2 * (C) GraÅžvydas "notaz" Ignotas, 2011
4 * This work is licensed under the terms of GNU GPL version 2 or later.
5 * See the COPYING file in the top-level directory.
20 @ XXX: gteMAC calc shouldn't be saturating, but it is here
22 @ approximate gteMAC|123 flags
23 @ in: rr 123 as gteMAC|123
25 .macro do_mac_flags rr1 rr2 rr3
27 orrvs lr, #(1<<31)|(1<<27)
29 orrvs lr, #(1<<31)|(1<<26)
31 orrvs lr, #(1<<31)|(1<<25)
32 cmn \rr1, #1 @ same as adds ...
40 @ approximate 3x gteMACn flags
41 @ in: rr 123 as 3 instances gteMACn, *flags
43 .macro do_mac_flags3x rr1 rr2 rr3 nflags pflags
48 cmn \rr1, #1 @ adds ...
54 @ get gteIR|123 flags from gteMAC|123
55 @ in: rr 123 as gteMAC|123
57 .macro do_irs_flags rr1 rr2 rr3
61 orrne lr, #(1<<31)|(1<<24) @ IR1/limB1
65 orrne lr, #(1<<23) @ IR2/limB2
67 orrne lr, #(1<<22) @ IR3/limB3
72 * RTPS/RTPT register map:
74 * q | d | c code / phase 1 phase 2 scratch
75 * 0 0 gteR1* [s16] gteMAC3 = gteMAC3 \ v=0 *
76 * 1 gteR2* gteIR1-3 = gteIR1-3 / *
77 * 1 2 gteR3* gteMAC3 = gteMAC3 \ v=1
78 * 3 * gteIR1-3 = gteIR1-3 /
79 * 2 4 gteTRX<<12 [s64] gteOFX [s64] gteMAC3 \ v=2
80 * 5 gteTRY<<12 gteOFY [s64] gteIR1-3 /
81 * 3 6 gteTRZ<<12 gteDQA [s64] min gteMAC|12 v=012
82 * 7 0 gteDQB [s64] max gteMAC|12
83 * 4 8 VXYZ(v) / gteMAC1,2 [s32] min gteIR|123
84 * 9 * / gteMAC3 max gteIR|123
85 * 5 10 gteIR1-3 [s16] gteIR1-3 v=2 quotients 12
87 * 6 12 gteH (adj. for cmp)
88 * 13 gteH (float for div)
94 @ load gteR*, gteTR* and gteH (see map above), clear q15
99 vldmia r3, {d0-d2} @ gteR* [16*9]
101 add r3, r0, #4*(32+5)
102 vldmia r3, {d4-d5} @ gteTR*
103 vext.16 d2, d1, d2, #2 @ xxx3 -> x321
104 vext.16 d1, d0, d1, #3 @ xx32 -> x321
105 add r3, r0, #4*(32+26)
106 vld1.32 d11[0], [r3] @ gteH
107 vshll.s32 q3, d5, #12 @ gteTRZ
108 vshll.s32 q2, d4, #12 @ gteTR|XY
109 vmovl.s16 q6, d11 @ gteH
112 @ do RTP* gteMAC* calculation
113 @ in: gteR*, gteTR* as in map, d8 - VXYZ, r12 - 0
114 @ out: d8,d9 - gteMAC|123, d10 - gteIR|123
119 vmull.s16 q10, d2, d8
123 vadd.s64 d16, d17 @ d16=d0.16[2]*d8.16[2], as
124 vadd.s64 d18, d19 @ d8[3]==0, so won't affect
125 vadd.s64 d20, d21 @ QC
129 vqshrn.s64 d8, q8, #12 @ gteMAC1
130 vqshrn.s64 d18, q9, #12 @ gteMAC2
131 vqshrn.s64 d9, q10, #12 @ gteMAC3
132 vsli.u64 d8, d18, #32 @ gteMAC|12
134 vqmovn.s32 d10, q4 @ gteIR|123; losing 2 cycles?
137 .global gteRTPS_neon @ r0=CP2 (d,c),
141 @ fmrx r4, fpscr @ vmrs? at least 40 cycle hit
142 movw r1, #:lower16:scratch
143 movt r1, #:upper16:scratch
146 vldmia r0, {d8} @ VXYZ(0)
149 @ rtpx_mac @ slower here, faster in RTPT?
