | |
5758 | tst r12, #(RENDER_STATE_MASK_EVALUATE | RENDER_FLAGS_MODULATE_TEXELS \ | |
5759 | | RENDER_FLAGS_BLEND) | |
5760 | beq setup_sprite_untextured_simple | |
5761 | ||
5762 | stmdb sp!, { r4 - r11, r14 } | |
5763 | ||
5764 | ldr width, [ sp, #40 ] | |
5765 | ldr fb_ptr, [ psx_gpu, #psx_gpu_vram_out_ptr_offset ] | |
5766 | ||
5767 | ldr height, [ sp, #44 ] | |
5768 | add fb_ptr, fb_ptr, y, lsl #11 | |
5769 | ||
5770 | add fb_ptr, fb_ptr, x, lsl #1 | |
5771 | sub right_width, width, #1 | |
5772 | ||
5773 | ldr color, [ sp, #48 ] | |
5774 | and right_width, #7 | |
5775 | ||
5776 | add block_width, width, #7 | |
5777 | add right_width, #1 | |
5778 | ||
5779 | lsr block_width, #3 | |
5780 | mov right_mask_bits, #0xff | |
5781 | ||
5782 | sub fb_ptr_pitch, block_width, #1 | |
5783 | lsl right_mask_bits, right_width | |
5784 | ||
5785 | lsl fb_ptr_pitch, #3+1 | |
5786 | ubfx color_r, color, #3, #5 | |
5787 | ||
5788 | rsb fb_ptr_pitch, #1024*2 | |
5789 | ubfx color_g, color, #11, #5 | |
5790 | ||
5791 | vld1.u32 { test_mask }, [ psx_gpu, :128 ] | |
5792 | ubfx color_b, color, #19, #5 | |
5793 | ||
5794 | vdup.u16 right_mask, right_mask_bits | |
5795 | orr color, color_r, color_b, lsl #10 | |
5796 | ||
5797 | ldrh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5798 | orr color, color, color_g, lsl #5 | |
5799 | ||
5800 | vtst.u16 right_mask, right_mask, test_mask | |
5801 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5802 | ||
5803 | vdup.u16 colors, color | |
5804 | add block, block, num_blocks, lsl #6 | |
5805 | ||
5806 | ||
5807 | setup_sprite_untextured_height_loop: | |
5808 | add num_blocks, block_width | |
5809 | sub blocks_remaining, block_width, #1 | |
5810 | ||
5811 | cmp num_blocks, #MAX_BLOCKS | |
5812 | blgt setup_sprites_16bpp_flush | |
5813 | ||
5814 | cmp blocks_remaining, #0 | |
5815 | ble 1f | |
5816 | ||
5817 | vmov.u8 draw_mask, #0 /* zero_mask */ | |
5818 | vmov.u8 draw_mask_bits_fb_ptr, #0 | |
5819 | ||
5820 | 0: | |
5821 | vst1.u32 { draw_mask }, [ block, :128 ]! | |
5822 | subs blocks_remaining, #1 | |
5823 | ||
5824 | vst1.u32 { colors }, [ block, :128 ] | |
5825 | add block, block, #24 | |
5826 | ||
5827 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
5828 | vst1.u32 { draw_mask_bits_fb_ptr }, [ block, :64 ] | |
5829 | ||
5830 | add block, block, #24 | |
5831 | add fb_ptr, #8*2 | |
5832 | bgt 0b | |
5833 | ||
5834 | 1: | |
5835 | vst1.u32 { right_mask }, [ block, :128 ]! | |
5836 | subs height, #1 | |
5837 | ||
5838 | vst1.u32 { colors }, [ block, :128 ] | |
5839 | add block, block, #24 | |
5840 | ||
5841 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
5842 | vst1.u32 { draw_mask_bits_fb_ptr }, [ block, :64 ] | |
5843 | ||
5844 | add block, block, #24 | |
5845 | add fb_ptr, fb_ptr_pitch | |
5846 | ||
5847 | strh num_blocks, [ psx_gpu, #psx_gpu_num_blocks_offset ] | |
5848 | bgt setup_sprite_untextured_height_loop | |
5849 | ||
5850 | ldmia sp!, { r4 - r11, pc } | |
5851 | ||
5852 | ||
5853 | ||
75e28f62 E |
5854 | #undef texture_page_ptr |
5855 | #undef vram_ptr | |
5856 | #undef dirty_textures_mask | |
5857 | #undef current_texture_mask | |
5858 | ||
5859 | #define psx_gpu r0 | |
5860 | #define current_texture_page r1 | |
5861 | #define texture_page_ptr r2 | |
5862 | #define vram_ptr_a r3 | |
5863 | #define current_texture_page_x r12 | |
5864 | #define current_texture_page_y r4 | |
5865 | #define dirty_textures_mask r5 | |
5866 | #define tile_y r6 | |
5867 | #define tile_x r7 | |
5868 | #define sub_y r8 | |
5869 | #define current_texture_mask r9 | |
5870 | #define c_4096 r10 | |
5871 | #define vram_ptr_b r11 | |
5872 | ||
5873 | #define texel_block_a d0 | |
