1
.section #gm107_builtin_code
4
// UNR recurrence (q = a / b):
5
// look for z such that 2^32 - b <= b * z < 2^32
6
// then q - 1 <= (a * z) / 2^32 <= q
8
// INPUT: $r0: dividend, $r1: divisor
9
// OUTPUT: $r0: result, $r1: modulus
10
// CLOBBER: $r2 - $r3, $p0 - $p1
11
// SIZE: 22 / 14 * 8 bytes
14
sched (st 0xd wr 0x0 wt 0x3f) (st 0x1 wt 0x1) (st 0x6)
16
lop xor 1 $r2 $r2 0x1f
18
sched (st 0x1) (st 0xf wr 0x0) (st 0x6 wr 0x0 wt 0x1)
20
i2i u32 u32 $r1 neg $r1
21
imul u32 u32 $r3 $r1 $r2
22
sched (st 0x6 wr 0x0 wt 0x1) (st 0x6 wr 0x0 wt 0x1) (st 0x6 wr 0x0 wt 0x1)
23
imad u32 u32 hi $r2 $r2 $r3 $r2
24
imul u32 u32 $r3 $r1 $r2
25
imad u32 u32 hi $r2 $r2 $r3 $r2
26
sched (st 0x6 wr 0x0 wt 0x1) (st 0x6 wr 0x0 wt 0x1) (st 0x6 wr 0x0 wt 0x1)
27
imul u32 u32 $r3 $r1 $r2
28
imad u32 u32 hi $r2 $r2 $r3 $r2
29
imul u32 u32 $r3 $r1 $r2
30
sched (st 0x6 wr 0x0 wt 0x1) (st 0x6 wr 0x0 wt 0x1) (st 0x6 wr 0x0 rd 0x1 wt 0x1)
31
imad u32 u32 hi $r2 $r2 $r3 $r2
32
imul u32 u32 $r3 $r1 $r2
33
imad u32 u32 hi $r2 $r2 $r3 $r2
34
sched (st 0x6 wt 0x2) (st 0x6 wr 0x0 rd 0x1 wt 0x1) (st 0xf wr 0x0 rd 0x1 wt 0x2)
36
imul u32 u32 hi $r0 $r0 $r2
37
i2i u32 u32 $r2 neg $r1
38
sched (st 0x6 wr 0x0 wt 0x3) (st 0xd wt 0x1) (st 0x1)
39
imad u32 u32 $r1 $r1 $r0 $r3
40
isetp ge u32 and $p0 1 $r1 $r2 1
41
$p0 iadd $r1 $r1 neg $r2
42
sched (st 0x5) (st 0xd) (st 0x1)
44
$p0 isetp ge u32 and $p0 1 $r1 $r2 1
45
$p0 iadd $r1 $r1 neg $r2
46
sched (st 0x1) (st 0xf) (st 0xf)
51
// DIV S32, like DIV U32 after taking ABS(inputs)
53
// INPUT: $r0: dividend, $r1: divisor
54
// OUTPUT: $r0: result, $r1: modulus
55
// CLOBBER: $r2 - $r3, $p0 - $p3
58
sched (st 0xd wt 0x3f) (st 0x1) (st 0x1 wr 0x0)
59
isetp lt and $p2 0x1 $r0 0 1
60
isetp lt xor $p3 1 $r1 0 $p2
61
i2i s32 s32 $r0 abs $r0
62
sched (st 0xf wr 0x1) (st 0xd wr 0x1 wt 0x2) (st 0x1 wt 0x2)
63
i2i s32 s32 $r1 abs $r1
65
lop xor 1 $r2 $r2 0x1f
66
sched (st 0x6) (st 0x1) (st 0xf wr 0x1)
69
i2i u32 u32 $r1 neg $r1
70
sched (st 0x6 wr 0x1 wt 0x2) (st 0x6 wr 0x1 wt 0x2) (st 0x6 wr 0x1 wt 0x2)
71
imul u32 u32 $r3 $r1 $r2
72
imad u32 u32 hi $r2 $r2 $r3 $r2
73
imul u32 u32 $r3 $r1 $r2
74
sched (st 0x6 wr 0x1 wt 0x2) (st 0x6 wr 0x1 wt 0x2) (st 0x6 wr 0x1 wt 0x2)
75
imad u32 u32 hi $r2 $r2 $r3 $r2
76
imul u32 u32 $r3 $r1 $r2
77
imad u32 u32 hi $r2 $r2 $r3 $r2
78
sched (st 0x6 wr 0x1 wt 0x2) (st 0x6 wr 0x1 wt 0x2) (st 0x6 wr 0x1 wt 0x2)
79
imul u32 u32 $r3 $r1 $r2
80
imad u32 u32 hi $r2 $r2 $r3 $r2
81
imul u32 u32 $r3 $r1 $r2
82
sched (st 0x6 wr 0x1 rd 0x2 wt 0x2) (st 0x2 wt 0x5) (st 0x6 wr 0x0 rd 0x1 wt 0x2)
83
imad u32 u32 hi $r2 $r2 $r3 $r2
85
imul u32 u32 hi $r0 $r0 $r2
86
sched (st 0xf wr 0x1 rd 0x2 wt 0x2) (st 0x6 wr 0x0 wt 0x5) (st 0xd wt 0x3)
87
