Commit | Line | Data |
---|---|---|
75e28f62 E |
1 | /* |
2 | * Copyright (C) 2011 Gilead Kutnick "Exophase" <exophase@gmail.com> | |
59d15d23 | 3 | * Copyright (C) 2012 GraÅžvydas Ignotas "notaz" <notasas@gmail.com> |
75e28f62 E |
4 | * |
5 | * This program is free software; you can redistribute it and/or | |
6 | * modify it under the terms of the GNU General Public License as | |
7 | * published by the Free Software Foundation; either version 2 of | |
8 | * the License, or (at your option) any later version. | |
9 | * | |
10 | * This program is distributed in the hope that it will be useful, | |
11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
13 | * General Public License for more details. | |
14 | */ | |
15 | ||
16 | #define MAX_SPANS 512 | |
17 | #define MAX_BLOCKS 64 | |
18 | #define MAX_BLOCKS_PER_ROW 128 | |
19 | ||
f0931e56 | 20 | #define RENDER_STATE_MASK_EVALUATE 0x20 |
21 | #define RENDER_FLAGS_MODULATE_TEXELS 0x1 | |
22 | #define RENDER_FLAGS_BLEND 0x2 | |
d5c08ed3 | 23 | #define RENDER_INTERLACE_ENABLED 0x1 |
f0931e56 | 24 | |
cb88320b | 25 | #include "psx_gpu_offsets.h" |
75e28f62 | 26 | |
cb88320b | 27 | #define psx_gpu_b_dx_offset (psx_gpu_b_block_span_offset + 4) |
75e28f62 | 28 | |
75e28f62 E |
29 | #define edge_data_left_x_offset 0 |
30 | #define edge_data_num_blocks_offset 2 | |
31 | #define edge_data_right_mask_offset 4 | |
32 | #define edge_data_y_offset 6 | |
33 | ||
ed0fd81d | 34 | .syntax unified |
35 | .text | |
75e28f62 | 36 | |
9775fc33 | 37 | #if 0 |
38 | #define save_abi_regs() \ | |
39 | vpush {q4-q7} | |
40 | #define restore_abi_regs() \ | |
41 | vpop {q4-q7} | |
42 | #else | |
43 | #define save_abi_regs() | |
44 | #define restore_abi_regs() | |
45 | #endif | |
46 | ||
75e28f62 E |
47 | #define psx_gpu r0 |
48 | #define v_a r1 | |
49 | #define v_b r2 | |
50 | #define v_c r3 | |
51 | ||
52 | #define x0 r4 | |
53 | #define x1 r5 | |
54 | #define x2 r6 | |
55 | #define x0_x1 r5 | |
56 | #define x1_x2 r6 | |
57 | #define y0 r7 | |
58 | #define y1 r8 | |
59 | #define y2 r9 | |
60 | #define y0_y1 r7 | |
61 | #define y1_y2 r8 | |
62 | #define b0 r9 | |
63 | #define b1 r10 | |
64 | #define b2 r11 | |
65 | #define b0_b1 r10 | |
66 | #define b1_b2 r11 | |
67 | ||
68 | ||
69 | #define area_r_s r5 | |
70 | ||
71 | #define g_bx0 r2 | |
72 | #define g_bx r3 | |
73 | #define g_bx2 r4 | |
74 | #define g_bx3 r5 | |
75 | #define b_base r6 | |
76 | #define g_by r8 | |
77 | ||
78 | #define gs_bx r7 | |
79 | #define gs_by r10 | |
80 | ||
81 | #define ga_bx g_bx | |
82 | #define ga_by g_by | |
83 | ||
84 | #define gw_bx_h g_bx | |
85 | #define gw_by_h g_by | |
86 | ||
87 | #define gw_bx_l r11 | |
88 | #define gw_by_l gw_bx_l | |
89 | ||
90 | #define store_a r0 | |
91 | #define store_b r1 | |
92 | #define store_inc r5 | |
93 | ||
94 | ||
95 | #define v0 q0 | |
96 | #define uvrgb0 d0 | |
97 | #define x0_y0 d1 | |
98 | ||
99 | #define v1 q1 | |
100 | #define uvrgb1 d2 | |
101 | #define x1_y1 d3 | |
102 | ||
103 | #define v2 q2 | |
104 | #define uvrgb2 d4 | |
105 | #define x2_y2 d5 | |
106 | ||
107 | #define x0_ab q3 | |
108 | #define uvrg_xxxx0 q3 | |
109 | #define uvrg0 d6 | |
110 | #define xxxx0 d7 | |
111 | ||
112 | #define x1_ab q4 | |
113 | #define uvrg_xxxx1 q4 | |
114 | #define uvrg1 d8 | |
115 | #define xxxx1 d9 | |
116 | ||
117 | #define x2_ab q5 | |
118 | #define uvrg_xxxx2 q5 | |
119 | #define uvrg2 d10 | |
120 | #define xxxx2 d11 | |
121 | ||
122 | #define y0_ab q6 | |
123 | #define yyyy_uvrg0 q6 | |
124 | #define yyyy0 d12 | |
125 | #define uvrg0b d13 | |
126 | ||
127 | #define y1_ab q7 | |
128 | #define yyyy_uvrg1 q7 | |
129 | #define yyyy1 d14 | |
130 | #define uvrg1b d15 | |
131 | ||
132 | #define y2_ab q8 | |
133 | #define yyyy_uvrg2 q8 | |
134 | #define yyyy2 d16 | |
135 | #define uvrg2b d17 | |
136 | ||
137 | #define d0_ab q9 | |
138 | #define d0_a d18 | |
139 | #define d0_b d19 | |
140 | ||
141 | #define d1_ab q10 | |
142 | #define d1_a d20 | |
143 | #define d1_b d21 | |
144 | ||
145 | #define d2_ab q11 | |
146 | #define d2_a d22 | |
147 | #define d2_b d23 | |
148 | ||
149 | #define d3_ab q12 | |
150 | #define d3_a d24 | |
151 | #define d3_b d25 | |
152 | ||
153 | #define ga_uvrg_x q1 | |
154 | #define ga_uvrg_y q4 | |
155 | ||
156 | #define dx x0_x1 | |
157 | #define dy y0_y1 | |
158 | #define db b0_b1 | |
159 | ||
160 | #define uvrg_base q11 | |
161 | ||
162 | #define gs_uvrg_x q5 | |
163 | #define gs_uvrg_y q6 | |
164 | ||
165 | #define g_uvrg_x q1 | |
166 | #define ga_uv_x d2 | |
167 | #define g_uv_x d2 | |
168 | #define ga_rg_x d3 | |
169 | #define g_rg_x d3 | |
170 | ||
171 | #define g_uvrg_y q4 | |
172 | #define ga_uv_y d8 | |
173 | #define g_uv_y d8 | |
174 | #define ga_rg_y d9 | |
175 | #define g_rg_y d9 | |
176 | ||
177 | #define gw_uv_x q1 | |
178 | #define gw_rg_x q2 | |
179 | #define gw_uv_y q4 | |
180 | #define gw_rg_y q3 | |
181 | ||
182 | #define w_mask q9 | |
183 | #define w_mask_l d18 | |
184 | ||
185 | #define r_shift q10 | |
186 | ||
187 | #define uvrg_dx0 q0 | |
188 | #define uvrg_dx0l d0 | |
189 | #define uvrg_dx0h d1 | |
190 | ||
191 | #define uvrg_dx1 q1 | |
192 | #define uvrg_dx1l d2 | |
193 | #define uvrg_dx1h d3 | |
194 | ||
195 | #define uvrg_dx2 q2 | |
196 | #define uvrg_dx2l d4 | |
197 | #define uvrg_dx2h d5 | |
198 | ||
199 | #define uvrg_dx3 q3 | |
200 | #define uvrg_dx3l d6 | |
201 | #define uvrg_dx3h d7 | |
202 | ||
c6063f89 | 203 | #define uvrgb_phase q13 |
75e28f62 E |
204 | |
205 | .align 4 | |
206 | ||
0e4ad319 | 207 | #include "arm_features.h" |
8184d7c5 | 208 | |
0e4ad319 | 209 | #define function(name) FUNCTION(name): |
210 | ||
211 | #ifndef TEXRELS_FORBIDDEN | |
75e28f62 | 212 | |
8184d7c5 | 213 | #define JT_OP_REL(table_label, index_reg, temp) |
214 | #define JT_OP(x...) x | |
215 | #define JTE(start, target) target | |
216 | ||
217 | #else | |
218 | ||
8184d7c5 | 219 | #define JT_OP_REL(table_label, index_reg, temp) \ |
220 | adr temp, table_label; \ | |
e1f6de8f | 221 | ldr temp, [temp, index_reg, lsl #2]; \ |
8184d7c5 | 222 | add pc, pc, temp \ |
223 | ||
224 | #define JT_OP(x...) | |
225 | #define JTE(start, target) (target - start) | |
226 | ||
0e4ad319 | 227 | #endif |
4d646738 | 228 | |
0e4ad319 | 229 | #ifdef __MACH__ |
8184d7c5 | 230 | #define flush_render_block_buffer _flush_render_block_buffer |
231 | #define setup_sprite_untextured_simple _setup_sprite_untextured_simple | |
232 | #define update_texture_8bpp_cache _update_texture_8bpp_cache | |
8184d7c5 | 233 | #endif |
234 | ||
75e28f62 E |
235 | @ r0: psx_gpu |
236 | @ r1: v_a | |
237 | @ r2: v_b | |
238 | @ r3: v_c | |
239 | ||
240 | function(compute_all_gradients) | |
241 | // First compute the triangle area reciprocal and shift. The division will | |
242 | // happen concurrently with much of the work which follows. | |
243 | @ r12 = psx_gpu->triangle_area | |
e1f6de8f | 244 | ldr r12, [psx_gpu, #psx_gpu_triangle_area_offset] |
75e28f62 | 245 | stmdb sp!, { r4 - r11, lr } |
9775fc33 | 246 | save_abi_regs() |
75e28f62 E |
247 | |
248 | @ load exponent of 62 into upper half of double | |
249 | movw r4, #0 | |
250 | clz r14, r12 @ r14 = shift | |
251 | ||
252 | movt r4, #((62 + 1023) << 4) | |
253 | mov r12, r12, lsl r14 @ r12 = triangle_area_normalized | |
254 | ||
255 | @ load area normalized into lower half of double | |
256 | mov r5, r12, lsr #10 | |
257 | vmov.f64 d30, r5, r4 @ d30 = (1 << 62) + ta_n | |
258 | ||
259 | movt r4, #((1022 + 31) << 4) | |
260 | mov r5, r12, lsl #20 | |
261 | ||
262 | add r4, r4, r12, lsr #11 | |
263 | vmov.f64 d31, r5, r4 | |
264 | ||
265 | vdiv.f64 d30, d30, d31 @ d30 = ((1 << 62) + ta_n) / ta_n | |
266 | ||
267 | // ((x1 - x0) * (y2 - y1)) - ((x2 - x1) * (y1 - y0)) = | |
268 | // ( d0 * d1 ) - ( d2 * d3 ) = | |
269 | // ( m0 ) - ( m1 ) = gradient | |
270 | ||
271 | // This is split to do 12 elements at a time over three sets: a, b, and c. | |
272 | // Technically we only need to do 10 elements (uvrgb_x and uvrgb_y), so | |
273 | // two of the slots are unused. | |
274 | ||
275 | // Inputs are all 16-bit signed. The m0/m1 results are 32-bit signed, as | |
276 | // is g. | |
277 | ||
278 | // First type is: uvrg bxxx xxxx | |
279 | // Second type is: yyyy ybyy uvrg | |
280 | // Since x_a and y_c are the same the same variable is used for both. | |
281 | ||
e1f6de8f | 282 | vld1.u32 { v0 }, [v_a, :128] @ v0 = { uvrg0, b0, x0, y0 } |
283 | ldrsh x0, [v_a, #8] @ load x0 | |
75e28f62 | 284 | |
e1f6de8f | 285 | vld1.u32 { v1 }, [v_b, :128] @ v1 = { uvrg1, b1, x1, y1} |
286 | ldrh x1, [v_b, #8] @ load x1 | |
75e28f62 | 287 | |
e1f6de8f | 288 | vld1.u32 { v2 }, [v_c, :128] @ v2 = { uvrg2, b2, x2, y2 } |
289 | ldrh x2, [v_c, #8] @ load x2 | |
75e28f62 E |
290 | |
291 | vmovl.u8 uvrg_xxxx0, uvrgb0 @ uvrg_xxxx0 = { uv0, rg0, b0-, -- } | |
e1f6de8f | 292 | ldrh y0, [v_a, #10] @ load y0 |
75e28f62 E |
293 | |
294 | vmovl.u8 uvrg_xxxx1, uvrgb1 @ uvrg_xxxx1 = { uv1, rg1, b1-, -- } | |
e1f6de8f | 295 | ldrh y1, [v_b, #10] @ load y1 |
75e28f62 E |
296 | |
297 | vmovl.u8 uvrg_xxxx2, uvrgb2 @ uvrg_xxxx2 = { uv2, rg2, b2-, -- } | |
e1f6de8f | 298 | ldrh y2, [v_c, #10] @ load y2 |
75e28f62 E |
299 | |
300 | vmov.u8 uvrg0b, uvrg0 @ uvrg0b = { uv0, rg0 } | |
301 | vdup.u16 xxxx0, x0_y0[0] @ xxxx0 = { xx0, xx0 } | |
302 | ||
303 | orr x1_x2, x1, x2, lsl #16 @ x1_x2 = { x1, x2 } | |
304 | pkhbt x0_x1, x0, x1, lsl #16 @ x0_x1 = { x0, x1 } | |
305 | ||
306 | vmov.u8 uvrg1b, uvrg1 @ uvrg1b = { uv1, rg1 } | |
307 | vdup.u16 xxxx1, x1_y1[0] @ xxxx1 = { xx1, xx1 } | |
308 | ||
309 | vmov.u8 uvrg2b, uvrg2 @ uvrg2b = { uv2, rg2 } | |
310 | vdup.u16 xxxx2, x2_y2[0] @ xxxx2 = { xx2, xx2 } | |
311 | ||
e1f6de8f | 312 | ldrb b2, [v_c, #4] @ load b2 |
75e28f62 E |
313 | orr y0_y1, y0, y1, lsl #16 @ y0_y1 = { y0, y1 } |
314 | ||
e1f6de8f | 315 | ldrb b1, [v_b, #4] @ load b1 |
75e28f62 E |
316 | orr y1_y2, y1, y2, lsl #16 @ y1_y2 = { y1, y2 } |
317 | ||
318 | vdup.u16 yyyy0, x0_y0[1] @ yyyy0 = { yy0, yy0 } | |
319 | vsub.s16 d0_ab, x1_ab, x0_ab | |
320 | ||
e1f6de8f | 321 | ldrb b0, [v_a, #4] @ load b0 |
75e28f62 E |
322 | orr b1_b2, b1, b2, lsl #16 @ b1_b2 = { b1, b2 } |
323 | ||
324 | vdup.u16 yyyy1, x1_y1[1] @ yyyy1 = { yy1, yy1 } | |
325 | vsub.s16 d2_ab, x2_ab, x1_ab | |
326 | ||
327 | vdup.u16 yyyy2, x2_y2[1] @ yyyy2 = { yy2, yy2 } | |
328 | vsub.s16 d1_ab, y2_ab, y1_ab | |
329 | ||
330 | orr b0_b1, b0, b1, lsl #16 @ b1_b2 = { b1, b2 } | |
331 | ssub16 dx, x1_x2, x0_x1 @ dx = { x1 - x0, x2 - x1 } | |
332 | ||
333 | ssub16 dy, y1_y2, y0_y1 @ dy = { y1 - y0, y2 - y1 } | |
334 | ssub16 db, b1_b2, b0_b1 @ db = { b1 - b0, b2 - b1 } | |
335 | ||
336 | vsub.s16 d3_ab, y1_ab, y0_ab | |
337 | smusdx ga_by, dx, db @ ga_by = ((x1 - x0) * (b2 - b1)) - | |
338 | @ ((x2 - X1) * (b1 - b0)) | |
339 | vmull.s16 ga_uvrg_x, d0_a, d1_a | |
340 | smusdx ga_bx, db, dy @ ga_bx = ((b1 - b0) * (y2 - y1)) - | |
341 | @ ((b2 - b1) * (y1 - y0)) | |
342 | vmlsl.s16 ga_uvrg_x, d2_a, d3_a | |
343 | movs gs_bx, ga_bx, asr #31 | |
344 | ||
345 | vmull.s16 ga_uvrg_y, d0_b, d1_b | |
346 | rsbmi ga_bx, ga_bx, #0 | |
347 | ||
c6063f89 | 348 | @ r12 = psx_gpu->uvrgb_phase |
e1f6de8f | 349 | ldr r12, [psx_gpu, #psx_gpu_uvrgb_phase_offset] |
c6063f89 | 350 | |
75e28f62 E |
351 | vmlsl.s16 ga_uvrg_y, d2_b, d3_b |
352 | movs gs_by, ga_by, asr #31 | |
353 | ||
354 | vshr.u64 d0, d30, #22 | |
c6063f89 | 355 | add b_base, r12, b0, lsl #16 |
356 | ||
357 | vdup.u32 uvrgb_phase, r12 | |
75e28f62 E |
358 | |
359 | rsbmi ga_by, ga_by, #0 | |
360 | vclt.s32 gs_uvrg_x, ga_uvrg_x, #0 @ gs_uvrg_x = ga_uvrg_x < 0 | |
361 | ||
362 | @ r12 = psx_gpu->triangle_winding_offset | |
e1f6de8f | 363 | ldrb r12, [psx_gpu, #psx_gpu_triangle_winding_offset] |
75e28f62 E |
364 | vclt.s32 gs_uvrg_y, ga_uvrg_y, #0 @ gs_uvrg_y = ga_uvrg_y < 0 |
365 | ||
75e28f62 E |
366 | rsb r12, r12, #0 @ r12 = -(triangle->winding) |
367 | ||
368 | vdup.u32 w_mask, r12 @ w_mask = { -w, -w, -w, -w } | |
369 | sub r14, r14, #(62 - 12) @ r14 = shift - (62 - FIXED_BITS) | |
370 | ||
371 | vshll.u16 uvrg_base, uvrg0, #16 @ uvrg_base = uvrg0 << 16 | |
afcb5133 | 372 | vdup.u32 r_shift, r14 @ r_shift = { shift, shift*, shift, shift* } |
373 | @ * - vshl.u64: ignored by hw | |
c6063f89 | 374 | vadd.u32 uvrg_base, uvrgb_phase |
75e28f62 E |
375 | vabs.s32 ga_uvrg_x, ga_uvrg_x @ ga_uvrg_x = abs(ga_uvrg_x) |
376 | ||
377 | vmov area_r_s, s0 @ area_r_s = triangle_reciprocal | |
378 | vabs.s32 ga_uvrg_y, ga_uvrg_y @ ga_uvrg_y = abs(ga_uvrg_y) | |
379 | ||
380 | vmull.u32 gw_rg_x, ga_rg_x, d0[0] | |
381 | vmull.u32 gw_uv_x, ga_uv_x, d0[0] | |
382 | vmull.u32 gw_rg_y, ga_rg_y, d0[0] | |
383 | vmull.u32 gw_uv_y, ga_uv_y, d0[0] | |
384 | ||
385 | vshl.u64 gw_rg_x, gw_rg_x, r_shift | |
386 | vshl.u64 gw_uv_x, gw_uv_x, r_shift | |
387 | vshl.u64 gw_rg_y, gw_rg_y, r_shift | |
388 | vshl.u64 gw_uv_y, gw_uv_y, r_shift | |
389 | ||
390 | veor.u32 gs_uvrg_x, gs_uvrg_x, w_mask | |
391 | vmovn.u64 g_uv_x, gw_uv_x | |
392 | ||
393 | veor.u32 gs_uvrg_y, gs_uvrg_y, w_mask | |
394 | vmovn.u64 g_rg_x, gw_rg_x | |
395 | ||
396 | veor.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
397 | vmovn.u64 g_uv_y, gw_uv_y | |
398 | ||
399 | vsub.u32 g_uvrg_x, g_uvrg_x, gs_uvrg_x | |
400 | vmovn.u64 g_rg_y, gw_rg_y | |
401 | ||
402 | veor.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
403 | mov ga_bx, ga_bx, lsl #13 | |
404 | ||
405 | vsub.u32 g_uvrg_y, g_uvrg_y, gs_uvrg_y | |
406 | mov ga_by, ga_by, lsl #13 | |
407 | ||
408 | vdup.u32 x0_y0, x0 | |
409 | umull gw_bx_l, gw_bx_h, ga_bx, area_r_s | |
410 | ||
411 | vshl.u32 g_uvrg_x, g_uvrg_x, #4 | |
412 | vshl.u32 g_uvrg_y, g_uvrg_y, #4 | |
413 | ||
414 | umull gw_by_l, gw_by_h, ga_by, area_r_s | |
415 | vmls.s32 uvrg_base, ga_uvrg_x, x0_y0[0] | |
416 | ||
417 | eor gs_bx, gs_bx, r12 | |
418 | vadd.u32 uvrg_dx2, uvrg_dx1, uvrg_dx1 | |
419 | ||
420 | veor.u32 uvrg_dx0, uvrg_dx0, uvrg_dx0 | |
421 | eor gs_by, gs_by, r12 | |
422 | ||
423 | rsb r11, r14, #0 @ r11 = negative shift for scalar lsr | |
424 | add store_a, psx_gpu, #psx_gpu_uvrg_offset | |
425 | ||
426 | sub r11, r11, #(32 - 13) | |
427 | ||
428 | add store_b, store_a, #16 | |
429 | mov store_inc, #32 | |
430 | ||
431 | vadd.u32 uvrg_dx3, uvrg_dx2, uvrg_dx1 | |
e1f6de8f | 432 | vst1.u32 { uvrg_base }, [store_a, :128], store_inc |
75e28f62 | 433 | |
e1f6de8f | 434 | vst1.u32 { uvrg_dx1 }, [store_b, :128], store_inc |
75e28f62 E |
435 | mov g_bx, gw_bx_h, lsr r11 |
436 | ||
e1f6de8f | 437 | vst1.u32 { g_uvrg_y }, [store_a, :128], store_inc |
75e28f62 E |
438 | mov g_by, gw_by_h, lsr r11 |
439 | ||
440 | vst4.u32 { uvrg_dx0l, uvrg_dx1l, uvrg_dx2l, uvrg_dx3l }, \ | |
e1f6de8f | 441 | [store_b, :128], store_inc |
75e28f62 E |
442 | eor g_bx, g_bx, gs_bx |
443 | ||
444 | vst4.u32 { uvrg_dx0h, uvrg_dx1h, uvrg_dx2h, uvrg_dx3h }, \ | |
e1f6de8f | 445 | [store_b, :128], store_inc |
75e28f62 E |
446 | sub g_bx, g_bx, gs_bx |
447 | ||
448 | lsl g_bx, g_bx, #4 | |
449 | eor g_by, g_by, gs_by | |
450 | ||
451 | mls b_base, g_bx, x0, b_base | |
452 | sub g_by, g_by, gs_by | |
453 | ||
454 | lsl g_by, g_by, #4 | |
455 | mov g_bx0, #0 | |
456 | ||
457 | add g_bx2, g_bx, g_bx | |
458 | add g_bx3, g_bx, g_bx2 | |
459 | ||
460 | stmia store_b, { g_bx0, g_bx, g_bx2, g_bx3, b_base, g_by } | |
461 | ||
9775fc33 | 462 | restore_abi_regs() |
75e28f62 E |
463 | ldmia sp!, { r4 - r11, pc } |
464 | ||
465 | ||
466 | #define psx_gpu r0 | |
467 | #define v_a r1 | |
468 | #define v_b r2 | |
469 | #define v_c r3 | |
470 | ||
471 | #define temp r14 | |
472 | ||
473 | #define x_a r4 | |
474 | #define x_b r5 | |
475 | #define x_c r6 | |
476 | #define y_a r1 | |
477 | #define y_b r2 | |
478 | #define y_c r3 | |
479 | ||
480 | #define height_minor_a r7 | |
481 | #define height_minor_b r8 | |
482 | #define height_major r9 | |
483 | #define height r9 | |
484 | ||
485 | #define reciprocal_table_ptr r10 | |
486 | ||
487 | #define edge_alt_low r4 | |
488 | #define edge_alt_high r5 | |
489 | #define edge_dx_dy_alt r6 | |
490 | #define edge_shift_alt r10 | |
491 | ||
492 | #define edge_dx_dy_alt_low r4 | |
493 | #define edge_dx_dy_alt_high r5 | |
494 | ||
495 | #define span_edge_data r4 | |
496 | #define span_uvrg_offset r5 | |
497 | #define span_b_offset r6 | |
498 | ||
499 | #define clip r14 | |
500 | ||
501 | #define b r11 | |
502 | #define b_dy r12 | |
503 | ||
504 | ||
505 | #define alternate_x q0 | |
506 | #define alternate_dx_dy q1 | |
507 | #define alternate_x_32 q2 | |
508 | ||
509 | #define alternate_x_low d0 | |
510 | #define alternate_x_high d1 | |
511 | #define alternate_dx_dy_low d2 | |
512 | #define alternate_dx_dy_high d3 | |
513 | #define alternate_x_32_low d4 | |
514 | #define alternate_x_32_high d5 | |
515 | ||
516 | #define left_x q3 | |
517 | #define right_x q4 | |
518 | #define left_dx_dy q5 | |
519 | #define right_dx_dy q6 | |
520 | #define left_edge q7 | |
521 | #define right_edge q8 | |
522 | ||
523 | #define left_x_low d6 | |
524 | #define left_x_high d7 | |
525 | #define right_x_low d8 | |
526 | #define right_x_high d9 | |
527 | #define left_dx_dy_low d10 | |
528 | #define left_dx_dy_high d11 | |
529 | #define right_dx_dy_low d12 | |
530 | #define right_dx_dy_high d13 | |
531 | #define left_edge_low d14 | |
532 | #define left_edge_high d15 | |
533 | #define right_edge_low d16 | |
534 | #define right_edge_high d17 | |
535 | ||
536 | #define y_mid_point d18 | |
537 | #define c_0x0004 d19 | |
538 | ||
539 | #define left_right_x_16 q11 | |
540 | #define span_shifts_y q12 | |
541 | #define c_0x0001 q13 | |
542 | ||
543 | #define span_shifts d24 | |
544 | #define y_x4 d25 | |
545 | #define c_0xFFFE d26 | |
546 | #define c_0x0007 d27 | |
547 | ||
548 | #define left_right_x_16_low d22 | |
549 | #define left_right_x_16_high d23 | |
550 | ||
551 | #define uvrg q14 | |
552 | #define uvrg_dy q15 | |
553 | ||
554 | #define alternate_x_16 d4 | |
555 | ||
556 | #define v_clip q3 | |
557 | #define v_clip_low d6 | |
558 | ||
559 | #define right_x_32 q10 | |
560 | #define left_x_32 q11 | |
561 | #define alternate_select d24 | |
562 | ||
563 | #define right_x_32_low d20 | |
564 | #define right_x_32_high d21 | |
565 | #define left_x_32_low d22 | |
566 | #define left_x_32_high d23 | |
567 | ||
568 | #define edges_xy q0 | |
569 | #define edges_dx_dy d2 | |
570 | #define edge_shifts d3 | |
571 | #define edge_shifts_64 q2 | |
572 | ||
573 | #define edges_xy_left d0 | |
574 | #define edges_xy_right d1 | |
575 | ||
576 | #define height_reciprocals d6 | |
577 | #define heights d7 | |
578 | ||
579 | #define widths d8 | |
580 | #define c_0x01 d9 | |
581 | #define x_starts d10 | |
582 | #define x_ends d11 | |
583 | ||
584 | #define heights_b d12 | |
585 | #define edges_dx_dy_64 q10 | |
586 | ||
587 | #define edges_dx_dy_64_left d20 | |
588 | #define edges_dx_dy_64_right d21 | |
589 | ||
590 | ||
591 | #define setup_spans_prologue() \ | |
592 | stmdb sp!, { r4 - r11, lr }; \ | |
9775fc33 | 593 | save_abi_regs(); \ |
75e28f62 | 594 | \ |
e1f6de8f | 595 | ldrsh x_a, [v_a, #8]; \ |
596 | ldrsh x_b, [v_b, #8]; \ | |
597 | ldrsh x_c, [v_c, #8]; \ | |
598 | ldrsh y_a, [v_a, #10]; \ | |
599 | ldrsh y_b, [v_b, #10]; \ | |
600 | ldrsh y_c, [v_c, #10]; \ | |
75e28f62 E |
601 | \ |
602 | add temp, psx_gpu, #psx_gpu_uvrg_offset; \ | |
e1f6de8f | 603 | vld1.32 { uvrg }, [temp]; \ |
75e28f62 | 604 | add temp, psx_gpu, #psx_gpu_uvrg_dy_offset; \ |
e1f6de8f | 605 | vld1.32 { uvrg_dy }, [temp]; \ |
606 | ldr reciprocal_table_ptr, [psx_gpu, #psx_gpu_reciprocal_table_ptr_offset]; \ | |
75e28f62 E |
607 | \ |
608 | vmov.u32 c_0x01, #0x01 \ | |
609 | ||
610 | #define setup_spans_load_b() \ | |
e1f6de8f | 611 | ldr b, [psx_gpu, #psx_gpu_b_offset]; \ |
612 | ldr b_dy, [psx_gpu, #psx_gpu_b_dy_offset] \ | |
75e28f62 E |
613 | |
614 | #define setup_spans_prologue_b() \ | |
615 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
616 | add temp, psx_gpu, #psx_gpu_viewport_start_x_offset; \ | |
617 | \ | |
618 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
619 | vmov.u16 c_0x0004, #0x0004; \ | |
620 | \ | |
621 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
622 | vmov.u16 c_0x0001, #0x0001; \ | |
623 | \ | |
e1f6de8f | 624 | vld1.u16 { left_edge_low[], left_edge_high[] }, [temp]; \ |
75e28f62 E |
625 | add temp, psx_gpu, #psx_gpu_viewport_end_x_offset; \ |
626 | \ | |
e1f6de8f | 627 | vld1.u16 { right_edge_low[], right_edge_high[] }, [temp]; \ |
75e28f62 E |
628 | vadd.u16 right_edge, right_edge, c_0x0001; \ |
629 | \ | |
630 | vmov.u16 c_0x0007, #0x0007; \ | |
631 | vmvn.u16 c_0xFFFE, #0x0001 \ | |
632 | ||
633 | ||
634 | #define compute_edge_delta_x2() \ | |
e1f6de8f | 635 | ldr temp, [reciprocal_table_ptr, height, lsl #2]; \ |
75e28f62 E |
636 | \ |
637 | vdup.u32 heights, height; \ | |
638 | vsub.u32 widths, x_ends, x_starts; \ | |
639 | \ | |
640 | vdup.u32 edge_shifts, temp; \ | |
641 | vsub.u32 heights_b, heights, c_0x01; \ | |
7d5140f5 | 642 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
75e28f62 E |
643 | \ |
644 | vmla.s32 heights_b, x_starts, heights; \ | |
645 | vbic.u16 edge_shifts, #0xE0; \ | |
646 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
647 | vmull.s32 edges_xy, heights_b, height_reciprocals \ | |
648 | ||
649 | #define width_alt r6 | |
650 | #define height_reciprocal_alt r11 | |
651 | #define height_b_alt r12 | |
652 | ||
653 | #define compute_edge_delta_x3(start_c, height_a, height_b) \ | |
ed0fd81d | 654 | vmov heights, height_a, height_b; \ |
e1f6de8f | 655 | ldr temp, [reciprocal_table_ptr, height_a, lsl #2]; \ |
75e28f62 | 656 | vmov.u32 edge_shifts[0], temp; \ |
e1f6de8f | 657 | ldr temp, [reciprocal_table_ptr, height_b, lsl #2]; \ |
75e28f62 | 658 | vmov.u32 edge_shifts[1], temp; \ |
e1f6de8f | 659 | ldr edge_shift_alt, [reciprocal_table_ptr, height_minor_b, lsl #2]; \ |
75e28f62 E |
660 | \ |
661 | vsub.u32 widths, x_ends, x_starts; \ | |
662 | sub width_alt, x_c, start_c; \ | |
663 | \ | |
664 | vsub.u32 heights_b, heights, c_0x01; \ | |
665 | sub height_b_alt, height_minor_b, #1; \ | |
666 | \ | |
7d5140f5 E |
667 | vshr.u32 height_reciprocals, edge_shifts, #10; \ |
668 | lsr height_reciprocal_alt, edge_shift_alt, #10; \ | |
75e28f62 E |
669 | \ |
670 | vmla.s32 heights_b, x_starts, heights; \ | |
671 | mla height_b_alt, height_minor_b, start_c, height_b_alt; \ | |
672 | \ | |
673 | vbic.u16 edge_shifts, #0xE0; \ | |
674 | and edge_shift_alt, edge_shift_alt, #0x1F; \ | |
675 | \ | |
676 | vmul.s32 edges_dx_dy, widths, height_reciprocals; \ | |
677 | mul edge_dx_dy_alt, width_alt, height_reciprocal_alt; \ | |
678 | \ | |
679 | vmull.s32 edges_xy, heights_b, height_reciprocals; \ | |
680 | smull edge_alt_low, edge_alt_high, height_b_alt, height_reciprocal_alt \ | |
681 | ||
682 | ||
683 | #define setup_spans_adjust_y_up() \ | |
684 | vsub.u32 y_x4, y_x4, c_0x0004 \ | |
685 | ||
686 | #define setup_spans_adjust_y_down() \ | |
687 | vadd.u32 y_x4, y_x4, c_0x0004 \ | |
688 | ||
689 | #define setup_spans_adjust_interpolants_up() \ | |
690 | vsub.u32 uvrg, uvrg, uvrg_dy; \ | |
691 | sub b, b, b_dy \ | |
692 | ||
693 | #define setup_spans_adjust_interpolants_down() \ | |
694 | vadd.u32 uvrg, uvrg, uvrg_dy; \ | |
695 | add b, b, b_dy \ | |
696 | ||
697 | ||
698 | #define setup_spans_clip_interpolants_increment() \ | |
699 | mla b, b_dy, clip, b; \ | |
700 | vmla.s32 uvrg, uvrg_dy, v_clip \ | |
701 | ||
702 | #define setup_spans_clip_interpolants_decrement() \ | |
703 | mls b, b_dy, clip, b; \ | |
704 | vmls.s32 uvrg, uvrg_dy, v_clip \ | |
705 | ||
706 | #define setup_spans_clip_alternate_yes() \ | |
707 | smlal edge_alt_low, edge_alt_high, edge_dx_dy_alt, clip \ | |
708 | ||
709 | #define setup_spans_clip_alternate_no() \ | |
710 | ||
711 | #define setup_spans_clip(direction, alternate_active) \ | |
712 | vdup.u32 v_clip, clip; \ | |
713 | setup_spans_clip_alternate_##alternate_active(); \ | |
714 | setup_spans_clip_interpolants_##direction(); \ | |
715 | vmlal.s32 edges_xy, edges_dx_dy, v_clip_low \ | |
716 | ||
717 | ||
718 | #define setup_spans_adjust_edges_alternate_no(left_index, right_index) \ | |
719 | vmovl.s32 edge_shifts_64, edge_shifts; \ | |
720 | vmovl.s32 edges_dx_dy_64, edges_dx_dy; \ | |
721 | \ | |
722 | vshl.s64 edges_xy, edges_xy, edge_shifts_64; \ | |
723 | vshl.s64 edges_dx_dy_64, edges_dx_dy_64, edge_shifts_64; \ | |
724 | \ | |
725 | vmov left_x_low, edges_xy_##left_index; \ | |
726 | vmov right_x_low, edges_xy_##right_index; \ | |
727 | \ | |
728 | vmov left_dx_dy_low, edges_dx_dy_64_##left_index; \ | |
729 | vmov left_dx_dy_high, edges_dx_dy_64_##left_index; \ | |
730 | vmov right_dx_dy_low, edges_dx_dy_64_##right_index; \ | |
731 | vmov right_dx_dy_high, edges_dx_dy_64_##right_index; \ | |
732 | \ | |
733 | vadd.u64 left_x_high, left_x_low, left_dx_dy_low; \ | |
734 | vadd.u64 right_x_high, right_x_low, right_dx_dy_low; \ | |
735 | \ | |
736 | vadd.u64 left_dx_dy, left_dx_dy, left_dx_dy; \ | |
737 | vadd.u64 right_dx_dy, right_dx_dy, right_dx_dy \ | |
738 | ||
739 | ||
740 | #define setup_spans_adjust_edges_alternate_yes(left_index, right_index) \ | |
741 | setup_spans_adjust_edges_alternate_no(left_index, right_index); \ | |
742 | \ | |
743 | vdup.u16 y_mid_point, y_b; \ | |
744 | rsb temp, edge_shift_alt, #32; \ | |
745 | \ | |
746 | lsl edge_alt_high, edge_alt_high, edge_shift_alt; \ | |
747 | orr edge_alt_high, edge_alt_high, edge_alt_low, lsr temp; \ | |
748 | lsl edge_alt_low, edge_alt_low, edge_shift_alt; \ | |
749 | vmov alternate_x_low, edge_alt_low, edge_alt_high; \ | |
750 | \ | |
751 | asr edge_dx_dy_alt_high, edge_dx_dy_alt, temp; \ | |
752 | lsl edge_dx_dy_alt_low, edge_dx_dy_alt, edge_shift_alt; \ | |
753 | vmov alternate_dx_dy_low, edge_dx_dy_alt_low, edge_dx_dy_alt_high; \ | |
754 | vmov alternate_dx_dy_high, alternate_dx_dy_low; \ | |
755 | \ | |
756 | vadd.u64 alternate_x_high, alternate_x_low, alternate_dx_dy_low; \ | |
757 | vadd.u64 alternate_dx_dy, alternate_dx_dy, alternate_dx_dy \ | |
758 | ||
759 | ||
760 | #define setup_spans_y_select_up() \ | |
761 | vclt.s16 alternate_select, y_x4, y_mid_point \ | |
762 | ||
763 | #define setup_spans_y_select_down() \ | |
764 | vcgt.s16 alternate_select, y_x4, y_mid_point \ | |
765 | ||
766 | ||
767 | #define setup_spans_alternate_select_left() \ | |
768 | vbit.u16 left_right_x_16_low, alternate_x_16, alternate_select \ | |
769 | ||
770 | #define setup_spans_alternate_select_right() \ | |
771 | vbit.u16 left_right_x_16_high, alternate_x_16, alternate_select \ | |
772 | ||
773 | ||
774 | #define setup_spans_set_x4_alternate_yes(alternate, direction) \ | |
775 | vshrn.s64 alternate_x_32_low, alternate_x, #32; \ | |
776 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
777 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
778 | \ | |
779 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
780 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
781 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
782 | \ | |
783 | vshrn.s64 alternate_x_32_high, alternate_x, #32; \ | |
784 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
785 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
786 | \ | |
787 | vadd.u64 alternate_x, alternate_x, alternate_dx_dy; \ | |
788 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
789 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
790 | \ | |
791 | vmovn.u32 alternate_x_16, alternate_x_32; \ | |
792 | setup_spans_y_select_##direction(); \ | |
793 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
794 | \ | |
795 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
796 | setup_spans_alternate_select_##alternate(); \ | |
797 | \ | |
e1f6de8f | 798 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
799 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
800 | setup_spans_adjust_interpolants_##direction(); \ |
801 | \ | |
802 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
803 | \ | |
e1f6de8f | 804 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
805 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
806 | setup_spans_adjust_interpolants_##direction(); \ |
807 | \ | |
808 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
809 | \ | |
e1f6de8f | 810 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
811 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
812 | setup_spans_adjust_interpolants_##direction(); \ |
813 | \ | |
814 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
815 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
816 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
817 | \ | |
e1f6de8f | 818 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
819 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
820 | setup_spans_adjust_interpolants_##direction(); \ |
821 | \ | |
822 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
823 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
824 | \ | |
e1f6de8f | 825 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
