8cfbda97 |
1 | /* |
2 | * (C) GraÅžvydas "notaz" Ignotas, 2011 |
3 | * |
4 | * This work is licensed under the terms of any of these licenses |
5 | * (at your option): |
6 | * - GNU GPL, version 2 or later. |
7 | * - GNU LGPL, version 2.1 or later. |
8 | * See the COPYING file in the top-level directory. |
9 | */ |
10 | |
11 | |
12 | .bss |
13 | .align 6 @ cacheline |
14 | |
15 | scratch: |
16 | .rept 8*8*2/4 |
17 | .word 0 |
18 | .endr |
19 | |
20 | .text |
21 | .align 2 |
22 | |
5d8e3bf8 |
23 | @ approximate gteMAC|123 flags |
24 | @ in: rr 123 as gteMAC|123 |
25 | @ trash: nothing |
26 | .macro do_mac_flags rr1 rr2 rr3 |
27 | cmp \rr1, #1 |
28 | orrvs lr, #(1<<31)|(1<<27) |
29 | cmp \rr2, #1 |
30 | orrvs lr, #(1<<31)|(1<<26) |
31 | cmp \rr3, #1 |
32 | orrvs lr, #(1<<31)|(1<<25) |
33 | cmn \rr1, #1 @ same as adds ... |
34 | orrvs lr, #(1<<30) |
35 | cmn \rr2, #1 |
36 | orrvs lr, #(1<<29) |
37 | cmn \rr3, #1 |
38 | orrvs lr, #(1<<28) |
39 | .endm |
40 | |
41 | @ approximate 3x gteMACn flags |
42 | @ in: rr 123 as 3 instances gteMACn, *flags |
43 | @ trash: nothing |
44 | .macro do_mac_flags3x rr1 rr2 rr3 nflags pflags |
45 | cmp \rr1, #1 |
46 | cmpvc \rr2, #1 |
47 | cmpvc \rr3, #1 |
48 | orrvs lr, #\nflags |
49 | cmn \rr1, #1 @ adds ... |
50 | cmnvc \rr2, #1 |
51 | cmnvc \rr3, #1 |
52 | orrvs lr, #\pflags |
53 | .endm |
54 | |
17ed0d69 |
55 | @ get gteIR|123 flags from gteMAC|123 |
56 | @ in: rr 123 as gteMAC|123 |
5d8e3bf8 |
57 | @ trash: r2,r3 |
58 | .macro do_irs_flags rr1 rr2 rr3 |
17ed0d69 |
59 | add r2, \rr1, #0x8000 |
60 | add r3, \rr2, #0x8000 |
61 | lsrs r2, #16 |
62 | orrne lr, #(1<<31)|(1<<24) @ IR1/limB1 |
63 | lsrs r3, #16 |
64 | add r2, \rr3, #0x8000 |
65 | orrne lr, #(1<<31) |
66 | orrne lr, #(1<<23) @ IR2/limB2 |
67 | lsrs r2, #16 |
68 | orrne lr, #(1<<22) @ IR3/limB3 |
8cfbda97 |
69 | .endm |
70 | |
71 | |
72 | /* |
5d8e3bf8 |
73 | * RTPS/RTPT register map: |
74 | * |
8cfbda97 |
75 | * q | d | c code / phase 1 phase 2 scratch |
76 | * 0 0 gteR1* [s16] gteMAC3 = gteMAC3 \ v=0 * |
77 | * 1 gteR2* gteIR1-3 = gteIR1-3 / * |
78 | * 1 2 gteR3* gteMAC3 = gteMAC3 \ v=1 |
79 | * 3 * gteIR1-3 = gteIR1-3 / |
80 | * 2 4 gteTRX<<12 [s64] gteOFX [s64] gteMAC3 \ v=2 |
81 | * 5 gteTRY<<12 gteOFY [s64] gteIR1-3 / |
82 | * 3 6 gteTRZ<<12 gteDQA [s64] min gteMAC|12 v=012 |
83 | * 7 0 gteDQB [s64] max gteMAC|12 |
84 | * 4 8 VXYZ(v) / gteMAC1,2 [s32] min gteIR|123 |
85 | * 9 * / gteMAC3 max gteIR|123 |
86 | * 5 10 gteIR1-3 [s16] gteIR1-3 v=2 quotients 12 |
87 | * 11 0 quotient 3 |
88 | * 6 12 gteH (adj. for cmp) |
89 | * 13 gteH (float for div) |
90 | * ... <scratch> |
91 | * 15 30 0 |
92 | * 31 0 |
93 | */ |
