code wiki / _hdl_build / nx_intfp_gradcheck_gate.nx

nx_intfp_gradcheck_gate.nx

buildroot/runtime/_hdl_build/nx_intfp_gradcheck_gate.nx

5018 B80 linesdepth 2pulls 2 transitivereach 0 importersview sourcekind gate/prooftopic intfp
docsdependenciesstructsconstsfunctions

about

nx_intfp_gradcheck_gate.nx -- FIRST verified brick of FIXED-POINT INTEGER TRAINING (the sovereign fix for the measured 25x software-float tax / 8x training-speed loss vs PyTorch). A Q16 linear layer y = W@x with loss L = sum(y^2), done ENTIRELY in integer fixed-point (scale S=2^16): forward, analytic backprop, and a gold-standard INTEGER finite-difference gradcheck (numeric grad from perturbing the loss vs analytic grad). No float anywhere -- if analytic == numeric within tolerance, integer fwd+bwd is numerically SOUND and the full integer autograd tape can be built on it (=fast path, 358+ MMAC/s scalar / 8.4 Gop/s SIMD, deterministic). Q-format bookkeeping (S=65536): y_q[o] = (sum_k W_q*x_q) >> 16 (Q32 accum -> Q16) L_q32 = sum_o y_q*y_q (Q16^2 = Q32, kept unshifted for precision) analytic dL/dW[o,k] = 2*y*x -> gW_q = (2*y_q*x_q) >> 16 (Q16 of the real gradient) numeric dL/dW[o,k] -> (L+_q32 - L-_q32) / (2*delta_q16) (also Q16 -- units cancel to match analytic) license_tier: ORIGINAL

dependencies 1 imports · 0 importers

nx_syscalls.nx nx_intfp_gradcheck_gate.nx

imports: nx_syscalls.nx

imported by: nobody (leaf or entry point)

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

main w sys_write sys_mmap fwd_loss wn sys_write ↻ sys_mmap ↻ iabs

structs

none

consts

19const S: i64 = 65536 // Q16 scale
20const O: i64 = 3
21const K: i64 = 4

functions

15func w(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 }
called by 1: main calls 1: sys_write
16func wn(v: i64) -> i64 { if v==0 { sys_write(1,"0" as *u8,1); return 0 } var m: i64=v; if m<0{sys_write(1,"-" as *u8,1);m=0-m} let t: *u8=sys_mmap(24); var k: i64=0; while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} let o: *u8=sys_mmap(24); var q: i64=k-1; var i: i64=0; while q>=0{o[i]=t[q];i=i+1;q=q-1} sys_write(1,o,i); return 0 }
called by 1: main calls 2: sys_writesys_mmap
17func iabs(v: i64) -> i64 { if v<0 { return 0-v } return v }
called by 1: main
24func fwd_loss(W: *i64, x: *i64, ybuf: *i64) -> i64
called by 1: main
35func main() -> i64