5874 | #define texel_block_b d1 | |
5875 | #define texel_block_expanded_a q1 | |
5876 | #define texel_block_expanded_b q2 | |
5877 | #define texel_block_expanded_ab q2 | |
5878 | #define texel_block_expanded_c q3 | |
5879 | #define texel_block_expanded_d q4 | |
5880 | #define texel_block_expanded_cd q3 | |
5881 | ||
5882 | function(update_texture_4bpp_cache) | |
5883 | stmdb sp!, { r4 - r11, r14 } | |
5884 | vpush { q0 - q3 } | |
5885 | ||
5886 | ldrb current_texture_page, [ psx_gpu, #psx_gpu_current_texture_page_offset ] | |
5887 | ||
3867c6ef | 5888 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_base_offset ] |
75e28f62 E |
5889 | ldr vram_ptr_a, [ psx_gpu, #psx_gpu_vram_ptr_offset ] |
5890 | ||
5891 | and current_texture_page_x, current_texture_page, #0xF | |
5892 | ldr current_texture_mask, [ psx_gpu, #psx_gpu_current_texture_mask_offset ] | |
5893 | ||
5894 | mov current_texture_page_y, current_texture_page, lsr #4 | |
5895 | ldr dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ] | |
5896 | ||
5897 | add vram_ptr_a, vram_ptr_a, current_texture_page_y, lsl #19 | |
5898 | mov tile_y, #16 | |
5899 | ||
5900 | add vram_ptr_a, vram_ptr_a, current_texture_page_x, lsl #7 | |
5901 | bic dirty_textures_mask, current_texture_mask | |
5902 | ||
5903 | mov tile_x, #16 | |
5904 | str dirty_textures_mask, [ psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset ] | |
5905 | ||
5906 | mov sub_y, #8 | |
5907 | movw c_4096, #4096 | |
5908 | ||
5909 | add vram_ptr_b, vram_ptr_a, #2048 | |
5910 | ||
5911 | 0: | |
5912 | vld1.u32 { texel_block_a }, [ vram_ptr_a, :64 ], c_4096 | |
5913 | vld1.u32 { texel_block_b }, [ vram_ptr_b, :64 ], c_4096 | |
5914 | ||
5915 | vmovl.u8 texel_block_expanded_a, texel_block_a | |
5916 | vshll.u8 texel_block_expanded_b, texel_block_a, #4 | |
5917 | vmovl.u8 texel_block_expanded_c, texel_block_b | |
5918 | vshll.u8 texel_block_expanded_d, texel_block_b, #4 | |
5919 | ||
5920 | vbic.u16 texel_block_expanded_a, #0x00F0 | |
5921 | vbic.u16 texel_block_expanded_b, #0x00F0 | |
5922 | vbic.u16 texel_block_expanded_c, #0x00F0 | |
5923 | vbic.u16 texel_block_expanded_d, #0x00F0 | |
5924 | ||
5925 | vorr.u16 texel_block_expanded_ab, texel_block_expanded_a, \ | |
5926 | texel_block_expanded_b | |
5927 | vorr.u16 texel_block_expanded_cd, texel_block_expanded_c, \ | |
5928 | texel_block_expanded_d | |
5929 | ||
5930 | vst1.u32 { texel_block_expanded_ab, texel_block_expanded_cd }, \ | |
5931 | [ texture_page_ptr, :256 ]! | |
5932 | ||
5933 | subs sub_y, sub_y, #1 | |
5934 | bne 0b | |
5935 | ||
5936 | mov sub_y, #8 | |
5937 | add vram_ptr_a, vram_ptr_a, #8 | |
5938 | add vram_ptr_b, vram_ptr_b, #8 | |
5939 | ||
5940 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
5941 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
5942 | ||
5943 | subs tile_x, tile_x, #1 | |
5944 | bne 0b | |
5945 | ||
5946 | mov tile_x, #16 | |
5947 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
5948 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
5949 | ||
5950 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
5951 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
5952 | ||
5953 | subs tile_y, tile_y, #1 | |
5954 | bne 0b | |
5955 | ||
5956 | vpop { q0 - q3 } | |
5957 | ldmia sp!, { r4 - r11, pc } | |
5958 | ||
5959 | ||
5960 | #undef current_texture_page | |
5961 | ||
5962 | #define psx_gpu r0 | |
5963 | #define texture_page r1 | |
5964 | #define texture_page_ptr r2 | |
5965 | #define vram_ptr_a r3 | |
5966 | #define texture_page_x r12 | |
5967 | #define texture_page_y r4 | |
5968 | #define current_texture_page r5 | |
5969 | #define tile_y r6 | |
5970 | #define tile_x r7 | |
5971 | #define sub_y r8 | |
