code wiki / (root) / nx_fma_unroll_matmul_gate.nx

nx_fma_unroll_matmul_gate.nx

buildroot/runtime/nx_fma_unroll_matmul_gate.nx

7511 B150 linesdepth 2pulls 2 transitivereach 0 importersview sourcekind gate/proof
docsdependenciesstructsconstsfunctions

about

nx_fma_unroll_matmul_gate.nx -- unlock the FMA throughput win: MULTIPLE independent accumulators. The single-accumulator FMA (nx_fma_matmul_gate) was only ~1.03x over AVX2-dot because each vfmadd231ps depends on the previous (store->load->fma serial chain) = LATENCY-bound. FMA has ~4-5 cycle latency but ~0.5 cycle throughput, so to fill the pipeline you need >=4 INDEPENDENT accumulators in flight. This unrolls the k-loop by 4 (a0..a3, each its own 8-wide accumulator, disjoint k-chunks) so the 4 FMAs have NO inter-dependency -> the CPU pipelines them -> hides the latency. NishiLang-only (reuses __f32x8_fma/ __f32x8_hsum) -- NO compiler change. Same data, BIT-EXACT vs scalar. No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL

dependencies 1 imports · 0 importers

nx_syscalls.nx nx_fma_unroll_matmul_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 gx_puts sys_write sys_mmap pack4 mm_scalar mm_fma1 mm_fma4 sys_now_us sys_mmap ↻ sys_clock_gettime_mono gx_num sys_mmap ↻ sys_write ↻ sys_exit

structs

none

consts

none

functions

11func gx_puts(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
12func gx_num(v: i64) -> i64 { let b: *u8=sys_mmap(28); var m: i64=v; if m<0{m=0-m;sys_write(1,"-" as *u8,1)} let t: *u8=sys_mmap(28); var k: i64=0; if m==0{t[0]=48 as u8;k=1} while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} var i: i64=0; while i<k{b[i]=t[k-1-i];i=i+1} sys_write(1,b,k); return 0 }
called by 1: main calls 2: sys_mmapsys_write
13func pack4(buf: *u8, idx: i64, bits: i64) -> i64 { buf[idx*4+0]=(bits) as u8; buf[idx*4+1]=(bits>>8) as u8; buf[idx*4+2]=(bits>>16) as u8; buf[idx*4+3]=(bits>>24) as u8; return 0 }
called by 1: main
15func mm_scalar(sa: *i64, sb: *i64, c: *i64, M: i64, N: i64, K: i64) -> i64
called by 1: main
25func mm_fma1(pa: *u8, pb: *u8, c: *i64, acc: *u8, M: i64, N: i64, K: i64) -> i64
called by 1: main
38func mm_fma4(pa: *u8, pb: *u8, c: *i64, a0: *u8, a1: *u8, a2: *u8, a3: *u8, M: i64, N: i64, K: i64) -> i64
called by 1: main
69func main() -> i64