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.
22 movw \rd, #:lower16:scratch
23 movt \rd, #:upper16:scratch
29 @ XXX: gteMAC calc shouldn't be saturating, but it is here
31 @ approximate gteMAC|123 flags
32 @ in: rr 123 as gteMAC|123
34 .macro do_mac_flags rr1 rr2 rr3
36 orrvs lr, #(1<<31)|(1<<27)
38 orrvs lr, #(1<<31)|(1<<26)
40 orrvs lr, #(1<<31)|(1<<25)
41 cmn \rr1, #1 @ same as adds ...
49 @ approximate 3x gteMACn flags
50 @ in: rr 123 as 3 instances gteMACn, *flags
52 .macro do_mac_flags3x rr1 rr2 rr3 nflags pflags
57 cmn \rr1, #1 @ adds ...
63 @ get gteIR|123 flags from gteMAC|123
64 @ in: rr 123 as gteMAC|123
66 .macro do_irs_flags rr1 rr2 rr3
70 orrne lr, #(1<<31)|(1<<24) @ IR1/limB1
74 orrne lr, #(1<<23) @ IR2/limB2
76 orrne lr, #(1<<22) @ IR3/limB3
81 * RTPS/RTPT register map:
83 * q | d | c code / phase 1 phase 2 scratch
84 * 0 0 gteR1* [s16] gteMAC3 = gteMAC3 \ v=0 *
85 * 1 gteR2* gteIR1-3 = gteIR1-3 / *
86 * 1 2 gteR3* gteMAC3 = gteMAC3 \ v=1
87 * 3 * gteIR1-3 = gteIR1-3 /
88 * 2 4 gteTRX<<12 [s64] gteOFX [s64] gteMAC3 \ v=2
89 * 5 gteTRY<<12 gteOFY [s64] gteIR1-3 /
90 * 3 6 gteTRZ<<12 gteDQA [s64] min gteMAC|12 v=012
91 * 7 0 gteDQB [s64] max gteMAC|12
92 * 4 8 VXYZ(v) / gteMAC1,2 [s32] min gteIR|123
93 * 9 * / gteMAC3 max gteIR|123
94 * 5 10 gteIR1-3 [s16] gteIR1-3 v=2 quotients 12
96 * 6 12 gteH (adj. for cmp)
97 * 13 gteH (float for div)
103 @ load gteR*, gteTR* and gteH (see map above), clear q15
108 vldmia r3, {d0-d2} @ gteR* [16*9]
110 add r3, r0, #4*(32+5)
111 vldmia r3, {d4-d5} @ gteTR*
112 vext.16 d2, d1, d2, #2 @ xxx3 -> x321
113 vext.16 d1, d0, d1, #3 @ xx32 -> x321
114 add r3, r0, #4*(32+26)
115 vld1.32 d11[0], [r3] @ gteH
116 vshll.s32 q3, d5, #12 @ gteTRZ
117 vshll.s32 q2, d4, #12 @ gteTR|XY
118 vmovl.s16 q6, d11 @ gteH
121 @ do RTP* gteMAC* calculation
122 @ in: gteR*, gteTR* as in map, d8 - VXYZ, r12 - 0
123 @ out: d8,d9 - gteMAC|123, d10 - gteIR|123
128 vmull.s16 q10, d2, d8
132 vadd.s64 d16, d17 @ d16=d0.16[2]*d8.16[2], as
133 vadd.s64 d18, d19 @ d8[3]==0, so won't affect
134 vadd.s64 d20, d21 @ QC
138 vqshrn.s64 d8, q8, #12 @ gteMAC1
139 vqshrn.s64 d18, q9, #12 @ gteMAC2
140 vqshrn.s64 d9, q10, #12 @ gteMAC3
141 vsli.u64 d8, d18, #32 @ gteMAC|12
143 vqmovn.s32 d10, q4 @ gteIR|123; losing 2 cycles?
