// Fold (´) and Insert (˝) // Fold optimized operands: // ⊣⊢ on all types // +-∧∨=≠ and synonyms on booleans // ≤<>≥ on booleans, monadic only, with a search // +⌈⌊× on numbers // Integer +: sum blocks associatively as long as sum can't exceed +-2⋆53 // COULD implement fast numeric -´ // ∨ on boolean-valued integers, stopping at 1 // •math.Sum: +´ with faster and more precise SIMD code for i32, f64 #include "../core.h" #include "../builtins.h" #include "../utils/calls.h" #include "../utils/mut.h" #if SINGELI_SIMD #define SINGELI_FILE fold #include "../utils/includeSingeli.h" #endif static bool fold_ne(u64* x, u64 am) { u64 r = 0; for (u64 i = 0; i < (am>>6); i++) r^= x[i]; if (am&63) r^= x[am>>6]<<(64-am & 63); return POPC(r) & 1; } static i64 bit_diff(u64* x, u64 am) { i64 r = 0; u64 a = 0xAAAAAAAAAAAAAAAA; for (u64 i = 0; i < (am>>6); i++) r+= POPC(x[i]^a); if (am&63) r+= POPC((x[am>>6]^a)<<(64-am & 63)); return r - (i64)(am/2); } // It's safe to sum a block of integers as long as the current total // is far enough from +-1ull<<53 (and integer, in dyadic fold). static const usz sum_small_max = 1<<16; #define DEF_INT_SUM(T,W,M,A) \ static i64 sum_small_##T(void* xv, usz ia) { \ i##A s=0; for (usz i=0; i0 && -lim<=c && c<=lim) { \ usz e=i0; i0=(i0-1)&~(b-1); \ c+=sum_small_##T(xp+i0, e-i0); \ } \ r = c; \ } \ while (i0--) r+=xp[i0]; \ return r; \ } DEF_INT_SUM(i8 ,8 ,32,32) DEF_INT_SUM(i16,16,32,32) DEF_INT_SUM(i32,32,52,64) #undef DEF_SUM static f64 sum_f64(void* xv, usz i, f64 r) { while (i--) r += ((f64*)xv)[i]; return r; } static i64 (*const sum_small_fns[])(void*, usz) = { sum_small_i8, sum_small_i16, sum_small_i32 }; static f64 (*const sum_fns[])(void*, usz, f64) = { sum_i8, sum_i16, sum_i32, sum_f64 }; B sum_c1(B t, B x) { if (isAtm(x) || RNK(x)!=1) thrF("•math.Sum: Argument must be a list (%H ≡ ≢𝕩)", x); usz ia = IA(x); if (ia==0) return m_f64(0); u8 xe = TI(x,elType); if (!elNum(xe)) { x = any_squeeze(x); xe = TI(x,elType); if (!elNum(xe)) thrF("•math.Sum: Argument elements must be numbers", x); } f64 r; void* xv = tyany_ptr(x); if (xe == el_bit) { r = bit_sum(xv, ia); } else if (xe <= el_i32) { u8 sel = xe - el_i8; i64 s = 0; r = 0; i64 m = 1ull<<48; usz b = sum_small_max; for (usz i=0; i= m) { r+=m; s-=m; } if (s <= -m) { r-=m; s+=m; } } r += s; } else { #if SINGELI_SIMD r = simd_sum_f64(xv, ia); #else r=0; for (usz i=0; i) } DEF_MIN_MAX(i8) DEF_MIN_MAX(i16) DEF_MIN_MAX(i32) #if SINGELI_SIMD static f64 min_f64(void* xv, usz ia) { return simd_fold_min_f64(xv,ia); } static f64 max_f64(void* xv, usz ia) { return simd_fold_max_f64(xv,ia); } #else DEF_MIN_MAX(f64) #endif #undef DEF_MIN_MAX #undef MIN_MAX static f64 (*const min_fns[])(void*, usz) = { min_i8, min_i16, min_i32, min_f64 }; static f64 (*const max_fns[])(void*, usz) = { max_i8, max_i16, max_i32, max_f64 }; B fold_c1(Md1D* d, B x) { B f = d->f; if (isAtm(x) || RNK(x)!=1) thrF("´: Argument must be a list (%H ≡ ≢𝕩)", x); usz ia = IA(x); if (ia<=2) { if (ia==2) { SGet(x) B x0 = Get(x,0); B x1 = Get(x,1); decG(x); return c2(f, x0, x1); } else if (ia==1) { B r = IGet(x,0); decG(x); return r; } else { decG(x); if (isFun(f)) { B r = TI(f,identity)(f); if (!q_N(r)) return inc(r); } thrM("´: No identity found"); } } u8 xe = TI(x,elType); if (isFun(f) && v(f)->flags) { u8 rtid = v(f)->flags-1; if (rtid==n_ltack) { B r = IGet(x, 0 ); decG(x); return r; } if (rtid==n_rtack) { B r = IGet(x, ia-1); decG(x); return r; } if (xe>el_f64) goto base; if (xe==el_bit) { u64* xp = bitarr_ptr(x); f64 r; switch (rtid) { default: goto base; case n_add: r = bit_sum (xp, ia); break; case