75e28f62 E |
826 | \ |
827 | setup_spans_adjust_y_##direction() \ | |
828 | ||
829 | ||
830 | #define setup_spans_set_x4_alternate_no(alternate, direction) \ | |
831 | vshrn.s64 left_x_32_low, left_x, #32; \ | |
832 | vshrn.s64 right_x_32_low, right_x, #32; \ | |
833 | \ | |
834 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
835 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
836 | \ | |
837 | vshrn.s64 left_x_32_high, left_x, #32; \ | |
838 | vshrn.s64 right_x_32_high, right_x, #32; \ | |
839 | \ | |
840 | vadd.u64 left_x, left_x, left_dx_dy; \ | |
841 | vadd.u64 right_x, right_x, right_dx_dy; \ | |
842 | \ | |
843 | vmovn.u32 left_right_x_16_low, left_x_32; \ | |
844 | vmovn.u32 left_right_x_16_high, right_x_32; \ | |
845 | \ | |
e1f6de8f | 846 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
847 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
848 | setup_spans_adjust_interpolants_##direction(); \ |
849 | \ | |
850 | vmax.s16 left_right_x_16, left_right_x_16, left_edge; \ | |
851 | \ | |
e1f6de8f | 852 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
853 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
854 | setup_spans_adjust_interpolants_##direction(); \ |
855 | \ | |
856 | vmin.s16 left_right_x_16, left_right_x_16, right_edge; \ | |
857 | \ | |
e1f6de8f | 858 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
859 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
860 | setup_spans_adjust_interpolants_##direction(); \ |
861 | \ | |
862 | vsub.u16 left_right_x_16_high, left_right_x_16_high, left_right_x_16_low; \ | |
863 | vadd.u16 left_right_x_16_high, left_right_x_16_high, c_0x0007; \ | |
864 | vand.u16 span_shifts, left_right_x_16_high, c_0x0007; \ | |
865 | \ | |
e1f6de8f | 866 | vst1.u32 { uvrg }, [span_uvrg_offset, :128]!; \ |
867 | str b, [span_b_offset], #4; \ | |
75e28f62 E |
868 | setup_spans_adjust_interpolants_##direction(); \ |
869 | \ | |
870 | vshl.u16 span_shifts, c_0xFFFE, span_shifts; \ | |
871 | vshr.u16 left_right_x_16_high, left_right_x_16_high, #3; \ | |
872 | \ | |
e1f6de8f | 873 | vst4.u16 { left_right_x_16, span_shifts_y }, [span_edge_data]!; \ |
75e28f62 E |
874 | \ |
875 | setup_spans_adjust_y_##direction() \ | |
876 | ||
877 | ||
878 | #define edge_adjust_low r11 | |
879 | #define edge_adjust_high r12 | |
880 | ||
881 | #define setup_spans_alternate_adjust_yes() \ | |
882 | smull edge_adjust_low, edge_adjust_high, edge_dx_dy_alt, height_minor_a; \ | |
883 | subs edge_alt_low, edge_alt_low, edge_adjust_low; \ | |
884 | sbc edge_alt_high, edge_alt_high, edge_adjust_high \ | |
885 | ||
886 | #define setup_spans_alternate_adjust_no() \ | |
887 | ||
888 | ||
889 | #define setup_spans_down(left_index, right_index, alternate, alternate_active) \ | |
890 | setup_spans_alternate_adjust_##alternate_active(); \ | |
891 | setup_spans_load_b(); \ | |
892 | \ | |
e1f6de8f | 893 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
75e28f62 E |
894 | subs y_c, y_c, temp; \ |
895 | subgt height, height, y_c; \ | |
896 | addgt height, height, #1; \ | |
897 | \ | |
e1f6de8f | 898 | ldrsh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
75e28f62 E |
899 | subs clip, temp, y_a; \ |
900 | ble 0f; \ | |
901 | \ | |
902 | sub height, height, clip; \ | |
903 | add y_a, y_a, clip; \ | |
904 | setup_spans_clip(increment, alternate_active); \ | |
905 | \ | |
906 | 0: \ | |
907 | cmp height, #0; \ | |
908 | ble 1f; \ | |
909 | \ | |
910 | orr temp, y_a, y_a, lsl #16; \ | |
911 | add temp, temp, #(1 << 16); \ | |
912 | add y_a, temp, #2; \ | |
913 | add y_a, y_a, #(2 << 16); \ | |
ed0fd81d | 914 | vmov y_x4, temp, y_a; \ |
75e28f62 E |
915 | \ |
916 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
917 | right_index); \ | |
918 | setup_spans_prologue_b(); \ | |
919 | \ | |
e1f6de8f | 920 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
921 | \ |
922 | 2: \ | |
923 | setup_spans_set_x4_alternate_##alternate_active(alternate, down); \ | |
924 | subs height, height, #4; \ | |
925 | bhi 2b; \ | |
926 | \ | |
927 | 1: \ | |
928 | ||
929 | ||
930 | #define setup_spans_alternate_pre_increment_yes() \ | |
931 | adds edge_alt_low, edge_alt_low, edge_dx_dy_alt; \ | |
932 | adc edge_alt_high, edge_alt_high, edge_dx_dy_alt, asr #31 \ | |
933 | ||
934 | #define setup_spans_alternate_pre_increment_no() \ | |
935 | ||
936 | ||
937 | #define setup_spans_up_decrement_yes() \ | |
938 | suble height, height, #1 \ | |
939 | ||
940 | #define setup_spans_up_decrement_no() \ | |
941 | ||
942 | ||
943 | #define setup_spans_up(left_index, right_index, alternate, alternate_active) \ | |
944 | setup_spans_alternate_adjust_##alternate_active(); \ | |
945 | setup_spans_load_b(); \ | |
946 | sub y_a, y_a, #1; \ | |
947 | \ | |
e1f6de8f | 948 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset]; \ |
75e28f62 E |
949 | subs temp, temp, y_c; \ |
950 | subgt height, height, temp; \ | |
951 | setup_spans_up_decrement_##alternate_active(); \ | |
952 | \ | |
e1f6de8f | 953 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset]; \ |
75e28f62 E |
954 | subs clip, y_a, temp; \ |
955 | ble 0f; \ | |
956 | \ | |
957 | sub height, height, clip; \ | |
958 | sub y_a, y_a, clip; \ | |
959 | setup_spans_clip(decrement, alternate_active); \ | |
960 | \ | |
961 | 0: \ | |
962 | cmp height, #0; \ | |
963 | ble 1f; \ | |
964 | \ | |
965 | orr temp, y_a, y_a, lsl #16; \ | |
966 | sub temp, temp, #(1 << 16); \ | |
967 | sub y_a, temp, #2; \ | |
968 | sub y_a, y_a, #(2 << 16); \ | |
ed0fd81d | 969 | vmov y_x4, temp, y_a; \ |
75e28f62 E |
970 | \ |
971 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy; \ | |
972 | \ | |
973 | setup_spans_alternate_pre_increment_##alternate_active(); \ | |
974 | setup_spans_adjust_edges_alternate_##alternate_active(left_index, \ | |
975 | right_index); \ | |
976 | setup_spans_adjust_interpolants_up(); \ | |
977 | setup_spans_prologue_b(); \ | |
978 | \ | |
e1f6de8f | 979 | strh height, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
980 | \ |
981 | 2: \ | |
982 | setup_spans_set_x4_alternate_##alternate_active(alternate, up); \ | |
983 | subs height, height, #4; \ | |
984 | bhi 2b; \ | |
985 | \ | |
986 | 1: \ | |
987 | ||
988 | ||
989 | #define setup_spans_epilogue() \ | |
9775fc33 | 990 | restore_abi_regs(); \ |
75e28f62 E |
991 | ldmia sp!, { r4 - r11, pc } \ |
992 | ||
993 | ||
994 | #define setup_spans_up_up(minor, major) \ | |
995 | setup_spans_prologue(); \ | |
996 | sub height_minor_a, y_a, y_b; \ | |
997 | sub height_minor_b, y_b, y_c; \ | |
998 | sub height, y_a, y_c; \ | |
999 | \ | |
1000 | vdup.u32 x_starts, x_a; \ | |
ed0fd81d | 1001 | vmov x_ends, x_c, x_b; \ |
75e28f62 E |
1002 | \ |
1003 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
1004 | setup_spans_up(major, minor, minor, yes); \ | |
1005 | setup_spans_epilogue() \ | |
1006 | ||
1007 | function(setup_spans_up_left) | |
1008 | setup_spans_up_up(left, right) | |
1009 | ||
1010 | function(setup_spans_up_right) | |
1011 | setup_spans_up_up(right, left) | |
1012 | ||
75e28f62 E |
1013 | #define setup_spans_down_down(minor, major) \ |
1014 | setup_spans_prologue(); \ | |
1015 | sub height_minor_a, y_b, y_a; \ | |
1016 | sub height_minor_b, y_c, y_b; \ | |
1017 | sub height, y_c, y_a; \ | |
1018 | \ | |
1019 | vdup.u32 x_starts, x_a; \ | |
ed0fd81d | 1020 | vmov x_ends, x_c, x_b; \ |
75e28f62 E |
1021 | \ |
1022 | compute_edge_delta_x3(x_b, height_major, height_minor_a); \ | |
1023 | setup_spans_down(major, minor, minor, yes); \ | |
1024 | setup_spans_epilogue() \ | |
1025 | ||
1026 | function(setup_spans_down_left) | |
1027 | setup_spans_down_down(left, right) | |
1028 | ||
1029 | function(setup_spans_down_right) | |
1030 | setup_spans_down_down(right, left) | |
1031 | ||
1032 | ||
1033 | #define setup_spans_up_flat() \ | |
1034 | sub height, y_a, y_c; \ | |
1035 | \ | |
1036 | compute_edge_delta_x2(); \ | |
1037 | setup_spans_up(left, right, none, no); \ | |
1038 | setup_spans_epilogue() \ | |
1039 | ||
1040 | function(setup_spans_up_a) | |
1041 | setup_spans_prologue() | |
1042 | ||
ed0fd81d | 1043 | vmov x_starts, x_a, x_b |
75e28f62 E |
1044 | vdup.u32 x_ends, x_c |
1045 | ||
1046 | setup_spans_up_flat() | |
1047 | ||
1048 | function(setup_spans_up_b) | |
1049 | setup_spans_prologue() | |
1050 | ||
1051 | vdup.u32 x_starts, x_a | |
ed0fd81d | 1052 | vmov x_ends, x_b, x_c |
75e28f62 E |
1053 | |
1054 | setup_spans_up_flat() | |
1055 | ||
1056 | #define setup_spans_down_flat() \ | |
1057 | sub height, y_c, y_a; \ | |
1058 | \ | |
1059 | compute_edge_delta_x2(); \ | |
1060 | setup_spans_down(left, right, none, no); \ | |
1061 | setup_spans_epilogue() \ | |
1062 | ||
1063 | function(setup_spans_down_a) | |
1064 | setup_spans_prologue() | |
1065 | ||
ed0fd81d | 1066 | vmov x_starts, x_a, x_b |
75e28f62 E |
1067 | vdup.u32 x_ends, x_c |
1068 | ||
1069 | setup_spans_down_flat() | |
1070 | ||
1071 | function(setup_spans_down_b) | |
1072 | setup_spans_prologue() | |
1073 | ||
1074 | vdup.u32 x_starts, x_a | |
ed0fd81d | 1075 | vmov x_ends, x_b, x_c |
75e28f62 E |
1076 | |
1077 | setup_spans_down_flat() | |
1078 | ||
1079 | ||
1080 | #define middle_y r9 | |
1081 | ||
1082 | #define edges_xy_b q11 | |
1083 | #define edges_dx_dy_b d26 | |
1084 | #define edge_shifts_b d27 | |
1085 | #define edges_dx_dy_and_shifts_b q13 | |
1086 | #define height_increment d20 | |
1087 | ||
1088 | #define edges_dx_dy_and_shifts q1 | |
1089 | ||
1090 | #define edges_xy_b_left d22 | |
1091 | #define edges_xy_b_right d23 | |
1092 | ||
1093 | #define setup_spans_up_down_load_edge_set_b() \ | |
1094 | vmov edges_xy, edges_xy_b; \ | |
1095 | vmov edges_dx_dy_and_shifts, edges_dx_dy_and_shifts_b \ | |
1096 | ||
1097 | ||
1098 | function(setup_spans_up_down) | |
1099 | setup_spans_prologue() | |
1100 | ||
1101 | // s32 middle_y = y_a; | |
1102 | sub height_minor_a, y_a, y_b | |
1103 | sub height_minor_b, y_c, y_a | |
1104 | sub height_major, y_c, y_b | |
1105 | ||
ed0fd81d | 1106 | vmov x_starts, x_a, x_c |
75e28f62 E |
1107 | vdup.u32 x_ends, x_b |
1108 | ||
1109 | compute_edge_delta_x3(x_a, height_minor_a, height_major) | |
1110 | ||
1111 | mov temp, #0 | |
ed0fd81d | 1112 | vmov height_increment, temp, height_minor_b |
75e28f62 E |
1113 | vmlal.s32 edges_xy, edges_dx_dy, height_increment |
1114 | ||
1115 | vmov edges_xy_b_left, edge_alt_low, edge_alt_high | |
1116 | vmov edges_xy_b_right, edges_xy_right | |
1117 | ||
1118 | vmov edge_shifts_b, edge_shifts | |
1119 | vmov.u32 edge_shifts_b[0], edge_shift_alt | |
1120 | ||
1121 | vneg.s32 edges_dx_dy_b, edges_dx_dy | |
1122 | vmov.u32 edges_dx_dy_b[0], edge_dx_dy_alt | |
1123 | ||
1124 | mov middle_y, y_a | |
1125 | ||
1126 | setup_spans_load_b() | |
1127 | sub y_a, y_a, #1 | |
1128 | ||
e1f6de8f | 1129 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
75e28f62 E |
1130 | subs temp, temp, y_b |
1131 | subgt height_minor_a, height_minor_a, temp | |
1132 | ||
e1f6de8f | 1133 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
75e28f62 E |
1134 | subs clip, y_a, temp |
1135 | ble 0f | |
1136 | ||
1137 | sub height_minor_a, height_minor_a, clip | |
1138 | sub y_a, y_a, clip | |
1139 | setup_spans_clip(decrement, no) | |
1140 | ||
1141 | 0: | |
1142 | cmp height_minor_a, #0 | |
1143 | ble 3f | |
1144 | ||
1145 | orr temp, y_a, y_a, lsl #16 | |
1146 | sub temp, temp, #(1 << 16) | |
1147 | sub y_a, temp, #2 | |
1148 | sub y_a, y_a, #(2 << 16) | |
ed0fd81d | 1149 | vmov y_x4, temp, y_a |
75e28f62 E |
1150 | |
1151 | vaddw.s32 edges_xy, edges_xy, edges_dx_dy | |
1152 | ||
e1f6de8f | 1153 | strh height_minor_a, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1154 | |
1155 | setup_spans_adjust_edges_alternate_no(left, right); | |
1156 | setup_spans_adjust_interpolants_up() | |
1157 | setup_spans_up_down_load_edge_set_b() | |
1158 | ||
1159 | setup_spans_prologue_b() | |
1160 | ||
1161 | ||
1162 | 2: | |
1163 | setup_spans_set_x4_alternate_no(none, up) | |
1164 | subs height_minor_a, height_minor_a, #4 | |
1165 | bhi 2b | |
1166 | ||
1167 | add span_edge_data, span_edge_data, height_minor_a, lsl #3 | |
1168 | add span_uvrg_offset, span_uvrg_offset, height_minor_a, lsl #4 | |
1169 | add span_b_offset, span_b_offset, height_minor_a, lsl #2 | |
1170 | ||
1171 | 4: | |
1172 | add temp, psx_gpu, #psx_gpu_uvrg_offset | |
e1f6de8f | 1173 | vld1.32 { uvrg }, [temp] |
75e28f62 E |
1174 | mov y_a, middle_y |
1175 | ||
1176 | setup_spans_load_b() | |
1177 | ||
e1f6de8f | 1178 | ldrh temp, [psx_gpu, #psx_gpu_viewport_end_y_offset] |
75e28f62 E |
1179 | subs y_c, y_c, temp |
1180 | subgt height_minor_b, height_minor_b, y_c | |
1181 | addgt height_minor_b, height_minor_b, #1 | |
1182 | ||
e1f6de8f | 1183 | ldrh temp, [psx_gpu, #psx_gpu_viewport_start_y_offset] |
75e28f62 E |
1184 | subs clip, temp, y_a |
1185 | ble 0f | |
1186 | ||
1187 | sub height_minor_b, height_minor_b, clip | |
1188 | add y_a, y_a, clip | |
1189 | setup_spans_clip(increment, no) | |
1190 | ||
1191 | 0: | |
1192 | cmp height_minor_b, #0 | |
1193 | ble 1f | |
1194 | ||
1195 | orr temp, y_a, y_a, lsl #16 | |
1196 | add temp, temp, #(1 << 16) | |
1197 | add y_a, temp, #2 | |
1198 | add y_a, y_a, #(2 << 16) | |
ed0fd81d | 1199 | vmov y_x4, temp, y_a |
75e28f62 E |
1200 | |
1201 | setup_spans_adjust_edges_alternate_no(left, right) | |
1202 | ||
e1f6de8f | 1203 | ldrh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 | 1204 | add temp, temp, height_minor_b |
b7569147 | 1205 | |
1206 | cmp temp, #MAX_SPANS | |
1207 | beq 5f | |
1208 | ||
e1f6de8f | 1209 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1210 | |
1211 | 2: | |
1212 | setup_spans_set_x4_alternate_no(none, down) | |
1213 | subs height_minor_b, height_minor_b, #4 | |
1214 | bhi 2b | |
1215 | ||
1216 | 1: | |
1217 | setup_spans_epilogue() | |
1218 | ||
1219 | 3: | |
1220 | setup_spans_up_down_load_edge_set_b() | |
1221 | setup_spans_prologue_b() | |
1222 | bal 4b | |
1223 | ||
b7569147 | 1224 | 5: |
1225 | // FIXME: overflow corner case | |
1226 | sub temp, temp, height_minor_b | |
1227 | bics height_minor_b, #3 | |
1228 | add temp, temp, height_minor_b | |
e1f6de8f | 1229 | strh temp, [psx_gpu, #psx_gpu_num_spans_offset] |
b7569147 | 1230 | bne 2b |
1231 | bal 1b | |
1232 | ||
75e28f62 E |
1233 | #undef span_uvrg_offset |
1234 | #undef span_edge_data | |
1235 | #undef span_b_offset | |
1236 | #undef left_x | |
1237 | #undef b | |
1238 | ||
1239 | #define psx_gpu r0 | |
1240 | #define num_spans r1 | |
1241 | #define span_uvrg_offset r2 | |
1242 | #define span_edge_data r3 | |
1243 | #define span_b_offset r4 | |
1244 | #define b_dx r5 | |
1245 | #define span_num_blocks r6 | |
1246 | #define y r7 | |
1247 | #define left_x r8 | |
1248 | #define b r9 | |
1249 | #define dither_offset_ptr r10 | |
1250 | #define block_ptr_a r11 | |
1251 | #define fb_ptr r12 | |
1252 | #define num_blocks r14 | |
1253 | ||
1254 | #define uvrg_dx_ptr r2 | |
1255 | #define texture_mask_ptr r3 | |
1256 | #define dither_shift r8 | |
1257 | #define dither_row r10 | |
1258 | ||
1259 | #define c_32 r7 | |
1260 | #define b_dx4 r8 | |
1261 | #define b_dx8 r9 | |
1262 | #define block_ptr_b r10 | |
1263 | ||
1264 | #define block_span_ptr r10 | |
1265 | #define right_mask r8 | |
1266 | ||
1267 | #define color r2 | |
1268 | #define color_r r3 | |
1269 | #define color_g r4 | |
1270 | #define color_b r5 | |
1271 | ||
1272 | #undef uvrg | |
1273 | ||
1274 | #define u_block q0 | |
1275 | #define v_block q1 | |
1276 | #define r_block q2 | |
1277 | #define g_block q3 | |
1278 | #define b_block q4 | |
1279 | ||
1280 | #define uv_dx4 d10 | |
1281 | #define rg_dx4 d11 | |
1282 | #define uv_dx8 d12 | |
1283 | #define rg_dx8 d13 | |
1284 | #define b_whole_8 d14 | |
1285 | #define fb_mask_ptrs d15 | |
1286 | ||
1287 | #define uvrg_dx4 q5 | |
1288 | #define uvrg_dx8 q6 | |
1289 | #define uv_dx8 d12 | |
1290 | #define rg_dx8 d13 | |
1291 | ||
1292 | #define u_whole q8 | |
1293 | #define v_whole q9 | |
1294 | #define r_whole q10 | |
1295 | #define g_whole q11 | |
1296 | #define b_whole q12 | |
1297 | ||
1298 | #define u_whole_low d16 | |
1299 | #define u_whole_high d17 | |
1300 | #define v_whole_low d18 | |
1301 | #define v_whole_high d19 | |
1302 | #define r_whole_low d20 | |
1303 | #define r_whole_high d21 | |
1304 | #define g_whole_low d22 | |
1305 | #define g_whole_high d23 | |
1306 | #define b_whole_low d24 | |
1307 | #define b_whole_high d25 | |
1308 | ||
1309 | #define dx4 q13 | |
1310 | #define dx8 q13 | |
1311 | ||
1312 | #define u_whole_8 d26 | |
1313 | #define v_whole_8 d27 | |
1314 | #define u_whole_8b d24 | |
1315 | #define r_whole_8 d24 | |
1316 | #define g_whole_8 d25 | |
1317 | ||
1318 | #define uv_whole_8 q13 | |
1319 | #define uv_whole_8b q14 | |
1320 | ||
1321 | #define dither_offsets q14 | |
1322 | #define texture_mask q15 | |
1323 | #define texture_mask_u d30 | |
1324 | #define texture_mask_v d31 | |
1325 | ||
1326 | #define dither_offsets_short d28 | |
1327 | ||
1328 | #define v_left_x q8 | |
1329 | #define uvrg q9 | |
1330 | #define block_span q10 | |
1331 | ||
1332 | #define uv d18 | |
1333 | #define rg d19 | |
1334 | ||
1335 | #define draw_mask q1 | |
1336 | #define draw_mask_edge q13 | |
1337 | #define test_mask q0 | |
1338 | ||
1339 | #define uvrg_dx q3 | |
1340 | ||
1341 | #define colors q2 | |
1342 | ||
1343 | #define setup_blocks_texture_swizzled() \ | |
1344 | vand.u8 u_whole_8b, u_whole_8, texture_mask_u; \ | |
1345 | vsli.u8 u_whole_8, v_whole_8, #4; \ | |
1346 | vsri.u8 v_whole_8, u_whole_8b, #4 \ | |
1347 | ||
1348 | #define setup_blocks_texture_unswizzled() \ | |
1349 | ||
1350 | ||
1351 | #define setup_blocks_shaded_textured_builder(swizzling) \ | |
1352 | .align 3; \ | |
1353 | \ | |
1354 | function(setup_blocks_shaded_textured_dithered_##swizzling##_indirect) \ | |
e1f6de8f | 1355 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
1356 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
1357 | \ | |
e1f6de8f | 1358 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
75e28f62 E |
1359 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
1360 | \ | |
1361 | cmp num_spans, #0; \ | |
1362 | bxeq lr; \ | |
1363 | \ | |
1364 | stmdb sp!, { r4 - r11, r14 }; \ | |
9775fc33 | 1365 | save_abi_regs(); \ |
75e28f62 E |
1366 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
1367 | \ | |
e1f6de8f | 1368 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
75e28f62 E |
1369 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ |
1370 | \ | |
e1f6de8f | 1371 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
75e28f62 E |
1372 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
1373 | \ | |
e1f6de8f | 1374 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1375 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
1376 | \ | |
1377 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
1378 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1379 | \ | |
1380 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
1381 | \ | |
1382 | 0: \ | |
1383 | vmov.u8 fb_mask_ptrs, #0; \ | |
1384 | \ | |
e1f6de8f | 1385 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
1386 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
1387 | \ | |
e1f6de8f | 1388 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
1389 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
1390 | \ |
1391 | cmp span_num_blocks, #0; \ | |
1392 | beq 1f; \ | |
1393 | \ | |
e1f6de8f | 1394 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
1395 | add num_blocks, span_num_blocks, num_blocks; \ |
1396 | \ | |
1397 | cmp num_blocks, #MAX_BLOCKS; \ | |
1398 | bgt 2f; \ | |
1399 | \ | |
1400 | 3: \ | |
e1f6de8f | 1401 | ldr b, [span_b_offset]; \ |
75e28f62 E |
1402 | add fb_ptr, fb_ptr, y, lsl #11; \ |
1403 | \ | |
1404 | vdup.u32 v_left_x, left_x; \ | |
1405 | and y, y, #0x3; \ | |
1406 | \ | |
e1f6de8f | 1407 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
1408 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
1409 | \ | |
1410 | mla b, b_dx, left_x, b; \ | |
1411 | and dither_shift, left_x, #0x03; \ | |
1412 | \ | |
e1f6de8f | 1413 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
1414 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
1415 | \ | |
1416 | mov dither_shift, dither_shift, lsl #3; \ | |
1417 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
1418 | \ | |
1419 | mov c_32, #32; \ | |
1420 | subs span_num_blocks, span_num_blocks, #1; \ | |
1421 | \ | |
1422 | mov dither_row, dither_row, ror dither_shift; \ | |
1423 | mov b_dx4, b_dx, lsl #2; \ | |
1424 | \ | |
1425 | vdup.u32 dither_offsets_short, dither_row; \ | |
1426 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
1427 | \ | |
1428 | vdup.u32 b_block, b; \ | |
1429 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
1430 | \ | |
1431 | vdup.u32 u_block, uv[0]; \ | |
1432 | mov b_dx8, b_dx, lsl #3; \ | |
1433 | \ | |
1434 | vdup.u32 v_block, uv[1]; \ | |
1435 | vdup.u32 r_block, rg[0]; \ | |
1436 | vdup.u32 g_block, rg[1]; \ | |
1437 | \ | |
e1f6de8f | 1438 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1439 | \ |
1440 | vadd.u32 u_block, u_block, block_span; \ | |
e1f6de8f | 1441 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1442 | \ |
1443 | vadd.u32 v_block, v_block, block_span; \ | |
e1f6de8f | 1444 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1445 | \ |
1446 | vadd.u32 r_block, r_block, block_span; \ | |
e1f6de8f | 1447 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1448 | \ |
1449 | vadd.u32 g_block, g_block, block_span; \ | |
e1f6de8f | 1450 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
75e28f62 E |
1451 | \ |
1452 | vadd.u32 b_block, b_block, block_span; \ | |
1453 | add block_ptr_b, block_ptr_a, #16; \ | |
1454 | \ | |
1455 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1456 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1457 | vshrn.u32 r_whole_low, r_block, #16; \ | |
1458 | vshrn.u32 g_whole_low, g_block, #16; \ | |
1459 | \ | |
1460 | vdup.u32 dx4, uv_dx4[0]; \ | |
1461 | vshrn.u32 b_whole_low, b_block, #16; \ | |
1462 | \ | |
1463 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1464 | vdup.u32 dx4, uv_dx4[1]; \ | |
1465 | \ | |
1466 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1467 | vdup.u32 dx4, rg_dx4[0]; \ | |
1468 | \ | |
1469 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
1470 | vdup.u32 dx4, rg_dx4[1]; \ | |
1471 | \ | |
1472 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
1473 | vdup.u32 dx4, b_dx4; \ | |
1474 | \ | |
1475 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
1476 | vdup.u32 dx8, uv_dx8[0]; \ | |
1477 | \ | |
1478 | vadd.u32 u_block, u_block, dx8; \ | |
1479 | vdup.u32 dx8, uv_dx8[1]; \ | |
1480 | \ | |
1481 | vadd.u32 v_block, v_block, dx8; \ | |
1482 | vdup.u32 dx8, rg_dx8[0]; \ | |
1483 | \ | |
1484 | vadd.u32 r_block, r_block, dx8; \ | |
1485 | vdup.u32 dx8, rg_dx8[1]; \ | |
1486 | \ | |
1487 | vadd.u32 g_block, g_block, dx8; \ | |
1488 | vdup.u32 dx8, b_dx8; \ | |
1489 | \ | |
1490 | vadd.u32 b_block, b_block, dx8; \ | |
1491 | vmovn.u16 u_whole_8, u_whole; \ | |
1492 | \ | |
1493 | vmovn.u16 v_whole_8, v_whole; \ | |
1494 | \ | |
1495 | vmovn.u16 b_whole_8, b_whole; \ | |
e1f6de8f | 1496 | pld [fb_ptr]; \ |
75e28f62 E |
1497 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
1498 | \ | |
1499 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1500 | setup_blocks_texture_##swizzling(); \ | |
1501 | \ | |
1502 | vmovn.u16 r_whole_8, r_whole; \ | |
1503 | beq 5f; \ | |
1504 | \ | |
1505 | 4: \ | |
1506 | vmovn.u16 g_whole_8, g_whole; \ | |
1507 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1508 | \ | |
e1f6de8f | 1509 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1510 | vshrn.u32 v_whole_low, v_block, #16; \ |
1511 | \ | |
e1f6de8f | 1512 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
75e28f62 E |
1513 | vshrn.u32 r_whole_low, r_block, #16; \ |
1514 | \ | |
e1f6de8f | 1515 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1516 | vshrn.u32 g_whole_low, g_block, #16; \ |
1517 | \ | |
1518 | vdup.u32 dx4, uv_dx4[0]; \ | |
1519 | vshrn.u32 b_whole_low, b_block, #16; \ | |
1520 | \ | |
1521 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1522 | vdup.u32 dx4, uv_dx4[1]; \ | |
1523 | \ | |
1524 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1525 | vdup.u32 dx4, rg_dx4[0]; \ | |
1526 | \ | |
1527 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
1528 | vdup.u32 dx4, rg_dx4[1]; \ | |
1529 | \ | |
1530 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
1531 | vdup.u32 dx4, b_dx4; \ | |
1532 | \ | |
1533 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
1534 | vdup.u32 dx8, uv_dx8[0]; \ | |
1535 | \ | |
1536 | vadd.u32 u_block, u_block, dx8; \ | |
1537 | vdup.u32 dx8, uv_dx8[1]; \ | |
1538 | \ | |
1539 | vadd.u32 v_block, v_block, dx8; \ | |
1540 | vdup.u32 dx8, rg_dx8[0]; \ | |
1541 | \ | |
1542 | vadd.u32 r_block, r_block, dx8; \ | |
1543 | vdup.u32 dx8, rg_dx8[1]; \ | |
1544 | \ | |
1545 | vadd.u32 g_block, g_block, dx8; \ | |
1546 | vdup.u32 dx8, b_dx8; \ | |
1547 | \ | |
1548 | vadd.u32 b_block, b_block, dx8; \ | |
1549 | vmovn.u16 u_whole_8, u_whole; \ | |
1550 | \ | |
1551 | add fb_ptr, fb_ptr, #16; \ | |
1552 | vmovn.u16 v_whole_8, v_whole; \ | |
1553 | \ | |
e1f6de8f | 1554 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
75e28f62 E |
1555 | vmovn.u16 b_whole_8, b_whole; \ |
1556 | \ | |
e1f6de8f | 1557 | pld [fb_ptr]; \ |
75e28f62 E |
1558 | \ |
1559 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1560 | subs span_num_blocks, span_num_blocks, #1; \ | |
1561 | \ | |
1562 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1563 | setup_blocks_texture_##swizzling(); \ | |
1564 | \ | |
1565 | vmovn.u16 r_whole_8, r_whole; \ | |
1566 | bne 4b; \ | |
1567 | \ | |
1568 | 5: \ | |
1569 | vmovn.u16 g_whole_8, g_whole; \ | |
e1f6de8f | 1570 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 | 1571 | \ |
e1f6de8f | 1572 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
1573 | vdup.u8 draw_mask, right_mask; \ |
1574 | \ | |
1575 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
1576 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
1577 | vzip.u8 u_whole_8, v_whole_8; \ | |
1578 | \ | |
1579 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
e1f6de8f | 1580 | vst1.u32 { r_whole_8, g_whole_8 }, [block_ptr_b, :128], c_32; \ |
1581 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ | |
1582 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
1583 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
75e28f62 E |
1584 | \ |
1585 | 1: \ | |
1586 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
1587 | add span_b_offset, span_b_offset, #4; \ | |
1588 | \ | |
1589 | add span_edge_data, span_edge_data, #8; \ | |
1590 | subs num_spans, num_spans, #1; \ | |
1591 | \ | |
e1f6de8f | 1592 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1593 | bne 0b; \ |
1594 | \ | |
9775fc33 | 1595 | restore_abi_regs(); \ |
75e28f62 E |
1596 | ldmia sp!, { r4 - r11, pc }; \ |
1597 | \ | |
1598 | 2: \ | |
1599 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
1600 | vpush { texture_mask }; \ | |
1601 | vpush { uvrg_dx4 }; \ | |
1602 | \ | |
d38927c1 | 1603 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; /* r14=num_blocks */ \ |
75e28f62 | 1604 | bl flush_render_block_buffer; \ |
d38927c1 | 1605 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ |
75e28f62 E |
1606 | \ |
1607 | vpop { uvrg_dx4 }; \ | |
1608 | vpop { texture_mask }; \ | |
1609 | \ | |
1610 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
1611 | vmov.u8 fb_mask_ptrs, #0; \ | |
1612 | \ | |
1613 | mov num_blocks, span_num_blocks; \ | |
1614 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1615 | bal 3b \ | |
1616 | ||
1617 | ||
1618 | setup_blocks_shaded_textured_builder(swizzled) | |
1619 | setup_blocks_shaded_textured_builder(unswizzled) | |
1620 | ||
1621 | ||
1622 | #define setup_blocks_unshaded_textured_builder(swizzling) \ | |
1623 | .align 3; \ | |
1624 | \ | |
1625 | function(setup_blocks_unshaded_textured_dithered_##swizzling##_indirect) \ | |
e1f6de8f | 1626 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
1627 | add uvrg_dx_ptr, psx_gpu, #psx_gpu_uvrg_dx_offset; \ |
1628 | \ | |
e1f6de8f | 1629 | vld1.u32 { uvrg_dx }, [uvrg_dx_ptr, :128]; \ |
75e28f62 E |
1630 | add texture_mask_ptr, psx_gpu, #psx_gpu_texture_mask_width_offset; \ |
1631 | \ | |
1632 | cmp num_spans, #0; \ | |
1633 | bxeq lr; \ | |
1634 | \ | |
1635 | stmdb sp!, { r4 - r11, r14 }; \ | |
9775fc33 | 1636 | save_abi_regs(); \ |
75e28f62 E |
1637 | vshl.u32 uvrg_dx4, uvrg_dx, #2; \ |
1638 | \ | |
1639 | vshl.u32 uvrg_dx8, uvrg_dx, #3; \ | |
1640 | \ | |
e1f6de8f | 1641 | vld2.u8 { texture_mask_u[], texture_mask_v[] }, [texture_mask_ptr, :16]; \ |
75e28f62 E |
