code wiki / _hdl_build / nx_f32_embed_lmhead_gate.nx

nx_f32_embed_lmhead_gate.nx

buildroot/runtime/_hdl_build/nx_f32_embed_lmhead_gate.nx

12706 B161 linesdepth 4pulls 5 transitivereach 0 importersview sourcekind gate/prooftopic f32
docsdependenciesstructsconstsfunctions

dependencies 4 imports · 0 importers

nx_gate_gn.nx nx_gate_base.nx nx_f32_hw.nx nx_syscalls.nx nx_f32_embed_lmhead_gate.nx

imports: nx_gate_gn.nxnx_gate_base.nxnx_f32_hw.nxnx_syscalls.nx

imported by: nobody (leaf or entry point)

call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown

main gw sys_write f32_div f32_of f32_mul sys_mmap f32_int gm gn f32_int ↻ f32_mul ↻ f32_of ↻ fwd_L sys_mmap ↻ f32_of ↻ f32_add f32_mul ↻ softmax3 f32_max3 f32_le f32_sub f32_neg f32_exp f32_div ↻ f32_of ↻ f32_mul ↻ f32_le ↻ f32_int ↻ f32_add ↻ f32_sub ↻ f32_sub ↻ f32_add ↻ f32_div ↻ f32_neg ↻ f32_log f32_div ↻ f32_sub ↻ f32_of ↻ f32_add ↻

structs

none

consts

34const V: i64 = 3
35const D: i64 = 3 // D >= V so logits=E*W has rank to represent the V-way transition (D=2 underfit a 3-cycle)

functions

18func grow(name: *u8, ok: i64) -> i64 { if ok==1 { gw(" PASS " as *u8) } else { gw(" FAIL " as *u8) } gw(name); gw("
calls 1: gw
20func gm(x: i64) -> i64 { return gn(f32_int(f32_mul(x, f32_of(1000)))) }
called by 3: f32_max3f32_expmain calls 1: f32_sub
21func f32_le(x: i64, y: i64) -> i64 { let d: i64=f32_sub(x,y) & 0xFFFFFFFF; if ((d>>31)&1)==1 { return 1 } if (d & 0x7FFFFFFF)==0 { return 1 } return 0 }
called by 1: main
22func f32_abs(x: i64) -> i64 { return x & 0x7FFFFFFF }
called by 1: softmax3 calls 1: f32_le
23func f32_max3(a: i64,b: i64,c: i64) -> i64 { var m: i64=a; if f32_le(m,b)==1 { m=b } if f32_le(m,c)==1 { m=c } return m }
called by 1: main calls 3: f32_off32_divf32_add
24func f32_sqrt(x: i64) -> i64 { if (x & 0x7FFFFFFF)==0 { return f32_of(0) } var y: i64=x; var i: i64=0; while i<16 { y=f32_div(f32_add(y, f32_div(x,y)), f32_of(2)); i=i+1 } return y }
called by 1: softmax3 calls 7: f32_divf32_off32_mulf32_lef32_intf32_add+1
25func f32_exp(x: i64) -> i64
32func f32_log(x: i64) -> i64 { let b: i64=x & 0xFFFFFFFF; let e: i64=((b>>23)&0xFF)-127; let m: i64=(b & 0x7FFFFF)|0x3F800000; let u: i64=f32_div(f32_sub(m,f32_of(1)),f32_add(m,f32_of(1))); let u2: i64=f32_mul(u,u); var t: i64=u; var s: i64=u; var k: i64=1; while k<=7 { t=f32_mul(t,u2); s=f32_add(s,f32_div(t,f32_of((2*k)+1))); k=k+1 } let ln2: i64=f32_div(f32_of(693147),f32_of(1000000)); return f32_add(f32_mul(f32_of(e),ln2),f32_mul(f32_of(2),s)) }
36func softmax3(z: *i64, out: *i64) -> i64 { let mx: i64=f32_max3(z[0],z[1],z[2]); let e0: i64=f32_exp(f32_sub(z[0],mx)); let e1: i64=f32_exp(f32_sub(z[1],mx)); let e2: i64=f32_exp(f32_sub(z[2],mx)); let s: i64=f32_add(f32_add(e0,e1),e2); out[0]=f32_div(e0,s); out[1]=f32_div(e1,s); out[2]=f32_div(e2,s); return 0 }
38func fwd_L(E: *i64, W: *i64, t: i64, tgt: i64) -> i64
45func main() -> i64