146 .global gteRTPS_neon @ r0=CP2 (d,c),
150 @ fmrx r4, fpscr @ vmrs? at least 40 cycle hit
154 vldmia r0, {d8} @ VXYZ(0)
157 @ rtpx_mac @ slower here, faster in RTPT?
158 vmov.16 d8[3], r12 @ kill unused upper vector
161 vmull.s16 q10, d2, d8
162 vpadd.s32 d16, d16, d17
163 vpadd.s32 d17, d18, d19
164 vpadd.s32 d18, d20, d21
166 vpadal.s32 q3, q9 @ d6, d18 is slow?
167 vqshrn.s64 d8, q2, #12 @ gteMAC|12
168 vqshrn.s64 d9, q3, #12 @ gteMAC3
172 vst1.32 d9[0], [r3] @ wb gteMAC|123
173 vqmovn.s32 d10, q4 @ gteIR|123
175 add r3, r0, #4*17 @ gteSZ*
176 vldmia r3, {q7} @ d14,d15 gteSZ|123x
177 vmov.i32 d28, #0xffff @ 0xffff[32]
178 vmax.s32 d11, d9, d31
179 vshr.s32 d16, d12, #1 @ | gteH/2 (adjust for cmp)
181 vmin.u32 d11, d28 @ saturate to 0..0xffff limD/fSZ3
182 vmovl.s16 q9, d10 @ || expand gteIR|123
183 vshl.u32 d13, d12, #16 @ | preparing gteH
188 vsli.u64 d15, d11, #32 @ new gteSZ|0123 in q7
189 vclt.u32 d16, d16, d11 @ gteH/2 < fSZ3?
191 add r3, r0, #4*(32+24)
192 vld1.32 d4, [r3] @ || gteOF|XY
193 add r3, r0, #4*(32+27)
194 vld1.32 d6, [r3] @ || gteDQ|AB
197 vmovl.s32 q2, d4 @ || gteOF|XY [64]
198 vmax.u32 d11, d26 @ make divisor 1 if not
199 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
200 add r3, r0, #4*16 @ | gteSZ*
201 vstmia r3, {q7} @ | d14,d15 gteSZ|123x
203 vcvt.f32.u32 d13, d13 @ gteH (float for div)
204 vcvt.f32.u32 d11, d11 @ divisor
206 @ divide.. it's not worth messing with reciprocals here
207 @ just for 1 value, let's just use VFP divider here
208 vdiv.f32 s22, s26, s22
212 vcvt.u32.f32 d11, d11 @ quotient
214 @ while NEON's busy we calculate some flags on ARM
217 ldmia r3, {r4-r6} @ gteMAC|123
219 vst1.32 d11, [r1, :64] @ wb quotient for flags (pre-limE)
222 do_mac_flags r4, r5, r6
224 vshr.u32 d11, #15 @ quotient (limE)
226 do_irs_flags r4, r5, r6
228 vmlal.s32 q2, d18, d11[0]@ gteOF|XY + gteIR|12 * quotient
230 vld1.32 d16, [r3] @ || load fS|XY12, new 01
231 vqmovn.s64 d18, q2 @ saturate to 32
232 vmull.s32 q10, d6, d11[0]@ | d20 = gteDQA * quotient
233 vqshl.s32 d19, d18, #5 @ 11bit precision
235 ldr r4, [r1] @ quotient
238 orrne lr, #(1<<18) @ fSZ (limD)
240 vst1.32 d18, [r1, :64] @ writeback fS|XY2 before limG
242 vshr.s32 d18, d19, #16+5@ can't vqshrn because of insn
243 vadd.s64 d20, d7 @ | gteDQB + gteDQA * quotient
244 vmovn.s32 d18, q9 @ fS|XY2 [s16]