1642 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ |
1643 | \ | |
e1f6de8f | 1644 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1645 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
1646 | \ | |
1647 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1648 | \ | |
1649 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
1650 | \ | |
1651 | 0: \ | |
1652 | vmov.u8 fb_mask_ptrs, #0; \ | |
1653 | \ | |
e1f6de8f | 1654 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
1655 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
1656 | \ | |
e1f6de8f | 1657 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
1658 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
1659 | \ |
1660 | cmp span_num_blocks, #0; \ | |
1661 | beq 1f; \ | |
1662 | \ | |
e1f6de8f | 1663 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
1664 | add num_blocks, span_num_blocks, num_blocks; \ |
1665 | \ | |
1666 | cmp num_blocks, #MAX_BLOCKS; \ | |
1667 | bgt 2f; \ | |
1668 | \ | |
1669 | 3: \ | |
1670 | add fb_ptr, fb_ptr, y, lsl #11; \ | |
1671 | \ | |
1672 | vdup.u32 v_left_x, left_x; \ | |
1673 | and y, y, #0x3; \ | |
1674 | \ | |
e1f6de8f | 1675 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
1676 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
1677 | \ | |
1678 | and dither_shift, left_x, #0x03; \ | |
1679 | \ | |
e1f6de8f | 1680 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
1681 | vshr.u32 uvrg_dx, uvrg_dx4, #2; \ |
1682 | \ | |
1683 | mov dither_shift, dither_shift, lsl #3; \ | |
1684 | vmla.u32 uvrg, uvrg_dx, v_left_x; \ | |
1685 | \ | |
1686 | mov c_32, #32; \ | |
1687 | subs span_num_blocks, span_num_blocks, #1; \ | |
1688 | \ | |
1689 | mov dither_row, dither_row, ror dither_shift; \ | |
1690 | \ | |
1691 | vdup.u32 dither_offsets_short, dither_row; \ | |
1692 | add block_span_ptr, psx_gpu, #psx_gpu_u_block_span_offset; \ | |
1693 | \ | |
1694 | vshll.s8 dither_offsets, dither_offsets_short, #4; \ | |
1695 | \ | |
1696 | vdup.u32 u_block, uv[0]; \ | |
1697 | \ | |
1698 | vdup.u32 v_block, uv[1]; \ | |
e1f6de8f | 1699 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1700 | \ |
1701 | vadd.u32 u_block, u_block, block_span; \ | |
e1f6de8f | 1702 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
1703 | \ |
1704 | vadd.u32 v_block, v_block, block_span; \ | |
1705 | add block_ptr_b, block_ptr_a, #16; \ | |
1706 | \ | |
1707 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1708 | vshrn.u32 v_whole_low, v_block, #16; \ | |
1709 | \ | |
1710 | vdup.u32 dx4, uv_dx4[0]; \ | |
1711 | \ | |
1712 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1713 | vdup.u32 dx4, uv_dx4[1]; \ | |
1714 | \ | |
1715 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1716 | vdup.u32 dx8, uv_dx8[0]; \ | |
1717 | \ | |
1718 | vadd.u32 u_block, u_block, dx8; \ | |
1719 | vdup.u32 dx8, uv_dx8[1]; \ | |
1720 | \ | |
1721 | vadd.u32 v_block, v_block, dx8; \ | |
1722 | vmovn.u16 u_whole_8, u_whole; \ | |
1723 | \ | |
1724 | vmovn.u16 v_whole_8, v_whole; \ | |
1725 | \ | |
e1f6de8f | 1726 | pld [fb_ptr]; \ |
75e28f62 E |
1727 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ |
1728 | \ | |
1729 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1730 | setup_blocks_texture_##swizzling(); \ | |
1731 | \ | |
1732 | beq 5f; \ | |
1733 | \ | |
1734 | 4: \ | |
1735 | vshrn.u32 u_whole_low, u_block, #16; \ | |
1736 | \ | |
e1f6de8f | 1737 | vst2.u8 { u_whole_8, v_whole_8 }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1738 | vshrn.u32 v_whole_low, v_block, #16; \ |
1739 | \ | |
1740 | add block_ptr_b, block_ptr_b, #32; \ | |
e1f6de8f | 1741 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ |
75e28f62 E |
1742 | \ |
1743 | vdup.u32 dx4, uv_dx4[0]; \ | |
1744 | vaddhn.u32 u_whole_high, u_block, dx4; \ | |
1745 | vdup.u32 dx4, uv_dx4[1]; \ | |
1746 | \ | |
1747 | vaddhn.u32 v_whole_high, v_block, dx4; \ | |
1748 | vdup.u32 dx8, uv_dx8[0]; \ | |
1749 | \ | |
1750 | vadd.u32 u_block, u_block, dx8; \ | |
1751 | vdup.u32 dx8, uv_dx8[1]; \ | |
1752 | \ | |
1753 | vadd.u32 v_block, v_block, dx8; \ | |
1754 | vmovn.u16 u_whole_8, u_whole; \ | |
1755 | \ | |
1756 | add fb_ptr, fb_ptr, #16; \ | |
1757 | vmovn.u16 v_whole_8, v_whole; \ | |
1758 | \ | |
e1f6de8f | 1759 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ |
1760 | pld [fb_ptr]; \ | |
75e28f62 E |
1761 | \ |
1762 | vmov.u32 fb_mask_ptrs[1], fb_ptr; \ | |
1763 | subs span_num_blocks, span_num_blocks, #1; \ | |
1764 | \ | |
1765 | vand.u8 uv_whole_8, uv_whole_8, texture_mask; \ | |
1766 | setup_blocks_texture_##swizzling(); \ | |
1767 | \ | |
1768 | bne 4b; \ | |
1769 | \ | |
1770 | 5: \ | |
e1f6de8f | 1771 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 | 1772 | \ |
e1f6de8f | 1773 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
1774 | vdup.u8 draw_mask, right_mask; \ |
1775 | \ | |
1776 | vmov.u32 fb_mask_ptrs[0], right_mask; \ | |
1777 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
1778 | vzip.u8 u_whole_8, v_whole_8; \ | |
1779 | \ | |
1780 | vbic.u16 uv_whole_8, uv_whole_8, draw_mask; \ | |
1781 | add block_ptr_b, block_ptr_b, #32; \ | |
e1f6de8f | 1782 | vst1.u32 { uv_whole_8 }, [block_ptr_a, :128], c_32; \ |
1783 | vst1.u32 { dither_offsets }, [block_ptr_b, :128], c_32; \ | |
1784 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32; \ | |
75e28f62 E |
1785 | \ |
1786 | 1: \ | |
1787 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
1788 | add span_edge_data, span_edge_data, #8; \ | |
1789 | subs num_spans, num_spans, #1; \ | |
1790 | \ | |
e1f6de8f | 1791 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
1792 | bne 0b; \ |
1793 | \ | |
9775fc33 | 1794 | restore_abi_regs(); \ |
75e28f62 E |
1795 | ldmia sp!, { r4 - r11, pc }; \ |
1796 | \ | |
1797 | 2: \ | |
1798 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
1799 | vpush { texture_mask }; \ | |
1800 | vpush { uvrg_dx4 }; \ | |
1801 | \ | |
d38927c1 | 1802 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; /* r14=num_blocks */ \ |
75e28f62 | 1803 | bl flush_render_block_buffer; \ |
d38927c1 | 1804 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12 }; \ |
75e28f62 E |
1805 | \ |
1806 | vpop { uvrg_dx4 }; \ | |
1807 | vpop { texture_mask }; \ | |
1808 | \ | |
1809 | vadd.u32 uvrg_dx8, uvrg_dx4, uvrg_dx4; \ | |
1810 | vmov.u8 fb_mask_ptrs, #0; \ | |
1811 | \ | |
1812 | mov num_blocks, span_num_blocks; \ | |
1813 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
1814 | bal 3b \ | |
1815 | ||
1816 | ||
1817 | setup_blocks_unshaded_textured_builder(swizzled) | |
1818 | setup_blocks_unshaded_textured_builder(unswizzled) | |
1819 | ||
1820 | ||
1821 | .align 3 | |
1822 | ||
1823 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_indirect) | |
e1f6de8f | 1824 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1825 | veor.u32 draw_mask, draw_mask, draw_mask |
1826 | ||
1827 | cmp num_spans, #0 | |
1828 | bxeq lr | |
1829 | ||
1830 | stmdb sp!, { r4 - r11, r14 } | |
9775fc33 | 1831 | save_abi_regs() |
e1f6de8f | 1832 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 | 1833 | |
e1f6de8f | 1834 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
75e28f62 E |
1835 | |
1836 | ubfx color_r, color, #3, #5 | |
1837 | ubfx color_g, color, #11, #5 | |
1838 | ubfx color_b, color, #19, #5 | |
1839 | ||
1840 | orr color, color_r, color_b, lsl #10 | |
1841 | orr color, color, color_g, lsl #5 | |
1842 | ||
1843 | vdup.u16 colors, color | |
1844 | ||
e1f6de8f | 1845 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
1846 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset |
1847 | ||
1848 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
1849 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6 | |
1850 | ||
1851 | 0: | |
e1f6de8f | 1852 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
1853 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
75e28f62 | 1854 | |
e1f6de8f | 1855 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
75e28f62 E |
1856 | |
1857 | cmp span_num_blocks, #0 | |
1858 | beq 1f | |
1859 | ||
e1f6de8f | 1860 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
75e28f62 E |
1861 | add num_blocks, span_num_blocks, num_blocks |
1862 | ||
1863 | cmp num_blocks, #MAX_BLOCKS | |
1864 | bgt 2f | |
1865 | ||
1866 | 3: | |
1867 | add fb_ptr, fb_ptr, y, lsl #11 | |
1868 | and y, y, #0x3 | |
1869 | ||
1870 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
1871 | mov c_32, #32 | |
1872 | ||
1873 | subs span_num_blocks, span_num_blocks, #1 | |
1874 | ||
1875 | add block_ptr_b, block_ptr_a, #16 | |
e1f6de8f | 1876 | pld [fb_ptr] |
75e28f62 E |
1877 | |
1878 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
1879 | beq 5f | |
1880 | ||
1881 | 4: | |
e1f6de8f | 1882 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_32 |
1883 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 | |
1884 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 | |
75e28f62 E |
1885 | |
1886 | add fb_ptr, fb_ptr, #16 | |
1887 | add block_ptr_b, block_ptr_b, #32 | |
1888 | ||
e1f6de8f | 1889 | pld [fb_ptr] |
75e28f62 E |
1890 | |
1891 | vmov.u32 fb_mask_ptrs[1], fb_ptr | |
1892 | subs span_num_blocks, span_num_blocks, #1 | |
1893 | ||
1894 | bne 4b | |
1895 | ||
1896 | 5: | |
e1f6de8f | 1897 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset] |
75e28f62 E |
1898 | |
1899 | vdup.u8 draw_mask_edge, right_mask | |
1900 | vtst.u16 draw_mask_edge, draw_mask_edge, test_mask | |
1901 | ||
e1f6de8f | 1902 | vst1.u32 { colors }, [block_ptr_b, :128], c_32 |
1903 | vst1.u32 { draw_mask_edge }, [block_ptr_a, :128], c_32 | |
75e28f62 | 1904 | add block_ptr_b, block_ptr_b, #32 |
e1f6de8f | 1905 | vst1.u32 { b_whole_8, fb_mask_ptrs }, [block_ptr_a, :128], c_32 |
75e28f62 E |
1906 | |
1907 | 1: | |
1908 | add span_edge_data, span_edge_data, #8 | |
1909 | subs num_spans, num_spans, #1 | |
1910 | ||
e1f6de8f | 1911 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
1912 | bne 0b |
1913 | ||
9775fc33 | 1914 | restore_abi_regs() |
75e28f62 E |
1915 | ldmia sp!, { r4 - r11, pc } |
1916 | ||
1917 | 2: | |
1918 | vpush { colors } | |
1919 | ||
4d646738 | 1920 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 1921 | bl flush_render_block_buffer |
4d646738 | 1922 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 E |
1923 | |
1924 | vpop { colors } | |
1925 | ||
e1f6de8f | 1926 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 E |
1927 | veor.u32 draw_mask, draw_mask, draw_mask |
1928 | ||
1929 | mov num_blocks, span_num_blocks | |
1930 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset | |
1931 | bal 3b | |
1932 | ||
1933 | ||
1934 | #define mask_msb_scalar r14 | |
1935 | ||
1936 | #define msb_mask q15 | |
1937 | ||
1938 | #define pixels_low d16 | |
1939 | ||
1940 | #define msb_mask_low d30 | |
1941 | #define msb_mask_high d31 | |
1942 | ||
1943 | ||
1944 | .align 3 | |
1945 | ||
1946 | function(setup_blocks_unshaded_untextured_undithered_unswizzled_direct) | |
e1f6de8f | 1947 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset] |
75e28f62 E |
1948 | |
1949 | cmp num_spans, #0 | |
1950 | bxeq lr | |
1951 | ||
1952 | stmdb sp!, { r4 - r11, r14 } | |
1953 | ||
e1f6de8f | 1954 | ldr color, [psx_gpu, #psx_gpu_triangle_color_offset] |
75e28f62 E |
1955 | |
1956 | ubfx color_r, color, #3, #5 | |
1957 | ubfx color_g, color, #11, #5 | |
1958 | ||
e1f6de8f | 1959 | ldrh mask_msb_scalar, [psx_gpu, #psx_gpu_mask_msb_offset] |
75e28f62 E |
1960 | ubfx color_b, color, #19, #5 |
1961 | ||
1962 | orr color, color_r, color_b, lsl #10 | |
1963 | orr color, color, color_g, lsl #5 | |
1964 | orr color, color, mask_msb_scalar | |
1965 | ||
1966 | vdup.u16 colors, color | |
1967 | ||
1968 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset | |
ed0fd81d | 1969 | orr color, color, color, lsl #16 |
3867c6ef | 1970 | |
75e28f62 E |
1971 | |
1972 | 0: | |
e1f6de8f | 1973 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset] |
1974 | ldrh y, [span_edge_data, #edge_data_y_offset] | |
75e28f62 | 1975 | |
e1f6de8f | 1976 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
75e28f62 E |
1977 | |
1978 | cmp span_num_blocks, #0 | |
1979 | beq 1f | |
1980 | ||
e1f6de8f | 1981 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset] |
75e28f62 E |
1982 | |
1983 | add fb_ptr, fb_ptr, y, lsl #11 | |
1984 | subs span_num_blocks, span_num_blocks, #1 | |
1985 | ||
1986 | add fb_ptr, fb_ptr, left_x, lsl #1 | |
1987 | beq 3f | |
1988 | ||
1989 | 2: | |
e1f6de8f | 1990 | vst1.u32 { colors }, [fb_ptr]! |
75e28f62 E |
1991 | subs span_num_blocks, span_num_blocks, #1 |
1992 | ||
1993 | bne 2b | |
1994 | ||
1995 | 3: | |
e1f6de8f | 1996 | ldrb right_mask, [span_edge_data, #edge_data_right_mask_offset] |
75e28f62 | 1997 | |
3867c6ef E |
1998 | cmp right_mask, #0x0 |
1999 | beq 5f | |
2000 | ||
2001 | tst right_mask, #0xF | |
e1f6de8f | 2002 | streq color, [fb_ptr], #4 |
3867c6ef | 2003 | moveq right_mask, right_mask, lsr #4 |
e1f6de8f | 2004 | streq color, [fb_ptr], #4 |
3867c6ef E |
2005 | |
2006 | tst right_mask, #0x3 | |
e1f6de8f | 2007 | streq color, [fb_ptr], #4 |
3867c6ef E |
2008 | moveq right_mask, right_mask, lsr #2 |
2009 | ||
2010 | tst right_mask, #0x1 | |
e1f6de8f | 2011 | strheq color, [fb_ptr] |
75e28f62 E |
2012 | |
2013 | 1: | |
2014 | add span_edge_data, span_edge_data, #8 | |
2015 | subs num_spans, num_spans, #1 | |
75e28f62 E |
2016 | bne 0b |
2017 | ||
2018 | ldmia sp!, { r4 - r11, pc } | |
2019 | ||
3867c6ef | 2020 | 5: |
e1f6de8f | 2021 | vst1.u32 { colors }, [fb_ptr] |
3867c6ef | 2022 | bal 1b |
75e28f62 E |
2023 | |
2024 | ||
2025 | #undef c_64 | |
2026 | ||
2027 | #define c_64 r7 | |
2028 | #define rg_dx_ptr r2 | |
2029 | ||
2030 | ||
2031 | #undef r_block | |
2032 | #undef g_block | |
2033 | #undef b_block | |
2034 | #undef r_whole | |
2035 | #undef g_whole | |
2036 | #undef b_whole | |
2037 | #undef r_whole_low | |
2038 | #undef r_whole_high | |
2039 | #undef g_whole_low | |
2040 | #undef g_whole_high | |
2041 | #undef b_whole_low | |
2042 | #undef b_whole_high | |
2043 | #undef r_whole_8 | |
2044 | #undef g_whole_8 | |
2045 | #undef b_whole_8 | |
2046 | #undef dither_offsets | |
2047 | #undef rg_dx4 | |
2048 | #undef rg_dx8 | |
2049 | #undef dx4 | |
2050 | #undef dx8 | |
2051 | #undef v_left_x | |
2052 | #undef uvrg | |
2053 | #undef block_span | |
2054 | #undef rg | |
2055 | #undef draw_mask | |
2056 | #undef test_mask | |
2057 | ||
2058 | #define r_block q0 | |
2059 | #define g_block q1 | |
2060 | #define b_block q2 | |
2061 | ||
2062 | #define r_whole q3 | |
2063 | #define g_whole q4 | |
2064 | #define b_whole q5 | |
2065 | ||
2066 | #define r_whole_low d6 | |
2067 | #define r_whole_high d7 | |
2068 | #define g_whole_low d8 | |
2069 | #define g_whole_high d9 | |
2070 | #define b_whole_low d10 | |
2071 | #define b_whole_high d11 | |
2072 | ||
2073 | #define gb_whole_8 q6 | |
2074 | ||
2075 | #define g_whole_8 d12 | |
2076 | #define b_whole_8 d13 | |
2077 | ||
2078 | #define r_whole_8 d14 | |
2079 | ||
2080 | #define pixels q8 | |
2081 | ||
2082 | #define rg_dx4 d18 | |
2083 | #define rg_dx8 d19 | |
2084 | ||
2085 | #define dx4 q10 | |
2086 | #define dx8 q10 | |
2087 | ||
2088 | #define v_left_x d6 | |
2089 | #define uvrg q4 | |
2090 | #define block_span q5 | |
2091 | ||
2092 | #define rg d9 | |
2093 | ||
2094 | #define d64_1 d22 | |
2095 | #define d64_128 d23 | |
2096 | ||
2097 | #define d128_4 q12 | |
2098 | #define d128_0x7 q13 | |
2099 | ||
2100 | #define d64_4 d24 | |
2101 | ||
2102 | #define dither_offsets q14 | |
2103 | #define draw_mask q15 | |
2104 | ||
2105 | #define dither_offsets_low d28 | |
2106 | ||
2107 | #define rg_dx d0 | |
2108 | #define test_mask q10 | |
2109 | ||
2110 | ||
2111 | #define setup_blocks_shaded_untextured_dither_a_dithered() \ | |
2112 | vqadd.u8 r_whole_8, r_whole_8, dither_offsets_low; \ | |
2113 | vqadd.u8 gb_whole_8, gb_whole_8, dither_offsets; \ | |
2114 | ||
2115 | #define setup_blocks_shaded_untextured_dither_b_dithered() \ | |
2116 | vqsub.u8 r_whole_8, r_whole_8, d64_4; \ | |
2117 | vqsub.u8 gb_whole_8, gb_whole_8, d128_4 \ | |
2118 | ||
2119 | #define setup_blocks_shaded_untextured_dither_a_undithered() \ | |
2120 | ||
2121 | #define setup_blocks_shaded_untextured_dither_b_undithered() \ | |
2122 | ||
2123 | ||
2124 | #define setup_blocks_shaded_untextured_indirect_builder(dithering) \ | |
2125 | .align 3; \ | |
2126 | \ | |
2127 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_indirect) \ | |
e1f6de8f | 2128 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
2129 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
2130 | \ | |
e1f6de8f | 2131 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
75e28f62 E |
2132 | \ |
2133 | cmp num_spans, #0; \ | |
2134 | bxeq lr; \ | |
2135 | \ | |
2136 | stmdb sp!, { r4 - r11, r14 }; \ | |
9775fc33 | 2137 | save_abi_regs(); \ |
75e28f62 E |
2138 | vshl.u32 rg_dx4, rg_dx, #2; \ |
2139 | \ | |
e1f6de8f | 2140 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
75e28f62 E |
2141 | vshl.u32 rg_dx8, rg_dx, #3; \ |
2142 | \ | |
2143 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
2144 | \ | |
e1f6de8f | 2145 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
2146 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ |
2147 | \ | |
2148 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
2149 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
2150 | \ | |
2151 | add block_ptr_a, block_ptr_a, num_blocks, lsl #6; \ | |
2152 | vmov.u8 d64_1, #1; \ | |
2153 | \ | |
2154 | vmov.u8 d128_4, #4; \ | |
2155 | vmov.u8 d64_128, #128; \ | |
2156 | \ | |
2157 | vmov.u8 d128_0x7, #0x7; \ | |
2158 | \ | |
2159 | 0: \ | |
e1f6de8f | 2160 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
2161 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
2162 | \ | |
e1f6de8f | 2163 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
2164 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
2165 | \ |
2166 | cmp span_num_blocks, #0; \ | |
2167 | beq 1f; \ | |
2168 | \ | |
e1f6de8f | 2169 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
2170 | add num_blocks, span_num_blocks, num_blocks; \ |
2171 | \ | |
2172 | cmp num_blocks, #MAX_BLOCKS; \ | |
2173 | bgt 2f; \ | |
2174 | \ | |
2175 | 3: \ | |
e1f6de8f | 2176 | ldr b, [span_b_offset]; \ |
75e28f62 E |
2177 | add fb_ptr, fb_ptr, y, lsl #11; \ |
2178 | \ | |
2179 | vdup.u32 v_left_x, left_x; \ | |
2180 | and y, y, #0x3; \ | |
2181 | \ | |
e1f6de8f | 2182 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
2183 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
2184 | \ | |
2185 | mla b, b_dx, left_x, b; \ | |
2186 | and dither_shift, left_x, #0x03; \ | |
2187 | \ | |
e1f6de8f | 2188 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
2189 | vshr.u32 rg_dx, rg_dx4, #2; \ |
2190 | \ | |
2191 | mov dither_shift, dither_shift, lsl #3; \ | |
2192 | vmla.u32 rg, rg_dx, v_left_x; \ | |
2193 | \ | |
2194 | mov c_64, #64; \ | |
2195 | subs span_num_blocks, span_num_blocks, #1; \ | |
2196 | \ | |
2197 | mov dither_row, dither_row, ror dither_shift; \ | |
2198 | mov b_dx4, b_dx, lsl #2; \ | |
2199 | \ | |
2200 | vdup.u32 dither_offsets, dither_row; \ | |
2201 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
2202 | \ | |
2203 | vdup.u32 b_block, b; \ | |
2204 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
2205 | \ | |
2206 | mov b_dx8, b_dx, lsl #3; \ | |
2207 | vdup.u32 r_block, rg[0]; \ | |
2208 | vdup.u32 g_block, rg[1]; \ | |
2209 | \ | |
e1f6de8f | 2210 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2211 | \ |
2212 | vadd.u32 r_block, r_block, block_span; \ | |
e1f6de8f | 2213 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2214 | \ |
2215 | vadd.u32 g_block, g_block, block_span; \ | |
e1f6de8f | 2216 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
75e28f62 E |
2217 | \ |
2218 | vadd.u32 b_block, b_block, block_span; \ | |
2219 | add block_ptr_b, block_ptr_a, #16; \ | |
2220 | \ | |
2221 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2222 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2223 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2224 | vdup.u32 dx4, rg_dx4[0]; \ | |
2225 | \ | |
2226 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2227 | vdup.u32 dx4, rg_dx4[1]; \ | |
2228 | \ | |
2229 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2230 | vdup.u32 dx4, b_dx4; \ | |
2231 | \ | |
2232 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2233 | vdup.u32 dx8, rg_dx8[0]; \ | |
2234 | \ | |
2235 | vadd.u32 r_block, r_block, dx8; \ | |
2236 | vdup.u32 dx8, rg_dx8[1]; \ | |
2237 | \ | |
2238 | vadd.u32 g_block, g_block, dx8; \ | |
2239 | vdup.u32 dx8, b_dx8; \ | |
2240 | \ | |
2241 | vadd.u32 b_block, b_block, dx8; \ | |
2242 | \ | |
2243 | vmovn.u16 r_whole_8, r_whole; \ | |
2244 | vmovn.u16 g_whole_8, g_whole; \ | |
2245 | vmovn.u16 b_whole_8, b_whole; \ | |
2246 | \ | |
2247 | beq 5f; \ | |
2248 | veor.u32 draw_mask, draw_mask, draw_mask; \ | |
2249 | \ | |
2250 | 4: \ | |
2251 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2252 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2253 | \ | |
2254 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
2255 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2256 | \ | |
2257 | vshrn.u32 b_whole_low, b_block, #16; \ | |
e1f6de8f | 2258 | str fb_ptr, [block_ptr_a, #44]; \ |
75e28f62 E |
2259 | \ |
2260 | vdup.u32 dx4, rg_dx4[0]; \ | |
2261 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2262 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
2263 | \ | |
2264 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2265 | vdup.u32 dx4, rg_dx4[1]; \ | |
2266 | \ | |
2267 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2268 | vdup.u32 dx4, b_dx4; \ | |
2269 | \ | |
2270 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2271 | vdup.u32 dx8, rg_dx8[0]; \ | |
2272 | \ | |
2273 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
2274 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2275 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2276 | \ | |
2277 | vadd.u32 r_block, r_block, dx8; \ | |
2278 | vdup.u32 dx8, rg_dx8[1]; \ | |
2279 | \ | |
2280 | vadd.u32 g_block, g_block, dx8; \ | |
2281 | vdup.u32 dx8, b_dx8; \ | |
2282 | \ | |
2283 | vadd.u32 b_block, b_block, dx8; \ | |
2284 | add fb_ptr, fb_ptr, #16; \ | |
2285 | \ | |
2286 | vmovn.u16 r_whole_8, r_whole; \ | |
2287 | vmovn.u16 g_whole_8, g_whole; \ | |
2288 | vmovn.u16 b_whole_8, b_whole; \ | |
2289 | \ | |
e1f6de8f | 2290 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
2291 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
75e28f62 | 2292 | \ |
e1f6de8f | 2293 | pld [fb_ptr]; \ |
75e28f62 E |
2294 | \ |
2295 | subs span_num_blocks, span_num_blocks, #1; \ | |
2296 | bne 4b; \ | |
2297 | \ | |
2298 | 5: \ | |
e1f6de8f | 2299 | str fb_ptr, [block_ptr_a, #44]; \ |
75e28f62 E |
2300 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ |
2301 | \ | |
e1f6de8f | 2302 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 E |
2303 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
2304 | \ | |
2305 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2306 | vdup.u8 draw_mask, right_mask; \ | |
2307 | \ | |
2308 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
e1f6de8f | 2309 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
2310 | \ |
2311 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
2312 | \ | |
2313 | vmull.u8 pixels, r_whole_8, d64_1; \ | |
2314 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2315 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2316 | \ | |
e1f6de8f | 2317 | vst1.u32 { draw_mask }, [block_ptr_a, :128], c_64; \ |
2318 | vst1.u32 { pixels }, [block_ptr_b, :128], c_64; \ | |
75e28f62 E |
2319 | \ |
2320 | 1: \ | |
2321 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
2322 | add span_b_offset, span_b_offset, #4; \ | |
2323 | \ | |
2324 | add span_edge_data, span_edge_data, #8; \ | |
2325 | subs num_spans, num_spans, #1; \ | |
2326 | \ | |
e1f6de8f | 2327 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
2328 | bne 0b; \ |
2329 | \ | |
9775fc33 | 2330 | restore_abi_regs(); \ |
75e28f62 E |
2331 | ldmia sp!, { r4 - r11, pc }; \ |
2332 | \ | |
2333 | 2: \ | |
2334 | /* TODO: Load from psx_gpu instead of saving/restoring these */\ | |
2335 | vpush { rg_dx4 }; \ | |
2336 | \ | |
4d646738 | 2337 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 | 2338 | bl flush_render_block_buffer; \ |
4d646738 | 2339 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 }; \ |
75e28f62 E |
2340 | \ |
2341 | vpop { rg_dx4 }; \ | |
2342 | \ | |
2343 | vmov.u8 d64_1, #1; \ | |
2344 | vmov.u8 d128_4, #4; \ | |
2345 | vmov.u8 d64_128, #128; \ | |
2346 | vmov.u8 d128_0x7, #0x7; \ | |
2347 | \ | |
2348 | vadd.u32 rg_dx8, rg_dx4, rg_dx4; \ | |
2349 | \ | |
2350 | mov num_blocks, span_num_blocks; \ | |
2351 | add block_ptr_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
2352 | bal 3b \ | |
2353 | ||
2354 | ||
2355 | setup_blocks_shaded_untextured_indirect_builder(undithered) | |
2356 | setup_blocks_shaded_untextured_indirect_builder(dithered) | |
2357 | ||
2358 | ||
2359 | #undef draw_mask | |
2360 | ||
2361 | #define mask_msb_ptr r14 | |
2362 | ||
2363 | #define draw_mask q0 | |
2364 | #define pixels_low d16 | |
3867c6ef | 2365 | #define pixels_high d17 |
75e28f62 E |
2366 | |
2367 | ||
2368 | ||
2369 | #define setup_blocks_shaded_untextured_direct_builder(dithering) \ | |
2370 | .align 3; \ | |
2371 | \ | |
2372 | function(setup_blocks_shaded_untextured_##dithering##_unswizzled_direct) \ | |
e1f6de8f | 2373 | ldrh num_spans, [psx_gpu, #psx_gpu_num_spans_offset]; \ |
75e28f62 E |
2374 | add rg_dx_ptr, psx_gpu, #(psx_gpu_uvrg_dx_offset + 8); \ |
2375 | \ | |
e1f6de8f | 2376 | vld1.u32 { rg_dx }, [rg_dx_ptr, :64]; \ |
75e28f62 E |
2377 | \ |
2378 | cmp num_spans, #0; \ | |
2379 | bxeq lr; \ | |
2380 | \ | |
2381 | stmdb sp!, { r4 - r11, r14 }; \ | |
9775fc33 | 2382 | save_abi_regs(); \ |
75e28f62 E |
2383 | vshl.u32 rg_dx4, rg_dx, #2; \ |
2384 | \ | |
e1f6de8f | 2385 | ldr b_dx, [psx_gpu, #psx_gpu_b_dx_offset]; \ |
75e28f62 E |
2386 | vshl.u32 rg_dx8, rg_dx, #3; \ |
2387 | \ | |
2388 | add span_uvrg_offset, psx_gpu, #psx_gpu_span_uvrg_offset_offset; \ | |
2389 | add span_edge_data, psx_gpu, #psx_gpu_span_edge_data_offset; \ | |
2390 | \ | |
2391 | add span_b_offset, psx_gpu, #psx_gpu_span_b_offset_offset; \ | |
2392 | vmov.u8 d64_1, #1; \ | |
2393 | \ | |
2394 | vmov.u8 d128_4, #4; \ | |
2395 | vmov.u8 d64_128, #128; \ | |
2396 | \ | |
2397 | vmov.u8 d128_0x7, #0x7; \ | |
2398 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
e1f6de8f | 2399 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
2400 | \ |
2401 | 0: \ | |
e1f6de8f | 2402 | ldrh span_num_blocks, [span_edge_data, #edge_data_num_blocks_offset]; \ |
75e28f62 E |
2403 | add dither_offset_ptr, psx_gpu, #psx_gpu_dither_table_offset; \ |
2404 | \ | |
e1f6de8f | 2405 | ldrh y, [span_edge_data, #edge_data_y_offset]; \ |
2406 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 E |
2407 | \ |
2408 | cmp span_num_blocks, #0; \ | |
2409 | beq 1f; \ | |
2410 | \ | |
e1f6de8f | 2411 | ldrh left_x, [span_edge_data, #edge_data_left_x_offset]; \ |
75e28f62 E |
2412 | add fb_ptr, fb_ptr, y, lsl #11; \ |
2413 | \ | |
e1f6de8f | 2414 | ldr b, [span_b_offset]; \ |
75e28f62 E |
2415 | vdup.u32 v_left_x, left_x; \ |
2416 | and y, y, #0x3; \ | |
2417 | \ | |
e1f6de8f | 2418 | ldr dither_row, [dither_offset_ptr, y, lsl #2]; \ |
75e28f62 E |
2419 | add fb_ptr, fb_ptr, left_x, lsl #1; \ |
2420 | \ | |
2421 | mla b, b_dx, left_x, b; \ | |
2422 | and dither_shift, left_x, #0x03; \ | |
2423 | \ | |
e1f6de8f | 2424 | vld1.u32 { uvrg }, [span_uvrg_offset, :128]; \ |
75e28f62 E |
2425 | vshr.u32 rg_dx, rg_dx4, #2; \ |
2426 | \ | |
2427 | mov dither_shift, dither_shift, lsl #3; \ | |
2428 | vmla.u32 rg, rg_dx, v_left_x; \ | |
2429 | \ | |
2430 | subs span_num_blocks, span_num_blocks, #1; \ | |
2431 | \ | |
2432 | mov dither_row, dither_row, ror dither_shift; \ | |
2433 | mov b_dx4, b_dx, lsl #2; \ | |
2434 | \ | |
2435 | vdup.u32 dither_offsets, dither_row; \ | |
2436 | add block_span_ptr, psx_gpu, #psx_gpu_r_block_span_offset; \ | |
2437 | \ | |
2438 | vdup.u32 b_block, b; \ | |
2439 | vadd.u8 dither_offsets, dither_offsets, d128_4; \ | |
2440 | \ | |
2441 | mov b_dx8, b_dx, lsl #3; \ | |
2442 | vdup.u32 r_block, rg[0]; \ | |
2443 | vdup.u32 g_block, rg[1]; \ | |
2444 | \ | |
e1f6de8f | 2445 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2446 | \ |
2447 | vadd.u32 r_block, r_block, block_span; \ | |
e1f6de8f | 2448 | vld1.u32 { block_span }, [block_span_ptr, :128]!; \ |
75e28f62 E |
2449 | \ |
2450 | vadd.u32 g_block, g_block, block_span; \ | |
e1f6de8f | 2451 | vld1.u32 { block_span }, [block_span_ptr, :128]; \ |
75e28f62 E |
2452 | \ |
2453 | vadd.u32 b_block, b_block, block_span; \ | |
