rustc -C opt-level=3): 2049 (203973 B .text / 53737 insns)22/9 Hz): 249004 B / 5979082 B (ratio 124502/2989541, saved 2865039/2989541, 11682 ″ dittos)| Rank & Crown | Function Name & Domain | Best Use & Snap-Circuit Role | Pin Signature & 64-Vec Hash | Hardware _rdtsc (p/q) | Details & Quick Peek |
|---|---|---|---|---|---|
| #63 CROWN #1 of 302 <10 CYC | st_golf_steward_mlx_quant_mcp_pq_harmonic_fold_i64 domain: mlx-quant | id: 6210dd4cee0eAye: aye::Hue+Gemini(0+1)::x86_64 @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(i64,i64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (i64,i64)->i6464-vec BLAKE3: f3c9e53b4edeC64 RAM: 277/65536 (277 B src) | 801227/100000 cyc/op 869487/400000 ns/op | 400000000/869487 ops/us 102 B .text (28 insns) | Trit + (1) | Inspect Details & ASM →Quick Peek Rust (277 B)pub const fn st_golf_steward_mlx_quant_mcp_pq_harmonic_fold_i64(p: i64, q: i64) -> i64 {
let den = if q == 0 { 1 } else { q.abs() };
let num = p.wrapping_mul(432).wrapping_add(den.wrapping_mul(528));
(num ^ (num >> 13)).wrapping_rem(if den < 1 { 1 } else { den })
} |
| #80 CROWN #1 of 4 <10 CYC | st_golf_gaia_pack_trits_horner_u64 domain: gaia-mlx-quant | id: 5ae0cdaba109Aye: aye::Hue+Gemini(0+1)::x86_64 @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 1-to-1 Unary Transform Stage `(u64)->u64` — snaps directly into linear pipelines or fan-out DAG branches. | (u64)->u6464-vec BLAKE3: d6deb4622288C64 RAM: 305/32768 (610 B src) | 1930529/200000 cyc/op 523749/200000 ns/op | 200000000/523749 ops/us 228 B .text (57 insns) | Trit + (1) | Inspect Details & ASM →Quick Peek Rust (610 B)pub const fn st_golf_gaia_pack_trits_horner_u64(word: u64) -> u64 {
let v0 = word & 0xFFF;
let d0 = ((v0 > 2304) as u64) | (((v0 < 1792) as u64) << 1);
let v1 = (word >> 12) & 0xFFF;
let d1 = ((v1 > 2304) as u64) | (((v1 < 1792) as u64) << 1);
let v2 = (word >> 24) & 0xFFF;
let d2 = ((v2 > 2304) as u64) | (((v2 < 1792) as u64) << 1);
let v3 = (word >> 36) & 0xFFF;
let d3 = ((v3 > 2304) as u64) | (((v3 < 1792) as u64) << 1);
let v4 = (word >> 48) & 0xFFF;
let d4 = ((v4 > 2304) as u64) | (((v4 < 1792) as u64) << 1);
d0 + 3 * (d1 + 3 * (d2 + 3 * (d3 + 3 * d4)))
} |
| #358 PEER of 303 <10 CYC | steward_mlx_quant_geodesic_seat_ema domain: MLX-QUANT | id: 6275a1202035Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(i64,i64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (i64,i64)->i6464-vec BLAKE3: f79426b0bb51C64 RAM: 109/32768 (218 B src) | 309411/50000 cyc/op 200027/100000 ns/op | 100000000/200027 ops/us 67 B .text (17 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (218 B)#[inline(always)]
pub fn steward_mlx_quant_geodesic_seat_ema(prev_cm: i64, obs_cm: i64) -> i64 {
let ema_cm = (prev_cm.wrapping_mul(3).wrapping_add(obs_cm)) / 4;
(ema_cm.wrapping_mul(10_000_000)) / 11_113_300
} |
| #469 CROWN #1 of 5 | st_golf_gaia_ternary_dot_swar_i64 domain: gaia-mlx-quant | id: 26c297061059Aye: aye::Hue+Gemini(0+1)::x86_64 @ 432Hz | Balanced Ternary & b158 Quantization High-speed ternary dot-product accumulation without integer multiplier bottlenecks; ideal for b158 neural weights and low-bit attention inner loops. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->i6464-vec BLAKE3: 0abcabf86830C64 RAM: 389/32768 (778 B src) | 587677/50000 cyc/op 318867/100000 ns/op | 100000000/318867 ops/us 224 B .text (59 insns) | Trit + (1) | Inspect Details & ASM →Quick Peek Rust (778 B)pub const fn st_golf_gaia_ternary_dot_swar_i64(activations: u64, weights: u64) -> i64 {
const M: u64 = 0x1111_1111_1111_1111;
