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.
24 movw \rd, #:lower16:scratch
25 movt \rd, #:upper16:scratch
31 @ XXX: gteMAC calc shouldn't be saturating, but it is here
33 @ approximate gteMAC|123 flags
34 @ in: rr 123 as gteMAC|123
36 .macro do_mac_flags rr1 rr2 rr3
38 orrvs lr, #(1<<31)|(1<<27)
40 orrvs lr, #(1<<31)|(1<<26)
42 orrvs lr, #(1<<31)|(1<<25)
43 cmn \rr1, #1 @ same as adds ...
51 @ approximate 3x gteMACn flags
52 @ in: rr 123 as 3 instances gteMACn, *flags
54 .macro do_mac_flags3x rr1 rr2 rr3 nflags pflags
59 cmn \rr1, #1 @ adds ...
65 @ get gteIR|123 flags from gteMAC|123
66 @ in: rr 123 as gteMAC|123
68 .macro do_irs_flags rr1 rr2 rr3
72 orrne lr, #(1<<31)|(1<<24) @ IR1/limB1
76 orrne lr, #(1<<23) @ IR2/limB2
78 orrne lr, #(1<<22) @ IR3/limB3
83 * RTPS/RTPT register map:
85 * q | d | c code / phase 1 phase 2 scratch
86 * 0 0 gteR1* [s16] gteMAC3 = gteMAC3 \ v=0 *
87 * 1 gteR2* gteIR1-3 = gteIR1-3 / *
88 * 1 2 gteR3* gteMAC3 = gteMAC3 \ v=1
89 * 3 * gteIR1-3 = gteIR1-3 /
90 * 2 4 gteTRX<<12 [s64] gteOFX [s64] gteMAC3 \ v=2
91 * 5 gteTRY<<12 gteOFY [s64] gteIR1-3 /
92 * 3 6 gteTRZ<<12 gteDQA [s64] min gteMAC|12 v=012
93 * 7 0 gteDQB [s64] max gteMAC|12
94 * 4 8 VXYZ(v) / gteMAC1,2 [s32] min gteIR|123
95 * 9 * / gteMAC3 max gteIR|123
96 * 5 10 gteIR1-3 [s16] gteIR1-3 v=2 quotients 12
98 * 6 12 gteH (adj. for cmp)
99 * 13 gteH (float for div)
105 @ load gteR*, gteTR* and gteH (see map above), clear q15
110 vldmia r3, {d0-d2} @ gteR* [16*9]
112 add r3, r0, #4*(32+5)
113 vldmia r3, {d4-d5} @ gteTR*
114 vext.16 d2, d1, d2, #2 @ xxx3 -> x321
115 vext.16 d1, d0, d1, #3 @ xx32 -> x321
116 add r3, r0, #4*(32+26)
117 vld1.32 d11[0], [r3] @ gteH
118 vshll.s32 q3, d5, #12 @ gteTRZ
119 vshll.s32 q2, d4, #12 @ gteTR|XY
120 vmovl.s16 q6, d11 @ gteH
123 @ do RTP* gteMAC* calculation
124 @ in: gteR*, gteTR* as in map, d8 - VXYZ, r12 - 0
125 @ out: d8,d9 - gteMAC|123, d10 - gteIR|123
130 vmull.s16 q10, d2, d8
134 vadd.s64 d16, d17 @ d16=d0.16[2]*d8.16[2], as
135 vadd.s64 d18, d19 @ d8[3]==0, so won't affect
136 vadd.s64 d20, d21 @ QC
140 vqshrn.s64 d8, q8, #12 @ gteMAC1
141 vqshrn.s64 d18, q9, #12 @ gteMAC2
142 vqshrn.s64 d9, q10, #12 @ gteMAC3
143 vsli.u64 d8, d18, #32 @ gteMAC|12
145 vqmovn.s32 d10, q4 @ gteIR|123; losing 2 cycles?
148 .global gteRTPS_neon @ r0=CP2 (d,c),
152 @ fmrx r4, fpscr @ vmrs? at least 40 cycle hit
156 vldmia r0, {d8} @ VXYZ(0)
159 @ rtpx_mac @ slower here, faster in RTPT?
160 vmov.16 d8[3], r12 @ kill unused upper vector
163 vmull.s16 q10, d2, d8
164 vpadd.s32 d16, d16, d17
165 vpadd.s32 d17, d18, d19
166 vpadd.s32 d18, d20, d21
168 vpadal.s32 q3, q9 @ d6, d18 is slow?
