"""Draw the three structure figures of simd-hash.md (hand-laid-out SVG). Usage: python draw_figures.py (writes ../xxh3-accumulate-lanes.svg, ../xxh3-block-secret.svg, ../wyhash-three-chains.svg) """ import pathlib OUT = pathlib.Path(__file__).resolve().parent.parent def draw_acc(): W,H=980,600 o=[] a=o.append a(f'') a('' '' '') a(f'') a('XXH3_accumulate_512: one 64-byte stripe, 8 lanes x 64 bit') X0,LW,G=190,86,11 def lx(i): return X0+i*(LW+G) def label(y,t1,t2=None): a(f'{t1}') if t2: a(f'{t2}') def box(x,y,w,h,stroke,fill,txt,tc="#cdd9e5",fs=13): a(f'') a(f'{txt}') def arrow(x1,y1,x2,y2,m="ar",c="#768390"): a(f'') y1,y2,y3,y4,y5=60,120,190,265,350 label(y1,"input","64 B stripe") label(y2,"secret","+ 8 bytes * n") label(y3,"dk = d ^ k") label(y4,"p | swap(d)","lo32*hi32 | d[i^1]") label(y5,"acc[i] +=","p + d[i^1]") for i in range(8): x=lx(i) box(x,y1,LW,32,"#636e7b","#22272e",f"d{i}") box(x+24,y2,LW-24,32,"#a371f7","#2f2a3d",f"k{i}") arrow(x+10,y1+32,x+10,y3-3) arrow(x+24+(LW-24)/2,y2+32,x+24+(LW-24)/2,y3-3) box(x,y3,LW,32,"#636e7b","#22272e",f"dk{i}") hw=(LW-6)/2 box(x,y4,hw,32,"#f0883e","#3a2f24",f"p{i}") box(x+hw+6,y4,hw,32,"#3fb950","#233327",f"d{i^1}") arrow(x+hw/2,y3+32,x+hw/2,y4-3,"ao","#f0883e") arrow(x+hw/2,y4+32,x+LW/2-8,y5-3,"ao","#f0883e") arrow(x+hw+6+hw/2,y4+32,x+LW/2+8,y5-3,"ag","#3fb950") box(x,y5,LW,34,"#388bfd","#1f2d44",f"acc{i}") # register grouping def bracket(i0,i1,y,c,txt): xa=lx(i0); xb=lx(i1)+LW a(f'') a(f'{txt}') yb=410 a(f'AVX-512') bracket(0,7,yb,"#cdd9e5","1 x __m512i (zmm), 1 iteration per stripe") yb+=44 a(f'AVX2') bracket(0,3,yb,"#388bfd","__m256i (ymm) #0"); bracket(4,7,yb,"#388bfd","__m256i (ymm) #1") yb+=44 a(f'SSE2') for j in range(4): bracket(2*j,2*j+1,yb,"#a371f7",f"xmm #{j}") yb+=44 a(f'NEON') a(f'AArch64, non-Apple') bracket(0,5,yb,"#3fb950","3 x uint64x2_t (NEON)"); bracket(6,7,yb,"#f0883e","scalar") a('') open(OUT / 'xxh3-accumulate-lanes.svg', 'w').write('\n'.join(o)) def draw_block(): W,H=980,470 o=[];a=o.append a(f'') a(f'') a('XXH3 long input (len > 240): stripes, blocks, 192-byte default secret') def t(x,y,s,c="#adbac7",fs=12,anc="middle"): a(f'{s}') def rect(x,y,w,h,st,fi,sw=1.4,dash=None): d=f' stroke-dasharray="{dash}"' if dash else '' a(f'') # ---- input row ---- t(40,62,"input",anc="start",c="#cdd9e5",fs=13) y=72; h=34 bx=[40,300,560]; bw=240 for j,x in enumerate(bx[:2]): rect(x,y,bw,h,"#388bfd","#1f2d44") for k in range(1,16): xx=x+k*bw/16 a(f'') t(x+bw/2,y+h+18,f"block {j}: 16 stripes x 64 B = 1024 B",fs=11) # partial block px=560; pw=6*bw/16 rect(px,y,pw,h,"#388bfd","#1f2d44") for k in range(1,6): xx=px+k*bw/16 a(f'') rect(px+pw,y,50,h,"#636e7b","#22272e") t(px+pw/2+25,y+h+18,"last partial block",fs=11) t(px+pw/2+25,y+h+32,"(len-1-1024*nb)/64 stripes",fs=11,c="#768390") # last stripe overlapping end=px+pw+50 rect(end-bw/16*1.6,y-8,bw/16*1.6,h+16,"#f85149","none",sw=1.8,dash="4 3") t(end+8,y+14,"last stripe:",anc="start",c="#f85149",fs=11) t(end+8,y+28,"input[len-64, len)",anc="start",c="#f85149",fs=11) t(end+8,y+42,"(may