150 vmov.16 d8[3], r12 @ kill unused upper vector
153 vmull.s16 q10, d2, d8
154 vpadd.s32 d16, d16, d17
155 vpadd.s32 d17, d18, d19
156 vpadd.s32 d18, d20, d21
158 vpadal.s32 q3, q9 @ d6, d18 is slow?
159 vqshrn.s64 d8, q2, #12 @ gteMAC|12
160 vqshrn.s64 d9, q3, #12 @ gteMAC3
164 vst1.32 d9[0], [r3] @ wb gteMAC|123
165 vqmovn.s32 d10, q4 @ gteIR|123
167 add r3, r0, #4*17 @ gteSZ*
168 vldmia r3, {q7} @ d14,d15 gteSZ|123x
169 vmov.i32 d28, #0xffff @ 0xffff[32]
170 vmax.s32 d11, d9, d31
171 vshr.s32 d16, d12, #1 @ | gteH/2 (adjust for cmp)
173 vmin.u32 d11, d28 @ saturate to 0..0xffff limD/fSZ3
174 vmovl.s16 q9, d10 @ || expand gteIR|123
175 vshl.u32 d13, d12, #16 @ | preparing gteH
180 vsli.u64 d15, d11, #32 @ new gteSZ|0123 in q7
181 vclt.u32 d16, d16, d11 @ gteH/2 < fSZ3?
183 add r3, r0, #4*(32+24)
184 vld1.32 d4, [r3] @ || gteOF|XY
185 add r3, r0, #4*(32+27)
186 vld1.32 d6, [r3] @ || gteDQ|AB
189 vmovl.s32 q2, d4 @ || gteOF|XY [64]
190 vmax.u32 d11, d26 @ make divisor 1 if not
191 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
192 add r3, r0, #4*16 @ | gteSZ*
193 vstmia r3, {q7} @ | d14,d15 gteSZ|123x
195 vcvt.f32.u32 d13, d13 @ gteH (float for div)
196 vcvt.f32.u32 d11, d11 @ divisor
198 @ divide.. it's not worth messing with reciprocals here
199 @ just for 1 value, let's just use VFP divider here
200 vdiv.f32 s22, s26, s22
202 vcvt.u32.f32 d11, d11 @ quotient
204 @ while NEON's busy we calculate some flags on ARM
207 ldmia r3, {r4-r6} @ gteMAC|123
209 vst1.32 d11, [r1, :64] @ wb quotient for flags (pre-limE)
212 do_mac_flags r4, r5, r6
214 vshr.u32 d11, #15 @ quotient (limE)
216 do_irs_flags r4, r5, r6
218 vmlal.s32 q2, d18, d11[0]@ gteOF|XY + gteIR|12 * quotient
220 vld1.32 d16, [r3] @ || load fS|XY12, new 01
221 vqmovn.s64 d18, q2 @ saturate to 32
222 vmull.s32 q10, d6, d11[0]@ | d20 = gteDQA * quotient
223 vqshl.s32 d19, d18, #5 @ 11bit precision
225 ldr r4, [r1] @ quotient
228 orrne lr, #(1<<18) @ fSZ (limD)
230 vst1.32 d18, [r1, :64] @ writeback fS|XY2 before limG
232 vshr.s32 d18, d19, #16+5@ can't vqshrn because of insn
233 vadd.s64 d20, d7 @ | gteDQB + gteDQA * quotient
234 vmovn.s32 d18, q9 @ fS|XY2 [s16]
236 vqmovn.s64 d20, q10 @ | gteMAC0
238 vst1.32 d16, [r3]! @ writeback fS|XY01
239 vst1.32 d18[0], [r3] @ ...2
241 vshr.s32 d21, d20, #12
242 vst1.32 d20[0], [r3] @ gteMAC0
246 orrne lr, #(1<<17) @ limE
249 vmov.i32 d22, #0x1000
252 vst1.16 d21[0], [r3] @ gteIR0
254 ldmia r1, {r4,r5} @ fS|XY2 before limG, after 11bit sat
255 add r2, r4, #0x400<<16
256 add r3, r5, #0x400<<16
258 orrne lr, #(1<<14) @ limG1
261 orrne lr, #(1<<13) @ limG2
265 orrvs lr, #(1<<16) @ F
271 ldr r4, [r0, #4*24] @ gteMAC0
275 orrvs lr, #(1<<16) @ F
278 orrvs lr, #(1<<15) @ F
281 orrhi lr, #(1<<12) @ limH
283 str lr, [r0, #4*(32+31)] @ gteFLAG
286 .size gteRTPS_neon, .-gteRTPS_neon
290 .global gteRTPT_neon @ r0=CP2 (d,c),
294 movw r1, #:lower16:scratch
295 movt r1, #:upper16:scratch
300 vmov.i32 d22, #0x7fffffff
301 vmov.i32 d23, #0x80000000
305 vldmia r2!, {d8} @ VXYZ(v)
306 vmov.16 d8[3], r12 @ kill unused upper vector
309 vmin.s32 d22, d8 @ min gteMAC|12
310 vmax.s32 d23, d8 @ max gteMAC|12
312 vst1.32 {d9,d10}, [r1, :128]!