169 vqshrn.s64 d8, q2, #12 @ gteMAC|12
170 vqshrn.s64 d9, q3, #12 @ gteMAC3
174 vst1.32 d9[0], [r3] @ wb gteMAC|123
175 vqmovn.s32 d10, q4 @ gteIR|123
177 add r3, r0, #4*17 @ gteSZ*
178 vldmia r3, {q7} @ d14,d15 gteSZ|123x
179 vmov.i32 d28, #0xffff @ 0xffff[32]
180 vmax.s32 d11, d9, d31
181 vshr.s32 d16, d12, #1 @ | gteH/2 (adjust for cmp)
183 vmin.u32 d11, d28 @ saturate to 0..0xffff limD/fSZ3
184 vmovl.s16 q9, d10 @ || expand gteIR|123
185 vshl.u32 d13, d12, #16 @ | preparing gteH
190 vsli.u64 d15, d11, #32 @ new gteSZ|0123 in q7
191 vclt.u32 d16, d16, d11 @ gteH/2 < fSZ3?
193 add r3, r0, #4*(32+24)
194 vld1.32 d4, [r3] @ || gteOF|XY
195 add r3, r0, #4*(32+27)
196 vld1.32 d6, [r3] @ || gteDQ|AB
199 vmovl.s32 q2, d4 @ || gteOF|XY [64]
200 vmax.u32 d11, d26 @ make divisor 1 if not
201 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
202 add r3, r0, #4*16 @ | gteSZ*
203 vstmia r3, {q7} @ | d14,d15 gteSZ|123x
205 vcvt.f32.u32 d13, d13 @ gteH (float for div)
206 vcvt.f32.u32 d11, d11 @ divisor
208 @ divide.. it's not worth messing with reciprocals here
209 @ just for 1 value, let's just use VFP divider here
210 vdiv.f32 s22, s26, s22
214 vcvt.u32.f32 d11, d11 @ quotient
216 @ while NEON's busy we calculate some flags on ARM
219 ldmia r3, {r4-r6} @ gteMAC|123
221 vst1.32 d11, [r1, :64] @ wb quotient for flags (pre-limE)
224 do_mac_flags r4, r5, r6
226 vshr.u32 d11, #15 @ quotient (limE)
228 do_irs_flags r4, r5, r6
230 vmlal.s32 q2, d18, d11[0]@ gteOF|XY + gteIR|12 * quotient
232 vld1.32 d16, [r3] @ || load fS|XY12, new 01
233 vqmovn.s64 d18, q2 @ saturate to 32
234 vmull.s32 q10, d6, d11[0]@ | d20 = gteDQA * quotient
235 vqshl.s32 d19, d18, #5 @ 11bit precision
237 ldr r4, [r1] @ quotient
240 orrne lr, #(1<<18) @ fSZ (limD)
242 vst1.32 d18, [r1, :64] @ writeback fS|XY2 before limG
244 vshr.s32 d18, d19, #16+5@ can't vqshrn because of insn
245 vadd.s64 d20, d7 @ | gteDQB + gteDQA * quotient
246 vmovn.s32 d18, q9 @ fS|XY2 [s16]
248 vqmovn.s64 d20, q10 @ | gteMAC0
250 vst1.32 d16, [r3]! @ writeback fS|XY01
251 vst1.32 d18[0], [r3] @ ...2
253 vshr.s32 d21, d20, #12
254 vst1.32 d20[0], [r3] @ gteMAC0
258 orrne lr, #(1<<17) @ limE
261 vmov.i32 d22, #0x1000
264 vst1.16 d21[0], [r3] @ gteIR0
266 ldmia r1, {r4,r5} @ fS|XY2 before limG, after 11bit sat
267 add r2, r4, #0x400<<16
268 add r3, r5, #0x400<<16
270 orrne lr, #(1<<14) @ limG1
273 orrne lr, #(1<<13) @ limG2
277 orrvs lr, #(1<<16) @ F
283 ldr r4, [r0, #4*24] @ gteMAC0
287 orrvs lr, #(1<<16) @ F
290 orrvs lr, #(1<<15) @ F
293 orrhi lr, #(1<<12) @ limH
295 str lr, [r0, #4*(32+31)] @ gteFLAG
298 .size gteRTPS_neon, .-gteRTPS_neon
302 .global gteRTPT_neon @ r0=CP2 (d,c),
311 vmov.i32 d22, #0x7fffffff
312 vmov.i32 d23, #0x80000000
316 vldmia r2!, {d8} @ VXYZ(v)
317 vmov.16 d8[3], r12 @ kill unused upper vector
320 vmin.s32 d22, d8 @ min gteMAC|12
321 vmax.s32 d23, d8 @ max gteMAC|12
323 vst1.32 {d9,d10}, [r1, :128]!