2454 | add block_ptr_b, block_ptr_a, #16; \ | |
2455 | \ | |
2456 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2457 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2458 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2459 | vdup.u32 dx4, rg_dx4[0]; \ | |
2460 | \ | |
2461 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2462 | vdup.u32 dx4, rg_dx4[1]; \ | |
2463 | \ | |
2464 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2465 | vdup.u32 dx4, b_dx4; \ | |
2466 | \ | |
2467 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2468 | vdup.u32 dx8, rg_dx8[0]; \ | |
2469 | \ | |
2470 | vadd.u32 r_block, r_block, dx8; \ | |
2471 | vdup.u32 dx8, rg_dx8[1]; \ | |
2472 | \ | |
2473 | vadd.u32 g_block, g_block, dx8; \ | |
2474 | vdup.u32 dx8, b_dx8; \ | |
2475 | \ | |
2476 | vadd.u32 b_block, b_block, dx8; \ | |
2477 | \ | |
2478 | vmovn.u16 r_whole_8, r_whole; \ | |
2479 | vmovn.u16 g_whole_8, g_whole; \ | |
2480 | vmovn.u16 b_whole_8, b_whole; \ | |
2481 | \ | |
2482 | beq 3f; \ | |
2483 | \ | |
2484 | 2: \ | |
2485 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2486 | vshrn.u32 r_whole_low, r_block, #16; \ | |
2487 | \ | |
2488 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ | |
2489 | vshrn.u32 g_whole_low, g_block, #16; \ | |
2490 | \ | |
2491 | vshrn.u32 b_whole_low, b_block, #16; \ | |
2492 | \ | |
2493 | vdup.u32 dx4, rg_dx4[0]; \ | |
2494 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
2495 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
2496 | \ | |
2497 | vaddhn.u32 r_whole_high, r_block, dx4; \ | |
2498 | vdup.u32 dx4, rg_dx4[1]; \ | |
2499 | \ | |
2500 | vmov pixels, msb_mask; \ | |
2501 | vaddhn.u32 g_whole_high, g_block, dx4; \ | |
2502 | vdup.u32 dx4, b_dx4; \ | |
2503 | \ | |
2504 | vaddhn.u32 b_whole_high, b_block, dx4; \ | |
2505 | vdup.u32 dx8, rg_dx8[0]; \ | |
2506 | \ | |
2507 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
2508 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2509 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2510 | \ | |
2511 | vadd.u32 r_block, r_block, dx8; \ | |
2512 | vdup.u32 dx8, rg_dx8[1]; \ | |
2513 | \ | |
2514 | vadd.u32 g_block, g_block, dx8; \ | |
2515 | vdup.u32 dx8, b_dx8; \ | |
2516 | \ | |
2517 | vadd.u32 b_block, b_block, dx8; \ | |
2518 | \ | |
2519 | vmovn.u16 r_whole_8, r_whole; \ | |
2520 | vmovn.u16 g_whole_8, g_whole; \ | |
2521 | vmovn.u16 b_whole_8, b_whole; \ | |
2522 | \ | |
e1f6de8f | 2523 | vst1.u32 { pixels }, [fb_ptr]!; \ |
75e28f62 E |
2524 | subs span_num_blocks, span_num_blocks, #1; \ |
2525 | bne 2b; \ | |
2526 | \ | |
2527 | 3: \ | |
2528 | setup_blocks_shaded_untextured_dither_a_##dithering(); \ | |
2529 | \ | |
e1f6de8f | 2530 | ldrh right_mask, [span_edge_data, #edge_data_right_mask_offset]; \ |
75e28f62 E |
2531 | setup_blocks_shaded_untextured_dither_b_##dithering(); \ |
2532 | \ | |
2533 | vshr.u8 r_whole_8, r_whole_8, #3; \ | |
3867c6ef | 2534 | rbit right_mask, right_mask; \ |
75e28f62 E |
2535 | vmov pixels, msb_mask; \ |
2536 | vbic.u8 gb_whole_8, gb_whole_8, d128_0x7; \ | |
3867c6ef | 2537 | clz right_mask, right_mask; \ |
75e28f62 E |
2538 | \ |
2539 | vmlal.u8 pixels, r_whole_8, d64_1; \ | |
2540 | vmlal.u8 pixels, g_whole_8, d64_4; \ | |
2541 | vmlal.u8 pixels, b_whole_8, d64_128; \ | |
2542 | \ | |
8184d7c5 | 2543 | JT_OP_REL(100f, right_mask, temp); \ |
e1f6de8f | 2544 | JT_OP(ldr pc, [pc, right_mask, lsl #2]); \ |
3867c6ef | 2545 | nop; \ |
8184d7c5 | 2546 | 100: \ |
3867c6ef | 2547 | nop; \ |
8184d7c5 | 2548 | .word JTE(100b, 4f); \ |
2549 | .word JTE(100b, 5f); \ | |
2550 | .word JTE(100b, 6f); \ | |
2551 | .word JTE(100b, 7f); \ | |
2552 | .word JTE(100b, 8f); \ | |
2553 | .word JTE(100b, 9f); \ | |
2554 | .word JTE(100b, 10f); \ | |
2555 | .word JTE(100b, 11f); \ | |
3867c6ef | 2556 | \ |
75e28f62 | 2557 | 4: \ |
e1f6de8f | 2558 | vst1.u16 { pixels_low[0] }, [fb_ptr]; \ |
3867c6ef E |
2559 | bal 1f; \ |
2560 | \ | |
2561 | 5: \ | |
e1f6de8f | 2562 | vst1.u32 { pixels_low[0] }, [fb_ptr]; \ |
3867c6ef E |
2563 | bal 1f; \ |
2564 | \ | |
2565 | 6: \ | |
e1f6de8f | 2566 | vst1.u32 { pixels_low[0] }, [fb_ptr]!; \ |
2567 | vst1.u16 { pixels_low[2] }, [fb_ptr]; \ | |
3867c6ef E |
2568 | bal 1f; \ |
2569 | \ | |
2570 | 7: \ | |
e1f6de8f | 2571 | vst1.u32 { pixels_low }, [fb_ptr]; \ |
3867c6ef E |
2572 | bal 1f; \ |
2573 | \ | |
2574 | 8: \ | |
e1f6de8f | 2575 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
2576 | vst1.u16 { pixels_high[0] }, [fb_ptr]; \ | |
3867c6ef E |
2577 | bal 1f; \ |
2578 | \ | |
2579 | 9: \ | |
e1f6de8f | 2580 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
2581 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
3867c6ef E |
2582 | bal 1f; \ |
2583 | \ | |
2584 | 10: \ | |
e1f6de8f | 2585 | vst1.u32 { pixels_low }, [fb_ptr]!; \ |
2586 | vst1.u32 { pixels_high[0] }, [fb_ptr]!; \ | |
2587 | vst1.u16 { pixels_high[2] }, [fb_ptr]; \ | |
3867c6ef E |
2588 | bal 1f; \ |
2589 | \ | |
2590 | 11: \ | |
e1f6de8f | 2591 | vst1.u32 { pixels }, [fb_ptr]; \ |
3867c6ef | 2592 | bal 1f; \ |
75e28f62 E |
2593 | \ |
2594 | 1: \ | |
2595 | add span_uvrg_offset, span_uvrg_offset, #16; \ | |
2596 | add span_b_offset, span_b_offset, #4; \ | |
2597 | \ | |
2598 | add span_edge_data, span_edge_data, #8; \ | |
2599 | subs num_spans, num_spans, #1; \ | |
2600 | \ | |
2601 | bne 0b; \ | |
2602 | \ | |
9775fc33 | 2603 | restore_abi_regs(); \ |
75e28f62 E |
2604 | ldmia sp!, { r4 - r11, pc } \ |
2605 | ||
2606 | setup_blocks_shaded_untextured_direct_builder(undithered) | |
2607 | setup_blocks_shaded_untextured_direct_builder(dithered) | |
2608 | ||
2609 | ||
2610 | #undef psx_gpu | |
2611 | #undef num_blocks | |
2612 | #undef triangle | |
2613 | #undef c_64 | |
2614 | ||
2615 | #define psx_gpu r0 | |
2616 | #define block_ptr r1 | |
2617 | #define num_blocks r2 | |
2618 | #define uv_01 r3 | |
2619 | #define uv_23 r4 | |
2620 | #define uv_45 r5 | |
2621 | #define uv_67 r6 | |
2622 | #define uv_0 r7 | |
2623 | #define uv_1 r3 | |
2624 | #define uv_2 r8 | |
2625 | #define uv_3 r4 | |
2626 | #define uv_4 r9 | |
2627 | #define uv_5 r5 | |
2628 | #define uv_6 r10 | |
2629 | #define uv_7 r6 | |
2630 | #define texture_ptr r11 | |
2631 | ||
2632 | #define pixel_0 r7 | |
2633 | #define pixel_1 r3 | |
2634 | #define pixel_2 r8 | |
2635 | #define pixel_3 r4 | |
2636 | #define pixel_4 r9 | |
2637 | #define pixel_5 r5 | |
2638 | #define pixel_6 r10 | |
2639 | #define pixel_7 r6 | |
2640 | ||
2641 | #define pixels_a r7 | |
2642 | #define pixels_b r9 | |
2643 | #define pixels_c r8 | |
2644 | #define pixels_d r10 | |
2645 | ||
2646 | #define c_64 r0 | |
2647 | ||
2648 | #define clut_ptr r12 | |
2649 | #define current_texture_mask r5 | |
2650 | #define dirty_textures_mask r6 | |
2651 | ||
2652 | #define texels d0 | |
2653 | ||
2654 | #define clut_low_a d2 | |
2655 | #define clut_low_b d3 | |
2656 | #define clut_high_a d4 | |
2657 | #define clut_high_b d5 | |
2658 | ||
2659 | #define clut_a q1 | |
2660 | #define clut_b q2 | |
2661 | ||
2662 | #define texels_low d6 | |
2663 | #define texels_high d7 | |
2664 | ||
2665 | .align 3 | |
2666 | ||
2667 | function(texture_blocks_untextured) | |
2668 | bx lr | |
2669 | ||
2670 | ||
2671 | .align 3 | |
2672 | ||
2673 | function(texture_blocks_4bpp) | |
2674 | stmdb sp!, { r3 - r11, r14 } | |
2675 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2676 | ||
e1f6de8f | 2677 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
2678 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
75e28f62 | 2679 | |
e1f6de8f | 2680 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
2681 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128] | |
75e28f62 | 2682 | |
e1f6de8f | 2683 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
75e28f62 E |
2684 | vuzp.u8 clut_a, clut_b |
2685 | ||
e1f6de8f | 2686 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
75e28f62 E |
2687 | tst dirty_textures_mask, current_texture_mask |
2688 | ||
2689 | bne 1f | |
2690 | mov c_64, #64 | |
2691 | ||
2692 | 0: | |
2693 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
2694 | ||
2695 | uxtah uv_0, texture_ptr, uv_01 | |
2696 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
2697 | ||
2698 | uxtah uv_2, texture_ptr, uv_23 | |
2699 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
2700 | ||
2701 | uxtah uv_4, texture_ptr, uv_45 | |
e1f6de8f | 2702 | ldrb pixel_0, [uv_0] |
75e28f62 E |
2703 | |
2704 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
e1f6de8f | 2705 | ldrb pixel_1, [uv_1] |
75e28f62 E |
2706 | |
2707 | uxtah uv_6, texture_ptr, uv_67 | |
e1f6de8f | 2708 | ldrb pixel_2, [uv_2] |
75e28f62 E |
2709 | |
2710 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
e1f6de8f | 2711 | ldrb pixel_3, [uv_3] |
75e28f62 | 2712 | |
e1f6de8f | 2713 | ldrb pixel_4, [uv_4] |
75e28f62 E |
2714 | subs num_blocks, num_blocks, #1 |
2715 | ||
e1f6de8f | 2716 | ldrb pixel_5, [uv_5] |
75e28f62 E |
2717 | orr pixels_a, pixel_0, pixel_1, lsl #8 |
2718 | ||
e1f6de8f | 2719 | ldrb pixel_6, [uv_6] |
75e28f62 E |
2720 | orr pixels_b, pixel_4, pixel_5, lsl #8 |
2721 | ||
e1f6de8f | 2722 | ldrb pixel_7, [uv_7] |
75e28f62 E |
2723 | orr pixels_a, pixels_a, pixel_2, lsl #16 |
2724 | ||
2725 | orr pixels_b, pixels_b, pixel_6, lsl #16 | |
2726 | orr pixels_a, pixels_a, pixel_3, lsl #24 | |
2727 | ||
2728 | orr pixels_b, pixels_b, pixel_7, lsl #24 | |
ed0fd81d | 2729 | vmov texels, pixels_a, pixels_b |
75e28f62 E |
2730 | |
2731 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels | |
2732 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels | |
2733 | ||
e1f6de8f | 2734 | vst2.u8 { texels_low, texels_high }, [block_ptr, :128], c_64 |
75e28f62 E |
2735 | bne 0b |
2736 | ||
2737 | ldmia sp!, { r3 - r11, pc } | |
2738 | ||
2739 | 1: | |
2740 | stmdb sp!, { r1 - r2 } | |
2741 | bl update_texture_4bpp_cache | |
2742 | ||
2743 | mov c_64, #64 | |
2744 | ldmia sp!, { r1 - r2 } | |
2745 | bal 0b | |
2746 | ||
2747 | ||
2748 | .align 3 | |
2749 | ||
2750 | function(texture_blocks_8bpp) | |
2751 | stmdb sp!, { r3 - r11, r14 } | |
2752 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2753 | ||
e1f6de8f | 2754 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
2755 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] | |
75e28f62 | 2756 | |
e1f6de8f | 2757 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] |
2758 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] | |
75e28f62 | 2759 | |
e1f6de8f | 2760 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset] |
75e28f62 E |
2761 | tst dirty_textures_mask, current_texture_mask |
2762 | ||
2763 | bne 1f | |
2764 | nop | |
2765 | ||
2766 | 0: | |
2767 | ldm block_ptr, { uv_01, uv_23, uv_45, uv_67 } | |
2768 | ||
2769 | uxtah uv_0, texture_ptr, uv_01 | |
2770 | uxtah uv_1, texture_ptr, uv_01, ror #16 | |
2771 | ||
2772 | uxtah uv_2, texture_ptr, uv_23 | |
2773 | uxtah uv_3, texture_ptr, uv_23, ror #16 | |
2774 | ||
2775 | uxtah uv_4, texture_ptr, uv_45 | |
e1f6de8f | 2776 | ldrb pixel_0, [uv_0] |
75e28f62 E |
2777 | |
2778 | uxtah uv_5, texture_ptr, uv_45, ror #16 | |
e1f6de8f | 2779 | ldrb pixel_1, [uv_1] |
75e28f62 E |
2780 | |
2781 | uxtah uv_6, texture_ptr, uv_67 | |
e1f6de8f | 2782 | ldrb pixel_2, [uv_2] |
75e28f62 E |
2783 | |
2784 | uxtah uv_7, texture_ptr, uv_67, ror #16 | |
e1f6de8f | 2785 | ldrb pixel_3, [uv_3] |
75e28f62 | 2786 | |
e1f6de8f | 2787 | ldrb pixel_4, [uv_4] |
75e28f62 E |
2788 | add pixel_0, pixel_0, pixel_0 |
2789 | ||
e1f6de8f | 2790 | ldrb pixel_5, [uv_5] |
75e28f62 E |
2791 | add pixel_1, pixel_1, pixel_1 |
2792 | ||
e1f6de8f | 2793 | ldrb pixel_6, [uv_6] |
75e28f62 E |
2794 | add pixel_2, pixel_2, pixel_2 |
2795 | ||
e1f6de8f | 2796 | ldrb pixel_7, [uv_7] |
75e28f62 E |
2797 | add pixel_3, pixel_3, pixel_3 |
2798 | ||
e1f6de8f | 2799 | ldrh pixel_0, [clut_ptr, pixel_0] |
75e28f62 E |
2800 | add pixel_4, pixel_4, pixel_4 |
2801 | ||
e1f6de8f | 2802 | ldrh pixel_1, [clut_ptr, pixel_1] |
75e28f62 E |
2803 | add pixel_5, pixel_5, pixel_5 |
2804 | ||
e1f6de8f | 2805 | ldrh pixel_2, [clut_ptr, pixel_2] |
75e28f62 E |
2806 | add pixel_6, pixel_6, pixel_6 |
2807 | ||
e1f6de8f | 2808 | ldrh pixel_3, [clut_ptr, pixel_3] |
75e28f62 E |
2809 | add pixel_7, pixel_7, pixel_7 |
2810 | ||
e1f6de8f | 2811 | ldrh pixel_4, [clut_ptr, pixel_4] |
75e28f62 E |
2812 | orr pixels_a, pixel_0, pixel_1, lsl #16 |
2813 | ||
e1f6de8f | 2814 | ldrh pixel_5, [clut_ptr, pixel_5] |
75e28f62 E |
2815 | orr pixels_c, pixel_2, pixel_3, lsl #16 |
2816 | ||
e1f6de8f | 2817 | ldrh pixel_6, [clut_ptr, pixel_6] |
75e28f62 E |
2818 | subs num_blocks, num_blocks, #1 |
2819 | ||
e1f6de8f | 2820 | ldrh pixel_7, [clut_ptr, pixel_7] |
75e28f62 E |
2821 | orr pixels_b, pixel_4, pixel_5, lsl #16 |
2822 | ||
2823 | orr pixels_d, pixel_6, pixel_7, lsl #16 | |
2824 | stm block_ptr, { pixels_a, pixels_c, pixels_b, pixels_d } | |
2825 | ||
2826 | add block_ptr, block_ptr, #64 | |
2827 | bne 0b | |
2828 | ||
2829 | ldmia sp!, { r3 - r11, pc } | |
2830 | ||
2831 | 1: | |
4d646738 | 2832 | stmdb sp!, { r1 - r2, EXTRA_UNSAVED_REGS r12 } |
75e28f62 E |
2833 | |
2834 | bl update_texture_8bpp_cache | |
2835 | ||
4d646738 | 2836 | ldmia sp!, { r1 - r2, EXTRA_UNSAVED_REGS r12 } |
75e28f62 E |
2837 | bal 0b |
2838 | ||
2839 | ||
2840 | #undef uv_0 | |
2841 | #undef uv_1 | |
2842 | #undef uv_2 | |
2843 | #undef uv_3 | |
2844 | #undef uv_4 | |
2845 | #undef uv_5 | |
2846 | #undef uv_6 | |
2847 | #undef uv_7 | |
2848 | ||
2849 | #undef pixel_0 | |
2850 | #undef pixel_1 | |
2851 | #undef pixel_2 | |
2852 | #undef pixel_3 | |
2853 | #undef pixel_4 | |
2854 | #undef pixel_5 | |
2855 | #undef pixel_6 | |
2856 | #undef pixel_7 | |
2857 | ||
2858 | #undef texture_ptr | |
2859 | ||
2860 | #undef pixels_a | |
2861 | #undef pixels_b | |
2862 | #undef pixels_c | |
2863 | #undef pixels_d | |
2864 | ||
2865 | #define psx_gpu r0 | |
2866 | #define block_ptr r1 | |
2867 | #define num_blocks r2 | |
2868 | ||
2869 | #define uv_0 r3 | |
2870 | #define uv_1 r4 | |
2871 | #define u_0 r3 | |
2872 | #define u_1 r4 | |
2873 | #define v_0 r5 | |
2874 | #define v_1 r6 | |
2875 | ||
2876 | #define uv_2 r5 | |
2877 | #define uv_3 r6 | |
2878 | #define u_2 r5 | |
2879 | #define u_3 r6 | |
2880 | #define v_2 r7 | |
2881 | #define v_3 r8 | |
2882 | ||
2883 | #define uv_4 r7 | |
2884 | #define uv_5 r8 | |
2885 | #define u_4 r7 | |
2886 | #define u_5 r8 | |
2887 | #define v_4 r9 | |
2888 | #define v_5 r10 | |
2889 | ||
2890 | #define uv_6 r9 | |
2891 | #define uv_7 r10 | |
2892 | #define u_6 r9 | |
2893 | #define u_7 r10 | |
2894 | #define v_6 r11 | |
2895 | #define v_7 r0 | |
2896 | ||
2897 | #define pixel_0 r3 | |
2898 | #define pixel_1 r4 | |
2899 | #define pixel_2 r5 | |
2900 | #define pixel_3 r6 | |
2901 | #define pixel_4 r7 | |
2902 | #define pixel_5 r8 | |
2903 | #define pixel_6 r9 | |
2904 | #define pixel_7 r10 | |
2905 | ||
2906 | #define pixels_a r3 | |
2907 | #define pixels_b r5 | |
2908 | #define pixels_c r7 | |
2909 | #define pixels_d r9 | |
2910 | ||
2911 | #define texture_ptr r12 | |
2912 | ||
2913 | ||
2914 | .align 3 | |
2915 | ||
2916 | function(texture_blocks_16bpp) | |
2917 | stmdb sp!, { r3 - r11, r14 } | |
2918 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
2919 | ||
e1f6de8f | 2920 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
2921 | ldr texture_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] | |
75e28f62 E |
2922 | |
2923 | 0: | |
e1f6de8f | 2924 | ldrh uv_0, [block_ptr] |
75e28f62 E |
2925 | subs num_blocks, num_blocks, #1 |
2926 | ||
e1f6de8f | 2927 | ldrh uv_1, [block_ptr, #2] |
75e28f62 E |
2928 | |
2929 | and v_0, uv_0, #0xFF00 | |
2930 | and v_1, uv_1, #0xFF00 | |
2931 | ||
2932 | and u_0, uv_0, #0xFF | |
2933 | and u_1, uv_1, #0xFF | |
2934 | ||
2935 | add uv_0, u_0, v_0, lsl #2 | |
e1f6de8f | 2936 | ldrh uv_2, [block_ptr, #4] |
75e28f62 E |
2937 | |
2938 | add uv_1, u_1, v_1, lsl #2 | |
e1f6de8f | 2939 | ldrh uv_3, [block_ptr, #6] |
75e28f62 E |
2940 | |
2941 | add uv_0, uv_0, uv_0 | |
2942 | add uv_1, uv_1, uv_1 | |
2943 | ||
2944 | and v_2, uv_2, #0xFF00 | |
2945 | and v_3, uv_3, #0xFF00 | |
2946 | ||
2947 | and u_2, uv_2, #0xFF | |
2948 | and u_3, uv_3, #0xFF | |
2949 | ||
2950 | add uv_2, u_2, v_2, lsl #2 | |
e1f6de8f | 2951 | ldrh uv_4, [block_ptr, #8] |
75e28f62 E |
2952 | |
2953 | add uv_3, u_3, v_3, lsl #2 | |
e1f6de8f | 2954 | ldrh uv_5, [block_ptr, #10] |
75e28f62 E |
2955 | |
2956 | add uv_2, uv_2, uv_2 | |
2957 | add uv_3, uv_3, uv_3 | |
2958 | ||
2959 | and v_4, uv_4, #0xFF00 | |
2960 | and v_5, uv_5, #0xFF00 | |
2961 | ||
2962 | and u_4, uv_4, #0xFF | |
2963 | and u_5, uv_5, #0xFF | |
2964 | ||
2965 | add uv_4, u_4, v_4, lsl #2 | |
e1f6de8f | 2966 | ldrh uv_6, [block_ptr, #12] |
75e28f62 E |
2967 | |
2968 | add uv_5, u_5, v_5, lsl #2 | |
e1f6de8f | 2969 | ldrh uv_7, [block_ptr, #14] |
75e28f62 E |
2970 | |
2971 | add uv_4, uv_4, uv_4 | |
e1f6de8f | 2972 | ldrh pixel_0, [texture_ptr, uv_0] |
75e28f62 E |
2973 | |
2974 | add uv_5, uv_5, uv_5 | |
e1f6de8f | 2975 | ldrh pixel_1, [texture_ptr, uv_1] |
75e28f62 E |
2976 | |
2977 | and v_6, uv_6, #0xFF00 | |
e1f6de8f | 2978 | ldrh pixel_2, [texture_ptr, uv_2] |
75e28f62 E |
2979 | |
2980 | and v_7, uv_7, #0xFF00 | |
e1f6de8f | 2981 | ldrh pixel_3, [texture_ptr, uv_3] |
75e28f62 E |
2982 | |
2983 | and u_6, uv_6, #0xFF | |
e1f6de8f | 2984 | ldrh pixel_4, [texture_ptr, uv_4] |
75e28f62 E |
2985 | |
2986 | and u_7, uv_7, #0xFF | |
e1f6de8f | 2987 | ldrh pixel_5, [texture_ptr, uv_5] |
75e28f62 E |
2988 | |
2989 | add uv_6, u_6, v_6, lsl #2 | |
2990 | add uv_7, u_7, v_7, lsl #2 | |
2991 | ||
2992 | add uv_6, uv_6, uv_6 | |
2993 | add uv_7, uv_7, uv_7 | |
2994 | ||
2995 | orr pixels_a, pixel_0, pixel_1, lsl #16 | |
2996 | orr pixels_b, pixel_2, pixel_3, lsl #16 | |
2997 | ||
e1f6de8f | 2998 | ldrh pixel_6, [texture_ptr, uv_6] |
75e28f62 E |
2999 | orr pixels_c, pixel_4, pixel_5, lsl #16 |
3000 | ||
e1f6de8f | 3001 | ldrh pixel_7, [texture_ptr, uv_7] |
75e28f62 E |
3002 | orr pixels_d, pixel_6, pixel_7, lsl #16 |
3003 | ||
3004 | stm block_ptr, { pixels_a, pixels_b, pixels_c, pixels_d } | |
3005 | add block_ptr, block_ptr, #64 | |
3006 | ||
3007 | bne 0b | |
3008 | ||
3009 | ldmia sp!, { r3 - r11, pc } | |
3010 | ||
3011 | ||
3012 | #undef num_blocks | |
3013 | ||
3014 | #undef test_mask | |
3015 | #undef texels | |
3016 | #undef pixels_b | |
3017 | #undef pixels | |
3018 | #undef d64_1 | |
3019 | #undef d64_4 | |
3020 | #undef d64_128 | |
3021 | #undef draw_mask | |
3022 | #undef msb_mask | |
3023 | #undef msb_mask_low | |
3024 | #undef msb_mask_high | |
3025 | #undef fb_pixels | |
3026 | ||
3027 | #undef c_32 | |
3028 | #undef fb_ptr | |
3029 | #undef mask_msb_ptr | |
3030 | ||
3031 | #define psx_gpu r0 | |
3032 | #define num_blocks r1 | |
3033 | #define color_ptr r2 | |
3867c6ef E |
3034 | #define colors_scalar r2 |
3035 | #define colors_scalar_compare r3 | |
75e28f62 E |
3036 | #define mask_msb_ptr r2 |
3037 | ||
3038 | #define block_ptr_load_a r0 | |
3039 | #define block_ptr_store r3 | |
3040 | #define block_ptr_load_b r12 | |
3041 | #define c_32 r2 | |
3042 | ||
3043 | #define c_48 r4 | |
3044 | #define fb_ptr r14 | |
3045 | #define draw_mask_bits_scalar r5 | |
3046 | ||
3047 | #define d128_0x07 q0 | |
3048 | #define d128_0x1F q1 | |
3049 | #define d128_0x8000 q2 | |
3050 | #define test_mask q3 | |
3051 | #define texels q4 | |
3052 | #define colors_rg q5 | |
3053 | #define colors_b_dm_bits q6 | |
3054 | #define texels_rg q7 | |
3055 | #define pixels_r q8 | |
3056 | #define pixels_g q9 | |
3057 | #define pixels_b q10 | |
3058 | #define pixels q11 | |
3059 | #define zero_mask q4 | |
3060 | #define draw_mask q12 | |
3061 | #define msb_mask q13 | |
3062 | ||
3063 | #define fb_pixels q8 | |
3064 | ||
3065 | #define pixels_gb_low q9 | |
3066 | ||
3067 | #define colors_r d10 | |
3068 | #define colors_g d11 | |
3069 | #define colors_b d12 | |
3070 | #define draw_mask_bits d13 | |
3071 | #define texels_r d14 | |
3072 | #define texels_g d15 | |
3073 | #define pixels_r_low d16 | |
3074 | #define pixels_g_low d18 | |
3075 | #define pixels_b_low d19 | |
3076 | #define msb_mask_low d26 | |
3077 | #define msb_mask_high d27 | |
3078 | ||
3079 | #define d64_1 d28 | |
3080 | #define d64_4 d29 | |
3081 | #define d64_128 d30 | |
3082 | #define texels_b d31 | |
3083 | ||
3084 | #define shade_blocks_textured_modulated_prologue_indirect() \ | |
3085 | mov c_48, #48; \ | |
3086 | add block_ptr_store, psx_gpu, #psx_gpu_blocks_offset \ | |
3087 | ||
3088 | #define shade_blocks_textured_modulated_prologue_direct() \ | |
3089 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ | |
e1f6de8f | 3090 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] \ |
75e28f62 | 3091 | |
75e28f62 | 3092 | |
3867c6ef E |
3093 | #define shade_blocks_textured_modulated_prologue_shaded(dithering, target) \ |
3094 | ||
3095 | #define shade_blocks_textured_false_modulation_check_undithered(target) \ | |
e1f6de8f | 3096 | ldr colors_scalar, [psx_gpu, #psx_gpu_triangle_color_offset]; \ |
3867c6ef E |
3097 | movw colors_scalar_compare, #0x8080; \ |
3098 | \ | |
3099 | movt colors_scalar_compare, #0x80; \ | |
3100 | cmp colors_scalar, colors_scalar_compare; \ | |
3101 | beq shade_blocks_textured_unmodulated_##target \ | |
3102 | ||
3103 | #define shade_blocks_textured_false_modulation_check_dithered(target) \ | |
3104 | ||
3105 | #define shade_blocks_textured_modulated_prologue_unshaded(dithering, target) \ | |
3106 | shade_blocks_textured_false_modulation_check_##dithering(target); \ | |
75e28f62 | 3107 | add color_ptr, psx_gpu, #psx_gpu_triangle_color_offset; \ |
e1f6de8f | 3108 | vld1.u32 { colors_r[] }, [color_ptr, :32]; \ |
75e28f62 E |
3109 | vdup.u8 colors_g, colors_r[1]; \ |
3110 | vdup.u8 colors_b, colors_r[2]; \ | |
3111 | vdup.u8 colors_r, colors_r[0] \ | |
3112 | ||
3113 | ||
3114 | #define shade_blocks_textured_modulated_load_dithered(target) \ | |
e1f6de8f | 3115 | vld1.u32 { target }, [block_ptr_load_b, :128] \ |
75e28f62 E |
3116 | |
3117 | #define shade_blocks_textured_modulated_load_last_dithered(target) \ | |
e1f6de8f | 3118 | vld1.u32 { target }, [block_ptr_load_b, :128], c_32 \ |
75e28f62 E |
3119 | |
3120 | #define shade_blocks_textured_modulated_load_undithered(target) \ | |
3121 | ||
3122 | #define shade_blocks_textured_modulated_load_last_undithered(target) \ | |
3123 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
3124 | ||
3125 | #define shade_blocks_textured_modulate_dithered(channel) \ | |
3126 | vmlal.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
3127 | ||
3128 | #define shade_blocks_textured_modulate_undithered(channel) \ | |
3129 | vmull.u8 pixels_##channel, texels_##channel, colors_##channel \ | |
3130 | ||
3131 | ||
3132 | #define shade_blocks_textured_modulated_store_draw_mask_indirect(offset) \ | |
e1f6de8f | 3133 | vst1.u32 { draw_mask }, [block_ptr_store, :128]! \ |
75e28f62 E |
3134 | |
3135 | #define shade_blocks_textured_modulated_store_draw_mask_direct(offset) \ | |
e1f6de8f | 3136 | ldr fb_ptr, [block_ptr_load_b, #(offset - 64)]; \ |
3137 | vld1.u32 { fb_pixels }, [fb_ptr]; \ | |
75e28f62 E |
3138 | vbit.u16 pixels, fb_pixels, draw_mask \ |
3139 | ||
3140 | #define shade_blocks_textured_modulated_store_pixels_indirect() \ | |
e1f6de8f | 3141 | vst1.u32 { pixels }, [block_ptr_store, :128], c_48 \ |
75e28f62 E |
3142 | |
3143 | #define shade_blocks_textured_modulated_store_pixels_direct() \ | |
e1f6de8f | 3144 | vst1.u32 { pixels }, [fb_ptr] \ |
75e28f62 E |
3145 | |
3146 | ||
3147 | #define shade_blocks_textured_modulated_load_rg_shaded() \ | |
e1f6de8f | 3148 | vld1.u32 { colors_r, colors_g }, [block_ptr_load_b, :128], c_32 \ |
75e28f62 E |
3149 | |
3150 | #define shade_blocks_textured_modulated_load_rg_unshaded() \ | |
3151 | add block_ptr_load_b, block_ptr_load_b, #32 \ | |
3152 | ||
3153 | #define shade_blocks_textured_modulated_load_bdm_shaded() \ | |
e1f6de8f | 3154 | vld1.u32 { colors_b, draw_mask_bits }, [block_ptr_load_a, :128], c_32 \ |
75e28f62 E |
3155 | |
3156 | #define shade_blocks_textured_modulated_load_bdm_unshaded() \ | |
e1f6de8f | 3157 | ldr draw_mask_bits_scalar, [block_ptr_load_a, #8]; \ |
75e28f62 E |
3158 | add block_ptr_load_a, block_ptr_load_a, #32 \ |
3159 | ||
3160 | #define shade_blocks_textured_modulated_expand_draw_mask_shaded() \ | |
3161 | vdup.u16 draw_mask, draw_mask_bits[0] \ | |
3162 | ||
3163 | #define shade_blocks_textured_modulated_expand_draw_mask_unshaded() \ | |
3164 | vdup.u16 draw_mask, draw_mask_bits_scalar \ | |
3165 | ||
3166 | ||
3167 | #define shade_blocks_textured_modulated_apply_msb_mask_indirect() \ | |
3168 | ||
3169 | #define shade_blocks_textured_modulated_apply_msb_mask_direct() \ | |
3170 | vorr.u16 pixels, pixels, msb_mask \ | |
3171 | ||
3172 | ||
3173 | #define shade_blocks_textured_modulated_builder(shading, dithering, target) \ | |
3174 | .align 3; \ | |
3175 | \ | |
3176 | function(shade_blocks_##shading##_textured_modulated_##dithering##_##target) \ | |
7255f1c2 | 3177 | save_abi_regs(); \ |
3867c6ef | 3178 | shade_blocks_textured_modulated_prologue_##shading(dithering, target); \ |
75e28f62 | 3179 | stmdb sp!, { r4 - r5, lr }; \ |
e1f6de8f | 3180 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 | 3181 | \ |
e1f6de8f | 3182 | vld1.u32 { test_mask }, [psx_gpu, :128]; \ |
75e28f62 E |
3183 | \ |
3184 | shade_blocks_textured_modulated_prologue_##target(); \ | |
75e28f62 E |
3185 | \ |
3186 | add block_ptr_load_a, psx_gpu, #psx_gpu_blocks_offset; \ | |
3187 | mov c_32, #32; \ | |
3188 | \ | |
3189 | add block_ptr_load_b, block_ptr_load_a, #16; \ | |
3190 | vmov.u8 d64_1, #1; \ | |
3191 | vmov.u8 d64_4, #4; \ | |
3192 | vmov.u8 d64_128, #128; \ | |
3193 | \ | |
e1f6de8f | 3194 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
75e28f62 E |
3195 | vmov.u8 d128_0x07, #0x07; \ |
3196 | \ | |
3197 | shade_blocks_textured_modulated_load_rg_##shading(); \ | |
3198 | vmov.u8 d128_0x1F, #0x1F; \ | |
3199 | \ | |
3200 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
3201 | vmov.u16 d128_0x8000, #0x8000; \ | |
3202 | \ | |
3203 | vmovn.u16 texels_r, texels; \ | |
3204 | vshrn.u16 texels_g, texels, #5; \ | |
3205 | \ | |
3206 | vshrn.u16 texels_b, texels, #7; \ | |
3207 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
3208 | \ | |
3209 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
3210 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
3211 | \ | |
3212 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
3213 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
3214 | \ | |
3215 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
3216 | vshr.u8 texels_b, texels_b, #3; \ | |
3217 | \ | |
3218 | shade_blocks_textured_modulate_##dithering(r); \ | |
3219 | shade_blocks_textured_modulate_##dithering(g); \ | |
3220 | shade_blocks_textured_modulate_##dithering(b); \ | |
3221 | \ | |
3222 | vand.u16 pixels, texels, d128_0x8000; \ | |
3223 | vceq.u16 zero_mask, texels, #0; \ | |
3224 | \ | |
3225 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
3226 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
3227 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
3228 | \ | |
3229 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
3230 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
3231 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
3232 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
3233 | \ | |
3234 | subs num_blocks, num_blocks, #1; \ | |
3235 | beq 1f; \ | |
3236 | \ | |
3237 | .align 3; \ | |
3238 | \ | |
3239 | 0: \ | |
e1f6de8f | 3240 | vld1.u32 { texels }, [block_ptr_load_a, :128], c_32; \ |
75e28f62 E |
3241 | shade_blocks_textured_modulated_load_rg_##shading(); \ |
3242 | vshrn.u16 texels_g, texels, #5; \ | |
3243 | \ | |
3244 | shade_blocks_textured_modulated_load_bdm_##shading(); \ | |
3245 | vshrn.u16 texels_b, texels, #7; \ | |
3246 | \ | |
e1f6de8f | 3247 | pld [block_ptr_load_a]; \ |
75e28f62 E |
3248 | vmovn.u16 texels_r, texels; \ |
3249 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
3250 | \ | |
3251 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
3252 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
3253 | shade_blocks_textured_modulated_store_draw_mask_##target(-4); \ | |
3254 | \ | |
3255 | shade_blocks_textured_modulated_load_##dithering(pixels_r); \ | |
3256 | shade_blocks_textured_modulated_expand_draw_mask_##shading(); \ | |
3257 | \ | |
3258 | shade_blocks_textured_modulated_load_##dithering(pixels_g); \ | |
3259 | vand.u8 texels_rg, texels_rg, d128_0x1F; \ | |
3260 | \ | |
3261 | shade_blocks_textured_modulated_load_last_##dithering(pixels_b); \ | |
3262 | vtst.u16 draw_mask, draw_mask, test_mask; \ | |
3263 | \ | |
3264 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
3265 | vshr.u8 texels_b, texels_b, #3; \ | |
3266 | \ | |