246 vqmovn.s64 d20, q10 @ | gteMAC0
248 vst1.32 d16, [r3]! @ writeback fS|XY01
249 vst1.32 d18[0], [r3] @ ...2
251 vshr.s32 d21, d20, #12
252 vst1.32 d20[0], [r3] @ gteMAC0
256 orrne lr, #(1<<17) @ limE
259 vmov.i32 d22, #0x1000
262 vst1.16 d21[0], [r3] @ gteIR0
264 ldmia r1, {r4,r5} @ fS|XY2 before limG, after 11bit sat
265 add r2, r4, #0x400<<16
266 add r3, r5, #0x400<<16
268 orrne lr, #(1<<14) @ limG1
271 orrne lr, #(1<<13) @ limG2
275 orrvs lr, #(1<<16) @ F
281 ldr r4, [r0, #4*24] @ gteMAC0
285 orrvs lr, #(1<<16) @ F
288 orrvs lr, #(1<<15) @ F
291 orrhi lr, #(1<<12) @ limH
293 str lr, [r0, #4*(32+31)] @ gteFLAG
296 .size gteRTPS_neon, .-gteRTPS_neon
300 .global gteRTPT_neon @ r0=CP2 (d,c),
309 vmov.i32 d22, #0x7fffffff
310 vmov.i32 d23, #0x80000000
314 vldmia r2!, {d8} @ VXYZ(v)
315 vmov.16 d8[3], r12 @ kill unused upper vector
318 vmin.s32 d22, d8 @ min gteMAC|12
319 vmax.s32 d23, d8 @ max gteMAC|12
321 vst1.32 {d9,d10}, [r1, :128]!
324 vst1.32 {d22,d23}, [r1, :128]! @ min/max gteMAC|12, for flags
328 vldmia r1, {d0-d3} @ note: d4,d5 is for gteOF|XY
330 vmov d20, d0 @ gteMAC3 v=0
331 vmin.s16 d24, d1, d3 @ | find min IR
332 vshr.s32 d22, d12, #1 @ || gteH/2 (adjust for cmp)
333 vmax.s16 d25, d1, d3 @ | .. also max, for flag gen
334 vsli.u64 d20, d2, #32 @ gteMAC3 v=1
335 vmov d21, d9 @ ... v=2
337 vmov.i32 q14, #0xffff @ 0xffff[32]
340 vdup.32 q11, d22[0] @ gteH/2
341 vmin.u32 q10, q14 @ saturate to 0..0xffff limD/fSZ(v)
342 vmin.s16 d24, d10 @ | find min/max IR
343 vmax.s16 d25, d10 @ |
345 add r3, r0, #4*19 @ ||
346 vld1.32 d14[0], [r3] @ || gteSZ3
348 vclt.u32 q11, q11, q10 @ gteH/2 < fSZ(v)?
350 vst1.32 d20, [r3]! @ | writeback fSZ(v)
352 vst1.32 d21[0], [r3] @ |
353 vmax.u32 q10, q11, q13 @ make divisor 1 if not
355 vstmia r3, {q12} @ min/max IR for flags
356 vcvt.f32.u32 q10, q10
357 vshl.u32 d13, d12, #16 @ | preparing gteH
359 @ while NEON's busy we calculate some flags on ARM
362 ldmia r2, {r4-r7} @ min/max gteMAC|12
364 orrvs lr, #(1<<31)|(1<<27)
366 orrvs lr, #(1<<31)|(1<<26)
371 ldr r4, [r1, #0] @ gteMAC3 v=0
372 ldr r5, [r1, #8*2] @ ... v=1
373 ldr r6, [r1, #8*4] @ ... v=2
375 add r3, r0, #4*(32+24)
376 vld1.32 d4, [r3] @ || gteOF|XY
377 add r3, r0, #4*(32+27)
378 vld1.32 d6, [r3] @ || gteDQ|AB
382 vrecpe.f32 q11, q10 @ inv
383 vmovl.s32 q2, d4 @ || gteOF|XY [64]