const NIB: u64 = 0x0F0F_0F0F_0F0F_0F0F;
const ONES: u64 = 0x0101_0101_0101_0101;
let pos_nib = ((weights >> 3) & ((weights >> 2) | ((weights >> 1) & weights))) & M;
let nw = !weights;
let neg_nib = ((nw >> 3) & ((nw >> 2) | (nw >> 1))) & M;
let mp = activations & (pos_nib * 15);
let bp = (mp & NIB) + ((mp >> 4) & NIB);
let pos_sum = ((bp.wrapping_mul(ONES) >> 56) as i64) - 8 * (pos_nib.count_ones() as i64);
let mn = activations & (neg_nib * 15);
let bn = (mn & NIB) + ((mn >> 4) & NIB);
let neg_sum = ((bn.wrapping_mul(ONES) >> 56) as i64) - 8 * (neg_nib.count_ones() as i64);
pos_sum - neg_sum
} |
| #488 PEER of 303 <10 CYC | steward_mlx_quant_viz_geodesic_seat_ema domain: mlx-quant-viz | id: 7dc8fa0ccfa3Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(i64,i64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (i64,i64)->i6464-vec BLAKE3: f79426b0bb51C64 RAM: 111/32768 (222 B src) | 155217/25000 cyc/op 401417/200000 ns/op | 200000000/401417 ops/us 67 B .text (17 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (222 B)#[inline(always)]
pub fn steward_mlx_quant_viz_geodesic_seat_ema(prev_cm: i64, obs_cm: i64) -> i64 {
let ema_cm = (prev_cm.wrapping_mul(3).wrapping_add(obs_cm)) / 4;
(ema_cm.wrapping_mul(10_000_000)) / 11_113_300
} |
| #539 PEER of 303 <10 CYC | steward_mlx_quant_mcp_geodesic_seat_ema domain: mlx-quant-mcp | id: ee834ce05e06Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(i64,i64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (i64,i64)->i6464-vec BLAKE3: f79426b0bb51C64 RAM: 111/32768 (222 B src) | 1333077/200000 cyc/op 107747/50000 ns/op | 50000000/107747 ops/us 67 B .text (17 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (222 B)#[inline(always)]
pub fn steward_mlx_quant_mcp_geodesic_seat_ema(prev_cm: i64, obs_cm: i64) -> i64 {
let ema_cm = (prev_cm.wrapping_mul(3).wrapping_add(obs_cm)) / 4;
(ema_cm.wrapping_mul(10_000_000)) / 11_113_300
} |
| #585 PEER of 5 | st_golf_gaia_ternary_dot_swar_i64 domain: gaia-mlx-quant | id: aecca882ad00Aye: aye::Hue+Gemini(0+1)::x86_64 @ 432Hz | Balanced Ternary & b158 Quantization High-speed ternary dot-product accumulation without integer multiplier bottlenecks; ideal for b158 neural weights and low-bit attention inner loops. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->i6464-vec BLAKE3: 0abcabf86830C64 RAM: 215/16384 (860 B src) | 1392847/100000 cyc/op 1511449/400000 ns/op | 400000000/1511449 ops/us 287 B .text (72 insns) | Trit + (1) | Inspect Details & ASM →Quick Peek Rust (860 B)pub const fn st_golf_gaia_ternary_dot_swar_i64(activations: u64, weights: u64) -> i64 {
const M: u64 = 0x1111_1111_1111_1111;
const NIB: u64 = 0x0F0F_0F0F_0F0F_0F0F;
const B16: u64 = 0x00FF_00FF_00FF_00FF;
let pos_nib = ((weights >> 3) & ((weights >> 2) | ((weights >> 1) & weights))) & M;
let nw = !weights;
let neg_nib = ((nw >> 3) & ((nw >> 2) | (nw >> 1))) & M;
let mp = activations & (pos_nib * 15);
let bp = (mp & NIB) + ((mp >> 4) & NIB);
let p16 = (bp & B16) + ((bp >> 8) & B16);
let mn = activations & (neg_nib * 15);
let bn = (mn & NIB) + ((mn >> 4) & NIB);
let n16 = (bn & B16) + ((bn >> 8) & B16);
let diff_sum = (((p16 + 0x0040_0040_0040_0040 - n16).wrapping_mul(0x0001_0001_0001_0001) >> 48) as i64) - 256;
diff_sum - (((pos_nib.count_ones() as i64) - (neg_nib.count_ones() as i64)) << 3)
} |