3267 | shade_blocks_textured_modulate_##dithering(r); \ | |
3268 | shade_blocks_textured_modulate_##dithering(g); \ | |
3269 | shade_blocks_textured_modulate_##dithering(b); \ | |
3270 | \ | |
3271 | vand.u16 pixels, texels, d128_0x8000; \ | |
3272 | vceq.u16 zero_mask, texels, #0; \ | |
3273 | \ | |
3274 | subs num_blocks, num_blocks, #1; \ | |
3275 | \ | |
3276 | vqshrun.s16 pixels_r_low, pixels_r, #4; \ | |
3277 | vqshrun.s16 pixels_g_low, pixels_g, #4; \ | |
3278 | vqshrun.s16 pixels_b_low, pixels_b, #4; \ | |
3279 | \ | |
3280 | shade_blocks_textured_modulated_apply_msb_mask_##target(); \ | |
3281 | vorr.u16 draw_mask, draw_mask, zero_mask; \ | |
3282 | vshr.u8 pixels_r_low, pixels_r_low, #3; \ | |
3283 | vbic.u8 pixels_gb_low, pixels_gb_low, d128_0x07; \ | |
3284 | \ | |
3285 | bne 0b; \ | |
3286 | \ | |
3287 | 1: \ | |
3288 | vmlal.u8 pixels, pixels_r_low, d64_1; \ | |
3289 | vmlal.u8 pixels, pixels_g_low, d64_4; \ | |
3290 | vmlal.u8 pixels, pixels_b_low, d64_128; \ | |
3291 | \ | |
3292 | shade_blocks_textured_modulated_store_draw_mask_##target(28); \ | |
3293 | shade_blocks_textured_modulated_store_pixels_##target(); \ | |
3294 | \ | |
7255f1c2 | 3295 | ldmia sp!, { r4 - r5, lr }; \ |
9775fc33 | 3296 | restore_abi_regs(); \ |
7255f1c2 | 3297 | bx lr \ |
75e28f62 E |
3298 | |
3299 | ||
3300 | shade_blocks_textured_modulated_builder(shaded, dithered, direct); | |
3301 | shade_blocks_textured_modulated_builder(shaded, undithered, direct); | |
3302 | shade_blocks_textured_modulated_builder(unshaded, dithered, direct); | |
3303 | shade_blocks_textured_modulated_builder(unshaded, undithered, direct); | |
3304 | ||
3305 | shade_blocks_textured_modulated_builder(shaded, dithered, indirect); | |
3306 | shade_blocks_textured_modulated_builder(shaded, undithered, indirect); | |
3307 | shade_blocks_textured_modulated_builder(unshaded, dithered, indirect); | |
3308 | shade_blocks_textured_modulated_builder(unshaded, undithered, indirect); | |
3309 | ||
3310 | ||
3311 | #undef c_64 | |
3312 | #undef fb_ptr | |
3313 | #undef color_ptr | |
3314 | ||
3315 | #undef color_r | |
3316 | #undef color_g | |
3317 | #undef color_b | |
3318 | ||
3319 | #undef test_mask | |
3320 | #undef pixels | |
3321 | #undef draw_mask | |
3322 | #undef zero_mask | |
3323 | #undef fb_pixels | |
3324 | #undef msb_mask | |
3325 | #undef msb_mask_low | |
3326 | #undef msb_mask_high | |
3327 | ||
3328 | #define psx_gpu r0 | |
3329 | #define num_blocks r1 | |
3330 | #define mask_msb_ptr r2 | |
3331 | #define color_ptr r3 | |
3332 | ||
3333 | #define block_ptr_load r0 | |
3334 | #define draw_mask_store_ptr r3 | |
3335 | #define draw_mask_bits_ptr r12 | |
3336 | #define draw_mask_ptr r12 | |
3337 | #define pixel_store_ptr r14 | |
3338 | ||
3339 | #define fb_ptr_cmp r4 | |
3340 | ||
3341 | #define fb_ptr r3 | |
3342 | #define fb_ptr_next r14 | |
3343 | ||
3344 | #define c_64 r2 | |
3345 | ||
3346 | #define test_mask q0 | |
3347 | #define pixels q1 | |
3348 | #define draw_mask q2 | |
3349 | #define zero_mask q3 | |
3350 | #define draw_mask_combined q4 | |
3351 | #define fb_pixels q5 | |
3352 | #define fb_pixels_next q6 | |
3353 | #define msb_mask q7 | |
3354 | ||
3355 | #define draw_mask_low d4 | |
3356 | #define draw_mask_high d5 | |
3357 | #define msb_mask_low d14 | |
3358 | #define msb_mask_high d15 | |
3359 | ||
3360 | .align 3 | |
3361 | function(shade_blocks_textured_unmodulated_indirect) | |
9775fc33 | 3362 | stmdb sp!, { r4, r14 } |
3363 | save_abi_regs() | |
75e28f62 E |
3364 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
3365 | ||
e1f6de8f | 3366 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
3367 | add pixel_store_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
3368 | ||
e1f6de8f | 3369 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 E |
3370 | add draw_mask_store_ptr, psx_gpu, #psx_gpu_blocks_offset |
3371 | ||
3372 | mov c_64, #64 | |
3373 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset | |
3374 | ||
e1f6de8f | 3375 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
75e28f62 | 3376 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ |
e1f6de8f | 3377 | [draw_mask_bits_ptr, :16], c_64 |
75e28f62 E |
3378 | vceq.u16 zero_mask, pixels, #0 |
3379 | ||
3380 | vtst.u16 draw_mask, draw_mask, test_mask | |
e1f6de8f | 3381 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
75e28f62 E |
3382 | |
3383 | subs num_blocks, num_blocks, #1 | |
3384 | beq 1f | |
3385 | ||
3386 | 0: | |
e1f6de8f | 3387 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
75e28f62 E |
3388 | vorr.u16 draw_mask_combined, draw_mask, zero_mask |
3389 | ||
3390 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
e1f6de8f | 3391 | [draw_mask_bits_ptr, :16], c_64 |
75e28f62 E |
3392 | vceq.u16 zero_mask, pixels, #0 |
3393 | ||
3394 | vtst.u16 draw_mask, draw_mask, test_mask | |
e1f6de8f | 3395 | vst1.u32 { pixels }, [pixel_store_ptr, :128], c_64 |
75e28f62 | 3396 | |
e1f6de8f | 3397 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
75e28f62 E |
3398 | subs num_blocks, num_blocks, #1 |
3399 | ||
3400 | bne 0b | |
3401 | ||
3402 | 1: | |
3403 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
e1f6de8f | 3404 | vst1.u32 { draw_mask_combined }, [draw_mask_store_ptr, :128], c_64 |
75e28f62 | 3405 | |
9775fc33 | 3406 | restore_abi_regs() |
3407 | ldmia sp!, { r4, pc } | |
75e28f62 E |
3408 | |
3409 | ||
3410 | .align 3 | |
3411 | ||
3412 | function(shade_blocks_textured_unmodulated_direct) | |
3413 | stmdb sp!, { r4, r14 } | |
9775fc33 | 3414 | save_abi_regs() |
75e28f62 E |
3415 | add draw_mask_bits_ptr, psx_gpu, #(psx_gpu_blocks_offset + 40) |
3416 | ||
e1f6de8f | 3417 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
3418 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
3419 | ||
e1f6de8f | 3420 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
75e28f62 E |
3421 | mov c_64, #64 |
3422 | ||
e1f6de8f | 3423 | vld1.u32 { test_mask }, [psx_gpu, :128] |
75e28f62 E |
3424 | add block_ptr_load, psx_gpu, #psx_gpu_blocks_offset |
3425 | ||
3426 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
e1f6de8f | 3427 | [draw_mask_bits_ptr, :16], c_64 |
3428 | ldr fb_ptr_next, [block_ptr_load, #44] | |
75e28f62 | 3429 | |
e1f6de8f | 3430 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
3431 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
75e28f62 E |
3432 | vceq.u16 zero_mask, pixels, #0 |
3433 | vtst.u16 draw_mask, draw_mask, test_mask | |
3434 | ||
3435 | subs num_blocks, num_blocks, #1 | |
3436 | beq 1f | |
3437 | ||
3438 | 0: | |
3439 | mov fb_ptr, fb_ptr_next | |
e1f6de8f | 3440 | ldr fb_ptr_next, [block_ptr_load, #44] |
75e28f62 E |
3441 | |
3442 | vorr.u16 pixels, pixels, msb_mask | |
3443 | ||
3444 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3445 | vmov fb_pixels, fb_pixels_next | |
3446 | ||
3447 | vld1.u16 { draw_mask_low[], draw_mask_high[] }, \ | |
e1f6de8f | 3448 | [draw_mask_bits_ptr, :16], c_64 |
75e28f62 E |
3449 | vbif.u16 fb_pixels, pixels, draw_mask_combined |
3450 | ||
75e28f62 | 3451 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr |
e1f6de8f | 3452 | pld [fb_ptr_next, #64] |
8438c3c7 | 3453 | |
75e28f62 | 3454 | add fb_ptr_cmp, fb_ptr_cmp, #14 |
e1f6de8f | 3455 | vld1.u32 { pixels }, [block_ptr_load, :128], c_64 |
8438c3c7 | 3456 | |
75e28f62 E |
3457 | cmp fb_ptr_cmp, #28 |
3458 | bls 4f | |
3459 | ||
e1f6de8f | 3460 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 E |
3461 | vceq.u16 zero_mask, pixels, #0 |
3462 | ||
e1f6de8f | 3463 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 E |
3464 | vtst.u16 draw_mask, draw_mask, test_mask |
3465 | ||
3466 | 3: | |
3467 | subs num_blocks, num_blocks, #1 | |
3468 | bne 0b | |
3469 | ||
3470 | 1: | |
3471 | vorr.u16 draw_mask_combined, draw_mask, zero_mask | |
3472 | vbif.u16 fb_pixels_next, pixels, draw_mask_combined | |
3473 | ||
e1f6de8f | 3474 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 | 3475 | |
9775fc33 | 3476 | restore_abi_regs() |
75e28f62 E |
3477 | ldmia sp!, { r4, pc } |
3478 | ||
3479 | 4: | |
e1f6de8f | 3480 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 E |
3481 | vceq.u16 zero_mask, pixels, #0 |
3482 | ||
e1f6de8f | 3483 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 E |
3484 | vtst.u16 draw_mask, draw_mask, test_mask |
3485 | ||
3486 | bal 3b | |
3487 | ||
3488 | ||
3489 | function(shade_blocks_unshaded_untextured_indirect) | |
3490 | bx lr | |
3491 | ||
3492 | .align 3 | |
3493 | ||
3494 | function(shade_blocks_unshaded_untextured_direct) | |
3495 | stmdb sp!, { r4, r14 } | |
9775fc33 | 3496 | save_abi_regs() |
75e28f62 E |
3497 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset |
3498 | ||
e1f6de8f | 3499 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
3500 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
3501 | ||
e1f6de8f | 3502 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
75e28f62 E |
3503 | add color_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) |
3504 | ||
3505 | add block_ptr_load, psx_gpu, #(psx_gpu_blocks_offset + 44) | |
e1f6de8f | 3506 | vld1.u16 { pixels }, [color_ptr, :128] |
75e28f62 E |
3507 | |
3508 | mov c_64, #64 | |
e1f6de8f | 3509 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
75e28f62 E |
3510 | |
3511 | vorr.u16 pixels, pixels, msb_mask | |
3512 | subs num_blocks, num_blocks, #1 | |
3513 | ||
e1f6de8f | 3514 | ldr fb_ptr_next, [block_ptr_load], #64 |
75e28f62 | 3515 | |
e1f6de8f | 3516 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 E |
3517 | beq 1f |
3518 | ||
3519 | 0: | |
3520 | vmov fb_pixels, fb_pixels_next | |
3521 | mov fb_ptr, fb_ptr_next | |
e1f6de8f | 3522 | ldr fb_ptr_next, [block_ptr_load], #64 |
75e28f62 E |
3523 | |
3524 | vbif.u16 fb_pixels, pixels, draw_mask | |
e1f6de8f | 3525 | vld1.u16 { draw_mask }, [draw_mask_ptr, :128], c_64 |
75e28f62 E |
3526 | |
3527 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr | |
3528 | add fb_ptr_cmp, fb_ptr_cmp, #14 | |
3529 | cmp fb_ptr_cmp, #28 | |
3530 | bls 4f | |
3531 | ||
e1f6de8f | 3532 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] |
3533 | vst1.u16 { fb_pixels }, [fb_ptr] | |
75e28f62 E |
3534 | |
3535 | 3: | |
3536 | subs num_blocks, num_blocks, #1 | |
3537 | bne 0b | |
3538 | ||
3539 | 1: | |
3540 | vbif.u16 fb_pixels_next, pixels, draw_mask | |
e1f6de8f | 3541 | vst1.u16 { fb_pixels_next }, [fb_ptr_next] |
75e28f62 | 3542 | |
9775fc33 | 3543 | restore_abi_regs() |
75e28f62 E |
3544 | ldmia sp!, { r4, pc } |
3545 | ||
3546 | 4: | |
e1f6de8f | 3547 | vst1.u16 { fb_pixels }, [fb_ptr] |
3548 | vld1.u16 { fb_pixels_next }, [fb_ptr_next] | |
75e28f62 E |
3549 | bal 3b |
3550 | ||
3551 | ||
3552 | #undef draw_mask_ptr | |
3553 | #undef c_64 | |
3554 | #undef fb_ptr | |
3555 | #undef fb_ptr_next | |
3556 | #undef fb_ptr_cmp | |
3557 | ||
3558 | #define psx_gpu r0 | |
3559 | #define num_blocks r1 | |
3560 | #define msb_mask_ptr r2 | |
3561 | #define pixel_ptr r3 | |
3562 | #define draw_mask_ptr r0 | |
3563 | #define c_64 r2 | |
3564 | #define fb_ptr r12 | |
3565 | #define fb_ptr_next r14 | |
3566 | #define fb_ptr_cmp r4 | |
3567 | ||
3568 | #undef msb_mask | |
3569 | #undef draw_mask | |
3570 | #undef pixels | |
3571 | #undef fb_pixels | |
3572 | #undef d128_0x8000 | |
3573 | #undef msb_mask_low | |
3574 | #undef msb_mask_high | |
3575 | #undef draw_mask_next | |
3576 | #undef pixels_g | |
3577 | #undef blend_pixels | |
3578 | #undef fb_pixels_next | |
3579 | ||
3580 | #define msb_mask q0 | |
3581 | #define draw_mask q1 | |
3582 | #define pixels q2 | |
3583 | #define fb_pixels q3 | |
3584 | #define blend_pixels q4 | |
3585 | #define pixels_no_msb q5 | |
3586 | #define blend_mask q6 | |
3587 | #define fb_pixels_no_msb q7 | |
3588 | #define d128_0x8000 q8 | |
3589 | #define d128_0x0421 q9 | |
3590 | #define fb_pixels_next q10 | |
3591 | #define blend_pixels_next q11 | |
3592 | #define pixels_next q12 | |
3593 | #define draw_mask_next q13 | |
3594 | #define write_mask q14 | |
3595 | ||
3596 | #define pixels_rb q5 | |
3597 | #define pixels_mg q7 | |
3598 | #define pixels_g q7 | |
3599 | #define d128_0x7C1F q8 | |
3600 | #define d128_0x03E0 q9 | |
3601 | #define fb_pixels_rb q10 | |
3602 | #define fb_pixels_g q11 | |
3603 | #define fb_pixels_masked q11 | |
3604 | #define d128_0x83E0 q15 | |
3605 | #define pixels_fourth q7 | |
3606 | #define d128_0x1C07 q12 | |
3607 | #define d128_0x00E0 q13 | |
3608 | #define d128_0x80E0 q13 | |
3609 | ||
3610 | #define msb_mask_low d0 | |
3611 | #define msb_mask_high d1 | |
3612 | ||
3613 | #define blend_blocks_average_set_blend_mask_textured(source) \ | |
3614 | vclt.s16 blend_mask, source, #0 \ | |
3615 | ||
3616 | #define blend_blocks_average_set_stp_bit_textured() \ | |
3617 | vorr.u16 blend_pixels, #0x8000 \ | |
3618 | ||
3619 | #define blend_blocks_average_combine_textured(source) \ | |
3620 | vbif.u16 blend_pixels, source, blend_mask \ | |
3621 | ||
3622 | #define blend_blocks_average_set_blend_mask_untextured(source) \ | |
3623 | ||
3624 | #define blend_blocks_average_set_stp_bit_untextured() \ | |
3625 | ||
3626 | #define blend_blocks_average_combine_untextured(source) \ | |
3627 | ||
3628 | #define blend_blocks_average_mask_set_on() \ | |
3629 | vclt.s16 write_mask, fb_pixels_next, #0 \ | |
3630 | ||
3631 | #define blend_blocks_average_mask_copy_on() \ | |
3632 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
3633 | ||
3634 | #define blend_blocks_average_mask_copy_b_on() \ | |
3635 | vorr.u16 draw_mask_next, draw_mask_next, write_mask \ | |
3636 | ||
3637 | #define blend_blocks_average_mask_set_off() \ | |
3638 | ||
3639 | #define blend_blocks_average_mask_copy_off() \ | |
3640 | vmov draw_mask, draw_mask_next \ | |
3641 | ||
3642 | #define blend_blocks_average_mask_copy_b_off() \ | |
3643 | ||
3644 | #define blend_blocks_average_builder(texturing, mask_evaluate) \ | |
3645 | .align 3; \ | |
3646 | \ | |
3647 | function(blend_blocks_##texturing##_average_##mask_evaluate) \ | |
3648 | stmdb sp!, { r4, r14 }; \ | |
9775fc33 | 3649 | save_abi_regs(); \ |
75e28f62 | 3650 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 3651 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
3652 | \ |
3653 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 3654 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
3655 | \ |
3656 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3657 | mov c_64, #64; \ | |
3658 | \ | |
3659 | vmov.u16 d128_0x8000, #0x8000; \ | |
e1f6de8f | 3660 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
3661 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
75e28f62 E |
3662 | \ |
3663 | vmov.u16 d128_0x0421, #0x0400; \ | |
e1f6de8f | 3664 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3665 | \ |
3666 | vorr.u16 d128_0x0421, #0x0021; \ | |
e1f6de8f | 3667 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 E |
3668 | \ |
3669 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ | |
3670 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3671 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3672 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3673 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
3674 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3675 | \ | |
3676 | subs num_blocks, num_blocks, #1; \ | |
3677 | beq 1f; \ | |
3678 | \ | |
3679 | 0: \ | |
3680 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 3681 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 E |
3682 | \ |
3683 | vmov pixels, pixels_next; \ | |
e1f6de8f | 3684 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3685 | \ |
3686 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
3687 | \ | |
3688 | blend_blocks_average_mask_copy_##mask_evaluate(); \ | |
e1f6de8f | 3689 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
3690 | \ |
3691 | blend_blocks_average_set_blend_mask_##texturing(pixels); \ | |
3692 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
3693 | vmov fb_pixels, fb_pixels_next; \ | |
3694 | blend_blocks_average_combine_##texturing(pixels); \ | |
3695 | \ | |
3696 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
3697 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
3698 | cmp fb_ptr_cmp, #28; \ | |
3699 | bls 2f; \ | |
3700 | \ | |
e1f6de8f | 3701 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 E |
3702 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
3703 | \ | |
3704 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3705 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3706 | \ | |
3707 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3708 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
3709 | \ | |
3710 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3711 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3712 | blend_blocks_average_mask_set_##mask_evaluate(); \ | |
e1f6de8f | 3713 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
75e28f62 E |
3714 | \ |
3715 | 3: \ | |
3716 | subs num_blocks, num_blocks, #1; \ | |
3717 | bne 0b; \ | |
3718 | \ | |
3719 | 1: \ | |
3720 | blend_blocks_average_mask_copy_b_##mask_evaluate(); \ | |
3721 | vhadd.u16 blend_pixels, fb_pixels_no_msb, blend_pixels_next; \ | |
3722 | \ | |
3723 | blend_blocks_average_set_blend_mask_##texturing(pixels_next); \ | |
3724 | blend_blocks_average_set_stp_bit_##texturing(); \ | |
3725 | blend_blocks_average_combine_##texturing(pixels_next); \ | |
3726 | \ | |
3727 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3728 | vbif.u16 fb_pixels_next, blend_pixels, draw_mask_next; \ | |
e1f6de8f | 3729 | vst1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 | 3730 | \ |
9775fc33 | 3731 | restore_abi_regs(); \ |
75e28f62 E |
3732 | ldmia sp!, { r4, pc }; \ |
3733 | \ | |
3734 | 2: \ | |
3735 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3736 | vbif.u16 fb_pixels, blend_pixels, draw_mask; \ | |
e1f6de8f | 3737 | vst1.u16 { fb_pixels }, [fb_ptr]; \ |
75e28f62 | 3738 | \ |
e1f6de8f | 3739 | vld1.u16 { fb_pixels_next }, [fb_ptr_next]; \ |
75e28f62 E |
3740 | veor.u16 blend_pixels_next, pixels_next, fb_pixels_next; \ |
3741 | vbic.u16 pixels_no_msb, pixels_next, d128_0x8000; \ | |
3742 | vand.u16 blend_pixels_next, blend_pixels_next, d128_0x0421; \ | |
3743 | vsub.u16 blend_pixels_next, pixels_no_msb, blend_pixels_next; \ | |
3744 | vbic.u16 fb_pixels_no_msb, fb_pixels_next, d128_0x8000; \ | |
3745 | \ | |
3746 | bal 3b \ | |
3747 | ||
3748 | blend_blocks_average_builder(textured, off) | |
3749 | blend_blocks_average_builder(untextured, off) | |
3750 | blend_blocks_average_builder(textured, on) | |
3751 | blend_blocks_average_builder(untextured, on) | |
3752 | ||
3753 | ||
3754 | #define blend_blocks_add_mask_set_on() \ | |
3755 | vclt.s16 write_mask, fb_pixels, #0 \ | |
3756 | ||
3757 | #define blend_blocks_add_mask_copy_on() \ | |
3758 | vorr.u16 draw_mask, draw_mask, write_mask \ | |
3759 | ||
3760 | #define blend_blocks_add_mask_set_off() \ | |
3761 | ||
3762 | #define blend_blocks_add_mask_copy_off() \ | |
3763 | ||
3764 | ||
3765 | #define blend_blocks_add_textured_builder(mask_evaluate) \ | |
3766 | .align 3; \ | |
3767 | \ | |
3768 | function(blend_blocks_textured_add_##mask_evaluate) \ | |
3769 | stmdb sp!, { r4, r14 }; \ | |
9775fc33 | 3770 | save_abi_regs(); \ |
75e28f62 | 3771 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 3772 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
3773 | \ |
3774 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 3775 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
3776 | \ |
3777 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3778 | mov c_64, #64; \ | |
3779 | \ | |
3780 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
3781 | vmov.u16 d128_0x03E0, #0x0300; \ | |
3782 | vmov.u16 d128_0x83E0, #0x8000; \ | |
3783 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
3784 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
3785 | vorr.u16 d128_0x83E0, d128_0x83E0, d128_0x03E0; \ | |
3786 | \ | |
e1f6de8f | 3787 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
3788 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
3789 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
75e28f62 | 3790 | vclt.s16 blend_mask, pixels, #0; \ |
e1f6de8f | 3791 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3792 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3793 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3794 | \ | |
3795 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3796 | vorr.u16 pixels, pixels, msb_mask; \ | |
3797 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3798 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
3799 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
3800 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
3801 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3802 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
3803 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3804 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
3805 | \ | |
3806 | subs num_blocks, num_blocks, #1; \ | |
3807 | beq 1f; \ | |
3808 | \ | |
3809 | 0: \ | |
3810 | mov fb_ptr, fb_ptr_next; \ | |
3811 | \ | |
e1f6de8f | 3812 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 3813 | \ |
e1f6de8f | 3814 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3815 | vclt.s16 blend_mask, pixels, #0; \ |
3816 | \ | |
3817 | vorr.u16 pixels, pixels, msb_mask; \ | |
3818 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3819 | vand.u16 pixels_mg, pixels, d128_0x83E0; \ | |
3820 | \ | |
8438c3c7 | 3821 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ |
e1f6de8f | 3822 | pld [fb_ptr_next, #64]; \ |
75e28f62 E |
3823 | \ |
3824 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
8438c3c7 | 3825 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
3826 | \ | |
75e28f62 | 3827 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ |
e1f6de8f | 3828 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
8438c3c7 | 3829 | \ |
75e28f62 E |
3830 | cmp fb_ptr_cmp, #28; \ |
3831 | bls 2f; \ | |
3832 | \ | |
e1f6de8f | 3833 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3834 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3835 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3836 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3837 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3838 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
e1f6de8f | 3839 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3840 | \ |
3841 | 3: \ | |
3842 | vand.u16 fb_pixels_g, fb_pixels_masked, d128_0x03E0; \ | |
3843 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3844 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_mg; \ | |
3845 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3846 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x83E0; \ | |
3847 | \ | |
3848 | subs num_blocks, num_blocks, #1; \ | |
3849 | bne 0b; \ | |
3850 | \ | |
3851 | 1: \ | |
3852 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3853 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 3854 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 3855 | \ |
9775fc33 | 3856 | restore_abi_regs(); \ |
75e28f62 E |
3857 | ldmia sp!, { r4, pc }; \ |
3858 | \ | |
3859 | 2: \ | |
e1f6de8f | 3860 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3861 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
3862 | \ | |
e1f6de8f | 3863 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3864 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3865 | vand.u16 fb_pixels_masked, fb_pixels, blend_mask; \ | |
3866 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3867 | vand.u16 fb_pixels_rb, fb_pixels_masked, d128_0x7C1F; \ | |
3868 | bal 3b \ | |
3869 | ||
3870 | ||
3871 | #define blend_blocks_add_untextured_builder(mask_evaluate) \ | |
3872 | .align 3; \ | |
3873 | \ | |
3874 | function(blend_blocks_untextured_add_##mask_evaluate) \ | |
3875 | stmdb sp!, { r4, r14 }; \ | |
9775fc33 | 3876 | save_abi_regs(); \ |
75e28f62 | 3877 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 3878 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
3879 | \ |
3880 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 3881 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
3882 | \ |
3883 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
3884 | mov c_64, #64; \ | |
3885 | \ | |
3886 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
3887 | vmov.u16 d128_0x03E0, #0x0300; \ | |
3888 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
3889 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
3890 | \ | |
e1f6de8f | 3891 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
3892 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
3893 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
3894 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
75e28f62 E |
3895 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3896 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3897 | \ | |
3898 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3899 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
3900 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3901 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
3902 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3903 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
3904 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3905 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
3906 | \ | |
3907 | subs num_blocks, num_blocks, #1; \ | |
3908 | beq 1f; \ | |
3909 | \ | |
3910 | 0: \ | |
3911 | mov fb_ptr, fb_ptr_next; \ | |
3912 | \ | |
e1f6de8f | 3913 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 3914 | \ |
e1f6de8f | 3915 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
3916 | \ |
3917 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3918 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3919 | vand.u16 pixels_g, pixels, d128_0x03E0; \ | |
3920 | \ | |
3921 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 3922 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
3923 | \ |
3924 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
3925 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
3926 | cmp fb_ptr_cmp, #28; \ | |
3927 | bls 2f; \ | |
3928 | \ | |
e1f6de8f | 3929 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3930 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3931 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3932 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ | |
3933 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
e1f6de8f | 3934 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3935 | \ |
3936 | 3: \ | |
3937 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
3938 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
3939 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
3940 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
3941 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
3942 | \ | |
3943 | subs num_blocks, num_blocks, #1; \ | |
3944 | bne 0b; \ | |
3945 | \ | |
3946 | 1: \ | |
3947 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
3948 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
3949 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 3950 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 3951 | \ |
9775fc33 | 3952 | restore_abi_regs(); \ |
75e28f62 E |
3953 | ldmia sp!, { r4, pc }; \ |
3954 | \ | |
3955 | 2: \ | |
e1f6de8f | 3956 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
3957 | vand.u16 pixels_rb, pixels, d128_0x7C1F; \ |
3958 | \ | |
e1f6de8f | 3959 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
3960 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
3961 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
3962 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
3963 | bal 3b \ | |
3964 | ||
3965 | ||
3966 | blend_blocks_add_textured_builder(off) | |
3967 | blend_blocks_add_textured_builder(on) | |
3968 | blend_blocks_add_untextured_builder(off) | |
3969 | blend_blocks_add_untextured_builder(on) | |
3970 | ||
3971 | #define blend_blocks_subtract_set_blend_mask_textured() \ | |
3972 | vclt.s16 blend_mask, pixels_next, #0 \ | |
3973 | ||
3974 | #define blend_blocks_subtract_combine_textured() \ | |
3975 | vbif.u16 blend_pixels, pixels, blend_mask \ | |
3976 | ||
259c6f59 | 3977 | #define blend_blocks_subtract_set_stp_textured() \ |
75e28f62 E |
3978 | vorr.u16 blend_pixels, #0x8000 \ |
3979 | ||
3980 | #define blend_blocks_subtract_msb_mask_textured() \ | |
3981 | vorr.u16 pixels, pixels_next, msb_mask \ | |
3982 | ||
3983 | #define blend_blocks_subtract_set_blend_mask_untextured() \ | |
3984 | ||
3985 | #define blend_blocks_subtract_combine_untextured() \ | |
3986 | ||
259c6f59 | 3987 | #define blend_blocks_subtract_set_stp_untextured() \ |
75e28f62 E |