8cfbda97 |
94 | |
5d8e3bf8 |
95 | @ load gteR*, gteTR* and gteH (see map above), clear q15 |
96 | @ in: r0 - context |
97 | @ trash: r3 |
98 | .macro rtpx_preload |
8cfbda97 |
99 | add r3, r0, #4*32 |
100 | vldmia r3, {d0-d2} @ gteR* [16*9] |
4cc3050c |
101 | vmov.i32 q15, #0 |
8cfbda97 |
102 | add r3, r0, #4*(32+5) |
103 | vldmia r3, {d4-d5} @ gteTR* |
4cc3050c |
104 | vext.16 d2, d1, d2, #2 @ xxx3 -> x321 |
105 | vext.16 d1, d0, d1, #3 @ xx32 -> x321 |
8cfbda97 |
106 | add r3, r0, #4*(32+26) |
107 | vld1.32 d11[0], [r3] @ gteH |
5d8e3bf8 |
108 | vshll.s32 q3, d5, #12 @ gteTRZ |
109 | vshll.s32 q2, d4, #12 @ gteTR|XY |
8cfbda97 |
110 | vmovl.s16 q6, d11 @ gteH |
5d8e3bf8 |
111 | .endm |
8cfbda97 |
112 | |
5d8e3bf8 |
113 | @ do RTP* gteMAC* calculation |
114 | @ in: gteR*, gteTR* as in map, d8 - VXYZ, r12 - 0 |
115 | @ out: d8,d9 - gteMAC|123, d10 - gteIR|123 |
116 | @ trash: d16-d21 |
117 | .macro rtpx_mac |
8cfbda97 |
118 | vmull.s16 q8, d0, d8 |
119 | vmull.s16 q9, d1, d8 |
120 | vmull.s16 q10, d2, d8 |
121 | vpaddl.s32 q8, q8 |
122 | vpaddl.s32 q9, q9 |
123 | vpaddl.s32 q10, q10 |
124 | vadd.s64 d16, d17 @ d16=d0.16[2]*d8.16[2], as |
125 | vadd.s64 d18, d19 @ d8[3]==0, so won't affect |
126 | vadd.s64 d20, d21 @ QC |
127 | vadd.s64 d16, d4 |
128 | vadd.s64 d18, d5 |
129 | vadd.s64 d20, d6 |
130 | vqshrn.s64 d8, q8, #12 @ gteMAC1 |
131 | vqshrn.s64 d18, q9, #12 @ gteMAC2 |
132 | vqshrn.s64 d9, q10, #12 @ gteMAC3 |
133 | vsli.u64 d8, d18, #32 @ gteMAC|12 |
134 | vmov.32 d9[1], r12 |
5d8e3bf8 |
135 | vqmovn.s32 d10, q4 @ gteIR|123; losing 2 cycles? |
136 | .endm |
137 | |
138 | .global gteRTPS_neon @ r0=CP2 (d,c), |
139 | gteRTPS_neon: |
17ed0d69 |
140 | push {r4-r6,lr} |
5d8e3bf8 |
141 | |
142 | @ fmrx r4, fpscr @ vmrs? at least 40 cycle hit |
143 | movw r1, #:lower16:scratch |
144 | movt r1, #:upper16:scratch |
145 | mov r12, #0 |
146 | |
4cc3050c |
147 | vldmia r0, {d8} @ VXYZ(0) |
5d8e3bf8 |
148 | rtpx_preload |
149 | |
4cc3050c |
150 | @ rtpx_mac @ slower here, faster in RTPT? |
5d8e3bf8 |
151 | vmov.16 d8[3], r12 @ kill unused upper vector |
4cc3050c |
152 | vmull.s16 q8, d0, d8 |
153 | vmull.s16 q9, d1, d8 |
154 | vmull.s16 q10, d2, d8 |
155 | vpadd.s32 d16, d16, d17 |
156 | vpadd.s32 d17, d18, d19 |
157 | vpadd.s32 d18, d20, d21 |
158 | vpadal.s32 q2, q8 |
159 | vpadal.s32 q3, q9 @ d6, d18 is slow? |
160 | vqshrn.s64 d8, q2, #12 @ gteMAC|12 |
161 | vqshrn.s64 d9, q3, #12 @ gteMAC3 |
5d8e3bf8 |
162 | |
163 | add r3, r0, #4*25 |
164 | vst1.32 d8, [r3]! |
165 | vst1.32 d9[0], [r3] @ wb gteMAC|123 |
4cc3050c |
166 | vqmovn.s32 d10, q4 @ gteIR|123 |
5d8e3bf8 |
167 | |
168 | add r3, r0, #4*17 @ gteSZ* |