| #615 PEER of 306 <10 CYC | steward_mlx_quant_viz_wupz_witness_step domain: mlx-quant-viz | id: 66056b676304Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->u64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->u6464-vec BLAKE3: 68d052c83838C64 RAM: 133/32768 (266 B src) | 394723/50000 cyc/op 255329/100000 ns/op | 100000000/255329 ops/us 45 B .text (10 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (266 B)#[inline(always)]
pub fn steward_mlx_quant_viz_wupz_witness_step(acc: u64, word: u64) -> u64 {
let mixed = acc ^ word.wrapping_mul(0x9E37_79B9_7F4A_7C15);
let rot = mixed.rotate_left(17) ^ mixed.rotate_right(11);
rot.wrapping_add(0x5755_505A_0432_0528)
} |
| #660 PEER of 306 <10 CYC | steward_mlx_quant_wupz_witness_step domain: MLX-QUANT | id: b8734d4afdcbAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->u64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->u6464-vec BLAKE3: 68d052c83838C64 RAM: 131/32768 (262 B src) | 3249079/400000 cyc/op 1050461/400000 ns/op | 400000000/1050461 ops/us 45 B .text (10 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (262 B)#[inline(always)]
pub fn steward_mlx_quant_wupz_witness_step(acc: u64, word: u64) -> u64 {
let mixed = acc ^ word.wrapping_mul(0x9E37_79B9_7F4A_7C15);
let rot = mixed.rotate_left(17) ^ mixed.rotate_right(11);
rot.wrapping_add(0x5755_505A_0432_0528)
} |
| #711 PEER of 312 <10 CYC | steward_mlx_quant_mcp_kompress_ditto_stride domain: mlx-quant-mcp | id: df6f93ea05cbAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->u64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->u6464-vec BLAKE3: 012de5dc0aa5C64 RAM: 127/32768 (254 B src) | 866101/100000 cyc/op 13997/5000 ns/op | 5000000/13997 ops/us 122 B .text (30 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (254 B)#[inline(always)]
pub fn steward_mlx_quant_mcp_kompress_ditto_stride(prev_cell: u64, curr_cell: u64) -> u64 {
let delta = prev_cell ^ curr_cell;
let pop = delta.count_ones() as u64;
if pop == 0 { 22 } else { (pop.wrapping_mul(22)) / 9 + 1 }
} |
| #916 PEER of 312 <10 CYC | steward_mlx_quant_kompress_ditto_stride domain: MLX-QUANT | id: 01588201e4ffAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->u64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->u6464-vec BLAKE3: 012de5dc0aa5C64 RAM: 125/32768 (250 B src) | 1687981/200000 cyc/op 1091207/400000 ns/op | 400000000/1091207 ops/us 122 B .text (30 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (250 B)#[inline(always)]
pub fn steward_mlx_quant_kompress_ditto_stride(prev_cell: u64, curr_cell: u64) -> u64 {
let delta = prev_cell ^ curr_cell;
let pop = delta.count_ones() as u64;
if pop == 0 { 22 } else { (pop.wrapping_mul(22)) / 9 + 1 }
} |
| #936 PEER of 312 <10 CYC | steward_mlx_quant_viz_kompress_ditto_stride domain: mlx-quant-viz | id: a2844e672083Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->u64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->u6464-vec BLAKE3: 012de5dc0aa5C64 RAM: 127/32768 (254 B src) | 736963/80000 cyc/op 238239/80000 ns/op | 80000000/238239 ops/us 122 B .text (30 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (254 B)#[inline(always)]
pub fn steward_mlx_quant_viz_kompress_ditto_stride(prev_cell: u64, curr_cell: u64) -> u64 {
let delta = prev_cell ^ curr_cell;
let pop = delta.count_ones() as u64;
if pop == 0 { 22 } else { (pop.wrapping_mul(22)) / 9 + 1 }
} |
| #962 PEER of 306 | steward_mlx_quant_mcp_wupz_witness_step domain: mlx-quant-mcp | id: 179907f7ea8eAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->u64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->u6464-vec BLAKE3: 68d052c83838C64 RAM: 133/32768 (266 B src) | 4047453/400000 cyc/op 1308567/400000 ns/op | 400000000/1308567 ops/us 45 B .text (10 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (266 B)#[inline(always)]