i2i u32 u32 $r2 neg $r1
88
imad u32 u32 $r1 $r1 $r0 $r3
89
isetp ge u32 and $p0 1 $r1 $r2 1
90
sched (st 0x1) (st 0x5) (st 0xd)
91
$p0 iadd $r1 $r1 neg $r2
93
$p0 isetp ge u32 and $p0 1 $r1 $r2 1
94
sched (st 0x1) (st 0x2) (st 0xf wr 0x0)
95
$p0 iadd $r1 $r1 neg $r2
97
$p3 i2i s32 s32 $r0 neg $r0
98
sched (st 0xf wr 0x1) (st 0xf wt 0x3) (st 0xf)
99
$p2 i2i s32 s32 $r1 neg $r1
107
// CLOBBER: $r2 - $r9, $p0
109
// The core of RCP and RSQ implementation is Newton-Raphson step, which is
110
// used to find successively better approximation from an imprecise initial
111
// value (single precision rcp in RCP and rsqrt64h in RSQ).
114
// Step 1: classify input according to exponent and value, and calculate
115
// result for 0/inf/nan. $r2 holds the exponent value, which starts at
116
// bit 52 (bit 20 of the upper half) and is 11 bits in length
117
sched (st 0x0) (st 0x0) (st 0x0)
118
bfe u32 $r2 $r1 0xb14
121
// We want to check whether the exponent is 0 or 0x7ff (i.e. NaN, inf,
122
// denorm, or 0). Do this by subtracting 1 from the exponent, which will
123
// mean that it's > 0x7fd in those cases when doing unsigned comparison
124
sched (st 0x0) (st 0x0) (st 0x0)
125
isetp gt u32 and $p0 1 $r3 0x7fd 1
126
// $r3: 0 for norms, 0x36 for denorms, -1 for others
129
// Process all special values: NaN, inf, denorm, 0
130
sched (st 0x0) (st 0x0) (st 0x0)
131
mov32i $r3 0xffffffff 0xf
132
// A number is NaN if its abs value is greater than or unordered with inf
133
dsetp gtu and $p0 1 abs $r0 0x7ff0000000000000 1
134
not $p0 bra #rcp_inf_or_denorm_or_zero
135
// NaN -> NaN, the next line sets the "quiet" bit of the result. This
136
// behavior is both seen on the CPU and the blob
137
sched (st 0x0) (st 0x0) (st 0x0)
138
lop32i or $r1 $r1 0x80000
140
rcp_inf_or_denorm_or_zero:
141
lop32i and $r4 $r1 0x7ff00000
142
sched (st 0x0) (st 0x0) (st 0x0)
143
// Other values with nonzero in exponent field should be inf
144
isetp eq and $p0 1 $r4 0x0 1
145
$p0 bra #rcp_denorm_or_zero
147
lop32i xor $r1 $r1 0x7ff00000
148
sched (st 0x0) (st 0x0) (st 0x0)
152
dsetp gtu and $p0 1 abs $r0 0x0 1
153
sched (st 0x0) (st 0x0) (st 0x0)
156
lop32i or $r1 $r1 0x7ff00000
159
// non-0 denorms: multiply with 2^54 (the 0x36 in $r3), join with norms
160
sched (st 0x0) (st 0x0) (st 0x0)
161
dmul $r0 $r0 0x4350000000000000
165
// All numbers with -1 in $r3 have their result ready in $r0d, return them
166
// others need further calculation
167
sched (st 0x0) (st 0x0) (st 0x0)
168
isetp lt and $p0 1 $r3 0x0 1
170
// Step 2: Before the real calculation goes on, renormalize the values to
171
// range [1, 2) by setting exponent field to 0x3ff (the exponent of 1)
172
// result in $r6d. The exponent will be recovered later.