384 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
385 vrecps.f32 q12, q10, q11 @ step
386 vcvt.f32.u32 d13, d13 @ | gteH (float for div)
387 vmov.f32 q8, #0.5 @ |||
388 vmul.f32 q11, q12, q11 @ better inv
390 vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3
391 vdup.32 q13, d13[0] @ |
392 @ vrecps.f32 q12, q10, q11 @ step
393 @ vmul.f32 q11, q12, q11 @ better inv
394 vmul.f32 q10, q13, q11 @ result
396 vmov.f32 q8, #0.5 @ |||
397 vmovl.s32 q2, d4 @ || gteOF|XY [64]
398 vmovl.s32 q3, d6 @ || gteDQ|AB [64]
399 vcvt.f32.u32 d13, d13 @ | gteH (float for div)
400 vdup.32 q13, d13[0] @ |
402 vst1.32 d14[0], [r3] @ gteSZ0 = gteSZ3
405 vmov q0, q10 @ to test against C code
413 do_mac_flags3x r4, r5, r6, (1<<31)|(1<<25), (1<<27) @ MAC3
417 ldmia r4, {r7,r8,r10,r11} @ min/max IR
421 orrne lr, #(1<<18) @ fSZ (limD)
423 vadd.f32 q10, q8 @ adjust for vcvt rounding mode
425 vmovl.s16 q9, d1 @ expand gteIR|12 v=0
426 vmovl.s16 q10, d3 @ expand gteIR|12 v=1
428 vstmia r6, {q8} @ wb quotients for flags (pre-limE)
430 vmovl.s16 q11, d10 @ expand gteIR|12 v=2
431 vshr.u32 q8, #15 @ quotients (limE)
434 vdup.32 d26, d17[0] @ quotient (dup)
436 @ flags for minIR012 (r7,r8), maxIR012 (r10,r11)
441 orrvs lr, #(1<<23) @ IR2/limB2
442 rsbs r2, r4, r7, lsl #16
443 cmnvc r4, r10, lsl #16
444 orrvs lr, #(1<<31)|(1<<24) @ IR1/limB1
445 rsbs r2, r4, r8, lsl #16
446 cmnvc r4, r11, lsl #16
447 orrvs lr, #(1<<22) @ IR3/limB3
449 vmull.s32 q9, d18, d24 @ gteIR|12 * quotient v=0
450 vmull.s32 q10, d20, d25 @ ... v=1
451 vmull.s32 q11, d22, d26 @ ... v=2
452 vadd.s64 q9, q2 @ gteOF|XY + gteIR|12 * quotient
453 vadd.s64 q10, q2 @ ... v=1
454 vadd.s64 q11, q2 @ ... v=2
455 vqmovn.s64 d18, q9 @ saturate to 32 v=0
456 vqmovn.s64 d19, q10 @ ... v=1
457 vqmovn.s64 d20, q11 @ ... v=2
458 vmin.s32 d14, d18, d19 @ || find min/max fS|XY(v) [32]
459 vmax.s32 d15, d18, d19 @ || for flags
462 vqshl.s32 q11, q9, #5 @ 11bit precision, v=0,1
463 vqshl.s32 d24, d20, #5 @ ... v=2
464 vmull.s32 q13, d6, d17 @ | gteDQA * quotient v=2
465 vpmin.s32 d16, d14, d31 @ || also find min/max in pair
466 vpmax.s32 d17, d15, d31 @ ||
467 vshr.s32 q11, #16+5 @ can't vqshrn because of insn
468 vshr.s32 d24, #16+5 @ encoding doesn't allow 21 :(
469 vsli.u64 d16, d17, #32 @ || pack in-pair min/max
470 vadd.s64 d26, d7 @ | gteDQB + gteDQA * quotient