169 | vldmia r3, {q7} @ d14,d15 gteSZ|123x |
170 | vmov.i32 d28, #0xffff @ 0xffff[32] |
171 | vmax.s32 d11, d9, d31 |
172 | vshr.s32 d16, d12, #1 @ | gteH/2 (adjust for cmp) |
173 | vmov.i32 d26, #1 |
174 | vmin.u32 d11, d28 @ saturate to 0..0xffff limD/fSZ3 |
4cc3050c |
175 | vmovl.s16 q9, d10 @ || expand gteIR|123 |
5d8e3bf8 |
176 | vshl.u32 d13, d12, #16 @ | preparing gteH |
177 | add r3, r0, #4*9 |
178 | vst1.32 d18, [r3]! |
179 | vst1.32 d19[0], [r3] |
180 | |
181 | vsli.u64 d15, d11, #32 @ new gteSZ|0123 in q7 |
182 | vclt.u32 d16, d16, d11 @ gteH/2 < fSZ3? |
183 | |
184 | add r3, r0, #4*(32+24) |
185 | vld1.32 d4, [r3] @ || gteOF|XY |
186 | add r3, r0, #4*(32+27) |
187 | vld1.32 d6, [r3] @ || gteDQ|AB |
188 | |
189 | vand d11, d16 |
190 | vmovl.s32 q2, d4 @ || gteOF|XY [64] |
191 | vmax.u32 d11, d26 @ make divisor 1 if not |
192 | vmovl.s32 q3, d6 @ || gteDQ|AB [64] |
193 | add r3, r0, #4*16 @ | gteSZ* |
194 | vstmia r3, {q7} @ | d14,d15 gteSZ|123x |
195 | |
196 | vcvt.f32.u32 d13, d13 @ gteH (float for div) |
197 | vcvt.f32.u32 d11, d11 @ divisor |
198 | |
199 | @ divide.. it's not worth messing with reciprocals here |
200 | @ just for 1 value, let's just use VFP divider here |
201 | vdiv.f32 s22, s26, s22 |
202 | |
203 | vcvt.u32.f32 d11, d11 @ quotient |
204 | |
205 | @ while NEON's busy we calculate some flags on ARM |
206 | add r3, r0, #4*25 |
207 | mov lr, #0 @ gteFLAG |
208 | ldmia r3, {r4-r6} @ gteMAC|123 |
209 | |
4cc3050c |
210 | vst1.32 d11, [r1, :64] @ wb quotient for flags (pre-limE) |
5d8e3bf8 |
211 | vqshl.u32 d11, #15 |
212 | |
213 | do_mac_flags r4, r5, r6 |
214 | |
215 | vshr.u32 d11, #15 @ quotient (limE) |
5d8e3bf8 |
216 | |
5d8e3bf8 |
217 | do_irs_flags r4, r5, r6 |
218 | |
219 | vmlal.s32 q2, d18, d11[0]@ gteOF|XY + gteIR|12 * quotient |
220 | add r3, r0, #4*13 |
221 | vld1.32 d16, [r3] @ || load fS|XY12, new 01 |
222 | vqmovn.s64 d18, q2 @ saturate to 32 |
223 | vmull.s32 q10, d6, d11[0]@ | d20 = gteDQA * quotient |
224 | vqshl.s32 d19, d18, #5 @ 11bit precision |
225 | |
226 | ldr r4, [r1] @ quotient |
5d8e3bf8 |
227 | movs r3, r6, lsr #16 |
228 | orrne lr, #(1<<31) |
229 | orrne lr, #(1<<18) @ fSZ (limD) |
230 | |
17ed0d69 |
231 | vst1.32 d18, [r1, :64] @ writeback fS|XY2 before limG |
232 | |
5d8e3bf8 |
233 | vshr.s32 d18, d19, #16+5@ can't vqshrn because of insn |
234 | vadd.s64 d20, d7 @ | gteDQB + gteDQA * quotient |
235 | vmovn.s32 d18, q9 @ fS|XY2 [s16] |
236 | |
237 | vqshrn.s64 d20, q10, #12 @ | gteMAC0 |
238 | add r3, r0, #4*12 |
239 | vst1.32 d16, [r3]! @ writeback fS|XY01 |
240 | vst1.32 d18[0], [r3] @ ...2 |
241 | add r3, r0, #4*24 |
242 | vst1.32 d20[0], [r3] @ gteMAC0 |