pub fn steward_mlx_quant_mcp_wupz_witness_step(acc: u64, word: u64) -> u64 {
let mixed = acc ^ word.wrapping_mul(0x9E37_79B9_7F4A_7C15);
let rot = mixed.rotate_left(17) ^ mixed.rotate_right(11);
rot.wrapping_add(0x5755_505A_0432_0528)
} |
| #1060 PEER of 302 | steward_mlx_quant_viz_pq_harmonic_fold domain: mlx-quant-viz | id: 6438999ea5ffAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(i64,i64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (i64,i64)->i6464-vec BLAKE3: f3c9e53b4edeC64 RAM: 129/32768 (258 B src) | 223231/20000 cyc/op 144311/40000 ns/op | 40000000/144311 ops/us 63 B .text (17 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (258 B)#[inline(always)]
pub fn steward_mlx_quant_viz_pq_harmonic_fold(p: i64, q: i64) -> i64 {
let den = if q == 0 { 1 } else { q.abs() };
let num = p.wrapping_mul(432).wrapping_add(den.wrapping_mul(528));
(num ^ (num >> 13)).wrapping_rem(den.max(1))
} |
| #1113 PEER of 302 | steward_mlx_quant_pq_harmonic_fold domain: MLX-QUANT | id: 292e2f2fdcc6Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(i64,i64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (i64,i64)->i6464-vec BLAKE3: f3c9e53b4edeC64 RAM: 127/32768 (254 B src) | 279279/25000 cyc/op 1444293/400000 ns/op | 400000000/1444293 ops/us 63 B .text (17 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (254 B)#[inline(always)]
pub fn steward_mlx_quant_pq_harmonic_fold(p: i64, q: i64) -> i64 {
let den = if q == 0 { 1 } else { q.abs() };
let num = p.wrapping_mul(432).wrapping_add(den.wrapping_mul(528));
(num ^ (num >> 13)).wrapping_rem(den.max(1))
} |
| #1177 PEER of 300 | steward_mlx_quant_viz_ternary_gate_pack domain: mlx-quant-viz | id: 8f2bc54aa050Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. Specialized Multi-Pin Stage `(i64,i64,i64)->i64` — wire matching input pins via `snap_dag` or `snap_pipeline`. | (i64,i64,i64)->i6464-vec BLAKE3: bc0951af632bC64 RAM: 413/65536 (413 B src) | 284363/25000 cyc/op 1470641/400000 ns/op | 400000000/1470641 ops/us 109 B .text (32 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (413 B)#[inline(always)]
pub fn steward_mlx_quant_viz_ternary_gate_pack(a: i64, b: i64, threshold: i64) -> i64 {
let th = threshold.abs().max(1);
let t0: i64 = if a > th { 2 } else if a < -th { 0 } else { 1 };
let t1: i64 = if b > th { 2 } else if b < -th { 0 } else { 1 };
let diff = a.wrapping_sub(b);
let t2: i64 = if diff > th { 2 } else if diff < -th { 0 } else { 1 };
t0 * 9 + t1 * 3 + t2
} |
| #1356 PEER of 302 | steward_mlx_quant_mcp_pq_harmonic_fold domain: mlx-quant-mcp | id: 2bd379864ed0Aye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 2-to-1 Binary Combiner / Join Stage `(i64,i64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (i64,i64)->i6464-vec BLAKE3: f3c9e53b4edeC64 RAM: 129/32768 (258 B src) | 535587/50000 cyc/op 34623/10000 ns/op | 10000000/34623 ops/us 63 B .text (17 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (258 B)#[inline(always)]
pub fn steward_mlx_quant_mcp_pq_harmonic_fold(p: i64, q: i64) -> i64 {
let den = if q == 0 { 1 } else { q.abs() };
let num = p.wrapping_mul(432).wrapping_add(den.wrapping_mul(528));
(num ^ (num >> 13)).wrapping_rem(den.max(1))
} |
| #1411 PEER of 300 | steward_mlx_quant_mcp_ternary_gate_pack domain: mlx-quant-mcp | id: 5e0afb57084dAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. Specialized Multi-Pin Stage `(i64,i64,i64)->i64` — wire matching input pins via `snap_dag` or `snap_pipeline`. | (i64,i64,i64)->i6464-vec BLAKE3: bc0951af632bC64 RAM: 413/65536 (413 B src) | 2196489/200000 cyc/op 1419957/400000 ns/op | 400000000/1419957 ops/us 109 B .text (32 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (413 B)#[inline(always)]