315 vst1.32 {d22,d23}, [r1, :128]! @ min/max gteMAC|12, for flags
319 vldmia r1, {d0-d3} @ note: d4,d5 is for gteOF|XY
321 vmov d20, d0 @ gteMAC3 v=0
322 vmin.s16 d24, d1, d3 @ | find min IR
323 vshr.s32 d22, d12, #1 @ || gteH/2 (adjust for cmp)
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
328 vmov.i32 q14, #0xffff @ 0xffff[32]
331 vdup.32 q11, d22[0] @ gteH/2
332 vmin.u32 q10, q14 @ saturate to 0..0xffff limD/fSZ(v)
333 vmin.s16 d24, d10 @ | find min/max IR
334 vmax.s16 d25, d10 @ |
336 add r3, r0, #4*19 @ ||
337 vld1.32 d14[0], [r3] @ || gteSZ3
339 vclt.u32 q11, q11, q10 @ gteH/2 < fSZ(v)?
341 vst1.32 d20, [r3]! @ | writeback fSZ(v)
343 vst1.32 d21[0], [r3] @ |
344 vmax.u32 q10, q11, q13 @ make divisor 1 if not
346 vstmia r3, {q12} @ min/max IR for flags
347 vcvt.f32.u32 q10, q10
348 vshl.u32 d13, d12, #16 @ | preparing gteH
350 @ while NEON's busy we calculate some flags on ARM
353 ldmia r2, {r4-r7} @ min/max gteMAC|12
355 orrvs lr, #(1<<31)|(1<<27)
357 orrvs lr, #(1<<31)|(1<<26)
362 ldr r4, [r1, #0] @ gteMAC3 v=0
363 ldr r5, [r1, #8*2] @ ... v=1
364 ldr r6, [r1, #8*4] @ ... v=2
366 add r3, r0, #4*(32+24)
367 vld1.32 d4, [r3] @ || gteOF|XY
368 add r3, r0, #4*(32+27)
369 vld1.32 d6, [r3] @ || gteDQ|AB
373 vrecpe.f32 q11, q10 @ inv
374 vmovl.s32 q2, d4 @ || gteOF|XY [64]
375 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
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
380 vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3
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
386 vmovl.s32 q2, d4 @ || gteOF|XY [64]
387 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
388 vcvt.f32.u32 d13, d13 @ | gteH (float for div)
389 vdup.32 q13, d13[0] @ |
391 vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3
394 vmov q0, q10 @ to test against C code
402 do_mac_flags3x r4, r5, r6, (1<<31)|(1<<25), (1<<27) @ MAC3
406 ldmia r4, {r7,r8,r10,r11} @ min/max IR
410 orrne lr, #(1<<18) @ fSZ (limD)
412 @ vadd.f32 q10, q @ adjust for vcvt rounding mode
414 vmovl.s16 q9, d1 @ expand gteIR|12 v=0
415 vmovl.s16 q10, d3 @ expand gteIR|12 v=1
417 vstmia r6, {q8} @ wb quotients for flags (pre-limE)
419 vmovl.s16 q11, d10 @ expand gteIR|12 v=2
420 vshr.u32 q8, #15 @ quotients (limE)
423 vdup.32 d26, d17[0] @ quotient (dup)
425 @ flags for minIR012 (r7,r8), maxIR012 (r10,r11)
430 orrvs lr, #(1<<23) @ IR2/limB2
431 rsbs r2, r4, r7, lsl #16
432 cmnvc r4, r10, lsl #16
433 orrvs lr, #(1<<31)|(1<<24) @ IR1/limB1
434 rsbs r2, r4, r8, lsl #16
435 cmnvc r4, r11, lsl #16
436 orrvs lr, #(1<<22) @ IR3/limB3
