nx_conv_gemm_gate.nx
buildroot/runtime/nx_conv_gemm_gate.nx
about
nx_conv_gemm_gate.nx -- wire the COMPUTE LEVER into the diffusion workhorse: f32 Conv2D via im2col->GEMM.
Convolution dominates diffusion U-Nets (what sdcpp/GGML run). The fast path (what GGML uses) is im2col
then GEMM -- so a 3x3 conv becomes Weight[Cout, Cin*9] x im2col_T[H*W, Cin*9]^T, reusing the FMA-4acc
matmul kernel built this session. This compares a NAIVE direct f32 conv (scalar __f32_mul/add) against
im2col + FMA-4acc GEMM, BIT-EXACT (small-int f32) + MEASURES the speedup. This is the first sovereign
f32 image-gen primitive accelerated by our AVX2/FMA stack -- the concrete start of the sdcpp benchmark path.
No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL
dependencies 2 imports · 0 importers
imports: nx_syscalls.nxnx_gate_verdict.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| none |
functions
| 11 | func cg_puts(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } |
| 12 | func cg_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 } |
| 13 | func 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 |
| 16 | func fma4dot(wr: i64, ir: i64, K: i64, a0: *u8, a1: *u8, a2: *u8, a3: *u8) -> i64 called by 1: main |
| 38 | func main() -> i64 |