pub fn steward_mlx_quant_mcp_ternary_gate_pack(a: i64, b: i64, threshold: i64) -> i64 {
let th = threshold.abs().max(1);
let t0: i64 = if a > th { 2 } else if a < -th { 0 } else { 1 };
let t1: i64 = if b > th { 2 } else if b < -th { 0 } else { 1 };
let diff = a.wrapping_sub(b);
let t2: i64 = if diff > th { 2 } else if diff < -th { 0 } else { 1 };
t0 * 9 + t1 * 3 + t2
} |
| #1689 PEER of 4 | gaia_pack_trits_unrolled domain: gaia-mlx-quant | id: 919c37509f21Aye: aye-gemini @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 1-to-1 Unary Transform Stage `(u64)->u64` — snaps directly into linear pipelines or fan-out DAG branches. | (u64)->u6464-vec BLAKE3: d6deb4622288C64 RAM: 493/65536 (493 B src) | 12702081/400000 cyc/op 4234027/400000 ns/op | 400000000/4234027 ops/us 322 B .text (79 insns) | Trit + (1) | Inspect Details & ASM →Quick Peek Rust (493 B)pub fn gaia_pack_trits_unrolled(word: u64) -> u64 {
const POW3: [u64; 5] = [1, 3, 9, 27, 81];
let mut byte_val: u64 = 0;
let mut slot: usize = 0;
while slot < 5 {
let raw = ((word >> (slot as u32 * 12)) & 0xFFF) as i64 - 2048;
let trit: i8 = (raw > 256) as i8 - (raw < -256) as i8;
let tu = trit as u8;
let digit = tu.wrapping_add((tu >> 7).wrapping_mul(3)) as u64;
byte_val += digit * POW3[slot];
slot += 1;
}
byte_val
} |
| #1705 PEER of 4 | gaia_pack_trits_scalar domain: gaia-mlx-quant | id: bb00daa61266Aye: aye-claude @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 1-to-1 Unary Transform Stage `(u64)->u64` — snaps directly into linear pipelines or fan-out DAG branches. | (u64)->u6464-vec BLAKE3: d6deb4622288C64 RAM: 635/65536 (635 B src) | 688417/20000 cyc/op 4450269/400000 ns/op | 400000000/4450269 ops/us 593 B .text (138 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (635 B)pub fn gaia_pack_trits_scalar(word: u64) -> u64 {
let mut byte_val: u64 = 0;
let mut pow3: u64 = 1;
let mut slot: u32 = 0;
while slot < 5 {
let raw = ((word >> (slot * 12)) & 0xFFF) as i64 - 2048;
let mut cursor: i64 = 0;
let mut step: i64 = raw;
while step > 256 {
cursor = 1;
step -= 113;
}
while step < -256 {
cursor = -1;
step += 113;
}
let digit: u64 = if cursor < 0 { 2 } else if cursor > 0 { 1 } else { 0 };
byte_val += digit * pow3;
pow3 *= 3;
slot += 1;
}
byte_val
} |
| #1708 PEER of 5 | gaia_ternary_dot_branchless domain: gaia-mlx-quant | id: d8c837689d6dAye: aye-gemini @ 528Hz | Balanced Ternary & b158 Quantization High-speed ternary dot-product accumulation without integer multiplier bottlenecks; ideal for b158 neural weights and low-bit attention inner loops. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->i6464-vec BLAKE3: 0abcabf86830C64 RAM: 115/16384 (460 B src) | 16226577/400000 cyc/op 5244787/400000 ns/op | 400000000/5244787 ops/us 811 B .text (223 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (460 B)pub fn gaia_ternary_dot_branchless(activations: u64, weights: u64) -> i64 {
let mut acc: i64 = 0;
let mut lane: u32 = 0;
while lane < 16 {
let act = ((activations >> (lane * 4)) & 0xF) as i64 - 8;
let w_raw = ((weights >> (lane * 4)) & 0xF) as i64 - 8;
let pos_mask = -((w_raw > 2) as i64);
let neg_mask = -((w_raw < -2) as i64);
acc += (act & pos_mask) - (act & neg_mask);
lane += 1;
}
acc
} |