5972 | #define c_4096 r10 | |
5973 | #define vram_ptr_b r11 | |
5974 | ||
5975 | ||
5976 | #undef texels_a | |
5977 | #undef texels_b | |
5978 | ||
5979 | #define texels_a q0 | |
5980 | #define texels_b q1 | |
5981 | #define texels_c q2 | |
5982 | #define texels_d q3 | |
5983 | ||
5984 | ||
5985 | function(update_texture_8bpp_cache_slice) | |
5986 | stmdb sp!, { r4 - r11, r14 } | |
5987 | vpush { q0 - q3 } | |
5988 | ||
5989 | ldrb current_texture_page, [ psx_gpu, #psx_gpu_current_texture_page_offset ] | |
5990 | ldr vram_ptr_a, [ psx_gpu, #psx_gpu_vram_ptr_offset ] | |
5991 | ||
3867c6ef | 5992 | ldr texture_page_ptr, [ psx_gpu, #psx_gpu_texture_page_base_offset ] |
75e28f62 E |
5993 | mov tile_y, #16 |
5994 | ||
5995 | and texture_page_x, texture_page, #0xF | |
5996 | mov texture_page_y, texture_page, lsr #4 | |
5997 | ||
5998 | add vram_ptr_a, vram_ptr_a, texture_page_x, lsl #7 | |
5999 | mov tile_x, #8 | |
6000 | ||
6001 | add vram_ptr_a, vram_ptr_a, texture_page_y, lsl #19 | |
6002 | eor current_texture_page, current_texture_page, texture_page | |
6003 | ||
6004 | ands current_texture_page, current_texture_page, #0x1 | |
6005 | mov sub_y, #4 | |
6006 | ||
6007 | addne texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
6008 | movw c_4096, #4096 | |
6009 | ||
6010 | add vram_ptr_b, vram_ptr_a, #2048 | |
6011 | ||
6012 | 0: | |
6013 | vld1.u32 { texels_a }, [ vram_ptr_a, :128 ], c_4096 | |
6014 | vld1.u32 { texels_b }, [ vram_ptr_b, :128 ], c_4096 | |
6015 | vld1.u32 { texels_c }, [ vram_ptr_a, :128 ], c_4096 | |
6016 | vld1.u32 { texels_d }, [ vram_ptr_b, :128 ], c_4096 | |
6017 | ||
6018 | vst1.u32 { texels_a, texels_b }, [ texture_page_ptr, :256 ]! | |
6019 | vst1.u32 { texels_c, texels_d }, [ texture_page_ptr, :256 ]! | |
6020 | ||
6021 | subs sub_y, sub_y, #1 | |
6022 | bne 0b | |
6023 | ||
6024 | mov sub_y, #4 | |
6025 | ||
6026 | add vram_ptr_a, vram_ptr_a, #16 | |
6027 | add vram_ptr_b, vram_ptr_b, #16 | |
6028 | ||
6029 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6030 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6031 | ||
6032 | subs tile_x, tile_x, #1 | |
6033 | bne 0b | |
6034 | ||
6035 | mov tile_x, #8 | |
6036 | ||
6037 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6038 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6039 | ||
6040 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
6041 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
6042 | ||
6043 | subs tile_y, tile_y, #1 | |
6044 | add texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
6045 | ||
6046 | bne 0b | |
6047 | ||
6048 | vpop { q0 - q3 } | |
6049 | ldmia sp!, { r4 - r11, pc } | |
6050 | ||
50f9355a | 6051 | |
6052 | /* void scale2x_tiles8(void *dst, const void *src, int w8, int h) */ | |
6053 | function(scale2x_tiles8) | |
6054 | push { r4, r14 } | |
6055 | ||
6056 | mov r4, r1 | |
6057 | add r12, r0, #1024*2 | |
6058 | mov r14, r2 | |
6059 | ||
6060 | 0: | |
6061 | vld1.u16 { q0 }, [ r1, :128 ]! | |
6062 | vld1.u16 { q2 }, [ r1, :128 ]! | |
6063 | vmov q1, q0 | |
6064 | vmov q3, q2 | |
6065 | vzip.16 q0, q1 | |
6066 | vzip.16 q2, q3 | |
6067 | subs r14, #2 | |
6068 | vst1.u16 { q0, q1 }, [ r0, :128 ]! | |
6069 | vst1.u16 { q0, q1 }, [ r12, :128 ]! | |
6070 | blt 1f | |
6071 | vst1.u16 { q2, q3 }, [ r0, :128 ]! | |
6072 | vst1.u16 { q2, q3 }, [ r12, :128 ]! | |
6073 | bgt 0b | |
6074 | 1: | |
6075 | subs r3, #1 | |
6076 | mov r14, r2 | |
6077 | add r0, #1024*2*2 | |
6078 | add r4, #1024*2 | |
6079 | sub r0, r2, lsl #4+1 | |
6080 | mov r1, r4 | |
6081 | add r12, r0, #1024*2 | |
6082 | bgt 0b | |
6083 | nop | |
6084 | ||
6085 | pop { r4, pc } | |
59d15d23 | 6086 | |
6087 | // vim:filetype=armasm |