| 1 |
.section #gf100_builtin_code |
| 2 |
// DIV U32 |
| 3 |
// |
| 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 |
| 7 |
// |
| 8 |
// INPUT: $r0: dividend, $r1: divisor |
| 9 |
// OUTPUT: $r0: result, $r1: modulus |
| 10 |
// CLOBBER: $r2 - $r3, $p0 - $p1 |
| 11 |
// SIZE: 22 / 14 * 8 bytes |
| 12 |
// |
| 13 |
gf100_div_u32: |
| 14 |
bfind u32 $r2 $r1 |
| 15 |
xor b32 $r2 $r2 0x1f |
| 16 |
mov b32 $r3 0x1 |
| 17 |
shl b32 $r2 $r3 clamp $r2 |
| 18 |
cvt u32 $r1 neg u32 $r1 |
| 19 |
mul $r3 u32 $r1 u32 $r2 |
| 20 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 21 |
mul $r3 u32 $r1 u32 $r2 |
| 22 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 23 |
mul $r3 u32 $r1 u32 $r2 |
| 24 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 25 |
mul $r3 u32 $r1 u32 $r2 |
| 26 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 27 |
mul $r3 u32 $r1 u32 $r2 |
| 28 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 29 |
mov b32 $r3 $r0 |
| 30 |
mul high $r0 u32 $r0 u32 $r2 |
| 31 |
cvt u32 $r2 neg u32 $r1 |
| 32 |
add $r1 (mul u32 $r1 u32 $r0) $r3 |
| 33 |
set $p0 0x1 ge u32 $r1 $r2 |
| 34 |
$p0 sub b32 $r1 $r1 $r2 |
| 35 |
$p0 add b32 $r0 $r0 0x1 |
| 36 |
$p0 set $p0 0x1 ge u32 $r1 $r2 |
| 37 |
$p0 sub b32 $r1 $r1 $r2 |
| 38 |
$p0 add b32 $r0 $r0 0x1 |
| 39 |
ret |
| 40 |
|
| 41 |
// DIV S32, like DIV U32 after taking ABS(inputs) |
| 42 |
// |
| 43 |
// INPUT: $r0: dividend, $r1: divisor |
| 44 |
// OUTPUT: $r0: result, $r1: modulus |
| 45 |
// CLOBBER: $r2 - $r3, $p0 - $p3 |
| 46 |
// |
| 47 |
gf100_div_s32: |
| 48 |
set $p2 0x1 lt s32 $r0 0x0 |
| 49 |
set $p3 0x1 lt s32 $r1 0x0 xor $p2 |
| 50 |
cvt s32 $r0 abs s32 $r0 |
| 51 |
cvt s32 $r1 abs s32 $r1 |
| 52 |
bfind u32 $r2 $r1 |
| 53 |
xor b32 $r2 $r2 0x1f |
| 54 |
mov b32 $r3 0x1 |
| 55 |
shl b32 $r2 $r3 clamp $r2 |
| 56 |
cvt u32 $r1 neg u32 $r1 |
| 57 |
mul $r3 u32 $r1 u32 $r2 |
| 58 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 59 |
mul $r3 u32 $r1 u32 $r2 |
| 60 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 61 |
mul $r3 u32 $r1 u32 $r2 |
| 62 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 63 |
mul $r3 u32 $r1 u32 $r2 |
| 64 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 65 |
mul $r3 u32 $r1 u32 $r2 |
| 66 |
add $r2 (mul high u32 $r2 u32 $r3) $r2 |
| 67 |
mov b32 $r3 $r0 |
| 68 |
mul high $r0 u32 $r0 u32 $r2 |
| 69 |
cvt u32 $r2 neg u32 $r1 |
| 70 |
add $r1 (mul u32 $r1 u32 $r0) $r3 |
| 71 |
set $p0 0x1 ge u32 $r1 $r2 |
| 72 |
$p0 sub b32 $r1 $r1 $r2 |
| 73 |
$p0 add b32 $r0 $r0 0x1 |
| 74 |
$p0 set $p0 0x1 ge u32 $r1 $r2 |
| 75 |
$p0 sub b32 $r1 $r1 $r2 |
| 76 |
$p0 add b32 $r0 $r0 0x1 |
| 77 |
$p3 cvt s32 $r0 neg s32 $r0 |
| 78 |
$p2 cvt s32 $r1 neg s32 $r1 |
| 79 |
ret |
| 80 |
|
| 81 |
// RCP F64: Newton Raphson reciprocal(x): r_{i+1} = r_i * (2.0 - x * r_i) |
| 82 |
// |
| 83 |
// INPUT: $r0d (x) |
| 84 |
// OUTPUT: $r0d (rcp(x)) |
| 85 |
// CLOBBER: $r2 - $r7 |
| 86 |
// SIZE: 9 * 8 bytes |
| 87 |
// |
| 88 |
gf100_rcp_f64: |
| 89 |
nop |
| 90 |
ret |
| 91 |
|
| 92 |
// RSQ F64: Newton Raphson rsqrt(x): r_{i+1} = r_i * (1.5 - 0.5 * x * r_i * r_i) |
| 93 |
// |
| 94 |
// INPUT: $r0d (x) |
| 95 |
// OUTPUT: $r0d (rsqrt(x)) |
| 96 |
// CLOBBER: $r2 - $r7 |
| 97 |
// SIZE: 14 * 8 bytes |
| 98 |
// |
| 99 |
gf100_rsq_f64: |
| 100 |
nop |
| 101 |
ret |
| 102 |
|
| 103 |
.section #gf100_builtin_offsets |
| 104 |
.b64 #gf100_div_u32 |
| 105 |
.b64 #gf100_div_s32 |
| 106 |
.b64 #gf100_rcp_f64 |
| 107 |
.b64 #gf100_rsq_f64 |