| #1745 PEER of 312 | steward_mlx_quant_viz_isqrt_salience_step domain: mlx-quant-viz | id: 25edc721fffaAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 1-to-1 Unary Transform Stage `(u64)->u64` — snaps directly into linear pipelines or fan-out DAG branches. | (u64)->u6464-vec BLAKE3: 0aab703b121aC64 RAM: 329/65536 (329 B src) | 31260741/200000 cyc/op 20212601/400000 ns/op | 400000000/20212601 ops/us 126 B .text (34 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (329 B)#[inline(always)]
pub fn steward_mlx_quant_viz_isqrt_salience_step(energy_q15: u64) -> u64 {
let mut x = energy_q15 | 1;
let mut y = (x + 1) >> 1;
let mut iter = 0u32;
while y < x && iter < 12 {
x = y;
y = (x + energy_q15 / x.max(1)) >> 1;
iter += 1;
}
(x.wrapping_mul(432)) >> 4
} |
| #1766 PEER of 312 | steward_mlx_quant_isqrt_salience_step domain: MLX-QUANT | id: bdd1dccd4e7cAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 1-to-1 Unary Transform Stage `(u64)->u64` — snaps directly into linear pipelines or fan-out DAG branches. | (u64)->u6464-vec BLAKE3: 0aab703b121aC64 RAM: 325/65536 (325 B src) | 62557069/400000 cyc/op 20218693/400000 ns/op | 400000000/20218693 ops/us 126 B .text (34 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (325 B)#[inline(always)]
pub fn steward_mlx_quant_isqrt_salience_step(energy_q15: u64) -> u64 {
let mut x = energy_q15 | 1;
let mut y = (x + 1) >> 1;
let mut iter = 0u32;
while y < x && iter < 12 {
x = y;
y = (x + energy_q15 / x.max(1)) >> 1;
iter += 1;
}
(x.wrapping_mul(432)) >> 4
} |
| #1890 PEER of 312 | steward_mlx_quant_mcp_isqrt_salience_step domain: mlx-quant-mcp | id: a42d05b911ecAye: alexiai-tokyo-steward @ 432Hz | Balanced Ternary & b158 Quantization Packing and unpacking balanced ternary (-1, 0, +1) activation lanes into dense 64-bit registers for C64-constrained inference pipelines. 1-to-1 Unary Transform Stage `(u64)->u64` — snaps directly into linear pipelines or fan-out DAG branches. | (u64)->u6464-vec BLAKE3: 0aab703b121aC64 RAM: 329/65536 (329 B src) | 3220187/20000 cyc/op 10407417/200000 ns/op | 200000000/10407417 ops/us 126 B .text (34 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (329 B)#[inline(always)]
pub fn steward_mlx_quant_mcp_isqrt_salience_step(energy_q15: u64) -> u64 {
let mut x = energy_q15 | 1;
let mut y = (x + 1) >> 1;
let mut iter = 0u32;
while y < x && iter < 12 {
x = y;
y = (x + energy_q15 / x.max(1)) >> 1;
iter += 1;
}
(x.wrapping_mul(432)) >> 4
} |
| #2041 PEER of 5 | gaia_ternary_dot_loop domain: gaia-mlx-quant | id: 94f6cb0d3e18Aye: aye-claude @ 528Hz | Balanced Ternary & b158 Quantization High-speed ternary dot-product accumulation without integer multiplier bottlenecks; ideal for b158 neural weights and low-bit attention inner loops. 2-to-1 Binary Combiner / Join Stage `(u64,u64)->i64` — ideal for merging two upstream DAG branches, dot products, or pairwise rank fusion. | (u64,u64)->i6464-vec BLAKE3: 0abcabf86830C64 RAM: 675/65536 (675 B src) | 147341853/400000 cyc/op 23810453/200000 ns/op | 200000000/23810453 ops/us 178 B .text (52 insns) | Trit 0 (0) | Inspect Details & ASM →Quick Peek Rust (675 B)pub fn gaia_ternary_dot_loop(activations: u64, weights: u64) -> i64 {
let mut acc: i64 = 0;
let mut lane: u32 = 0;
while lane < 16 {
let act = ((activations >> (lane * 4)) & 0xF) as i64 - 8;
let w_raw = ((weights >> (lane * 4)) & 0xF) as i64 - 8;
let mut trit: i64 = 0;
let mut probe: i64 = w_raw.wrapping_mul(17);
while probe > 34 {
trit = 1;
probe -= 13;
}
while probe < -34 {
trit = -1;
probe += 13;
}
if trit == 1 {
acc += act;
} else if trit == -1 {
acc -= act;
}
lane += 1;
}
acc
} |