code wiki / _hdl_build / nx_f32_train_ckpt_gate.nx

nx_f32_train_ckpt_gate.nx

buildroot/runtime/_hdl_build/nx_f32_train_ckpt_gate.nx

14018 B173 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_train_ckpt_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 sys_mmap f32_div f32_of mk_state sys_mmap ↻ f32_div ↻ f32_of ↻ train_run f32_div ↻ f32_of ↻ sys_mmap ↻ fwd f32_add f32_mul f32_le f32_sub f32_neg f32_of ↻ sys_mmap ↻ f32_exp f32_div ↻ f32_of ↻ f32_mul ↻ f32_le ↻ f32_int f32_add ↻ f32_sub ↻ f32_sub ↻ f32_div ↻ f32_neg ↻ f32_log f32_div ↻ f32_sub ↻ f32_of ↻ f32_add ↻ f32_mul ↻ bwd

structs

none

consts

14const NIN: i64 = 6
15const NH: i64 = 10
16const NOUT: i64 = 4
17const NSAMP: i64 = 8

functions

19func grow(name: *u8, ok: i64) -> i64 { if ok==1 { gw(" PASS " as *u8) } else { gw(" FAIL " as *u8) } gw(name); gw("
calls 1: gw
21func gm(x: i64) -> i64 { return gn(f32_int(f32_mul(x, f32_of(1000)))) }
called by 3: f32_expfwdbwd calls 1: f32_sub
22func 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: adam1 calls 3: f32_off32_divf32_add
23func 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: fwd calls 7: f32_divf32_off32_mulf32_lef32_intf32_add+1
24func f32_exp(x: i64) -> i64
31func 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 ck_wr4(fd: i64, v: i64) -> i64 { let b: *u8=sys_mmap(8); b[0]=(v & 0xff) as u8; b[1]=((v>>8)&0xff) as u8; b[2]=((v>>16)&0xff) as u8; b[3]=((v>>24)&0xff) as u8; sys_write(fd, b, 4); return 0 }
called by 1: ckpt_save calls 1: ck_wr4
37func ck_wr8(fd: i64, v: i64) -> i64 { ck_wr4(fd, v & 0xffffffff); ck_wr4(fd, (v>>32) & 0xffffffff); return 0 }
called by 1: main calls 4: sys_openat_wrck_wr4ck_wr8sys_close
38func ckpt_save(path: *u8, bufs: *i64, sizes: *i64, nbuf: i64, scalars: *i64, nscalar: i64) -> i64
48func ck_rd4(buf: *u8, o: i64) -> i64 { return (buf[o]&0xff)|((buf[o+1]&0xff)<<8)|((buf[o+2]&0xff)<<16)|((buf[o+3]&0xff)<<24) }
called by 1: ckpt_load calls 1: ck_rd4
49func ck_rd8(buf: *u8, o: i64) -> i64 { let lo: i64=ck_rd4(buf,o) & 0xffffffff; let hi: i64=ck_rd4(buf,o+4) & 0xffffffff; return lo | (hi<<32) }
51func ckpt_load(buf: *u8, blen: i64, bufs: *i64, sizes: *i64, nbuf: i64, scalars: *i64, nscalar: i64) -> i64
64func fwd(W1: *i64, b1: *i64, W2: *i64, b2: *i64, x: *i64, tgt: i64, hpre: *i64, h: *i64, p: *i64) -> i64
72func bwd(W2: *i64, tgt: i64, x: *i64, hpre: *i64, h: *i64, p: *i64, dW1: *i64, db1: *i64, dW2: *i64, db2: *i64) -> i64
79func adam1(P: *i64, G: *i64, Mo: *i64, Vo: *i64, cnt: i64, lr: i64, b1: i64, b2: i64, bc1: i64, bc2: i64, aeps: i64) -> i64
87func train_run(W1: *i64, b1: *i64, W2: *i64, b2: *i64, mW1: *i64, vW1: *i64, mb1: *i64, vb1: *i64, mW2: *i64, vW2: *i64, mb2: *i64, vb2: *i64, X: *i64, Y: *i64, scal: *i64, target_ep: i64) -> i64
111func mk_state(seed: i64) -> *i64
124func hstate(H: *i64) -> i64 { var acc: i64=1469598103; var b: i64=0; while b<4 { let P: *i64=H[b] as *i64; var sz: i64=NH*NIN; if b==1 { sz=NH } if b==3 { sz=NOUT } var c: i64=0; while c<sz { acc=(acc ^ (P[c] & 0xffffffff)) * 16777619; acc=acc & 0xffffffffffffff; c=c+1 } b=b+1 } return acc }
126func main() -> i64