243 | |
17ed0d69 |
244 | movs r4, r4, lsr #17 |
245 | orrne lr, #(1<<31) |
246 | orrne lr, #(1<<17) @ limE |
247 | |
5d8e3bf8 |
248 | vmax.s32 d21, d20, d31 |
249 | vmov.i32 d22, #0x1000 |
250 | vmin.s32 d21, d22 |
251 | add r3, r0, #4*8 |
252 | vst1.16 d21[0], [r3] @ gteIR0 |
253 | |
17ed0d69 |
254 | ldmia r1, {r4,r5} @ fS|XY2 before limG, after 11bit sat |
255 | add r2, r4, #0x400<<16 |
256 | add r3, r5, #0x400<<16 |
257 | lsrs r2, #16+11 |
258 | orrne lr, #(1<<14) @ limG1 |
259 | orrne lr, #(1<<31) |
260 | lsrs r3, #16+11 |
261 | orrne lr, #(1<<13) @ limG2 |
5d8e3bf8 |
262 | orrne lr, #(1<<31) |
5d8e3bf8 |
263 | adds r2, r4, #1 |
264 | addvcs r3, r5, #1 |
265 | orrvs lr, #(1<<16) @ F |
266 | orrvs lr, #(1<<31) |
267 | subs r2, r4, #1 |
268 | subvcs r3, r5, #1 |
269 | orrvs lr, #(1<<31) |
270 | |
271 | ldr r4, [r0, #4*24] @ gteMAC0 |
272 | orrvs lr, #(1<<15) |
273 | |
274 | adds r3, r4, #1 |
275 | orrvs lr, #(1<<16) @ F |
276 | orrvs lr, #(1<<31) |
277 | subs r2, r4, #1 |
278 | orrvs lr, #(1<<15) @ F |
279 | orrvs lr, #(1<<31) |
280 | cmp r4, #0x1000 |
281 | orrhi lr, #(1<<12) @ limH |
282 | |
283 | str lr, [r0, #4*(32+31)] @ gteFLAG |
284 | |
17ed0d69 |
285 | pop {r4-r6,pc} |
5d8e3bf8 |
286 | .size gteRTPS_neon, .-gteRTPS_neon |
287 | |
288 | |
289 | |
290 | .global gteRTPT_neon @ r0=CP2 (d,c), |
291 | gteRTPT_neon: |
292 | push {r4-r11,lr} |
293 | |
294 | movw r1, #:lower16:scratch |
295 | movt r1, #:upper16:scratch |
296 | mov r12, #0 |
297 | |
298 | rtpx_preload |
299 | |
300 | vmov.i32 d22, #0x7fffffff |
301 | vmov.i32 d23, #0x80000000 |
302 | mov r3, #3 @ counter |
303 | mov r2, r0 @ VXYZ(0) |
304 | 0: |
305 | vldmia r2!, {d8} @ VXYZ(v) |
306 | vmov.16 d8[3], r12 @ kill unused upper vector |
307 | |
308 | rtpx_mac |
8cfbda97 |
309 | vmin.s32 d22, d8 @ min gteMAC|12 |
310 | vmax.s32 d23, d8 @ max gteMAC|12 |
311 | subs r3, #1 |
4cc3050c |
312 | vst1.32 {d9,d10}, [r1, :128]! |
8cfbda97 |
313 | bgt 0b |
314 | |
4cc3050c |
315 | vst1.32 {d22,d23}, [r1, :128]! @ min/max gteMAC|12, for flags |
8cfbda97 |
316 | |
317 | @ - phase2 - |
318 | sub r1, r1, #8*2*4 |
319 | vldmia r1, {d0-d3} @ note: d4,d5 is for gteOF|XY |
320 | |
321 | vmov d20, d0 @ gteMAC3 v=0 |
322 | vmin.s16 d24, d1, d3 @ | find min IR |
5d8e3bf8 |
323 | vshr.s32 d22, d12, #1 @ || gteH/2 (adjust for cmp) |
8cfbda97 |
324 | vmax.s16 d25, d1, d3 @ | .. also max, for flag gen |
325 | vsli.u64 d20, d2, #32 @ gteMAC3 v=1 |
326 | vmov d21, d9 @ ... v=2 |
327 | |
328 | vmov.i32 q14, #0xffff @ 0xffff[32] |
329 | vmax.s32 q10, q15 |
330 | vmov.i32 q13, #1 |
331 | vdup.32 q11, d22[0] @ gteH/2 |
5d8e3bf8 |
332 | vmin.u32 q10, q14 @ saturate to 0..0xffff limD/fSZ(v) |
