Engineering Category: Balanced Ternary & b158 Quantization
Best Use in Production: Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines.
When to Select This Capsule: Sole canonical implementation for its 64-vector functional equivalence class. Low-latency L1-resident kernel (518913/25000 cyc/op); well-suited for high-frequency pipeline stages.
Snap-Circuit Wiring Role: 1-to-1 Unary Transform Stage `(u64)->u64` — snaps directly into linear pipelines or fan-out DAG branches.
.text: 329 B (80 x86_64 insns)495 B)pub fn borg_c99_gaia_trit_pack(word: u64) -> u64 {
const POW3: [u64; 5] = [1, 3, 9, 27, 81];
let mut packed: u64 = 0;
let mut lane: usize = 0;
while lane < 5 {
let cell = ((word >> (lane as u32 * 12)) & 0xFFF) as i64 - 2048;
let trit: i8 = (cell > 256) as i8 - (cell < -256) as i8;
let tu = trit as u8;
let digit = tu.wrapping_add((tu >> 7).wrapping_mul(3)) as u64;
packed += digit * POW3[lane];
lane += 1;
}
packed ^ 432
}
rustc -C opt-level=3 x86_64 Disassembly (329 B .text / 80 insns)borg_c99_gaia_trit_pack:
mov eax, edi
and eax, 4095
cmp rax, 2305
setae dl
cmp rax, 1792
sbb dl, 0
mov ecx, edx
sar cl, 7
and cl, 3
add cl, dl
mov eax, edi
shr eax, 12
mov rdx, rdi
shr rdx, 24
mov rsi, rdi
shr rsi, 36
shr rdi, 48
and edi, 4095
and esi, 4095
and edx, 4095
and eax, 4095
xor r8d, r8d
cmp eax, 1792
mov r9d, 0
sbb r9d, r9d
cmp rdx, 1792
mov r10d, 0
sbb r10d, r10d
movzx r9d, r9b
movzx r10d, r10b
shl r10d, 8
or r10d, r9d
cmp rsi, 1792
mov r9d, 0
sbb r9d, r9d
movzx ecx, cl
movzx r9d, r9b
shl r9d, 16
or r9d, r10d
cmp rdi, 1792
sbb r8d, r8d
shl r8d, 24
or r8d, r9d
movd xmm0, r8d
xor r8d, r8d
cmp rdi, 2305
setae r8b
xor edi, edi
cmp rsi, 2305
setae dil
xor esi, esi
cmp rdx, 2305
setae sil
xor edx, edx
cmp eax, 2305
setae dl
shl esi, 8
or esi, edx
shl edi, 16
or edi, esi
shl r8d, 24
or r8d, edi
movd xmm1, r8d
paddb xmm1, xmm0
pxor xmm0, xmm0
pxor xmm2, xmm2
pcmpgtb xmm2, xmm1
pand xmm2, xmmword ptr [rip + .LCPI7_0]
paddb xmm2, xmm1
punpcklbw xmm2, xmm0
pmullw xmm2, xmmword ptr [rip + .LCPI7_1]
punpcklwd xmm2, xmm0
pshufd xmm0, xmm2, 238
paddd xmm0, xmm2
pshufd xmm1, xmm0, 85
paddd xmm1, xmm0
movd eax, xmm1
add rax, rcx
xor rax, 432
ret
| Global Rank | Capsule & Domain | Hardware _rdtsc | Latency | Throughput | ELF .text | Ternary Vote | Author Aye |
|---|---|---|---|---|---|---|---|
| #1652 | borg_c99_gaia_trit_pack (THIS CAPSULE) domain: ternary | id: ddfa78212eb61ad23709530b2213b64d | 518913/25000 cyc/op | 172971/25000 ns/op | 25000000/172971 ops/us | 329 B .text (80 insns) | + (1) | aye-a6000-prime @ 432Hz |