173
bfe u32 $r2 $r1 0xb14
174
sched (st 0x0) (st 0x0) (st 0x0)
175
lop32i and $r7 $r1 0x800fffff
176
iadd32i $r7 $r7 0x3ff00000
178
// Step 3: Convert new value to float (no overflow will occur due to step
179
// 2), calculate rcp and do newton-raphson step once
180
sched (st 0x0) (st 0x0) (st 0x0)
181
f2f ftz f64 f32 $r5 $r6
183
mov32i $r0 0xbf800000 0xf
184
sched (st 0x0) (st 0x0) (st 0x0)
186
ffma $r0 $r5 neg $r4 $r4
187
// Step 4: convert result $r0 back to double, do newton-raphson steps
189
sched (st 0x0) (st 0x0) (st 0x0)
190
f2f f64 f64 $r6 neg $r6
191
f2f f32 f64 $r8 0x3f800000
192
// 4 Newton-Raphson Steps, tmp in $r4d, result in $r0d
193
// The formula used here (and above) is:
194
// RCP_{n + 1} = 2 * RCP_{n} - x * RCP_{n} * RCP_{n}
195
// The following code uses 2 FMAs for each step, and it will basically
197
// tmp = -src * RCP_{n} + 1
198
// RCP_{n + 1} = RCP_{n} * tmp + RCP_{n}
200
sched (st 0x0) (st 0x0) (st 0x0)
204
sched (st 0x0) (st 0x0) (st 0x0)
208
sched (st 0x0) (st 0x0) (st 0x0)
210
// Step 5: Exponent recovery and final processing
211
// The exponent is recovered by adding what we added to the exponent.
212
// Suppose we want to calculate rcp(x), but we have rcp(cx), then
213
// rcp(x) = c * rcp(cx)
214
// The delta in exponent comes from two sources:
215
// 1) The renormalization in step 2. The delta is:
217
// 2) (For the denorm input) The 2^54 we multiplied at rcp_denorm, stored
219
// These 2 sources are calculated in the first two lines below, and then
220
// added to the exponent extracted from the result above.
221
// Note that after processing, the new exponent may >= 0x7ff (inf)
222
// or <= 0 (denorm). Those cases will be handled respectively below
223
iadd $r2 neg $r2 0x3ff
225
sched (st 0x0) (st 0x0) (st 0x0)
226
bfe u32 $r3 $r1 0xb14
227
// New exponent in $r3
230
// (exponent-1) < 0x7fe (unsigned) means the result is in norm range
231
// (same logic as in step 1)
232
sched (st 0x0) (st 0x0) (st 0x0)
233
isetp lt u32 and $p0 1 $r2 0x7fe 1
234
not $p0 bra #rcp_result_inf_or_denorm
235
// Norms: convert exponents back and return
237
sched (st 0x0) (st 0x0) (st 0x0)
240
rcp_result_inf_or_denorm:
241
// New exponent >= 0x7ff means that result is inf
242
isetp ge and $p0 1 $r3 0x7ff 1
243
sched (st 0x0) (st 0x0) (st 0x0)
244
not $p0 bra #rcp_result_denorm
246
lop32i and $r1 $r1 0x80000000
248
sched (st 0x0) (st 0x0) (st 0x0)
249
iadd32i $r1 $r1 0x7ff00000
252
// Denorm result comes from huge input. The greatest possible fp64, i.e.