3988 | vorr.u16 blend_pixels, blend_pixels, msb_mask \ |
3989 | ||
3990 | #define blend_blocks_subtract_msb_mask_untextured() \ | |
3991 | ||
3992 | ||
3993 | #define blend_blocks_subtract_mask_set_on() \ | |
3994 | vclt.s16 write_mask, fb_pixels, #0 \ | |
3995 | ||
3996 | #define blend_blocks_subtract_mask_copy_on() \ | |
3997 | vorr.u16 draw_mask, draw_mask_next, write_mask \ | |
3998 | ||
3999 | #define blend_blocks_subtract_mask_set_off() \ | |
4000 | ||
4001 | #define blend_blocks_subtract_mask_copy_off() \ | |
4002 | vmov draw_mask, draw_mask_next \ | |
4003 | ||
4004 | ||
4005 | #define blend_blocks_subtract_builder(texturing, mask_evaluate) \ | |
4006 | .align 3; \ | |
4007 | \ | |
4008 | function(blend_blocks_##texturing##_subtract_##mask_evaluate) \ | |
4009 | stmdb sp!, { r4, r14 }; \ | |
9775fc33 | 4010 | save_abi_regs(); \ |
75e28f62 | 4011 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 4012 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
4013 | \ |
4014 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 4015 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
4016 | \ |
4017 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4018 | mov c_64, #64; \ | |
4019 | \ | |
4020 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4021 | vmov.u16 d128_0x03E0, #0x0300; \ | |
4022 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
4023 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
4024 | \ | |
e1f6de8f | 4025 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
4026 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
4027 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ | |
75e28f62 | 4028 | blend_blocks_subtract_set_blend_mask_##texturing(); \ |
e1f6de8f | 4029 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4030 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
4031 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
4032 | \ | |
4033 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ | |
4034 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4035 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4036 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4037 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4038 | \ | |
4039 | subs num_blocks, num_blocks, #1; \ | |
4040 | beq 1f; \ | |
4041 | \ | |
4042 | 0: \ | |
4043 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
4044 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 4045 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 4046 | \ |
e1f6de8f | 4047 | vld1.u32 { draw_mask_next }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
4048 | blend_blocks_subtract_msb_mask_##texturing(); \ |
4049 | \ | |
e1f6de8f | 4050 | vld1.u32 { pixels_next }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
4051 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
4052 | vand.u16 pixels_rb, pixels_next, d128_0x7C1F; \ | |
259c6f59 | 4053 | blend_blocks_subtract_set_stp_##texturing(); \ |
75e28f62 E |
4054 | vand.u16 pixels_g, pixels_next, d128_0x03E0; \ |
4055 | blend_blocks_subtract_combine_##texturing(); \ | |
4056 | blend_blocks_subtract_set_blend_mask_##texturing(); \ | |
4057 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
4058 | \ | |
4059 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4060 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4061 | cmp fb_ptr_cmp, #28; \ | |
4062 | bls 2f; \ | |
4063 | \ | |
e1f6de8f | 4064 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4065 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
4066 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4067 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4068 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
e1f6de8f | 4069 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4070 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
4071 | \ | |
4072 | 3: \ | |
4073 | subs num_blocks, num_blocks, #1; \ | |
4074 | bne 0b; \ | |
4075 | \ | |
4076 | 1: \ | |
4077 | blend_blocks_subtract_mask_copy_##mask_evaluate(); \ | |
4078 | \ | |
4079 | blend_blocks_subtract_msb_mask_##texturing(); \ | |
4080 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
259c6f59 | 4081 | blend_blocks_subtract_set_stp_##texturing(); \ |
75e28f62 E |
4082 | blend_blocks_subtract_combine_##texturing(); \ |
4083 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4084 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4085 | \ |
9775fc33 | 4086 | restore_abi_regs(); \ |
75e28f62 E |
4087 | ldmia sp!, { r4, pc }; \ |
4088 | \ | |
4089 | 2: \ | |
e1f6de8f | 4090 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
4091 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
75e28f62 E |
4092 | blend_blocks_subtract_mask_set_##mask_evaluate(); \ |
4093 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4094 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4095 | vqsub.u8 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4096 | vqsub.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4097 | bal 3b \ | |
4098 | ||
4099 | ||
4100 | blend_blocks_subtract_builder(textured, off) | |
4101 | blend_blocks_subtract_builder(textured, on) | |
4102 | blend_blocks_subtract_builder(untextured, off) | |
4103 | blend_blocks_subtract_builder(untextured, on) | |
4104 | ||
4105 | ||
4106 | #define blend_blocks_add_fourth_textured_builder(mask_evaluate) \ | |
4107 | .align 3; \ | |
4108 | \ | |
4109 | function(blend_blocks_textured_add_fourth_##mask_evaluate) \ | |
4110 | stmdb sp!, { r4, r14 }; \ | |
9775fc33 | 4111 | save_abi_regs(); \ |
75e28f62 | 4112 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 4113 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
4114 | \ |
4115 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 4116 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
4117 | \ |
4118 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4119 | mov c_64, #64; \ | |
4120 | \ | |
4121 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4122 | vmov.u16 d128_0x03E0, #0x0300; \ | |
75e28f62 | 4123 | vmov.u16 d128_0x1C07, #0x1C00; \ |
d1c75d1e | 4124 | vmov.u16 d128_0x00E0, #0x00E0; \ |
75e28f62 E |
4125 | vorr.u16 d128_0x7C1F, #0x001F; \ |
4126 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
75e28f62 | 4127 | vorr.u16 d128_0x1C07, #0x0007; \ |
75e28f62 | 4128 | \ |
e1f6de8f | 4129 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
4130 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
4131 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
75e28f62 | 4132 | vclt.s16 blend_mask, pixels, #0; \ |
e1f6de8f | 4133 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4134 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4135 | vshr.s16 pixels_fourth, pixels, #2; \ | |
d1c75d1e | 4136 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
75e28f62 E |
4137 | \ |
4138 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
d1c75d1e E |
4139 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4140 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4141 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
75e28f62 | 4142 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
d1c75d1e | 4143 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
75e28f62 | 4144 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
d1c75d1e | 4145 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
75e28f62 E |
4146 | \ |
4147 | subs num_blocks, num_blocks, #1; \ | |
4148 | beq 1f; \ | |
4149 | \ | |
4150 | 0: \ | |
4151 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 4152 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 4153 | \ |
d1c75d1e | 4154 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ |
259c6f59 | 4155 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
d1c75d1e E |
4156 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
4157 | \ | |
e1f6de8f | 4158 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 | 4159 | vclt.s16 blend_mask, pixels, #0; \ |
75e28f62 | 4160 | vshr.s16 pixels_fourth, pixels, #2; \ |
d1c75d1e | 4161 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
75e28f62 E |
4162 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ |
4163 | \ | |
4164 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4165 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
4166 | \ |
4167 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4168 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4169 | cmp fb_ptr_cmp, #28; \ | |
4170 | bls 2f; \ | |
4171 | \ | |
e1f6de8f | 4172 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4173 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
75e28f62 | 4174 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
d1c75d1e E |
4175 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4176 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
e1f6de8f | 4177 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4178 | \ |
4179 | 3: \ | |
d1c75d1e | 4180 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ |
75e28f62 | 4181 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ |
d1c75d1e | 4182 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ |
75e28f62 | 4183 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ |
d1c75d1e | 4184 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ |
75e28f62 E |
4185 | \ |
4186 | subs num_blocks, num_blocks, #1; \ | |
4187 | bne 0b; \ | |
4188 | \ | |
4189 | 1: \ | |
4190 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
259c6f59 | 4191 | vorr.u16 blend_pixels, #0x8000; /* stp */ \ |
d1c75d1e | 4192 | vbif.u16 blend_pixels, pixels, blend_mask; \ |
259c6f59 | 4193 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ |
75e28f62 | 4194 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ |
e1f6de8f | 4195 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4196 | \ |
9775fc33 | 4197 | restore_abi_regs(); \ |
75e28f62 E |
4198 | ldmia sp!, { r4, pc }; \ |
4199 | \ | |
4200 | 2: \ | |
e1f6de8f | 4201 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
d1c75d1e | 4202 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
75e28f62 | 4203 | \ |
e1f6de8f | 4204 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4205 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
75e28f62 | 4206 | blend_blocks_add_mask_copy_##mask_evaluate(); \ |
d1c75d1e | 4207 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ |
75e28f62 E |
4208 | bal 3b \ |
4209 | ||
4210 | ||
d1c75d1e | 4211 | |
75e28f62 E |
4212 | #define blend_blocks_add_fourth_untextured_builder(mask_evaluate) \ |
4213 | .align 3; \ | |
4214 | \ | |
4215 | function(blend_blocks_untextured_add_fourth_##mask_evaluate) \ | |
4216 | stmdb sp!, { r4, r14 }; \ | |
9775fc33 | 4217 | save_abi_regs(); \ |
75e28f62 | 4218 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset; \ |
e1f6de8f | 4219 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
75e28f62 E |
4220 | \ |
4221 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16); \ | |
e1f6de8f | 4222 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16]; \ |
75e28f62 E |
4223 | \ |
4224 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset; \ | |
4225 | mov c_64, #64; \ | |
4226 | \ | |
4227 | vmov.u16 d128_0x7C1F, #0x7C00; \ | |
4228 | vmov.u16 d128_0x03E0, #0x0300; \ | |
75e28f62 E |
4229 | vmov.u16 d128_0x1C07, #0x1C00; \ |
4230 | vmov.u16 d128_0x00E0, #0x00E0; \ | |
4231 | vorr.u16 d128_0x7C1F, #0x001F; \ | |
4232 | vorr.u16 d128_0x03E0, #0x00E0; \ | |
75e28f62 E |
4233 | vorr.u16 d128_0x1C07, #0x0007; \ |
4234 | \ | |
e1f6de8f | 4235 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
4236 | ldr fb_ptr_next, [pixel_ptr, #28]; \ | |
4237 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ | |
4238 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ | |
75e28f62 E |
4239 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4240 | vshr.s16 pixels_fourth, pixels, #2; \ | |
4241 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
4242 | \ | |
4243 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4244 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
4245 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4246 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4247 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4248 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4249 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
4250 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
4251 | \ | |
4252 | subs num_blocks, num_blocks, #1; \ | |
4253 | beq 1f; \ | |
4254 | \ | |
4255 | 0: \ | |
4256 | mov fb_ptr, fb_ptr_next; \ | |
e1f6de8f | 4257 | ldr fb_ptr_next, [pixel_ptr, #28]; \ |
75e28f62 | 4258 | \ |
e1f6de8f | 4259 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64; \ |
75e28f62 E |
4260 | \ |
4261 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
4262 | vshr.s16 pixels_fourth, pixels, #2; \ | |
4263 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
4264 | vand.u16 pixels_rb, pixels_fourth, d128_0x1C07; \ | |
4265 | \ | |
4266 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4267 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64; \ |
75e28f62 E |
4268 | \ |
4269 | sub fb_ptr_cmp, fb_ptr_next, fb_ptr; \ | |
4270 | add fb_ptr_cmp, fb_ptr_cmp, #14; \ | |
4271 | cmp fb_ptr_cmp, #28; \ | |
4272 | bls 2f; \ | |
4273 | \ | |
e1f6de8f | 4274 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4275 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4276 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4277 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ | |
4278 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
e1f6de8f | 4279 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4280 | \ |
4281 | 3: \ | |
4282 | vand.u16 fb_pixels_g, fb_pixels, d128_0x03E0; \ | |
4283 | vadd.u16 fb_pixels_rb, fb_pixels_rb, pixels_rb; \ | |
4284 | vadd.u16 fb_pixels_g, fb_pixels_g, pixels_g; \ | |
4285 | vmin.u8 fb_pixels_rb, fb_pixels_rb, d128_0x7C1F; \ | |
4286 | vmin.u16 fb_pixels_g, fb_pixels_g, d128_0x03E0; \ | |
4287 | \ | |
4288 | subs num_blocks, num_blocks, #1; \ | |
4289 | bne 0b; \ | |
4290 | \ | |
4291 | 1: \ | |
4292 | vorr.u16 blend_pixels, fb_pixels_rb, fb_pixels_g; \ | |
4293 | vorr.u16 blend_pixels, blend_pixels, msb_mask; \ | |
4294 | vbit.u16 blend_pixels, fb_pixels, draw_mask; \ | |
e1f6de8f | 4295 | vst1.u16 { blend_pixels }, [fb_ptr_next]; \ |
75e28f62 | 4296 | \ |
9775fc33 | 4297 | restore_abi_regs(); \ |
75e28f62 E |
4298 | ldmia sp!, { r4, pc }; \ |
4299 | \ | |
4300 | 2: \ | |
e1f6de8f | 4301 | vst1.u16 { blend_pixels }, [fb_ptr]; \ |
75e28f62 E |
4302 | vand.u16 pixels_g, pixels_fourth, d128_0x00E0; \ |
4303 | \ | |
e1f6de8f | 4304 | vld1.u16 { fb_pixels }, [fb_ptr_next]; \ |
75e28f62 E |
4305 | blend_blocks_add_mask_set_##mask_evaluate(); \ |
4306 | blend_blocks_add_mask_copy_##mask_evaluate(); \ | |
4307 | vand.u16 fb_pixels_rb, fb_pixels, d128_0x7C1F; \ | |
4308 | bal 3b \ | |
4309 | ||
4310 | ||
4311 | blend_blocks_add_fourth_textured_builder(off) | |
4312 | blend_blocks_add_fourth_textured_builder(on) | |
4313 | blend_blocks_add_fourth_untextured_builder(off) | |
4314 | blend_blocks_add_fourth_untextured_builder(on) | |
4315 | ||
4316 | // TODO: Optimize this more. Need a scene that actually uses it for | |
4317 | // confirmation.. | |
4318 | ||
4319 | .align 3 | |
4320 | ||
4321 | function(blend_blocks_textured_unblended_on) | |
4322 | stmdb sp!, { r4, r14 } | |
9775fc33 | 4323 | save_abi_regs() |
75e28f62 | 4324 | add mask_msb_ptr, psx_gpu, #psx_gpu_mask_msb_offset |
e1f6de8f | 4325 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
4326 | |
4327 | add pixel_ptr, psx_gpu, #(psx_gpu_blocks_offset + 16) | |
e1f6de8f | 4328 | vld1.u16 { msb_mask_low[], msb_mask_high[] }, [mask_msb_ptr, :16] |
75e28f62 E |
4329 | |
4330 | add draw_mask_ptr, psx_gpu, #psx_gpu_blocks_offset | |
4331 | mov c_64, #64 | |
4332 | ||
e1f6de8f | 4333 | ldr fb_ptr, [pixel_ptr, #28] |
4334 | vld1.u16 { fb_pixels }, [fb_ptr] | |
4335 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
75e28f62 | 4336 | vclt.s16 write_mask, fb_pixels, #0 |
e1f6de8f | 4337 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
75e28f62 E |
4338 | |
4339 | subs num_blocks, num_blocks, #1 | |
4340 | beq 1f | |
4341 | ||
4342 | 0: | |
134f81ec | 4343 | vorr.u16 pixels, pixels, msb_mask |
75e28f62 E |
4344 | vorr.u16 draw_mask, draw_mask, write_mask |
4345 | vbif.u16 fb_pixels, pixels, draw_mask | |
e1f6de8f | 4346 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 | 4347 | |
e1f6de8f | 4348 | ldr fb_ptr, [pixel_ptr, #28] |
4349 | vld1.u16 { fb_pixels }, [fb_ptr] | |
4350 | vld1.u32 { draw_mask }, [draw_mask_ptr, :128], c_64 | |
75e28f62 | 4351 | vclt.s16 write_mask, fb_pixels, #0 |
e1f6de8f | 4352 | vld1.u32 { pixels }, [pixel_ptr, :128], c_64 |
75e28f62 E |
4353 | |
4354 | subs num_blocks, num_blocks, #1 | |
4355 | bne 0b | |
4356 | ||
4357 | 1: | |
134f81ec | 4358 | vorr.u16 pixels, pixels, msb_mask |
75e28f62 E |
4359 | vorr.u16 draw_mask, draw_mask, write_mask |
4360 | vbif.u16 fb_pixels, pixels, draw_mask | |
e1f6de8f | 4361 | vst1.u16 { fb_pixels }, [fb_ptr] |
75e28f62 | 4362 | |
9775fc33 | 4363 | restore_abi_regs() |
75e28f62 E |
4364 | ldmia sp!, { r4, pc } |
4365 | ||
4366 | ||
4367 | function(blend_blocks_textured_unblended_off) | |
4368 | bx lr | |
4369 | ||
4370 | ||
4371 | function(warmup) | |
4372 | mov r3, #64 | |
4373 | cmp r0, #0 | |
4374 | bxeq lr | |
4375 | ||
4376 | 0: | |
e1f6de8f | 4377 | vld1.u32 { u_whole_8, v_whole_8 }, [r1, :128], r3 |
75e28f62 E |
4378 | |
4379 | subs r0, r0, #1 | |
4380 | bne 0b | |
4381 | ||
4382 | bx lr | |
4383 | ||
6c4a10c4 | 4384 | #undef vram_ptr |
75e28f62 | 4385 | #undef color |
6c4a10c4 | 4386 | #undef width |
75e28f62 | 4387 | #undef height |
6c4a10c4 | 4388 | #undef pitch |
75e28f62 E |
4389 | |
4390 | #define vram_ptr r0 | |
6c4a10c4 E |
4391 | #define color r1 |
4392 | #define width r2 | |
4393 | #define height r3 | |
75e28f62 | 4394 | |
6c4a10c4 | 4395 | #define pitch r1 |
75e28f62 | 4396 | |
6c4a10c4 | 4397 | #define num_width r12 |
75e28f62 | 4398 | |
87c45ad1 E |
4399 | #undef colors_a |
4400 | #undef colors_b | |
75e28f62 | 4401 | |
87c45ad1 E |
4402 | #define colors_a q0 |
4403 | #define colors_b q1 | |
75e28f62 E |
4404 | |
4405 | .align 3 | |
4406 | ||
4407 | function(render_block_fill_body) | |
87c45ad1 | 4408 | vdup.u16 colors_a, color |
6c4a10c4 | 4409 | mov pitch, #2048 |
75e28f62 | 4410 | |
87c45ad1 | 4411 | vmov colors_b, colors_a |
75e28f62 | 4412 | sub pitch, pitch, width, lsl #1 |
75e28f62 | 4413 | |
6c4a10c4 | 4414 | mov num_width, width |
75e28f62 | 4415 | |
6c4a10c4 | 4416 | 0: |
e1f6de8f | 4417 | vst1.u32 { colors_a, colors_b }, [vram_ptr, :256]! |
75e28f62 | 4418 | |
d1c75d1e | 4419 | subs num_width, num_width, #16 |
6c4a10c4 | 4420 | bne 0b |
75e28f62 | 4421 | |
75e28f62 | 4422 | add vram_ptr, vram_ptr, pitch |
6c4a10c4 E |
4423 | mov num_width, width |
4424 | ||
75e28f62 E |
4425 | subs height, height, #1 |
4426 | bne 0b | |
75e28f62 | 4427 | |
6c4a10c4 E |
4428 | bx lr |
4429 | ||
75e28f62 E |
4430 | |
4431 | #undef x | |
4432 | #undef y | |
4433 | #undef width | |
4434 | #undef height | |
4435 | #undef fb_ptr | |
4436 | #undef texture_mask | |
4437 | #undef num_blocks | |
4438 | #undef temp | |
4439 | #undef dirty_textures_mask | |
4440 | #undef clut_ptr | |
4441 | #undef current_texture_mask | |
4442 | ||
4443 | #define psx_gpu r0 | |
4444 | #define x r1 | |
4445 | #define y r2 | |
4446 | #define u r3 | |
4447 | #define v r4 | |
4448 | #define width r5 | |
4449 | #define height r6 | |
4450 | #define offset_u r8 | |
4451 | #define offset_v r9 | |
4452 | #define offset_u_right r10 | |
4453 | #define width_rounded r11 | |
4454 | #define height_rounded r12 | |
4455 | ||
4456 | #define texture_offset_base r1 | |
4457 | #define tile_width r2 | |
4458 | #define tile_height r3 | |
4459 | #define num_blocks r4 | |
4460 | #define block r5 | |
4461 | #define sub_tile_height r6 | |
4462 | #define fb_ptr r7 | |
4463 | #define texture_mask r8 | |
4464 | #define column_data r9 | |
4465 | #define texture_offset r10 | |
4466 | #define tiles_remaining r11 | |
4467 | #define fb_ptr_advance_column r12 | |
4468 | #define texture_block_ptr r14 | |
4469 | ||
8184d7c5 | 4470 | #define temp r14 |
4471 | ||
75e28f62 E |
4472 | #define texture_page_ptr r3 |
4473 | #define left_block_mask r4 | |
4474 | #define right_block_mask r5 | |
4475 | #define texture_mask_rev r10 | |
4476 | #define control_mask r11 | |
4477 | ||
4478 | #define dirty_textures_mask r4 | |
4479 | #define clut_ptr r5 | |
4480 | #define current_texture_mask r6 | |
4481 | ||
4482 | ||
4483 | #undef texels | |
4484 | #undef clut_low_a | |
4485 | #undef clut_low_b | |
4486 | #undef clut_high_a | |
4487 | #undef clut_high_b | |
4488 | #undef clut_a | |
4489 | #undef clut_b | |
4490 | #undef texels_low | |
4491 | #undef texels_high | |
4492 | ||
4493 | #define texels d0 | |
4494 | #define draw_masks_fb_ptrs q1 | |
4495 | ||
4496 | #define draw_mask_fb_ptr_left d2 | |
4497 | #define draw_mask_fb_ptr_right d3 | |
4498 | ||
59d15d23 | 4499 | #define draw_mask_fb_ptr_left_a d2 |
4500 | #define draw_mask_fb_ptr_left_b d3 | |
4501 | #define draw_mask_fb_ptr_right_a d10 | |
4502 | #define draw_mask_fb_ptr_right_b d11 | |
4503 | #define draw_masks_fb_ptrs2 q5 | |
4504 | ||
75e28f62 E |
4505 | #define clut_low_a d4 |
4506 | #define clut_low_b d5 | |
4507 | #define clut_high_a d6 | |
4508 | #define clut_high_b d7 | |
4509 | ||
4510 | #define block_masks d8 | |
4511 | #define block_masks_shifted d9 | |
4512 | ||
4513 | #define clut_a q2 | |
4514 | #define clut_b q3 | |
4515 | ||
59d15d23 | 4516 | #define texels_low d12 |
4517 | #define texels_high d13 | |
75e28f62 | 4518 | |
59d15d23 | 4519 | #define texels_wide_low d14 |
4520 | #define texels_wide_high d15 | |
4521 | #define texels_wide q7 | |
75e28f62 E |
4522 | |
4523 | ||
59d15d23 | 4524 | setup_sprite_flush_blocks: |
4525 | vpush { q1 - q5 } | |
75e28f62 | 4526 | |
4d646738 | 4527 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 4528 | bl flush_render_block_buffer |
4d646738 | 4529 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 4530 | |
59d15d23 | 4531 | vpop { q1 - q5 } |
75e28f62 E |
4532 | |
4533 | add block, psx_gpu, #psx_gpu_blocks_offset | |
75e28f62 E |
4534 | bx lr |
4535 | ||
4536 | ||
4537 | setup_sprite_update_texture_4bpp_cache: | |
4538 | stmdb sp!, { r0 - r3, r14 } | |
4539 | bl update_texture_4bpp_cache | |
4540 | ldmia sp!, { r0 - r3, pc } | |
4541 | ||
4542 | ||
4543 | setup_sprite_update_texture_8bpp_cache: | |
4d646738 | 4544 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r14 } |
75e28f62 | 4545 | bl update_texture_8bpp_cache |
4d646738 | 4546 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS pc } |
75e28f62 E |
4547 | |
4548 | ||
4549 | #define setup_sprite_tiled_initialize_4bpp() \ | |
4550 | ldr dirty_textures_mask, \ | |
e1f6de8f | 4551 | [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset]; \ |
4552 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset]; \ | |
75e28f62 | 4553 | \ |
e1f6de8f | 4554 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset]; \ |
4555 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128]; \ | |
75e28f62 E |
4556 | \ |
4557 | tst current_texture_mask, dirty_textures_mask; \ | |
4558 | vuzp.u8 clut_a, clut_b; \ | |
4559 | \ | |
4560 | blne setup_sprite_update_texture_4bpp_cache \ | |
4561 | ||
4562 | #define setup_sprite_tiled_initialize_8bpp() \ | |
4563 | ldr dirty_textures_mask, \ | |
e1f6de8f | 4564 | [psx_gpu, #psx_gpu_dirty_textures_8bpp_mask_offset]; \ |
4565 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset]; \ | |
75e28f62 E |
4566 | \ |
4567 | tst current_texture_mask, dirty_textures_mask; \ | |
4568 | blne setup_sprite_update_texture_8bpp_cache \ | |
4569 | ||
4570 | ||
75e28f62 E |
4571 | #define setup_sprite_block_count_single() \ |
4572 | sub_tile_height \ | |
4573 | ||
4574 | #define setup_sprite_block_count_double() \ | |
4575 | sub_tile_height, lsl #1 \ | |
4576 | ||
4577 | #define setup_sprite_tile_add_blocks(type) \ | |
4578 | add num_blocks, num_blocks, setup_sprite_block_count_##type(); \ | |
4579 | cmp num_blocks, #MAX_BLOCKS; \ | |
4580 | \ | |
59d15d23 | 4581 | movgt num_blocks, setup_sprite_block_count_##type(); \ |
4582 | blgt setup_sprite_flush_blocks \ | |
75e28f62 E |
4583 | |
4584 | ||
4585 | #define setup_sprite_tile_full_4bpp(edge) \ | |
4586 | setup_sprite_tile_add_blocks(double); \ | |
4587 | \ | |
4588 | 4: \ | |
4589 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4590 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
4591 | \ | |
e1f6de8f | 4592 | pld [fb_ptr]; \ |
75e28f62 | 4593 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
e1f6de8f | 4594 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4595 | \ |
4596 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4597 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4598 | \ | |
e1f6de8f | 4599 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
75e28f62 E |
4600 | add texture_block_ptr, texture_offset, #8; \ |
4601 | \ | |
4602 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4603 | add block, block, #40; \ | |
4604 | \ | |
4605 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4606 | add fb_ptr, fb_ptr, #16; \ | |
4607 | \ | |
e1f6de8f | 4608 | vst1.u32 { draw_mask_fb_ptr_left }, [block, :64]; \ |
75e28f62 E |
4609 | add block, block, #24; \ |
4610 | \ | |
e1f6de8f | 4611 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4612 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ |
4613 | \ | |
e1f6de8f | 4614 | pld [fb_ptr]; \ |
75e28f62 E |
4615 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ |
4616 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4617 | \ | |
e1f6de8f | 4618 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
75e28f62 E |
4619 | add block, block, #40; \ |
4620 | \ | |
4621 | add texture_offset, texture_offset, #0x10; \ | |
4622 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
4623 | \ | |
e1f6de8f | 4624 | vst1.u32 { draw_mask_fb_ptr_right }, [block, :64]; \ |
75e28f62 E |
4625 | add block, block, #24; \ |
4626 | \ | |
4627 | subs sub_tile_height, sub_tile_height, #1; \ | |
4628 | bne 4b; \ | |
4629 | \ | |
4630 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4631 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4632 | |
4633 | ||
4634 | #define setup_sprite_tile_half_4bpp(edge) \ | |
4635 | setup_sprite_tile_add_blocks(single); \ | |
4636 | \ | |
4637 | 4: \ | |
4638 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4639 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
4640 | \ | |
e1f6de8f | 4641 | pld [fb_ptr]; \ |
75e28f62 | 4642 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
e1f6de8f | 4643 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4644 | \ |
4645 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4646 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
4647 | \ | |
e1f6de8f | 4648 | vst2.u8 { texels_low, texels_high }, [block, :128]; \ |
75e28f62 E |
4649 | add block, block, #40; \ |
4650 | \ | |
4651 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 4652 | vst1.u32 { draw_mask_fb_ptr_##edge }, [block, :64]; \ |
75e28f62 E |
4653 | \ |
4654 | add block, block, #24; \ | |
4655 | add texture_offset, texture_offset, #0x10; \ | |
4656 | \ | |
4657 | add fb_ptr, fb_ptr, #2048; \ | |
4658 | subs sub_tile_height, sub_tile_height, #1; \ | |
4659 | \ | |
4660 | bne 4b; \ | |
4661 | \ | |
4662 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4663 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4664 | |
4665 | ||
4666 | #define setup_sprite_tile_full_8bpp(edge) \ | |
4667 | setup_sprite_tile_add_blocks(double); \ | |
4668 | add block, block, #16; \ | |
4669 | \ | |
4670 | 4: \ | |
4671 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4672 | vmov.u32 draw_mask_fb_ptr_left[1], fb_ptr; \ | |
4673 | \ | |
e1f6de8f | 4674 | pld [fb_ptr]; \ |
75e28f62 | 4675 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ |
e1f6de8f | 4676 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 E |
4677 | \ |
4678 | add texture_block_ptr, texture_offset, #8; \ | |
e1f6de8f | 4679 | vst1.u32 { texels }, [block, :64]; \ |
75e28f62 E |
4680 | \ |
4681 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
4682 | add block, block, #24; \ | |
4683 | \ | |
4684 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
4685 | \ | |
4686 | add fb_ptr, fb_ptr, #16; \ | |
e1f6de8f | 4687 | vst1.u32 { draw_mask_fb_ptr_left }, [block, :64]; \ |
75e28f62 E |
4688 | \ |
4689 | add block, block, #40; \ | |
e1f6de8f | 4690 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
4691 | pld [fb_ptr]; \ | |
75e28f62 E |
4692 | \ |
4693 | vmov.u32 draw_mask_fb_ptr_right[1], fb_ptr; \ | |
e1f6de8f | 4694 | vst1.u32 { texels }, [block, :64]; \ |
75e28f62 E |
4695 | add block, block, #24; \ |