overlap)",anc="start",c="#f85149",fs=11) # scramble markers for x in (300,560): a(f'') t(x-10,y-22,"scramble",c="#f0883e",fs=11) # ---- secret row ---- sy=250; sx=100; ppb=4 # px per byte t(40,sy-100,"secret (192 B)",anc="start",c="#cdd9e5",fs=13) rect(sx,sy,192*ppb,30,"#a371f7","#2f2a3d") for b in (0,64,128,192): X=sx+b*ppb a(f'') t(X,sy+52,str(b),c="#768390",fs=11) def span(b0,b1,yy,c,label,above=True): x0=sx+b0*ppb; x1=sx+b1*ppb d=6 if above else -6 a(f'') t((x0+x1)/2,yy-6 if above else yy+16,label,c=c,fs=11) span(0,64,sy-72,"#388bfd","stripe 0 key [0,64)") span(120,184,sy-72,"#388bfd","stripe 15 key [120,184)") span(8,72,sy-40,"#388bfd","stripe 1 [8,72)") t(sx+100*ppb,sy-38,"... +8 B per stripe ...",c="#768390",fs=11) span(11,75,sy+80,"#3fb950","merge [11,75)",above=False) span(128,192,sy+80,"#f0883e","scramble key [128,192)",above=False) span(121,185,sy+122,"#f85149","last-stripe key [121,185)",above=False) t(490,H-18,"16 stripes per block = (secretSize - 64) / 8; offsets 121 = 192-64-7 and 11 are deliberately not multiples of 8",c="#768390",fs=11) a('') open(OUT / 'xxh3-block-secret.svg', 'w').write('\n'.join(o)) def draw_wy(): W,H=1000,400 o=[];a=o.append a(f'') a('') a(f'') a('wyhash final4, len > 48: three independent _wymix chains') def t(x,y,s,c="#adbac7",fs=12,anc="middle"): a(f'{s}') def box(x,y,w,h,st,fi,lines,cols=None): a(f'') n=len(lines); y0=y+h/2-(n-1)*8+4 for i,l in enumerate(lines): t(x+w/2,y0+i*16,l,c=(cols[i] if cols else "#cdd9e5"),fs=12) def arrow(x1,y1,x2,y2): a(f'') rows=[("seed",1,0,"#388bfd"),("see1",2,16,"#3fb950"),("see2",3,32,"#a371f7")] ys=[70,150,230]; bh=60 for (name,si,off,c),y in zip(rows,ys): t(60,y+bh/2+5,name,c=c,fs=14,anc="end") for r,x in enumerate((80,300)): o0=off+48*r box(x,y,190,bh,c,"#22272e",[f"{name} = _wymix(",f"p[{o0}:{o0+8}] ^ s[{si}],",f"p[{o0+8}:{o0+16}] ^ {name})"],None) arrow(270,y+bh/2,298,y+bh/2) arrow(490,y+bh/2,512,y+bh/2) t(530,y+bh/2+5,"...",c="#768390",fs=14) arrow(550,y+bh/2,578,y+bh/2) t(175,62,"round 0 (bytes 0..47)",c="#768390",fs=11) t(395,62,"round 1 (bytes 48..95)",c="#768390",fs=11) box(580,70,130,220,"#f0883e","#3a2f24",["after loop:","seed ^=","see1 ^ see2"]) box(760,70,220,100,"#636e7b","#22272e",["tail, while i > 16:","seed = _wymix(p[0:8]^s[1],","p[8:16]^seed)","(one serial chain)"]) box(760,190,220,100,"#388bfd","#1f2d44",["a,b = last 16 bytes","a^=s[1]; b^=seed;","_wymum(&a,&b)","_wymix(a^s[0]^len, b^s[1])"]) arrow(710,120,758,120); arrow(870,170,870,188) t(500,330,"_wymix(A,B) = lo64(A*B) ^ hi64(A*B), one 64x64 -> 128-bit scalar multiply (x86-64 MUL / MULX, AArch64 MUL + UMULH)",c="#adbac7",fs=11) t(500,352,"the three chains read disjoint 16-byte slices and never feed each other inside the loop, so their multiplies overlap",c="#768390",fs=11) t(500,374,"s[0..3] = secret (default _wyp); len <= 16 skips the loop and builds a, b from 4-byte or 1..3-byte reads",c="#768390",fs=11) a('') open(OUT / 'wyhash-three-chains.svg', 'w').write('\n'.join(o)) if __name__ == "__main__": draw_acc() draw_block() draw_wy()