326 vst1.32 {d22,d23}, [r1, :128]! @ min/max gteMAC|12, for flags
330 vldmia r1, {d0-d3} @ note: d4,d5 is for gteOF|XY
332 vmov d20, d0 @ gteMAC3 v=0
333 vmin.s16 d24, d1, d3 @ | find min IR
334 vshr.s32 d22, d12, #1 @ || gteH/2 (adjust for cmp)
335 vmax.s16 d25, d1, d3 @ | .. also max, for flag gen
336 vsli.u64 d20, d2, #32 @ gteMAC3 v=1
337 vmov d21, d9 @ ... v=2
339 vmov.i32 q14, #0xffff @ 0xffff[32]
342 vdup.32 q11, d22[0] @ gteH/2
343 vmin.u32 q10, q14 @ saturate to 0..0xffff limD/fSZ(v)
344 vmin.s16 d24, d10 @ | find min/max IR
345 vmax.s16 d25, d10 @ |
347 add r3, r0, #4*19 @ ||
348 vld1.32 d14[0], [r3] @ || gteSZ3
350 vclt.u32 q11, q11, q10 @ gteH/2 < fSZ(v)?
352 vst1.32 d20, [r3]! @ | writeback fSZ(v)
354 vst1.32 d21[0], [r3] @ |
355 vmax.u32 q10, q11, q13 @ make divisor 1 if not
357 vstmia r3, {q12} @ min/max IR for flags
358 vcvt.f32.u32 q10, q10
359 vshl.u32 d13, d12, #16 @ | preparing gteH
361 @ while NEON's busy we calculate some flags on ARM
364 ldmia r2, {r4-r7} @ min/max gteMAC|12
366 orrvs lr, #(1<<31)|(1<<27)
368 orrvs lr, #(1<<31)|(1<<26)
373 ldr r4, [r1, #0] @ gteMAC3 v=0
374 ldr r5, [r1, #8*2] @ ... v=1
375 ldr r6, [r1, #8*4] @ ... v=2
377 add r3, r0, #4*(32+24)
378 vld1.32 d4, [r3] @ || gteOF|XY
379 add r3, r0, #4*(32+27)
380 vld1.32 d6, [r3] @ || gteDQ|AB
384 vrecpe.f32 q11, q10 @ inv
385 vmovl.s32 q2, d4 @ || gteOF|XY [64]
386 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
387 vrecps.f32 q12, q10, q11 @ step
388 vcvt.f32.u32 d13, d13 @ | gteH (float for div)
389 vmov.f32 q8, #0.5 @ |||
390 vmul.f32 q11, q12, q11 @ better inv
392 vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3
393 vdup.32 q13, d13[0] @ |
394 @ vrecps.f32 q12, q10, q11 @ step
395 @ vmul.f32 q11, q12, q11 @ better inv
396 vmul.f32 q10, q13, q11 @ result
398 vmov.f32 q8, #0.5 @ |||
399 vmovl.s32 q2, d4 @ || gteOF|XY [64]
400 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
401 vcvt.f32.u32 d13, d13 @ | gteH (float for div)
402 vdup.32 q13, d13[0] @ |
404 vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3
407 vmov q0, q10 @ to test against C code
415 do_mac_flags3x r4, r5, r6, (1<<31)|(1<<25), (1<<27) @ MAC3
419 ldmia r4, {r7,r8,r10,r11} @ min/max IR
423 orrne lr, #(1<<18) @ fSZ (limD)
425 vadd.f32 q10, q8 @ adjust for vcvt rounding mode
427 vmovl.s16 q9, d1 @ expand gteIR|12 v=0
428 vmovl.s16 q10, d3 @ expand gteIR|12 v=1
430 vstmia r6, {q8} @ wb quotients for flags (pre-limE)
432 vmovl.s16 q11, d10 @ expand gteIR|12 v=2
433 vshr.u32 q8, #15 @ quotients (limE)
436 vdup.32 d26, d17[0] @ quotient (dup)
438 @ flags for minIR012 (r7,r8), maxIR012 (r10,r11)
443 orrvs lr, #(1<<23) @ IR2/limB2
444 rsbs r2, r4, r7, lsl #16
445 cmnvc r4, r10, lsl #16
446 orrvs lr, #(1<<31)|(1<<24) @ IR1/limB1
447 rsbs r2, r4, r8, lsl #16
448 cmnvc r4, r11, lsl #16
449 orrvs lr, #(1<<22) @ IR3/limB3
451 vmull.s32 q9, d18, d24 @ gteIR|12 * quotient v=0
452 vmull.s32 q10, d20, d25 @ ... v=1
453 vmull.s32 q11, d22, d26 @ ... v=2
454 vadd.s64 q9, q2 @ gteOF|XY + gteIR|12 * quotient
455 vadd.s64 q10, q2 @ ... v=1