8cfbda97 |
333 | vmin.s16 d24, d10 @ | find min/max IR |
334 | vmax.s16 d25, d10 @ | |
335 | |
5d8e3bf8 |
336 | add r3, r0, #4*19 @ || |
337 | vld1.32 d14[0], [r3] @ || gteSZ3 |
338 | |
8cfbda97 |
339 | vclt.u32 q11, q11, q10 @ gteH/2 < fSZ(v)? |
340 | add r3, r0, #4*17 |
341 | vst1.32 d20, [r3]! @ | writeback fSZ(v) |
342 | vand q11, q10, q11 |
343 | vst1.32 d21[0], [r3] @ | |
344 | vmax.u32 q10, q11, q13 @ make divisor 1 if not |
345 | add r3, r1, #8*8 |
346 | vstmia r3, {q12} @ min/max IR for flags |
347 | vcvt.f32.u32 q10, q10 |
348 | vshl.u32 d13, d12, #16 @ | preparing gteH |
349 | |
350 | @ while NEON's busy we calculate some flags on ARM |
351 | add r2, r1, #8*2*3 |
352 | mov lr, #0 @ gteFLAG |
353 | ldmia r2, {r4-r7} @ min/max gteMAC|12 |
354 | subs r2, r4, #1 |
355 | orrvs lr, #(1<<31)|(1<<27) |
356 | subs r3, r5, #1 |
357 | orrvs lr, #(1<<31)|(1<<26) |
358 | adds r2, r6, #1 |
359 | orrvs lr, #(1<<30) |
360 | adds r3, r7, #1 |
361 | orrvs lr, #(1<<29) |
362 | ldr r4, [r1, #0] @ gteMAC3 v=0 |
363 | ldr r5, [r1, #8*2] @ ... v=1 |
364 | ldr r6, [r1, #8*4] @ ... v=2 |
365 | |
366 | add r3, r0, #4*(32+24) |
367 | vld1.32 d4, [r3] @ || gteOF|XY |
368 | add r3, r0, #4*(32+27) |
5d8e3bf8 |
369 | vld1.32 d6, [r3] @ || gteDQ|AB |
8cfbda97 |
370 | |
371 | @ divide |
372 | .if 1 |
373 | vrecpe.f32 q11, q10 @ inv |
5d8e3bf8 |
374 | vmovl.s32 q2, d4 @ || gteOF|XY [64] |
375 | vmovl.s32 q3, d6 @ || gteDQ|AB [64] |
8cfbda97 |
376 | vrecps.f32 q12, q10, q11 @ step |
377 | vcvt.f32.u32 d13, d13 @ | gteH (float for div) |
378 | vmul.f32 q11, q12, q11 @ better inv |
5d8e3bf8 |
379 | add r3, r0, #4*16 |
380 | vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3 |
8cfbda97 |
381 | vdup.32 q13, d13[0] @ | |
382 | @ vrecps.f32 q12, q10, q11 @ step |
383 | @ vmul.f32 q11, q12, q11 @ better inv |
384 | vmul.f32 q10, q13, q11 @ result |
385 | .else |
5d8e3bf8 |
386 | vmovl.s32 q2, d4 @ || gteOF|XY [64] |
387 | vmovl.s32 q3, d6 @ || gteDQ|AB [64] |
8cfbda97 |
388 | vcvt.f32.u32 d13, d13 @ | gteH (float for div) |
389 | vdup.32 q13, d13[0] @ | |
5d8e3bf8 |
390 | add r3, r0, #4*16 |
391 | vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3 |
8cfbda97 |
392 | |
393 | vpush {q0} |
394 | vmov q0, q10 @ to test against C code |
395 | vdiv.f32 s0, s26, s0 |
396 | vdiv.f32 s1, s26, s1 |
397 | vdiv.f32 s2, s26, s2 |
398 | vmov q10, q0 |
399 | vpop {q0} |
400 | .endif |
401 | |
5d8e3bf8 |
402 | do_mac_flags3x r4, r5, r6, (1<<31)|(1<<25), (1<<27) @ MAC3 |
8cfbda97 |
403 | orr r7, r4, r5 |
404 | add r4, r1, #8*8 |
405 | orr r3, r7, r6 |
406 | ldmia r4, {r7,r8,r10,r11} @ min/max IR |
407 | |