4696 | \ | |
4697 | add texture_offset, texture_offset, #0x10; \ | |
4698 | add fb_ptr, fb_ptr, #(2048 - 16); \ | |
4699 | \ | |
e1f6de8f | 4700 | vst1.u32 { draw_mask_fb_ptr_right }, [block, :64]; \ |
75e28f62 E |
4701 | add block, block, #40; \ |
4702 | \ | |
4703 | subs sub_tile_height, sub_tile_height, #1; \ | |
4704 | bne 4b; \ | |
4705 | \ | |
4706 | sub block, block, #16; \ | |
4707 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4708 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4709 | |
4710 | ||
4711 | #define setup_sprite_tile_half_8bpp(edge) \ | |
4712 | setup_sprite_tile_add_blocks(single); \ | |
4713 | add block, block, #16; \ | |
4714 | \ | |
4715 | 4: \ | |
4716 | and texture_block_ptr, texture_offset, texture_mask; \ | |
4717 | vmov.u32 draw_mask_fb_ptr_##edge[1], fb_ptr; \ | |
e1f6de8f | 4718 | pld [fb_ptr]; \ |
75e28f62 E |
4719 | \ |
4720 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 4721 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
75e28f62 | 4722 | \ |
e1f6de8f | 4723 | vst1.u32 { texels }, [block, :64]; \ |
75e28f62 E |
4724 | add block, block, #24; \ |
4725 | \ | |
e1f6de8f | 4726 | vst1.u32 { draw_mask_fb_ptr_##edge }, [block, :64]; \ |
75e28f62 E |
4727 | add block, block, #40; \ |
4728 | \ | |
4729 | add texture_offset, texture_offset, #0x10; \ | |
4730 | add fb_ptr, fb_ptr, #2048; \ | |
4731 | \ | |
4732 | subs sub_tile_height, sub_tile_height, #1; \ | |
4733 | bne 4b; \ | |
4734 | \ | |
4735 | sub block, block, #16; \ | |
4736 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 4737 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
75e28f62 E |
4738 | |
4739 | ||
4740 | #define setup_sprite_tile_column_edge_pre_adjust_half_right() \ | |
4741 | add texture_offset, texture_offset_base, #8; \ | |
4742 | add fb_ptr, fb_ptr, #16 \ | |
4743 | ||
4744 | #define setup_sprite_tile_column_edge_pre_adjust_half_left() \ | |
4745 | mov texture_offset, texture_offset_base \ | |
4746 | ||
4747 | #define setup_sprite_tile_column_edge_pre_adjust_half(edge) \ | |
4748 | setup_sprite_tile_column_edge_pre_adjust_half_##edge() \ | |
4749 | ||
4750 | #define setup_sprite_tile_column_edge_pre_adjust_full(edge) \ | |
4751 | mov texture_offset, texture_offset_base \ | |
4752 | ||
4753 | #define setup_sprite_tile_column_edge_post_adjust_half_right() \ | |
4754 | sub fb_ptr, fb_ptr, #16 \ | |
4755 | ||
4756 | #define setup_sprite_tile_column_edge_post_adjust_half_left() \ | |
4757 | ||
4758 | #define setup_sprite_tile_column_edge_post_adjust_half(edge) \ | |
4759 | setup_sprite_tile_column_edge_post_adjust_half_##edge() \ | |
4760 | ||
4761 | #define setup_sprite_tile_column_edge_post_adjust_full(edge) \ | |
4762 | ||
4763 | ||
59d15d23 | 4764 | #define setup_sprite_tile_column_height_single(edge_mode, edge, texture_mode, \ |
4765 | x4mode) \ | |
75e28f62 | 4766 | mov sub_tile_height, column_data; \ |
59d15d23 | 4767 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
4768 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
4769 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
75e28f62 | 4770 | |
59d15d23 | 4771 | #define setup_sprite_tile_column_height_multi(edge_mode, edge, texture_mode, \ |
4772 | x4mode) \ | |
75e28f62 E |
4773 | and sub_tile_height, column_data, #0xFF; \ |
4774 | mov tiles_remaining, column_data, lsr #16; \ | |
59d15d23 | 4775 | setup_sprite_tile_column_edge_pre_adjust_##edge_mode##x4mode(edge); \ |
4776 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ | |
75e28f62 E |
4777 | \ |
4778 | subs tiles_remaining, tiles_remaining, #1; \ | |
4779 | beq 2f; \ | |
4780 | \ | |
4781 | 3: \ | |
4782 | mov sub_tile_height, #16; \ | |
59d15d23 | 4783 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
75e28f62 E |
4784 | subs tiles_remaining, tiles_remaining, #1; \ |
4785 | bne 3b; \ | |
4786 | \ | |
4787 | 2: \ | |
4788 | uxtb sub_tile_height, column_data, ror #8; \ | |
59d15d23 | 4789 | setup_sprite_tile_##edge_mode##_##texture_mode##x4mode(edge); \ |
4790 | setup_sprite_tile_column_edge_post_adjust_##edge_mode##x4mode(edge) \ | |
75e28f62 E |
4791 | |
4792 | ||
4793 | #define setup_sprite_column_data_single() \ | |
4794 | mov column_data, height; \ | |
e1f6de8f | 4795 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] \ |
75e28f62 E |
4796 | |
4797 | #define setup_sprite_column_data_multi() \ | |
4798 | and height_rounded, height_rounded, #0xF; \ | |
4799 | rsb column_data, offset_v, #16; \ | |
4800 | \ | |
4801 | add height_rounded, height_rounded, #1; \ | |
4802 | sub tile_height, tile_height, #1; \ | |
4803 | \ | |
4804 | orr column_data, column_data, tile_height, lsl #16; \ | |
e1f6de8f | 4805 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset]; \ |
75e28f62 E |
4806 | \ |
4807 | orr column_data, column_data, height_rounded, lsl #8 \ | |
4808 | ||
59d15d23 | 4809 | #define setup_sprite_setup_left_draw_mask_fb_ptr() \ |
4810 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
4811 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ | |
4812 | ||
4813 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column() \ | |
4814 | mov fb_ptr_advance_column, #32; \ | |
4815 | vdup.u8 draw_mask_fb_ptr_left, block_masks[0]; \ | |
4816 | \ | |
ed0fd81d | 4817 | sub fb_ptr_advance_column, fb_ptr_advance_column, height, lsl #11; \ |
59d15d23 | 4818 | vdup.u8 draw_mask_fb_ptr_right, block_masks[1] \ |
4819 | ||
4820 | #define setup_sprite_setup_right_draw_mask_fb_ptr() \ | |
4821 | vdup.u8 draw_mask_fb_ptr_left, block_masks[4]; \ | |
4822 | vdup.u8 draw_mask_fb_ptr_right, block_masks[5] \ | |
4823 | ||
4824 | #define setup_sprite_tile_column_width_single(tm, multi_height, edge_mode, \ | |
4825 | edge, x4mode) \ | |
4826 | setup_sprite_##tm##_single_##multi_height##_##edge_mode##_##edge##x4mode: \ | |
75e28f62 E |
4827 | setup_sprite_column_data_##multi_height(); \ |
4828 | vext.32 block_masks_shifted, block_masks, block_masks, #1; \ | |
4829 | vorr.u32 block_masks, block_masks, block_masks_shifted; \ | |
59d15d23 | 4830 | setup_sprite_setup_left_draw_mask_fb_ptr##x4mode(); \ |
75e28f62 | 4831 | \ |
59d15d23 | 4832 | setup_sprite_tile_column_height_##multi_height(edge_mode, edge, tm, x4mode); \ |
9775fc33 | 4833 | restore_abi_regs(); \ |
75e28f62 E |
4834 | ldmia sp!, { r4 - r11, pc } \ |
4835 | ||
4836 | #define setup_sprite_tiled_advance_column() \ | |
4837 | add texture_offset_base, texture_offset_base, #0x100; \ | |
4838 | tst texture_offset_base, #0xF00; \ | |
4839 | subeq texture_offset_base, texture_offset_base, #(0x100 + 0xF00) \ | |
4840 | ||
4841 | #define setup_sprite_tile_column_width_multi(tm, multi_height, left_mode, \ | |
59d15d23 | 4842 | right_mode, x4mode) \ |
4843 | setup_sprite_##tm##_multi_##multi_height##_##left_mode##_##right_mode##x4mode:\ | |
75e28f62 | 4844 | setup_sprite_column_data_##multi_height(); \ |
75e28f62 | 4845 | \ |
59d15d23 | 4846 | setup_sprite_setup_left_draw_mask_fb_ptr_advance_column##x4mode(); \ |
75e28f62 | 4847 | \ |
59d15d23 | 4848 | setup_sprite_tile_column_height_##multi_height(left_mode, right, tm, x4mode);\ |
75e28f62 E |
4849 | \ |
4850 | subs tile_width, tile_width, #2; \ | |
4851 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ | |
4852 | \ | |
75e28f62 E |
4853 | beq 1f; \ |
4854 | \ | |
59d15d23 | 4855 | vmov.u8 draw_masks_fb_ptrs, #0; \ |
4856 | vmov.u8 draw_masks_fb_ptrs2, #0; \ | |
4857 | \ | |
75e28f62 E |
4858 | 0: \ |
4859 | setup_sprite_tiled_advance_column(); \ | |
59d15d23 | 4860 | setup_sprite_tile_column_height_##multi_height(full, none, tm, x4mode); \ |
75e28f62 E |
4861 | add fb_ptr, fb_ptr, fb_ptr_advance_column; \ |
4862 | subs tile_width, tile_width, #1; \ | |
4863 | bne 0b; \ | |
4864 | \ | |
4865 | 1: \ | |
59d15d23 | 4866 | setup_sprite_setup_right_draw_mask_fb_ptr##x4mode(); \ |
75e28f62 E |
4867 | \ |
4868 | setup_sprite_tiled_advance_column(); \ | |
59d15d23 | 4869 | setup_sprite_tile_column_height_##multi_height(right_mode, left, tm, x4mode);\ |
9775fc33 | 4870 | restore_abi_regs(); \ |
75e28f62 E |
4871 | ldmia sp!, { r4 - r11, pc } \ |
4872 | ||
4873 | ||
59d15d23 | 4874 | #define setup_sprite_offset_u_adjust() \ |
4875 | ||
4876 | #define setup_sprite_get_left_block_mask() \ | |
4877 | and left_block_mask, left_block_mask, #0xFF \ | |
4878 | ||
4879 | #define setup_sprite_compare_left_block_mask() \ | |
4880 | cmp left_block_mask, #0xFF \ | |
4881 | ||
4882 | #define setup_sprite_get_right_block_mask() \ | |
4883 | uxtb right_block_mask, right_block_mask, ror #8 \ | |
4884 | ||
4885 | #define setup_sprite_compare_right_block_mask() \ | |
4886 | cmp right_block_mask, #0xFF \ | |
4887 | ||
4888 | ||
4889 | ||
4890 | /* 4x stuff */ | |
4891 | #define fb_ptr2 column_data | |
4892 | ||
4893 | #define setup_sprite_offset_u_adjust_4x() \ | |
4894 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
4895 | lsl offset_u_right, #1; \ | |
4896 | lsl offset_u, #1; \ | |
4897 | add offset_u_right, #1 \ | |
4898 | ||
4899 | #define setup_sprite_get_left_block_mask_4x() \ | |
4900 | sxth left_block_mask, left_block_mask \ | |
4901 | ||
4902 | #define setup_sprite_compare_left_block_mask_4x() \ | |
4903 | cmp left_block_mask, #0xFFFFFFFF \ | |
4904 | ||
4905 | #define setup_sprite_get_right_block_mask_4x() \ | |
4906 | sxth right_block_mask, right_block_mask, ror #16 \ | |
4907 | ||
4908 | #define setup_sprite_compare_right_block_mask_4x() \ | |
4909 | cmp right_block_mask, #0xFFFFFFFF \ | |
4910 | ||
4911 | ||
4912 | #define widen_texels_16bpp(texels_) \ | |
4913 | vmov texels_wide_low, texels_; \ | |
4914 | vmov texels_wide_high, texels_; \ | |
4915 | vzip.16 texels_wide_low, texels_wide_high \ | |
4916 | ||
4917 | #define widen_texels_8bpp(texels_) \ | |
4918 | vmov texels_wide_low, texels_; \ | |
4919 | vmov texels_wide_high, texels_; \ | |
4920 | vzip.8 texels_wide_low, texels_wide_high \ | |
4921 | ||
4922 | #define write_block_16bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
e1f6de8f | 4923 | vst1.u32 { texels_ }, [block_, :128]; \ |
59d15d23 | 4924 | add block_, block_, #40; \ |
4925 | \ | |
4926 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
e1f6de8f | 4927 | vst1.u32 { draw_mask_fb_ptr_ }, [block_, :64]; \ |
59d15d23 | 4928 | add block_, block_, #24 \ |
4929 | ||
4930 | /* assumes 16-byte offset already added to block_ */ | |
4931 | #define write_block_8bpp(texels_, block_, draw_mask_fb_ptr_, fb_ptr_) \ | |
e1f6de8f | 4932 | vst1.u32 { texels_ }, [block_, :64]; \ |
59d15d23 | 4933 | add block_, block_, #24; \ |
4934 | \ | |
4935 | vmov.u32 draw_mask_fb_ptr_[1], fb_ptr_; \ | |
e1f6de8f | 4936 | vst1.u32 { draw_mask_fb_ptr_ }, [block_, :64]; \ |
59d15d23 | 4937 | add block_, block_, #40 \ |
4938 | ||
4939 | #define do_texture_block_16bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
4940 | draw_mask_fb_ptr_b_) \ | |
4941 | widen_texels_16bpp(texels_low); \ | |
4942 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
4943 | \ | |
4944 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
4945 | \ | |
4946 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
4947 | widen_texels_16bpp(texels_high); \ | |
4948 | \ | |
4949 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
4950 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
4951 | \ | |
4952 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
4953 | write_block_16bpp(texels_wide, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
4954 | ||
4955 | #define do_texture_block_8bpp_4x(fb_ptr_tmp, draw_mask_fb_ptr_a_, \ | |
4956 | draw_mask_fb_ptr_b_) \ | |
4957 | widen_texels_8bpp(texels); \ | |
4958 | add fb_ptr_tmp, fb_ptr, #1024*2; \ | |
4959 | \ | |
4960 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr); \ | |
4961 | write_block_8bpp(texels_wide_low, block, draw_mask_fb_ptr_a_, fb_ptr_tmp); \ | |
4962 | \ | |
4963 | add fb_ptr_tmp, fb_ptr, #8*2; \ | |
4964 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp); \ | |
4965 | \ | |
4966 | add fb_ptr_tmp, fb_ptr_tmp, #1024*2; \ | |
4967 | write_block_8bpp(texels_wide_high, block, draw_mask_fb_ptr_b_, fb_ptr_tmp) \ | |
4968 | ||
4969 | ||
4970 | #define setup_sprite_tiled_initialize_4bpp_4x() \ | |
e1f6de8f | 4971 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset]; \ |
4972 | vld1.u32 { clut_a, clut_b }, [clut_ptr, :128]; \ | |
59d15d23 | 4973 | \ |
4974 | vuzp.u8 clut_a, clut_b \ | |
4975 | ||
4976 | #define setup_sprite_tiled_initialize_8bpp_4x() \ | |
4977 | ||
4978 | ||
4979 | #define setup_sprite_block_count_single_4x() \ | |
4980 | sub_tile_height, lsl #2 \ | |
4981 | ||
4982 | #define setup_sprite_block_count_double_4x() \ | |
4983 | sub_tile_height, lsl #(1+2) \ | |
4984 | ||
4985 | #define setup_sprite_tile_full_4bpp_4x(edge) \ | |
4986 | setup_sprite_tile_add_blocks(double_4x); \ | |
4987 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
4988 | \ | |
4989 | 4: \ | |
4990 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 4991 | pld [fb_ptr]; \ |
59d15d23 | 4992 | \ |
4993 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 4994 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 4995 | \ |
4996 | add texture_block_ptr, texture_offset, #8; \ | |
4997 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
4998 | \ | |
4999 | and texture_block_ptr, texture_block_ptr, texture_mask; \ | |
5000 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
5001 | \ | |
5002 | vzip.8 texels_low, texels_high; \ | |
5003 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
5004 | draw_mask_fb_ptr_left_b); \ | |
5005 | \ | |
5006 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5007 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5008 | \ |
e1f6de8f | 5009 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
8438c3c7 | 5010 | add fb_ptr, fb_ptr, #16*2; \ |
59d15d23 | 5011 | \ |
8438c3c7 | 5012 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ |
59d15d23 | 5013 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ |
5014 | \ | |
5015 | vzip.8 texels_low, texels_high; \ | |
5016 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
5017 | draw_mask_fb_ptr_right_b); \ | |
5018 | \ | |
5019 | add texture_offset, texture_offset, #0x10; \ | |
5020 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
5021 | \ | |
5022 | subs sub_tile_height, sub_tile_height, #1; \ | |
5023 | bne 4b; \ | |
5024 | \ | |
5025 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5026 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5027 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5028 | |
5029 | ||
5030 | #define setup_sprite_tile_half_4bpp_4x(edge) \ | |
5031 | setup_sprite_tile_add_blocks(single_4x); \ | |
5032 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
5033 | \ | |
5034 | 4: \ | |
5035 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 5036 | pld [fb_ptr]; \ |
59d15d23 | 5037 | \ |
5038 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5039 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5040 | \ |
5041 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
5042 | vtbl.8 texels_low, { clut_low_a, clut_low_b }, texels; \ | |
5043 | \ | |
5044 | vtbl.8 texels_high, { clut_high_a, clut_high_b }, texels; \ | |
5045 | add texture_offset, texture_offset, #0x10; \ | |
5046 | \ | |
5047 | vzip.8 texels_low, texels_high; \ | |
5048 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ | |
5049 | draw_mask_fb_ptr_##edge##_b); \ | |
5050 | \ | |
e1f6de8f | 5051 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5052 | add fb_ptr, fb_ptr, #2048 * 2; \ |
59d15d23 | 5053 | \ |
8438c3c7 | 5054 | subs sub_tile_height, sub_tile_height, #1; \ |
59d15d23 | 5055 | bne 4b; \ |
5056 | \ | |
5057 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5058 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5059 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5060 | |
5061 | ||
5062 | #define setup_sprite_tile_full_8bpp_4x(edge) \ | |
5063 | setup_sprite_tile_add_blocks(double_4x); \ | |
5064 | add block, block, #16; \ | |
5065 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
5066 | \ | |
5067 | 4: \ | |
5068 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 5069 | pld [fb_ptr]; \ |
59d15d23 | 5070 | \ |
5071 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5072 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5073 | \ |
5074 | add texture_block_ptr, texture_offset, #8; \ | |
5075 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_left_a, \ | |
5076 | draw_mask_fb_ptr_left_b); \ | |
5077 | \ | |
e1f6de8f | 5078 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5079 | and texture_block_ptr, texture_block_ptr, texture_mask; \ |
5080 | \ | |
5081 | add fb_ptr, fb_ptr, #16*2; \ | |
5082 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
5083 | \ | |
e1f6de8f | 5084 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5085 | \ |
5086 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_right_a, \ | |
5087 | draw_mask_fb_ptr_right_b); \ | |
5088 | \ | |
5089 | add texture_offset, texture_offset, #0x10; \ | |
5090 | add fb_ptr, fb_ptr, #(2048 - 16) * 2; \ | |
5091 | \ | |
5092 | subs sub_tile_height, sub_tile_height, #1; \ | |
5093 | bne 4b; \ | |
5094 | \ | |
5095 | sub block, block, #16; \ | |
5096 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5097 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5098 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5099 | |
5100 | ||
5101 | #define setup_sprite_tile_half_8bpp_4x(edge) \ | |
5102 | setup_sprite_tile_add_blocks(single_4x); \ | |
5103 | add block, block, #16; \ | |
5104 | str column_data, [sp, #-8]!; /* fb_ptr2 */ \ | |
5105 | \ | |
5106 | 4: \ | |
5107 | and texture_block_ptr, texture_offset, texture_mask; \ | |
e1f6de8f | 5108 | pld [fb_ptr]; \ |
59d15d23 | 5109 | \ |
5110 | add texture_block_ptr, texture_page_ptr, texture_block_ptr; \ | |
e1f6de8f | 5111 | vld1.u32 { texels }, [texture_block_ptr, :64]; \ |
59d15d23 | 5112 | \ |
e1f6de8f | 5113 | pld [fb_ptr, #2048]; \ |
59d15d23 | 5114 | do_texture_block_8bpp_4x(fb_ptr2, draw_mask_fb_ptr_##edge##_a, \ |
5115 | draw_mask_fb_ptr_##edge##_b); \ | |
5116 | \ | |
5117 | add texture_offset, texture_offset, #0x10; \ | |
5118 | add fb_ptr, fb_ptr, #2048 * 2; \ | |
5119 | \ | |
5120 | subs sub_tile_height, sub_tile_height, #1; \ | |
5121 | bne 4b; \ | |
5122 | \ | |
5123 | sub block, block, #16; \ | |
5124 | ldr column_data, [sp], #8; /* fb_ptr2 */ \ | |
5125 | add texture_offset, texture_offset, #0xF00; \ | |
e1f6de8f | 5126 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] \ |
59d15d23 | 5127 | |
5128 | ||
5129 | #define setup_sprite_tile_column_edge_pre_adjust_half_right_4x() \ | |
5130 | add texture_offset, texture_offset_base, #8; \ | |
5131 | add fb_ptr, fb_ptr, #16 * 2 \ | |
5132 | ||
5133 | #define setup_sprite_tile_column_edge_pre_adjust_half_left_4x() \ | |
5134 | mov texture_offset, texture_offset_base \ | |
5135 | ||
5136 | #define setup_sprite_tile_column_edge_pre_adjust_half_4x(edge) \ | |
5137 | setup_sprite_tile_column_edge_pre_adjust_half_##edge##_4x() \ | |
5138 | ||
5139 | #define setup_sprite_tile_column_edge_pre_adjust_full_4x(edge) \ | |
5140 | mov texture_offset, texture_offset_base \ | |
5141 | ||
5142 | #define setup_sprite_tile_column_edge_post_adjust_half_right_4x() \ | |
5143 | sub fb_ptr, fb_ptr, #16 * 2 \ | |
5144 | ||
5145 | #define setup_sprite_tile_column_edge_post_adjust_half_left_4x() \ | |
5146 | ||
5147 | #define setup_sprite_tile_column_edge_post_adjust_half_4x(edge) \ | |
5148 | setup_sprite_tile_column_edge_post_adjust_half_##edge##_4x() \ | |
5149 | ||
5150 | #define setup_sprite_tile_column_edge_post_adjust_full_4x(edge) \ | |
5151 | ||
5152 | ||
5153 | #define setup_sprite_setup_left_draw_mask_fb_ptr_4x() \ | |
5154 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
5155 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
5156 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ | |
5157 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
5158 | ||
5159 | #define setup_sprite_setup_left_draw_mask_fb_ptr_advance_column_4x() \ | |
5160 | mov fb_ptr_advance_column, #32 * 2; \ | |
5161 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[0]; \ | |
5162 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[1]; \ | |
ed0fd81d | 5163 | sub fb_ptr_advance_column, fb_ptr_advance_column, height, lsl #11 + 1; \ |
59d15d23 | 5164 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[2]; \ |
5165 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[3] \ | |
5166 | ||
5167 | #define setup_sprite_setup_right_draw_mask_fb_ptr_4x() \ | |
5168 | vdup.u8 draw_mask_fb_ptr_left_a, block_masks[4]; \ | |
5169 | vdup.u8 draw_mask_fb_ptr_left_b, block_masks[5]; \ | |
5170 | vdup.u8 draw_mask_fb_ptr_right_a, block_masks[6]; \ | |
5171 | vdup.u8 draw_mask_fb_ptr_right_b, block_masks[7] \ | |
5172 | ||
5173 | ||
75e28f62 E |
5174 | // r0: psx_gpu |
5175 | // r1: x | |
5176 | // r2: y | |
5177 | // r3: u | |
e1f6de8f | 5178 | // [sp]: v |
5179 | // [sp + 4]: width | |
5180 | // [sp + 8]: height | |
5181 | // [sp + 12]: color (unused) | |
75e28f62 | 5182 | |
59d15d23 | 5183 | #define setup_sprite_tiled_builder(texture_mode, x4mode) \ |
5184 | \ | |
5185 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, full, \ | |
5186 | x4mode); \ | |
5187 | setup_sprite_tile_column_width_single(texture_mode, multi, full, none, \ | |
5188 | x4mode); \ | |
5189 | setup_sprite_tile_column_width_multi(texture_mode, single, full, full, \ | |
5190 | x4mode); \ | |
5191 | setup_sprite_tile_column_width_single(texture_mode, single, full, none, \ | |
5192 | x4mode); \ | |
5193 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, full, \ | |
5194 | x4mode); \ | |
5195 | setup_sprite_tile_column_width_single(texture_mode, multi, half, right, \ | |
5196 | x4mode); \ | |
5197 | setup_sprite_tile_column_width_multi(texture_mode, single, half, full, \ | |
5198 | x4mode); \ | |
5199 | setup_sprite_tile_column_width_single(texture_mode, single, half, right, \ | |
5200 | x4mode); \ | |
5201 | setup_sprite_tile_column_width_multi(texture_mode, multi, full, half, \ | |
5202 | x4mode); \ | |
5203 | setup_sprite_tile_column_width_single(texture_mode, multi, half, left, \ | |
5204 | x4mode); \ | |
5205 | setup_sprite_tile_column_width_multi(texture_mode, single, full, half, \ | |
5206 | x4mode); \ | |
5207 | setup_sprite_tile_column_width_single(texture_mode, single, half, left, \ | |
5208 | x4mode); \ | |
5209 | setup_sprite_tile_column_width_multi(texture_mode, multi, half, half, \ | |
5210 | x4mode); \ | |
5211 | setup_sprite_tile_column_width_multi(texture_mode, single, half, half, \ | |
5212 | x4mode); \ | |
75e28f62 E |
5213 | \ |
5214 | .align 4; \ | |
5215 | \ | |
59d15d23 | 5216 | function(setup_sprite_##texture_mode##x4mode) \ |
75e28f62 | 5217 | stmdb sp!, { r4 - r11, r14 }; \ |
59d15d23 | 5218 | setup_sprite_tiled_initialize_##texture_mode##x4mode(); \ |
75e28f62 | 5219 | \ |
e1f6de8f | 5220 | ldr v, [sp, #36]; \ |
75e28f62 E |
5221 | and offset_u, u, #0xF; \ |
5222 | \ | |
e1f6de8f | 5223 | ldr width, [sp, #40]; \ |
5224 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset]; \ | |
75e28f62 | 5225 | \ |
e1f6de8f | 5226 | ldr height, [sp, #44]; \ |
75e28f62 E |
5227 | add fb_ptr, fb_ptr, y, lsl #11; \ |
5228 | \ | |
9775fc33 | 5229 | save_abi_regs(); \ |
5230 | \ | |
75e28f62 E |
5231 | add fb_ptr, fb_ptr, x, lsl #1; \ |
5232 | and offset_v, v, #0xF; \ | |
5233 | \ | |
5234 | sub fb_ptr, fb_ptr, offset_u, lsl #1; \ | |
5235 | add width_rounded, offset_u, width; \ | |
5236 | \ | |
5237 | add height_rounded, offset_v, height; \ | |
5238 | add width_rounded, width_rounded, #15; \ | |
5239 | \ | |
5240 | add height_rounded, height_rounded, #15; \ | |
5241 | mov tile_width, width_rounded, lsr #4; \ | |
5242 | \ | |
5243 | /* texture_offset_base = VH-VL-00-00 */\ | |
5244 | mov texture_offset_base, v, lsl #8; \ | |
5245 | and offset_u_right, width_rounded, #0xF; \ | |
5246 | \ | |
5247 | /* texture_offset_base = VH-UH-UL-00 */\ | |
5248 | bfi texture_offset_base, u, #4, #8; \ | |
59d15d23 | 5249 | mov right_block_mask, #0xFFFFFFFE; \ |
5250 | \ | |
5251 | setup_sprite_offset_u_adjust##x4mode(); \ | |
75e28f62 E |
5252 | \ |
5253 | /* texture_offset_base = VH-UH-VL-00 */\ | |
5254 | bfi texture_offset_base, v, #4, #4; \ | |
59d15d23 | 5255 | mov left_block_mask, #0xFFFFFFFF; \ |
75e28f62 E |
5256 | \ |
5257 | mov tile_height, height_rounded, lsr #4; \ | |
5258 | mvn left_block_mask, left_block_mask, lsl offset_u; \ | |
5259 | \ | |
5260 | /* texture_mask = HH-HL-WH-WL */\ | |
e1f6de8f | 5261 | ldrh texture_mask, [psx_gpu, #psx_gpu_texture_mask_width_offset]; \ |
75e28f62 E |
5262 | mov right_block_mask, right_block_mask, lsl offset_u_right; \ |
5263 | \ | |
5264 | /* texture_mask_rev = WH-WL-HH-HL */\ | |
5265 | rev16 texture_mask_rev, texture_mask; \ | |
5266 | vmov block_masks, left_block_mask, right_block_mask; \ | |
5267 | \ | |
5268 | /* texture_mask = HH-HL-HL-WL */\ | |
5269 | bfi texture_mask, texture_mask_rev, #4, #4; \ | |
5270 | /* texture_mask_rev = 00-00-00-WH */\ | |
5271 | mov texture_mask_rev, texture_mask_rev, lsr #12; \ | |
5272 | \ | |
5273 | /* texture_mask = HH-WH-HL-WL */\ | |
5274 | bfi texture_mask, texture_mask_rev, #8, #4; \ | |
59d15d23 | 5275 | setup_sprite_get_left_block_mask##x4mode(); \ |
75e28f62 E |
5276 | \ |
5277 | mov control_mask, #0; \ | |
59d15d23 | 5278 | setup_sprite_compare_left_block_mask##x4mode(); \ |
75e28f62 | 5279 | \ |
59d15d23 | 5280 | setup_sprite_get_right_block_mask##x4mode(); \ |
75e28f62 E |
5281 | orreq control_mask, control_mask, #0x4; \ |
5282 | \ | |
e1f6de8f | 5283 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset]; \ |
59d15d23 | 5284 | setup_sprite_compare_right_block_mask##x4mode(); \ |
75e28f62 E |
5285 | \ |
5286 | orreq control_mask, control_mask, #0x8; \ | |
5287 | cmp tile_width, #1; \ | |
5288 | \ | |
5289 | add block, psx_gpu, #psx_gpu_blocks_offset; \ | |
5290 | orreq control_mask, control_mask, #0x1; \ | |
5291 | \ | |
5292 | cmp tile_height, #1; \ | |
5293 | add block, block, num_blocks, lsl #6; \ | |
5294 | \ | |
5295 | orreq control_mask, control_mask, #0x2; \ | |
8184d7c5 | 5296 | JT_OP_REL(9f, control_mask, temp); \ |
e1f6de8f | 5297 | JT_OP(ldr pc, [pc, control_mask, lsl #2]); \ |
75e28f62 E |
5298 | nop; \ |
5299 | \ | |
8184d7c5 | 5300 | 9: \ |
5301 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_full_full##x4mode); \ | |
5302 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_full_none##x4mode); \ | |
5303 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_full_full##x4mode); \ | |
5304 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_full_none##x4mode); \ | |
5305 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_half_full##x4mode); \ | |
5306 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_half_right##x4mode); \ | |
5307 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_half_full##x4mode); \ | |
5308 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_half_right##x4mode);\ | |
5309 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_full_half##x4mode); \ | |
5310 | .word JTE(9b, setup_sprite_##texture_mode##_single_multi_half_left##x4mode); \ | |
5311 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_full_half##x4mode); \ | |
5312 | .word JTE(9b, setup_sprite_##texture_mode##_single_single_half_left##x4mode); \ | |
5313 | .word JTE(9b, setup_sprite_##texture_mode##_multi_multi_half_half##x4mode); \ | |
75e28f62 | 5314 | .word 0x00000000; \ |
8184d7c5 | 5315 | .word JTE(9b, setup_sprite_##texture_mode##_multi_single_half_half##x4mode); \ |
59d15d23 | 5316 | |
5317 | ||
5318 | setup_sprite_tiled_builder(4bpp,); | |
5319 | setup_sprite_tiled_builder(8bpp,); | |
75e28f62 | 5320 | |
59d15d23 | 5321 | #undef draw_mask_fb_ptr_left |
5322 | #undef draw_mask_fb_ptr_right | |
75e28f62 | 5323 | |
59d15d23 | 5324 | setup_sprite_tiled_builder(4bpp, _4x); |
5325 | setup_sprite_tiled_builder(8bpp, _4x); | |
75e28f62 E |
5326 | |
5327 | ||
5328 | #undef block_ptr | |
5329 | #undef num_blocks | |
5330 | #undef clut_ptr | |
5331 | ||
5332 | #define psx_gpu r0 | |
5333 | #define block_ptr r0 | |
5334 | #define num_blocks r1 | |
5335 | #define clut_ptr r2 | |
5336 | #define texel_shift_mask r3 | |
5337 | #define block_pixels_a r4 | |
5338 | #define block_pixels_b r5 | |
5339 | #define texel_0 r6 | |
5340 | #define texel_2 r7 | |
5341 | #define texel_4 r8 | |
5342 | #define texel_6 r9 | |
5343 | #define texel_1 r10 | |
5344 | #define texel_3 r11 | |
5345 | #define texel_5 r12 | |
5346 | #define texel_7 r14 | |
5347 | #define texels_01 r6 | |
5348 | #define texels_23 r7 | |
5349 | #define texels_45 r8 | |
5350 | #define texels_67 r9 | |
5351 | ||
5352 | function(texture_sprite_blocks_8bpp) | |
5353 | stmdb sp!, { r4 - r11, r14 } | |
5354 | movw texel_shift_mask, #(0xFF << 1) | |
5355 | ||