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
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
454 vpmin.s32 d16, d14, d31 @ || also find min/max in pair
455 vpmax.s32 d17, d15, d31 @ ||
456 vshr.s32 q11, #16+5 @ can't vqshrn because of insn
457 vshr.s32 d24, #16+5 @ encoding doesn't allow 21 :(
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
464 vst1.32 d12, [r3]! @ writeback fS|XY v=0,1
467 vqmovn.s64 d26, q13 @ | gteMAC0
468 vmovl.u16 q5, d10 @ expand gteIR|123 v=2
470 vmov.i32 d13, #0x1000
471 vshr.s32 d12, d26, #12
474 vst1.32 d26[0], [r3]! @ gteMAC0
476 vst1.32 d8, [r3]! @ gteMAC123 (last iteration)
479 vmin.s32 d12, d13 @ | gteIR0
481 ldmia r6, {r4-r6} @ quotients
487 vst1.32 d12[0], [r3]! @ gteIR0
488 vst1.32 d10, [r3]! @ gteIR12
489 vst1.32 d11[0], [r3] @ ..3
492 orrne lr, #(1<<31) @ limE
493 orrne lr, #(1<<17) @ limE
495 add r2, r4, #0x400<<16 @ min fSX
496 add r3, r6, #0x400<<16 @ max fSX
499 orrne lr, #(1<<31) @ limG1
501 add r2, r5, #0x400<<16 @ min fSY
502 add r3, r7, #0x400<<16 @ max fSY
505 orrne lr, #(1<<31) @ limG2
508 orrvs lr, #(1<<16) @ F (31 already done by above)
511 ldr r4, [r0, #4*24] @ gteMAC0
516 orrvs lr, #(1<<31) @ F
519 orrvs lr, #(1<<31) @ F
521 orrhi lr, #(1<<12) @ limH
523 str lr, [r0, #4*(32+31)] @ gteFLAG
526 .size gteRTPT_neon, .-gteRTPT_neon
530 @ note: non-std calling convention used
531 @ r0 = CP2 (d,c) (must preserve)
533 @ r4,r5 = VXYZ(v) packed
536 .global gteMVMVA_part_neon
540 vmov.32 d8[1], r5 @ VXYZ(v)
541 vldmia r6, {d0-d2} @ MXxy/gteR* [16*9]
542 vldmia r7, {d4-d5} @ CVx/gteTR*
545 vext.16 d2, d1, d2, #2 @ xxx3 -> x321
546 vext.16 d1, d0, d1, #3 @ xx32 -> x321
547 vshll.s32 q3, d5, #12 @ gteTRZ/CV3
548 vshll.s32 q2, d4, #12 @ gteTR|XY/CV12
552 vmull.s16 q10, d2, d8
553 vpadd.s32 d16, d16, d17
554 vpadd.s32 d17, d18, d19
555 vpadd.s32 d18, d20, d21
563 vqmovn.s64 d8, q2 @ gteMAC|12
564 vqmovn.s64 d9, q3 @ gteMAC3
568 vqmovn.s32 d10, q4 @ gteIR|123
570 vst1.32 d9[0], [r3] @ wb gteMAC|123
575 vmovl.s16 q9, d10 @ expand gteIR|123
580 .size gteMVMVA_part_neon, .-gteMVMVA_part_neon
583 @ get flags after gteMVMVA_part_neon operation
584 .global gteMACtoIR_flags_neon @ r0=CP2 (d,c), r1=lm
585 gteMACtoIR_flags_neon:
591 moveq r2, #0x8000 @ adj
592 moveq r12, #16 @ shift
595 ldmia r3, {r3-r5} @ gteMAC|123
597 do_mac_flags r3, r4, r5
603 orrne lr, #(1<<31)|(1<<24) @ IR1/limB1
606 orrne lr, #(1<<23) @ IR2/limB2
608 orrne lr, #(1<<22) @ IR3/limB3
609 str lr, [r0, #4*(32+31)] @ gteFLAG
612 .size gteMACtoIR_flags_neon, .-gteMACtoIR_flags_neon
616 @ vim:filetype=armasm