408 | movs r3, r3, lsr #16 |
409 | orrne lr, #(1<<31) |
410 | orrne lr, #(1<<18) @ fSZ (limD) |
411 | |
412 | @ vadd.f32 q10, q @ adjust for vcvt rounding mode |
413 | vcvt.u32.f32 q8, q10 |
414 | vmovl.s16 q9, d1 @ expand gteIR|12 v=0 |
415 | vmovl.s16 q10, d3 @ expand gteIR|12 v=1 |
416 | add r6, r1, #8*10 |
417 | vstmia r6, {q8} @ wb quotients for flags (pre-limE) |
418 | vqshl.u32 q8, #15 |
419 | vmovl.s16 q11, d10 @ expand gteIR|12 v=2 |
420 | vshr.u32 q8, #15 @ quotients (limE) |
421 | vdup.32 d24, d16[0] |
422 | vdup.32 d25, d16[1] |
423 | vdup.32 d26, d17[0] @ quotient (dup) |
424 | |
5d8e3bf8 |
425 | @ flags for minIR012 (r7,r8), maxIR012 (r10,r11) |
426 | mov r4, #0x10000 |
427 | cmp r7, #1<<16 |
428 | cmnvc r10, #1<<16 |
8cfbda97 |
429 | orrvs lr, #(1<<31) |
430 | orrvs lr, #(1<<23) @ IR2/limB2 |
5d8e3bf8 |
431 | rsbs r2, r4, r7, lsl #16 |
432 | cmnvc r4, r10, lsl #16 |
8cfbda97 |
433 | orrvs lr, #(1<<31)|(1<<24) @ IR1/limB1 |
5d8e3bf8 |
434 | rsbs r2, r4, r8, lsl #16 |
435 | cmnvc r4, r11, lsl #16 |
8cfbda97 |
436 | orrvs lr, #(1<<22) @ IR3/limB3 |
437 | |
438 | vmull.s32 q9, d18, d24 @ gteIR|12 * quotient v=0 |
439 | vmull.s32 q10, d20, d25 @ ... v=1 |
440 | vmull.s32 q11, d22, d26 @ ... v=2 |
441 | vadd.s64 q9, q2 @ gteOF|XY + gteIR|12 * quotient |
442 | vadd.s64 q10, q2 @ ... v=1 |
443 | vadd.s64 q11, q2 @ ... v=2 |
444 | vqmovn.s64 d18, q9 @ saturate to 32 v=0 |
445 | vqmovn.s64 d19, q10 @ ... v=1 |
446 | vqmovn.s64 d20, q11 @ ... v=2 |
447 | vmin.s32 d14, d18, d19 @ || find min/max fS|XY(v) [32] |
448 | vmax.s32 d15, d18, d19 @ || for flags |
449 | vmin.s32 d14, d20 |
450 | vmax.s32 d15, d20 |
451 | vqshl.s32 q11, q9, #5 @ 11bit precision, v=0,1 |
452 | vqshl.s32 d24, d20, #5 @ ... v=2 |
453 | vmull.s32 q13, d6, d17 @ | gteDQA * quotient v=2 |
5d8e3bf8 |
454 | vpmin.s32 d16, d14, d31 @ || also find min/max in pair |
455 | vpmax.s32 d17, d15, d31 @ || |
8cfbda97 |
456 | vshr.s32 q11, #16+5 @ can't vqshrn because of insn |
457 | vshr.s32 d24, #16+5 @ encoding doesn't allow 21 :( |
8cfbda97 |
458 | vsli.u64 d16, d17, #32 @ || pack in-pair min/max |
459 | vadd.s64 d26, d7 @ | gteDQB + gteDQA * quotient |
460 | vmovn.s32 d12, q11 @ fS|XY(v) [s16] v=0,1 |
461 | vmovn.s32 d13, q12 @ 3 |
462 | vstmia r1, {d14-d16} @ || other cacheline than quotients |
463 | add r3, r0, #4*12 |
464 | vst1.32 d12, [r3]! @ writeback fS|XY v=0,1 |
465 | vst1.32 d13[0], [r3] |
466 | |
467 | vqshrn.s64 d26, q13, #12 @ | gteMAC0 |
468 | vmovl.u16 q5, d10 @ expand gteIR|123 v=2 |
469 | |
470 | vmov.i32 d13, #0x1000 |
471 | vmax.s32 d12, d26, d30 |
472 | |
473 | add r3, r0, #4*24 |
474 | vst1.32 d26[0], [r3]! @ gteMAC0 |