e1f6de8f | 5356 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
5357 | ldr clut_ptr, [psx_gpu, #psx_gpu_clut_ptr_offset] | |
75e28f62 E |
5358 | |
5359 | add block_ptr, psx_gpu, #psx_gpu_blocks_offset | |
e1f6de8f | 5360 | ldr block_pixels_a, [block_ptr, #16] |
75e28f62 E |
5361 | |
5362 | 0: | |
5363 | and texel_0, texel_shift_mask, block_pixels_a, lsl #1 | |
e1f6de8f | 5364 | ldr block_pixels_b, [block_ptr, #20] |
75e28f62 E |
5365 | |
5366 | and texel_1, texel_shift_mask, block_pixels_a, lsr #7 | |
e1f6de8f | 5367 | ldrh texel_0, [clut_ptr, texel_0] |
75e28f62 E |
5368 | |
5369 | and texel_2, texel_shift_mask, block_pixels_a, lsr #15 | |
e1f6de8f | 5370 | ldrh texel_1, [clut_ptr, texel_1] |
75e28f62 E |
5371 | |
5372 | and texel_3, texel_shift_mask, block_pixels_a, lsr #23 | |
e1f6de8f | 5373 | ldr block_pixels_a, [block_ptr, #(64 + 16)] |
75e28f62 | 5374 | |
e1f6de8f | 5375 | ldrh texel_2, [clut_ptr, texel_2] |
75e28f62 E |
5376 | and texel_4, texel_shift_mask, block_pixels_b, lsl #1 |
5377 | ||
e1f6de8f | 5378 | ldrh texel_3, [clut_ptr, texel_3] |
75e28f62 E |
5379 | and texel_5, texel_shift_mask, block_pixels_b, lsr #7 |
5380 | ||
e1f6de8f | 5381 | ldrh texel_4, [clut_ptr, texel_4] |
75e28f62 E |
5382 | and texel_6, texel_shift_mask, block_pixels_b, lsr #15 |
5383 | ||
e1f6de8f | 5384 | ldrh texel_5, [clut_ptr, texel_5] |
75e28f62 E |
5385 | and texel_7, texel_shift_mask, block_pixels_b, lsr #23 |
5386 | ||
e1f6de8f | 5387 | ldrh texel_6, [clut_ptr, texel_6] |
75e28f62 E |
5388 | orr texels_01, texel_0, texel_1, lsl #16 |
5389 | ||
e1f6de8f | 5390 | ldrh texel_7, [clut_ptr, texel_7] |
75e28f62 E |
5391 | orr texels_23, texel_2, texel_3, lsl #16 |
5392 | ||
5393 | orr texels_45, texel_4, texel_5, lsl #16 | |
e1f6de8f | 5394 | str texels_01, [block_ptr, #0] |
75e28f62 E |
5395 | |
5396 | orr texels_67, texel_6, texel_7, lsl #16 | |
e1f6de8f | 5397 | str texels_23, [block_ptr, #4] |
75e28f62 E |
5398 | |
5399 | subs num_blocks, num_blocks, #1 | |
e1f6de8f | 5400 | str texels_45, [block_ptr, #8] |
75e28f62 | 5401 | |
e1f6de8f | 5402 | str texels_67, [block_ptr, #12] |
75e28f62 E |
5403 | add block_ptr, block_ptr, #64 |
5404 | ||
5405 | bne 0b | |
5406 | ||
5407 | ldmia sp!, { r4 - r11, pc } | |
5408 | ||
5409 | ||
5410 | #undef width_rounded | |
5411 | #undef texture_mask | |
5412 | #undef num_blocks | |
5413 | #undef texture_offset | |
59d15d23 | 5414 | #undef texels_low |
5415 | #undef texels_high | |
5416 | #undef texels_wide_low | |
5417 | #undef texels_wide_high | |
5418 | #undef texels_wide | |
5419 | #undef fb_ptr2 | |
8184d7c5 | 5420 | #undef temp |
75e28f62 E |
5421 | |
5422 | #define psx_gpu r0 | |
5423 | #define x r1 | |
5424 | #define y r2 | |
5425 | #define u r3 | |
5426 | #define v r4 | |
5427 | #define width r5 | |
5428 | #define height r6 | |
5429 | #define left_offset r8 | |
5430 | #define width_rounded r9 | |
5431 | #define right_width r10 | |
59d15d23 | 5432 | |
75e28f62 E |
5433 | #define block_width r11 |
5434 | ||
5435 | #define texture_offset_base r1 | |
5436 | #define texture_mask r2 | |
5437 | #define texture_page_ptr r3 | |
5438 | #define num_blocks r4 | |
5439 | #define block r5 | |
5440 | #define fb_ptr r7 | |
5441 | #define texture_offset r8 | |
5442 | #define blocks_remaining r9 | |
59d15d23 | 5443 | #define fb_ptr2 r10 |
75e28f62 E |
5444 | #define fb_ptr_pitch r12 |
5445 | #define texture_block_ptr r14 | |
5446 | ||
5447 | #define texture_mask_width r2 | |
5448 | #define texture_mask_height r3 | |
5449 | #define left_mask_bits r4 | |
5450 | #define right_mask_bits r5 | |
5451 | ||
5452 | ||
5453 | #undef block_masks | |
5454 | #undef block_masks_shifted | |
5455 | #undef texels | |
5456 | ||
5457 | #define block_masks d0 | |
5458 | #define block_masks_shifted d1 | |
5459 | #define draw_mask_fb_ptr d2 | |
5460 | #define texels q2 | |
5461 | ||
59d15d23 | 5462 | #define draw_mask_fb_ptr_a d2 |
5463 | #define draw_mask_fb_ptr_b d3 | |
5464 | #define texels_low d4 | |
5465 | #define texels_high d5 | |
5466 | #define texels_wide_low d6 | |
5467 | #define texels_wide_high d7 | |
5468 | #define texels_wide q3 | |
75e28f62 | 5469 | |
75e28f62 | 5470 | |
59d15d23 | 5471 | setup_sprites_16bpp_flush: |
5472 | vpush { d0 - d3 } | |
75e28f62 | 5473 | |
4d646738 | 5474 | stmdb sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 5475 | bl flush_render_block_buffer |
4d646738 | 5476 | ldmia sp!, { r0 - r3, EXTRA_UNSAVED_REGS r12, r14 } |
75e28f62 | 5477 | |
59d15d23 | 5478 | vpop { d0 - d3 } |
75e28f62 E |
5479 | |
5480 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5481 | mov num_blocks, block_width | |
5482 | ||
5483 | bx lr | |
5484 | ||
5485 | function(setup_sprite_16bpp) | |
5486 | stmdb sp!, { r4 - r11, r14 } | |
e1f6de8f | 5487 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
75e28f62 | 5488 | |
e1f6de8f | 5489 | ldr v, [sp, #36] |
75e28f62 E |
5490 | add fb_ptr, fb_ptr, y, lsl #11 |
5491 | ||
e1f6de8f | 5492 | ldr width, [sp, #40] |
75e28f62 E |
5493 | add fb_ptr, fb_ptr, x, lsl #1 |
5494 | ||
e1f6de8f | 5495 | ldr height, [sp, #44] |
75e28f62 E |
5496 | and left_offset, u, #0x7 |
5497 | ||
5498 | add texture_offset_base, u, u | |
5499 | add width_rounded, width, #7 | |
5500 | ||
ed0fd81d | 5501 | add texture_offset_base, texture_offset_base, v, lsl #11 |
75e28f62 E |
5502 | mov left_mask_bits, #0xFF |
5503 | ||
e1f6de8f | 5504 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
75e28f62 E |
5505 | add width_rounded, width_rounded, left_offset |
5506 | ||
e1f6de8f | 5507 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
75e28f62 E |
5508 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
5509 | ||
5510 | add texture_mask, texture_mask_width, texture_mask_width | |
5511 | mov right_mask_bits, #0xFE | |
5512 | ||
5513 | and right_width, width_rounded, #0x7 | |
5514 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
5515 | ||
ed0fd81d | 5516 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
75e28f62 E |
5517 | mov block_width, width_rounded, lsr #3 |
5518 | ||
5519 | mov right_mask_bits, right_mask_bits, lsl right_width | |
5520 | movw fb_ptr_pitch, #(2048 + 16) | |
5521 | ||
5522 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4 | |
5523 | vmov block_masks, left_mask_bits, right_mask_bits | |
5524 | ||
e1f6de8f | 5525 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
5526 | add block, psx_gpu, #psx_gpu_blocks_offset |
5527 | ||
6ea0f7bf | 5528 | bic texture_offset_base, texture_offset_base, #0xF |
75e28f62 E |
5529 | cmp block_width, #1 |
5530 | ||
e1f6de8f | 5531 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
75e28f62 E |
5532 | add block, block, num_blocks, lsl #6 |
5533 | ||
5534 | bne 0f | |
5535 | ||
5536 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
5537 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
5538 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
5539 | ||
5540 | 1: | |
5541 | add num_blocks, num_blocks, #1 | |
5542 | cmp num_blocks, #MAX_BLOCKS | |
59d15d23 | 5543 | blgt setup_sprites_16bpp_flush |
75e28f62 E |
5544 | |
5545 | and texture_block_ptr, texture_offset_base, texture_mask | |
5546 | subs height, height, #1 | |
5547 | ||
5548 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5549 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 | 5550 | |
e1f6de8f | 5551 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5552 | add block, block, #40 |
5553 | ||
5554 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5555 | pld [fb_ptr] |
75e28f62 | 5556 | |
e1f6de8f | 5557 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5558 | |
5559 | add block, block, #24 | |
5560 | add texture_offset_base, texture_offset_base, #2048 | |
5561 | add fb_ptr, fb_ptr, #2048 | |
e1f6de8f | 5562 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
5563 | bne 1b |
5564 | ||
5565 | ldmia sp!, { r4 - r11, pc } | |
5566 | ||
5567 | 0: | |
5568 | add num_blocks, num_blocks, block_width | |
5569 | mov texture_offset, texture_offset_base | |
5570 | ||
5571 | cmp num_blocks, #MAX_BLOCKS | |
59d15d23 | 5572 | blgt setup_sprites_16bpp_flush |
75e28f62 E |
5573 | |
5574 | add texture_offset_base, texture_offset_base, #2048 | |
5575 | and texture_block_ptr, texture_offset, texture_mask | |
5576 | ||
5577 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5578 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 | 5579 | |
e1f6de8f | 5580 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5581 | add block, block, #40 |
5582 | ||
5583 | vdup.u8 draw_mask_fb_ptr, block_masks[0] | |
5584 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5585 | pld [fb_ptr] |
75e28f62 | 5586 | |
e1f6de8f | 5587 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5588 | subs blocks_remaining, block_width, #2 |
5589 | ||
5590 | add texture_offset, texture_offset, #16 | |
5591 | add fb_ptr, fb_ptr, #16 | |
5592 | ||
5593 | vmov.u8 draw_mask_fb_ptr, #0 | |
5594 | ||
5595 | add block, block, #24 | |
5596 | beq 2f | |
5597 | ||
5598 | 1: | |
5599 | and texture_block_ptr, texture_offset, texture_mask | |
5600 | subs blocks_remaining, blocks_remaining, #1 | |
5601 | ||
5602 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5603 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 | 5604 | |
e1f6de8f | 5605 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5606 | add block, block, #40 |
5607 | ||
5608 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5609 | pld [fb_ptr] |
75e28f62 | 5610 | |
e1f6de8f | 5611 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5612 | |
5613 | add texture_offset, texture_offset, #16 | |
5614 | add fb_ptr, fb_ptr, #16 | |
5615 | ||
5616 | add block, block, #24 | |
5617 | bne 1b | |
5618 | ||
5619 | 2: | |
5620 | and texture_block_ptr, texture_offset, texture_mask | |
5621 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5622 | ||
e1f6de8f | 5623 | vld1.u32 { texels }, [texture_block_ptr, :128] |
75e28f62 E |
5624 | vdup.u8 draw_mask_fb_ptr, block_masks[4] |
5625 | ||
e1f6de8f | 5626 | vst1.u32 { texels }, [block, :128] |
75e28f62 E |
5627 | add block, block, #40 |
5628 | ||
5629 | vmov.u32 draw_mask_fb_ptr[1], fb_ptr | |
e1f6de8f | 5630 | vst1.u32 { draw_mask_fb_ptr }, [block, :64] |
75e28f62 E |
5631 | |
5632 | add block, block, #24 | |
5633 | subs height, height, #1 | |
5634 | ||
5635 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
e1f6de8f | 5636 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
75e28f62 E |
5637 | |
5638 | bne 0b | |
5639 | ||
5640 | ldmia sp!, { r4 - r11, pc } | |
5641 | ||
5642 | ||
59d15d23 | 5643 | // 4x version |
5644 | // FIXME: duplicate code with normal version :( | |
5645 | #undef draw_mask_fb_ptr | |
5646 | ||
5647 | function(setup_sprite_16bpp_4x) | |
5648 | stmdb sp!, { r4 - r11, r14 } | |
e1f6de8f | 5649 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] |
59d15d23 | 5650 | |
e1f6de8f | 5651 | ldr v, [sp, #36] |
59d15d23 | 5652 | add fb_ptr, fb_ptr, y, lsl #11 |
5653 | ||
e1f6de8f | 5654 | ldr width, [sp, #40] |
59d15d23 | 5655 | add fb_ptr, fb_ptr, x, lsl #1 |
5656 | ||
e1f6de8f | 5657 | ldr height, [sp, #44] |
59d15d23 | 5658 | and left_offset, u, #0x7 |
5659 | ||
5660 | add texture_offset_base, u, u | |
5661 | add width_rounded, width, #7 | |
5662 | ||
ed0fd81d | 5663 | add texture_offset_base, texture_offset_base, v, lsl #11 |
59d15d23 | 5664 | movw left_mask_bits, #0xFFFF |
5665 | ||
e1f6de8f | 5666 | ldrb texture_mask_width, [psx_gpu, #psx_gpu_texture_mask_width_offset] |
59d15d23 | 5667 | add width_rounded, width_rounded, left_offset |
5668 | ||
5669 | lsl left_offset, #1 | |
5670 | ||
e1f6de8f | 5671 | ldrb texture_mask_height, [psx_gpu, #psx_gpu_texture_mask_height_offset] |
59d15d23 | 5672 | sub fb_ptr, fb_ptr, left_offset, lsl #1 |
5673 | ||
5674 | add texture_mask, texture_mask_width, texture_mask_width | |
5675 | movw right_mask_bits, #0xFFFC | |
5676 | ||
5677 | and right_width, width_rounded, #0x7 | |
5678 | mvn left_mask_bits, left_mask_bits, lsl left_offset | |
5679 | ||
5680 | lsl right_width, #1 | |
5681 | ||
ed0fd81d | 5682 | add texture_mask, texture_mask, texture_mask_height, lsl #11 |
59d15d23 | 5683 | mov block_width, width_rounded, lsr #3 |
5684 | ||
5685 | mov right_mask_bits, right_mask_bits, lsl right_width | |
5686 | movw fb_ptr_pitch, #(2048 + 16) * 2 | |
5687 | ||
5688 | sub fb_ptr_pitch, fb_ptr_pitch, block_width, lsl #4+1 | |
5689 | vmov block_masks, left_mask_bits, right_mask_bits | |
5690 | ||
e1f6de8f | 5691 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
59d15d23 | 5692 | add block, psx_gpu, #psx_gpu_blocks_offset |
5693 | ||
5694 | bic texture_offset_base, texture_offset_base, #0xF | |
5695 | cmp block_width, #1 | |
5696 | ||
e1f6de8f | 5697 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_ptr_offset] |
59d15d23 | 5698 | add block, block, num_blocks, lsl #6 |
5699 | ||
5700 | lsl block_width, #2 | |
5701 | bne 0f | |
5702 | ||
5703 | vext.32 block_masks_shifted, block_masks, block_masks, #1 | |
5704 | vorr.u32 block_masks, block_masks, block_masks_shifted | |
5705 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] | |
5706 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
5707 | ||
5708 | 1: | |
5709 | add num_blocks, num_blocks, block_width | |
5710 | cmp num_blocks, #MAX_BLOCKS | |
5711 | blgt setup_sprites_16bpp_flush | |
5712 | ||
5713 | and texture_block_ptr, texture_offset_base, texture_mask | |
5714 | subs height, height, #1 | |
5715 | ||
5716 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5717 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5718 | |
5719 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5720 | ||
5721 | add texture_offset_base, texture_offset_base, #2048 | |
5722 | add fb_ptr, fb_ptr, #2048*2 | |
e1f6de8f | 5723 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
59d15d23 | 5724 | bne 1b |
5725 | ||
5726 | ldmia sp!, { r4 - r11, pc } | |
5727 | ||
5728 | 0: | |
5729 | add num_blocks, num_blocks, block_width | |
5730 | mov texture_offset, texture_offset_base | |
5731 | ||
5732 | vdup.u8 draw_mask_fb_ptr_a, block_masks[0] // left_mask_bits | |
5733 | vdup.u8 draw_mask_fb_ptr_b, block_masks[1] | |
5734 | ||
5735 | cmp num_blocks, #MAX_BLOCKS | |
5736 | blgt setup_sprites_16bpp_flush | |
5737 | ||
5738 | add texture_offset_base, texture_offset_base, #2048 | |
5739 | and texture_block_ptr, texture_offset, texture_mask | |
5740 | ||
5741 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5742 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5743 | |
5744 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5745 | ||
5746 | subs blocks_remaining, block_width, #2*4 | |
5747 | add texture_offset, texture_offset, #16 | |
5748 | ||
5749 | vmov.u8 draw_mask_fb_ptr_a, #0 | |
5750 | vmov.u8 draw_mask_fb_ptr_b, #0 | |
5751 | ||
5752 | add fb_ptr, fb_ptr, #16*2 | |
5753 | beq 2f | |
5754 | ||
5755 | 1: | |
5756 | and texture_block_ptr, texture_offset, texture_mask | |
5757 | subs blocks_remaining, blocks_remaining, #4 | |
5758 | ||
5759 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
e1f6de8f | 5760 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5761 | |
5762 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5763 | add texture_offset, texture_offset, #16 | |
5764 | ||
5765 | add fb_ptr, fb_ptr, #16*2 | |
5766 | bgt 1b | |
5767 | ||
5768 | 2: | |
5769 | vdup.u8 draw_mask_fb_ptr_a, block_masks[4] // right_mask_bits | |
5770 | vdup.u8 draw_mask_fb_ptr_b, block_masks[5] | |
5771 | ||
5772 | and texture_block_ptr, texture_offset, texture_mask | |
5773 | add texture_block_ptr, texture_page_ptr, texture_block_ptr | |
5774 | ||
e1f6de8f | 5775 | vld1.u32 { texels }, [texture_block_ptr, :128] |
59d15d23 | 5776 | |
5777 | do_texture_block_16bpp_4x(fb_ptr2, draw_mask_fb_ptr_a, draw_mask_fb_ptr_b) | |
5778 | subs height, height, #1 | |
5779 | ||
5780 | add fb_ptr, fb_ptr, fb_ptr_pitch | |
e1f6de8f | 5781 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
59d15d23 | 5782 | |
5783 | bne 0b | |
5784 | ||
5785 | ldmia sp!, { r4 - r11, pc } | |
5786 | ||
5787 | ||
f0931e56 | 5788 | #undef width |
5789 | #undef right_width | |
5790 | #undef right_mask_bits | |
5791 | #undef color | |
5792 | #undef height | |
5793 | #undef blocks_remaining | |
5794 | #undef colors | |
5795 | #undef right_mask | |
5796 | #undef test_mask | |
5797 | #undef draw_mask | |
5798 | ||
5799 | #define psx_gpu r0 | |
5800 | #define x r1 | |
5801 | #define y r2 | |
5802 | #define width r3 | |
5803 | #define right_width r5 | |
5804 | #define right_mask_bits r6 | |
5805 | #define fb_ptr r7 | |
5806 | #define color r8 | |
5807 | #define height r9 | |
5808 | #define fb_ptr_pitch r12 | |
5809 | ||
5810 | // referenced by setup_sprites_16bpp_flush | |
5811 | #define num_blocks r4 | |
5812 | #define block r5 | |
5813 | #define block_width r11 | |
5814 | ||
5815 | #define color_r r1 | |
5816 | #define color_g r2 | |
5817 | #define color_b r8 | |
5818 | #define blocks_remaining r6 | |
5819 | ||
5820 | #define colors q0 | |
5821 | #define right_mask q1 | |
5822 | #define test_mask q2 | |
5823 | #define draw_mask q2 | |
5824 | #define draw_mask_bits_fb_ptr d6 | |
5825 | ||
5826 | ||
5827 | .align 3 | |
5828 | ||
5829 | function(setup_sprite_untextured) | |
e1f6de8f | 5830 | ldrh r12, [psx_gpu, #psx_gpu_render_state_offset] |
f0931e56 | 5831 | tst r12, #(RENDER_STATE_MASK_EVALUATE | RENDER_FLAGS_MODULATE_TEXELS \ |
5832 | | RENDER_FLAGS_BLEND) | |
e1f6de8f | 5833 | ldrbeq r12, [psx_gpu, #psx_gpu_render_mode_offset] |
d5c08ed3 | 5834 | tsteq r12, #RENDER_INTERLACE_ENABLED |
f0931e56 | 5835 | beq setup_sprite_untextured_simple |
5836 | ||
5837 | stmdb sp!, { r4 - r11, r14 } | |
5838 | ||
e1f6de8f | 5839 | ldr width, [sp, #40] |
5840 | ldr fb_ptr, [psx_gpu, #psx_gpu_vram_out_ptr_offset] | |
f0931e56 | 5841 | |
e1f6de8f | 5842 | ldr height, [sp, #44] |
f0931e56 | 5843 | add fb_ptr, fb_ptr, y, lsl #11 |
5844 | ||
5845 | add fb_ptr, fb_ptr, x, lsl #1 | |
5846 | sub right_width, width, #1 | |
5847 | ||
e1f6de8f | 5848 | ldr color, [sp, #48] |
f0931e56 | 5849 | and right_width, #7 |
5850 | ||
5851 | add block_width, width, #7 | |
5852 | add right_width, #1 | |
5853 | ||
5854 | lsr block_width, #3 | |
5855 | mov right_mask_bits, #0xff | |
5856 | ||
5857 | sub fb_ptr_pitch, block_width, #1 | |
5858 | lsl right_mask_bits, right_width | |
5859 | ||
5860 | lsl fb_ptr_pitch, #3+1 | |
5861 | ubfx color_r, color, #3, #5 | |
5862 | ||
5863 | rsb fb_ptr_pitch, #1024*2 | |
5864 | ubfx color_g, color, #11, #5 | |
5865 | ||
e1f6de8f | 5866 | vld1.u32 { test_mask }, [psx_gpu, :128] |
f0931e56 | 5867 | ubfx color_b, color, #19, #5 |
5868 | ||
5869 | vdup.u16 right_mask, right_mask_bits | |
5870 | orr color, color_r, color_b, lsl #10 | |
5871 | ||
e1f6de8f | 5872 | ldrh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
f0931e56 | 5873 | orr color, color, color_g, lsl #5 |
5874 | ||
5875 | vtst.u16 right_mask, right_mask, test_mask | |
5876 | add block, psx_gpu, #psx_gpu_blocks_offset | |
5877 | ||
5878 | vdup.u16 colors, color | |
5879 | add block, block, num_blocks, lsl #6 | |
5880 | ||
5881 | ||
5882 | setup_sprite_untextured_height_loop: | |
5883 | add num_blocks, block_width | |
5884 | sub blocks_remaining, block_width, #1 | |
5885 | ||
5886 | cmp num_blocks, #MAX_BLOCKS | |
5887 | blgt setup_sprites_16bpp_flush | |
5888 | ||
5889 | cmp blocks_remaining, #0 | |
5890 | ble 1f | |
5891 | ||
5892 | vmov.u8 draw_mask, #0 /* zero_mask */ | |
5893 | vmov.u8 draw_mask_bits_fb_ptr, #0 | |
5894 | ||
5895 | 0: | |
e1f6de8f | 5896 | vst1.u32 { draw_mask }, [block, :128]! |
f0931e56 | 5897 | subs blocks_remaining, #1 |
5898 | ||
e1f6de8f | 5899 | vst1.u32 { colors }, [block, :128] |
f0931e56 | 5900 | add block, block, #24 |
5901 | ||
5902 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
e1f6de8f | 5903 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
f0931e56 | 5904 | |
5905 | add block, block, #24 | |
5906 | add fb_ptr, #8*2 | |
5907 | bgt 0b | |
5908 | ||
5909 | 1: | |
e1f6de8f | 5910 | vst1.u32 { right_mask }, [block, :128]! |
f0931e56 | 5911 | subs height, #1 |
5912 | ||
e1f6de8f | 5913 | vst1.u32 { colors }, [block, :128] |
f0931e56 | 5914 | add block, block, #24 |
5915 | ||
5916 | vmov.u32 draw_mask_bits_fb_ptr[1], fb_ptr | |
e1f6de8f | 5917 | vst1.u32 { draw_mask_bits_fb_ptr }, [block, :64] |
f0931e56 | 5918 | |
5919 | add block, block, #24 | |
5920 | add fb_ptr, fb_ptr_pitch | |
5921 | ||
e1f6de8f | 5922 | strh num_blocks, [psx_gpu, #psx_gpu_num_blocks_offset] |
f0931e56 | 5923 | bgt setup_sprite_untextured_height_loop |
5924 | ||
5925 | ldmia sp!, { r4 - r11, pc } | |
5926 | ||
5927 | ||
5928 | ||
75e28f62 E |
5929 | #undef texture_page_ptr |
5930 | #undef vram_ptr | |
5931 | #undef dirty_textures_mask | |
5932 | #undef current_texture_mask | |
5933 | ||
5934 | #define psx_gpu r0 | |
5935 | #define current_texture_page r1 | |
5936 | #define texture_page_ptr r2 | |
5937 | #define vram_ptr_a r3 | |
5938 | #define current_texture_page_x r12 | |
5939 | #define current_texture_page_y r4 | |
5940 | #define dirty_textures_mask r5 | |
5941 | #define tile_y r6 | |
5942 | #define tile_x r7 | |
5943 | #define sub_y r8 | |
5944 | #define current_texture_mask r9 | |
5945 | #define c_4096 r10 | |
5946 | #define vram_ptr_b r11 | |
5947 | ||
5948 | #define texel_block_a d0 | |
5949 | #define texel_block_b d1 | |
5950 | #define texel_block_expanded_a q1 | |
5951 | #define texel_block_expanded_b q2 | |
5952 | #define texel_block_expanded_ab q2 | |
5953 | #define texel_block_expanded_c q3 | |
7255f1c2 | 5954 | #define texel_block_expanded_d q0 |
75e28f62 E |
5955 | #define texel_block_expanded_cd q3 |
5956 | ||
5957 | function(update_texture_4bpp_cache) | |
5958 | stmdb sp!, { r4 - r11, r14 } | |
5959 | vpush { q0 - q3 } | |
5960 | ||
e1f6de8f | 5961 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
75e28f62 | 5962 | |
e1f6de8f | 5963 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
5964 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
75e28f62 E |
5965 | |
5966 | and current_texture_page_x, current_texture_page, #0xF | |
e1f6de8f | 5967 | ldr current_texture_mask, [psx_gpu, #psx_gpu_current_texture_mask_offset] |
75e28f62 E |
5968 | |
5969 | mov current_texture_page_y, current_texture_page, lsr #4 | |
e1f6de8f | 5970 | ldr dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
75e28f62 E |
5971 | |
5972 | add vram_ptr_a, vram_ptr_a, current_texture_page_y, lsl #19 | |
5973 | mov tile_y, #16 | |
5974 | ||
5975 | add vram_ptr_a, vram_ptr_a, current_texture_page_x, lsl #7 | |
5976 | bic dirty_textures_mask, current_texture_mask | |
5977 | ||
5978 | mov tile_x, #16 | |
e1f6de8f | 5979 | str dirty_textures_mask, [psx_gpu, #psx_gpu_dirty_textures_4bpp_mask_offset] |
75e28f62 E |
5980 | |
5981 | mov sub_y, #8 | |
5982 | movw c_4096, #4096 | |
5983 | ||
5984 | add vram_ptr_b, vram_ptr_a, #2048 | |
5985 | ||
5986 | 0: | |
e1f6de8f | 5987 | vld1.u32 { texel_block_a }, [vram_ptr_a, :64], c_4096 |
5988 | vld1.u32 { texel_block_b }, [vram_ptr_b, :64], c_4096 | |
75e28f62 E |
5989 | |
5990 | vmovl.u8 texel_block_expanded_a, texel_block_a | |
5991 | vshll.u8 texel_block_expanded_b, texel_block_a, #4 | |
5992 | vmovl.u8 texel_block_expanded_c, texel_block_b | |
5993 | vshll.u8 texel_block_expanded_d, texel_block_b, #4 | |
5994 | ||
5995 | vbic.u16 texel_block_expanded_a, #0x00F0 | |
5996 | vbic.u16 texel_block_expanded_b, #0x00F0 | |
5997 | vbic.u16 texel_block_expanded_c, #0x00F0 | |
5998 | vbic.u16 texel_block_expanded_d, #0x00F0 | |
5999 | ||
6000 | vorr.u16 texel_block_expanded_ab, texel_block_expanded_a, \ | |
6001 | texel_block_expanded_b | |
6002 | vorr.u16 texel_block_expanded_cd, texel_block_expanded_c, \ | |
6003 | texel_block_expanded_d | |
6004 | ||
6005 | vst1.u32 { texel_block_expanded_ab, texel_block_expanded_cd }, \ | |
e1f6de8f | 6006 | [texture_page_ptr, :256]! |
75e28f62 E |
6007 | |
6008 | subs sub_y, sub_y, #1 | |
6009 | bne 0b | |
6010 | ||
6011 | mov sub_y, #8 | |
6012 | add vram_ptr_a, vram_ptr_a, #8 | |
6013 | add vram_ptr_b, vram_ptr_b, #8 | |
6014 | ||
6015 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6016 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6017 | ||
6018 | subs tile_x, tile_x, #1 | |
6019 | bne 0b | |
6020 | ||
6021 | mov tile_x, #16 | |
6022 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6023 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6024 | ||
6025 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
6026 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
6027 | ||
6028 | subs tile_y, tile_y, #1 | |
6029 | bne 0b | |
6030 | ||
6031 | vpop { q0 - q3 } | |
6032 | ldmia sp!, { r4 - r11, pc } | |
6033 | ||
6034 | ||
6035 | #undef current_texture_page | |
6036 | ||
6037 | #define psx_gpu r0 | |
6038 | #define texture_page r1 | |
6039 | #define texture_page_ptr r2 | |
6040 | #define vram_ptr_a r3 | |
6041 | #define texture_page_x r12 | |
6042 | #define texture_page_y r4 | |
6043 | #define current_texture_page r5 | |
6044 | #define tile_y r6 | |
6045 | #define tile_x r7 | |
6046 | #define sub_y r8 | |
6047 | #define c_4096 r10 | |
6048 | #define vram_ptr_b r11 | |
6049 | ||
6050 | ||
6051 | #undef texels_a | |
6052 | #undef texels_b | |
6053 | ||
6054 | #define texels_a q0 | |
6055 | #define texels_b q1 | |
6056 | #define texels_c q2 | |
6057 | #define texels_d q3 | |
6058 | ||
6059 | ||
6060 | function(update_texture_8bpp_cache_slice) | |
6061 | stmdb sp!, { r4 - r11, r14 } | |
6062 | vpush { q0 - q3 } | |
6063 | ||
e1f6de8f | 6064 | ldrb current_texture_page, [psx_gpu, #psx_gpu_current_texture_page_offset] |
6065 | ldr vram_ptr_a, [psx_gpu, #psx_gpu_vram_ptr_offset] | |
75e28f62 | 6066 | |
e1f6de8f | 6067 | ldr texture_page_ptr, [psx_gpu, #psx_gpu_texture_page_base_offset] |
75e28f62 E |
6068 | mov tile_y, #16 |
6069 | ||
6070 | and texture_page_x, texture_page, #0xF | |
6071 | mov texture_page_y, texture_page, lsr #4 | |
6072 | ||
6073 | add vram_ptr_a, vram_ptr_a, texture_page_x, lsl #7 | |
6074 | mov tile_x, #8 | |
6075 | ||
6076 | add vram_ptr_a, vram_ptr_a, texture_page_y, lsl #19 | |
6077 | eor current_texture_page, current_texture_page, texture_page | |
6078 | ||
6079 | ands current_texture_page, current_texture_page, #0x1 | |
6080 | mov sub_y, #4 | |
6081 | ||
6082 | addne texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
6083 | movw c_4096, #4096 | |
6084 | ||
6085 | add vram_ptr_b, vram_ptr_a, #2048 | |
6086 | ||
6087 | 0: | |
e1f6de8f | 6088 | vld1.u32 { texels_a }, [vram_ptr_a, :128], c_4096 |
6089 | vld1.u32 { texels_b }, [vram_ptr_b, :128], c_4096 | |
6090 | vld1.u32 { texels_c }, [vram_ptr_a, :128], c_4096 | |
6091 | vld1.u32 { texels_d }, [vram_ptr_b, :128], c_4096 | |
75e28f62 | 6092 | |
e1f6de8f | 6093 | vst1.u32 { texels_a, texels_b }, [texture_page_ptr, :256]! |
6094 | vst1.u32 { texels_c, texels_d }, [texture_page_ptr, :256]! | |
75e28f62 E |
6095 | |
6096 | subs sub_y, sub_y, #1 | |
6097 | bne 0b | |
6098 | ||
6099 | mov sub_y, #4 | |
6100 | ||
6101 | add vram_ptr_a, vram_ptr_a, #16 | |
6102 | add vram_ptr_b, vram_ptr_b, #16 | |
6103 | ||
6104 | sub vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6105 | sub vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6106 | ||
6107 | subs tile_x, tile_x, #1 | |
6108 | bne 0b | |
6109 | ||
6110 | mov tile_x, #8 | |
6111 | ||
6112 | add vram_ptr_a, vram_ptr_a, #(16 * 2048) | |
6113 | add vram_ptr_b, vram_ptr_b, #(16 * 2048) | |
6114 | ||
6115 | sub vram_ptr_a, vram_ptr_a, #(8 * 16) | |
6116 | sub vram_ptr_b, vram_ptr_b, #(8 * 16) | |
6117 | ||
6118 | subs tile_y, tile_y, #1 | |
6119 | add texture_page_ptr, texture_page_ptr, #(8 * 16 * 16) | |
6120 | ||
6121 | bne 0b | |
6122 | ||
6123 | vpop { q0 - q3 } | |
6124 | ldmia sp!, { r4 - r11, pc } | |
6125 | ||
50f9355a | 6126 | |
6127 | /* void scale2x_tiles8(void *dst, const void *src, int w8, int h) */ | |
6128 | function(scale2x_tiles8) | |
6129 | push { r4, r14 } | |
6130 | ||
6131 | mov r4, r1 | |
6132 | add r12, r0, #1024*2 | |
6133 | mov r14, r2 | |
6134 | ||
6135 | 0: | |
e1f6de8f | 6136 | vld1.u16 { q0 }, [r1, :128]! |
6137 | vld1.u16 { q2 }, [r1, :128]! | |
50f9355a | 6138 | vmov q1, q0 |
6139 | vmov q3, q2 | |
6140 | vzip.16 q0, q1 | |
6141 | vzip.16 q2, q3 | |
6142 | subs r14, #2 | |
e1f6de8f | 6143 | vst1.u16 { q0, q1 }, [r0, :128]! |
6144 | vst1.u16 { q0, q1 }, [r12, :128]! | |
50f9355a | 6145 | blt 1f |
e1f6de8f | 6146 | vst1.u16 { q2, q3 }, [r0, :128]! |
6147 | vst1.u16 { q2, q3 }, [r12, :128]! | |
50f9355a | 6148 | bgt 0b |
6149 | 1: | |
6150 | subs r3, #1 | |
6151 | mov r14, r2 | |
6152 | add r0, #1024*2*2 | |
6153 | add r4, #1024*2 | |
ed0fd81d | 6154 | sub r0, r0, r2, lsl #4+1 |
50f9355a | 6155 | mov r1, r4 |
6156 | add r12, r0, #1024*2 | |
6157 | bgt 0b | |
6158 | nop | |
6159 | ||
6160 | pop { r4, pc } | |
59d15d23 | 6161 | |
6162 | // vim:filetype=armasm |