456 vadd.s64 q11, q2 @ ... v=2
457 vqmovn.s64 d18, q9 @ saturate to 32 v=0
458 vqmovn.s64 d19, q10 @ ... v=1
459 vqmovn.s64 d20, q11 @ ... v=2
460 vmin.s32 d14, d18, d19 @ || find min/max fS|XY(v) [32]
461 vmax.s32 d15, d18, d19 @ || for flags
464 vqshl.s32 q11, q9, #5 @ 11bit precision, v=0,1
465 vqshl.s32 d24, d20, #5 @ ... v=2
466 vmull.s32 q13, d6, d17 @ | gteDQA * quotient v=2
467 vpmin.s32 d16, d14, d31 @ || also find min/max in pair
468 vpmax.s32 d17, d15, d31 @ ||
469 vshr.s32 q11, #16+5 @ can't vqshrn because of insn
470 vshr.s32 d24, #16+5 @ encoding doesn't allow 21 :(
471 vsli.u64 d16, d17, #32 @ || pack in-pair min/max
472 vadd.s64 d26, d7 @ | gteDQB + gteDQA * quotient
473 vmovn.s32 d12, q11 @ fS|XY(v) [s16] v=0,1
474 vmovn.s32 d13, q12 @ 3
475 vstmia r1, {d14-d16} @ || other cacheline than quotients
477 vst1.32 d12, [r3]! @ writeback fS|XY v=0,1
480 vqmovn.s64 d26, q13 @ | gteMAC0
481 vmovl.u16 q5, d10 @ expand gteIR|123 v=2
483 vmov.i32 d13, #0x1000
484 vshr.s32 d12, d26, #12
487 vst1.32 d26[0], [r3]! @ gteMAC0
489 vst1.32 d8, [r3]! @ gteMAC123 (last iteration)
492 vmin.s32 d12, d13 @ | gteIR0
494 ldmia r6, {r4-r6} @ quotients
500 vst1.32 d12[0], [r3]! @ gteIR0
501 vst1.32 d10, [r3]! @ gteIR12
502 vst1.32 d11[0], [r3] @ ..3
505 orrne lr, #(1<<31) @ limE
506 orrne lr, #(1<<17) @ limE
508 add r2, r4, #0x400<<16 @ min fSX
509 add r3, r6, #0x400<<16 @ max fSX
512 orrne lr, #(1<<31) @ limG1
514 add r2, r5, #0x400<<16 @ min fSY
515 add r3, r7, #0x400<<16 @ max fSY
518 orrne lr, #(1<<31) @ limG2
521 orrvs lr, #(1<<16) @ F (31 already done by above)
524 ldr r4, [r0, #4*24] @ gteMAC0
529 orrvs lr, #(1<<31) @ F
532 orrvs lr, #(1<<31) @ F
534 orrhi lr, #(1<<12) @ limH
536 str lr, [r0, #4*(32+31)] @ gteFLAG
539 .size gteRTPT_neon, .-gteRTPT_neon
543 @ note: non-std calling convention used
544 @ r0 = CP2 (d,c) (must preserve)
546 @ r4,r5 = VXYZ(v) packed
549 .global gteMVMVA_part_neon
553 vmov.32 d8[1], r5 @ VXYZ(v)
554 vldmia r6, {d0-d2} @ MXxy/gteR* [16*9]
555 vldmia r7, {d4-d5} @ CVx/gteTR*
558 vext.16 d2, d1, d2, #2 @ xxx3 -> x321
559 vext.16 d1, d0, d1, #3 @ xx32 -> x321
560 vshll.s32 q3, d5, #12 @ gteTRZ/CV3
561 vshll.s32 q2, d4, #12 @ gteTR|XY/CV12
565 vmull.s16 q10, d2, d8
566 vpadd.s32 d16, d16, d17
567 vpadd.s32 d17, d18, d19
568 vpadd.s32 d18, d20, d21
576 vqmovn.s64 d8, q2 @ gteMAC|12
577 vqmovn.s64 d9, q3 @ gteMAC3
581 vqmovn.s32 d10, q4 @ gteIR|123
583 vst1.32 d9[0], [r3] @ wb gteMAC|123
588 vmovl.s16 q9, d10 @ expand gteIR|123
593 .size gteMVMVA_part_neon, .-gteMVMVA_part_neon
596 @ get flags after gteMVMVA_part_neon operation
597 .global gteMACtoIR_flags_neon @ r0=CP2 (d,c), r1=lm
598 gteMACtoIR_flags_neon:
604 moveq r2, #0x8000 @ adj
605 moveq r12, #16 @ shift
608 ldmia r3, {r3-r5} @ gteMAC|123
610 do_mac_flags r3, r4, r5
616 orrne lr, #(1<<31)|(1<<24) @ IR1/limB1
619 orrne lr, #(1<<23) @ IR2/limB2
621 orrne lr, #(1<<22) @ IR3/limB3
622 str lr, [r0, #4*(32+31)] @ gteFLAG
625 .size gteMACtoIR_flags_neon, .-gteMACtoIR_flags_neon
629 @ vim:filetype=armasm