471 vmovn.s32 d12, q11 @ fS|XY(v) [s16] v=0,1
472 vmovn.s32 d13, q12 @ 3
473 vstmia r1, {d14-d16} @ || other cacheline than quotients
475 vst1.32 d12, [r3]! @ writeback fS|XY v=0,1
478 vqmovn.s64 d26, q13 @ | gteMAC0
479 vmovl.u16 q5, d10 @ expand gteIR|123 v=2
481 vmov.i32 d13, #0x1000
482 vshr.s32 d12, d26, #12
485 vst1.32 d26[0], [r3]! @ gteMAC0
487 vst1.32 d8, [r3]! @ gteMAC123 (last iteration)
490 vmin.s32 d12, d13 @ | gteIR0
492 ldmia r6, {r4-r6} @ quotients
498 vst1.32 d12[0], [r3]! @ gteIR0
499 vst1.32 d10, [r3]! @ gteIR12
500 vst1.32 d11[0], [r3] @ ..3
503 orrne lr, #(1<<31) @ limE
504 orrne lr, #(1<<17) @ limE
506 add r2, r4, #0x400<<16 @ min fSX
507 add r3, r6, #0x400<<16 @ max fSX
510 orrne lr, #(1<<31) @ limG1
512 add r2, r5, #0x400<<16 @ min fSY
513 add r3, r7, #0x400<<16 @ max fSY
516 orrne lr, #(1<<31) @ limG2
519 orrvs lr, #(1<<16) @ F (31 already done by above)
522 ldr r4, [r0, #4*24] @ gteMAC0
527 orrvs lr, #(1<<31) @ F
530 orrvs lr, #(1<<31) @ F
532 orrhi lr, #(1<<12) @ limH
534 str lr, [r0, #4*(32+31)] @ gteFLAG
537 .size gteRTPT_neon, .-gteRTPT_neon
541 @ note: non-std calling convention used
542 @ r0 = CP2 (d,c) (must preserve)
544 @ r4,r5 = VXYZ(v) packed
547 .global gteMVMVA_part_neon
551 vmov.32 d8[1], r5 @ VXYZ(v)
552 vldmia r6, {d0-d2} @ MXxy/gteR* [16*9]
553 vldmia r7, {d4-d5} @ CVx/gteTR*
556 vext.16 d2, d1, d2, #2 @ xxx3 -> x321
557 vext.16 d1, d0, d1, #3 @ xx32 -> x321
558 vshll.s32 q3, d5, #12 @ gteTRZ/CV3
559 vshll.s32 q2, d4, #12 @ gteTR|XY/CV12
563 vmull.s16 q10, d2, d8
564 vpadd.s32 d16, d16, d17
565 vpadd.s32 d17, d18, d19
566 vpadd.s32 d18, d20, d21
574 vqmovn.s64 d8, q2 @ gteMAC|12
575 vqmovn.s64 d9, q3 @ gteMAC3
579 vqmovn.s32 d10, q4 @ gteIR|123
581 vst1.32 d9[0], [r3] @ wb gteMAC|123
586 vmovl.s16 q9, d10 @ expand gteIR|123
591 .size gteMVMVA_part_neon, .-gteMVMVA_part_neon
594 @ get flags after gteMVMVA_part_neon operation
595 .global gteMACtoIR_flags_neon @ r0=CP2 (d,c), r1=lm
596 gteMACtoIR_flags_neon:
602 moveq r2, #0x8000 @ adj
603 moveq r12, #16 @ shift
606 ldmia r3, {r3-r5} @ gteMAC|123
608 do_mac_flags r3, r4, r5
614 orrne lr, #(1<<31)|(1<<24) @ IR1/limB1
617 orrne lr, #(1<<23) @ IR2/limB2
619 orrne lr, #(1<<22) @ IR3/limB3
620 str lr, [r0, #4*(32+31)] @ gteFLAG
623 .size gteMACtoIR_flags_neon, .-gteMACtoIR_flags_neon
627 @ vim:filetype=armasm