475 | vst1.32 d8, [r3]! @ gteMAC123 (last iteration) |
476 | vst1.32 d9[0], [r3] |
477 | |
478 | vmin.s32 d12, d13 @ | gteIR0 |
479 | |
8cfbda97 |
480 | ldmia r6, {r4-r6} @ quotients |
481 | orr r4, r5 |
482 | orr r4, r6 |
5d8e3bf8 |
483 | add r3, r0, #4*8 |
8cfbda97 |
484 | movs r4, r4, lsr #17 |
8cfbda97 |
485 | |
8cfbda97 |
486 | vst1.32 d12[0], [r3]! @ gteIR0 |
487 | vst1.32 d10, [r3]! @ gteIR12 |
488 | vst1.32 d11[0], [r3] @ ..3 |
489 | |
17ed0d69 |
490 | @ ~23 cycles |
5d8e3bf8 |
491 | orrne lr, #(1<<31) @ limE |
492 | orrne lr, #(1<<17) @ limE |
8cfbda97 |
493 | ldmia r1, {r4-r9} |
17ed0d69 |
494 | add r2, r4, #0x400<<16 @ min fSX |
495 | add r3, r6, #0x400<<16 @ max fSX |
496 | lsrs r2, #16+11 |
497 | lsreqs r3, #16+11 |
498 | orrne lr, #(1<<31) @ limG1 |
499 | orrne lr, #(1<<14) |
500 | add r2, r5, #0x400<<16 @ min fSY |
501 | add r3, r7, #0x400<<16 @ max fSY |
502 | lsrs r2, #16+11 |
503 | lsreqs r3, #16+11 |
504 | orrne lr, #(1<<31) @ limG2 |
505 | orrne lr, #(1<<13) |
8cfbda97 |
506 | adds r2, r9, #1 |
17ed0d69 |
507 | orrvs lr, #(1<<16) @ F (31 already done by above) |
8cfbda97 |
508 | subs r3, r8, #1 |
8cfbda97 |
509 | |
510 | ldr r4, [r0, #4*24] @ gteMAC0 |
511 | orrvs lr, #(1<<15) |
512 | |
513 | adds r3, r4, #1 |
514 | orrvs lr, #(1<<16) |
515 | orrvs lr, #(1<<31) @ F |
516 | subs r2, r4, #1 |
517 | orrvs lr, #(1<<15) |
518 | orrvs lr, #(1<<31) @ F |
519 | cmp r4, #0x1000 |
5d8e3bf8 |
520 | orrhi lr, #(1<<12) @ limH |
8cfbda97 |
521 | |
522 | str lr, [r0, #4*(32+31)] @ gteFLAG |
523 | |
524 | pop {r4-r11,pc} |
5d8e3bf8 |
525 | .size gteRTPT_neon, .-gteRTPT_neon |
526 | |
527 | |
528 | |
17ed0d69 |
529 | .global gteMVMVA_neon @ r0=CP2 (d,c), op |
530 | gteMVMVA_neon: |
531 | push {r4-r5,lr} |
532 | |
533 | add r12, r0, #4*32 |
534 | |
535 | ubfx r2, r1, #15, #2 @ v |
536 | |
537 | vmov.i32 q0, #0 @ d0,d1 |
538 | vmov.i32 q1, #0 @ d2,d3 |
539 | vmov.i32 q2, #0 @ d4,d5 |
540 | cmp r2, #3 |
541 | addeq r4, r0, #4*9 |
542 | addne r3, r0, r2, lsl #3 |
543 | ldmeqia r4, {r3-r5} |
544 | ldmneia r3, {r4,r5} |
545 | pkhbteq r4, r3, r4, lsl #16 |
546 | uxth r5, r5 |
547 | vmov.32 d8[0], r4 |
548 | vmov.32 d8[1], r5 @ VXYZ(v) |
549 | ubfx r3, r1, #17, #2 @ mx |
550 | ubfx r2, r1, #13, #2 @ cv |
551 | cmp r3, #3 |
552 | beq 0f @ very rare case |
553 | add r3, r12, r3, lsl #5 |
554 | vldmia r3, {d0-d2} @ MXxy/gteR* [16*9] |
555 | 0: |
556 | cmp r2, #3 |
557 | add r3, r12, r2, lsl #5 |
558 | beq 0f |
559 | add r3, #4*5 |
560 | vldmia r3, {d4-d5} @ CVx/gteTR* |
561 | |
562 | 0: |
563 | vmov.i32 q15, #0 |
564 | vext.16 d2, d1, d2, #2 @ xxx3 -> x321 |
565 | vext.16 d1, d0, d1, #3 @ xx32 -> x321 |