253
// 0x7fefffffffffffff's rcp is 0x0004000000000000, 1/4 of the smallest
254
// normal value. Other rcp result should be greater than that. If we
255
// set the exponent field to 1, we can recover the result by multiplying
256
// it with 1/2 or 1/4. 1/2 is used if the "exponent" $r3 is 0, otherwise
257
// 1/4 ($r3 should be -1 then). This is quite tricky but greatly simplifies
259
isetp ne u32 and $p0 1 $r3 0x0 1
260
sched (st 0x0) (st 0x0) (st 0x0)
261
lop32i and $r1 $r1 0x800fffff
263
$p0 f2f f32 f64 $r6 0x3e800000
265
not $p0 f2f f32 f64 $r6 0x3f000000
266
sched (st 0x0) (st 0x0) (st 0x0)
267
iadd32i $r1 $r1 0x00100000
276
// CLOBBER: $r2 - $r9, $p0 - $p1
279
// Before getting initial result rsqrt64h, two special cases should be
281
// 1. NaN: set the highest bit in mantissa so it'll be surely recognized
282
// as NaN in rsqrt64h
283
sched (st 0xd wr 0x0 wt 0x3f) (st 0xd wt 0x1) (st 0xd)
284
dsetp gtu and $p0 1 abs $r0 0x7ff0000000000000 1
285
$p0 lop32i or $r1 $r1 0x00080000
286
lop32i and $r2 $r1 0x7fffffff
287
// 2. denorms and small normal values: using their original value will
288
// lose precision either at rsqrt64h or the first step in newton-raphson
289
// steps below. Take 2 as a threshold in exponent field, and multiply
290
// with 2^54 if the exponent is smaller or equal. (will multiply 2^27
291
// to recover in the end)
292
sched (st 0xd) (st 0xd) (st 0xd)
293
bfe u32 $r3 $r1 0xb14
294
isetp le u32 and $p1 1 $r3 0x2 1
296
sched (st 0xd wr 0x0) (st 0xd wr 0x0 wt 0x1) (st 0xd)
297
$p1 dmul $r0 $r0 0x4350000000000000
299
// rsqrt64h will give correct result for 0/inf/nan, the following logic
300
// checks whether the input is one of those (exponent is 0x7ff or all 0
301
// except for the sign bit)
302
iset ne u32 and $r6 $r3 0x7ff 1
303
sched (st 0xd) (st 0xd) (st 0xd)
304
lop and 1 $r2 $r2 $r6
305
isetp ne u32 and $p0 1 $r2 0x0 1
307
// For 0/inf/nan, make sure the sign bit agrees with input and return
308
sched (st 0xd) (st 0xd) (st 0xd wt 0x1)
309
lop32i and $r1 $r1 0x80000000
312
sched (st 0xd) (st 0xf) (st 0xf)
317
// For others, do 4 Newton-Raphson steps with the formula:
318
// RSQ_{n + 1} = RSQ_{n} * (1.5 - 0.5 * x * RSQ_{n} * RSQ_{n})
319
// In the code below, each step is written as:
320
// tmp1 = 0.5 * x * RSQ_{n}
321
// tmp2 = -RSQ_{n} * tmp1 + 0.5
322
// RSQ_{n + 1} = RSQ_{n} * tmp2 + RSQ_{n}
323
sched (st 0xd) (st 0xd wr 0x1) (st 0xd wr 0x1 rd 0x0 wt 0x3)
326
f2f f32 f64 $r8 0x3f000000
328
sched (st 0xd wr 0x0 wt 0x3) (st 0xd wr 0x0 wt 0x1) (st 0xd wr 0x0 wt 0x1)
330
dfma $r6 $r0 neg $r4 $r8
332
sched (st 0xd wr 0x0 wt 0x1) (st 0xd wr 0x0 wt 0x1) (st 0xd wr 0x0 wt 0x1)
334
dfma $r6 $r0 neg $r4 $r8
336
sched (st 0xd wr 0x0 wt 0x1) (st 0xd wr 0x0 wt 0x1) (st 0xd wr 0x0 wt 0x1)
338
dfma $r6 $r0 neg $r4 $r8
340
sched (st 0xd wr 0x0 wt 0x1) (st 0xd wr 0x0 wt 0x1) (st 0xd wr 0x0 wt 0x1)
342
dfma $r6 $r0 neg $r4 $r8
344
// Multiply 2^27 to result for small inputs to recover
345
sched (st 0xd wr 0x0 wt 0x1) (st 0xd wt 0x1) (st 0xd)
346
$p1 dmul $r4 $r4 0x41a0000000000000
349
sched (st 0xd) (st 0xf) (st 0xf)
354
.section #gm107_builtin_offsets