566 | vshll.s32 q3, d5, #12 @ gteTRZ/CV3 |
567 | vshll.s32 q2, d4, #12 @ gteTR|XY/CV12 |
568 | |
569 | vmull.s16 q8, d0, d8 |
570 | vmull.s16 q9, d1, d8 |
571 | vmull.s16 q10, d2, d8 |
572 | vpadd.s32 d16, d16, d17 |
573 | vpadd.s32 d17, d18, d19 |
574 | vpadd.s32 d18, d20, d21 |
575 | vpadal.s32 q2, q8 |
576 | vpadal.s32 q3, q9 |
577 | tst r1, #1<<19 |
578 | beq 0f |
579 | vshr.s64 q2, q2, #12 |
580 | vshr.s64 q3, q3, #12 |
581 | 0: |
582 | vqmovn.s64 d8, q2 @ gteMAC|12 |
583 | vqmovn.s64 d9, q3 @ gteMAC3 |
584 | |
585 | tst r1, #1<<10 |
586 | add r3, r0, #4*25 |
587 | vqmovn.s32 d10, q4 @ gteIR|123 |
588 | vst1.32 d8, [r3]! |
589 | vst1.32 d9[0], [r3] @ wb gteMAC|123 |
590 | |
591 | beq 0f |
592 | vmax.s16 d10, d31 |
593 | 0: |
594 | vmovl.s16 q9, d10 @ expand gteIR|123 |
595 | add r3, r0, #4*9 |
596 | vst1.32 d18, [r3]! |
597 | vst1.32 d19[0], [r3] |
598 | |
599 | tst r1, #1<<10 @ lm |
600 | mov r2, #0 |
601 | mov lr, #0 @ gteFLAG |
602 | mov r12, #15 |
603 | moveq r2, #0x8000 @ adj |
604 | moveq r12, #16 @ shift |
605 | |
606 | add r3, r0, #4*25 |
607 | ldmia r3, {r3-r5} @ gteMAC|123 |
608 | |
609 | do_mac_flags r3, r4, r5 |
610 | |
611 | add r3, r2 |
612 | add r4, r2 |
613 | add r5, r2 |
614 | asrs r3, r12 |
615 | orrne lr, #(1<<31)|(1<<24) @ IR1/limB1 |
616 | asrs r4, r12 |
617 | orrne lr, #(1<<31) |
618 | orrne lr, #(1<<23) @ IR2/limB2 |
619 | asrs r5, r12 |
620 | orrne lr, #(1<<22) @ IR3/limB3 |
621 | str lr, [r0, #4*(32+31)] @ gteFLAG |
622 | |
623 | pop {r4-r5,pc} |
624 | .size gteMVMVA_neon, .-gteMVMVA_neon |
625 | |
626 | |
627 | |
5d8e3bf8 |
628 | @ the name is misnormer, this doesn't use NEON but oh well.. |
629 | .global gteNCLIP_neon @ r0=CP2 (d,c), |
630 | gteNCLIP_neon: |
631 | push {r4-r6,lr} |
632 | |
633 | add r1, r0, #4*12 |
634 | ldmia r1, {r1-r3} |
635 | mov r4, r1, asr #16 |
636 | mov r5, r2, asr #16 |
637 | mov r6, r3, asr #16 |
638 | sub r12, r4, r5 @ 3: gteSY0 - gteSY1 |
639 | sub r5, r5, r6 @ 1: gteSY1 - gteSY2 |
640 | sxth r1, r1 |
641 | smull r1, r5, r1, r5 @ RdLo, RdHi |
642 | sub r6, r4 @ 2: gteSY2 - gteSY0 |
643 | sxth r2, r2 |
644 | smlal r1, r5, r2, r6 |
645 | mov lr, #0 @ gteFLAG |
646 | sxth r3, r3 |
647 | smlal r1, r5, r3, r12 |
648 | mov r6, #1<<31 |
649 | orr r6, #1<<15 |
650 | movs r2, r1, lsl #1 |
651 | adc r5, r5 |
652 | cmp r5, #0 |
653 | movtgt lr, #((1<<31)|(1<<16))>>16 |
654 | mvngt r1, #1<<31 @ maxint |
655 | cmn r5, #1 |
656 | movmi r1, #1<<31 @ minint |
657 | orrmi lr, r6 |
658 | str r1, [r0, #4*24] |
659 | str lr, [r0, #4*(32+31)] @ gteFLAG |
660 | |
661 | pop {r4-r6,pc} |
662 | .size gteNCLIP_neon, .-gteNCLIP_neon |
663 | |
8cfbda97 |
664 | |
665 | @ vim:filetype=armasm |