n_sub: r = bit_diff(xp, ia); break; case n_and: case n_mul: case n_floor: r = bit_has (xp, ia, 0) ^ 1; break; case n_or: case n_ceil: r = bit_has (xp, ia, 1) ; break; case n_ne: r = fold_ne (xp, ia) ; break; case n_eq: r = fold_ne (xp, ia) ^ (1&~ia); break; case n_lt: r = bit_find(xp, ia, 1) == ia-1; break; case n_le: r = bit_find(xp, ia, 0) != ia-1; break; case n_gt: r = bit_find(xp, ia, 0) & 1; break; case n_ge: r =~bit_find(xp, ia, 1) & 1; break; } decG(x); return m_f64(r); } if (rtid==n_add) { // + void *xv = tyany_ptr(x); bool small = xe<=el_i32 & ia<=sum_small_max; u8 sel = xe - el_i8; f64 r = small ? sum_small_fns[sel](xv, ia) : sum_fns[sel](xv, ia, 0); decG(x); return m_f64(r); } if (rtid==n_floor) { f64 r=min_fns[xe-el_i8](tyany_ptr(x), ia); decG(x); return m_f64(r); } // ⌊ if (rtid==n_ceil ) { f64 r=max_fns[xe-el_i8](tyany_ptr(x), ia); decG(x); return m_f64(r); } // ⌈ if (rtid==n_mul | rtid==n_and) { // ×/∧ void *xv = tyany_ptr(x); u8 sel = xe - el_i8; f64 r = xe<=el_i32 ? prod_int_fns[sel](xv, ia, 1) : prod_fns[sel](xv, ia, 1); decG(x); return m_f64(r); } if (rtid==n_or) { // ∨ if (xe==el_i8 ) { i8* xp = i8any_ptr (x); usz i=ia; while (i--) { i8 c=xp[i]; if (c==1) break; if (c!=0) goto base; } decG(x); return m_i32(i+1 > 0); } if (xe==el_i16) { i16* xp = i16any_ptr(x); usz i=ia; while (i--) { i16 c=xp[i]; if (c==1) break; if (c!=0) goto base; } decG(x); return m_i32(i+1 > 0); } if (xe==el_i32) { i32* xp = i32any_ptr(x); usz i=ia; while (i--) { i32 c=xp[i]; if (c==1) break; if (c!=0) goto base; } decG(x); return m_i32(i+1 > 0); } } } base:; SLOW2("𝕎´ 𝕩", f, x); SGet(x) FC2 fc2 = c2fn(f); B c; if (TI(x,elType)==el_i32) { i32* xp = i32any_ptr(x); c = m_i32(xp[ia-1]); for (usz i = ia-1; i>0; i--) c = fc2(f, m_i32(xp[i-1]), c); } else { c = Get(x, ia-1); for (usz i = ia-1; i>0; i--) c = fc2(f, Get(x, i-1), c); } decG(x); return c; } B fold_c2(Md1D* d, B w, B x) { B f = d->f; if (isAtm(x) || RNK(x)!=1) thrF("´: 𝕩 must be a list (%H ≡ ≢𝕩)", x); usz ia = IA(x); u8 xe = TI(x,elType); if (isFun(f) && v(f)->flags) { u8 rtid = v(f)->flags-1; if (rtid==n_ltack) { B r = w; if (ia) { dec(w); r=IGet(x, 0); } decG(x); return r; } if (rtid==n_rtack) { decG(x); return w; } if (!isF64(w) || xe>el_f64) goto base; f64 wf = o2fG(w); if (xe==el_bit) { i32 wi = wf; if (wi!=wf) goto base; u64* xp = bitarr_ptr(x); if (rtid==n_add) { B r = m_f64(wi + bit_sum (xp, ia)); decG(x); return r; } if (rtid==n_sub) { B r = m_f64((ia&1?-wi:wi) + bit_diff(xp, ia)); decG(x); return r; } if (wi!=(wi&1)) goto base; if (rtid==n_and | rtid==n_mul | rtid==n_floor) { B r = m_i32(wi && !bit_has(xp, ia, 0)); decG(x); return r; } if (rtid==n_or | rtid==n_ceil ) { B r = m_i32(wi || bit_has(xp, ia, 1)); decG(x); return r; } if (rtid==n_ne) { bool r=wi^fold_ne(xp, ia) ; decG(x); return m_i32(r); } if (rtid==n_eq) { bool r=wi^fold_ne(xp, ia) ^ (1&ia); decG(x); return m_i32(r); } goto base; } if (rtid==n_add) { // + u8 sel = xe - el_i8; f64 r = sum_fns[sel](tyany_ptr(x), ia, wf); decG(x); return m_f64(r); } if (rtid==n_floor) { f64 r=wf; if (ia>0) { f64 m=min_fns[xe-el_i8](tyany_ptr(x), ia); if (m0) { f64 m=max_fns[xe-el_i8](tyany_ptr(x), ia); if (m>r) r=m; } decG(x); return m_f64(r); } // ⌈ i32 wi = wf; if (rtid==n_mul | rtid==n_and) { // ×/∧ void *xv = tyany_ptr(x); bool isint = xe<=el_i32 && wi==wf; u8 sel = xe - el_i8; f64 r = isint ? prod_int_fns[sel](xv, ia, wi) : prod_fns[sel](xv, ia, wf); decG(x); return m_f64(r); } if (rtid==n_or && (wi&1)==wf) { // ∨ if (xe==el_i8 ) { i8* xp = i8any_ptr (x); usz i=ia; if (!wi) while (i--) { i8 c=xp[i]; if (c==1) break; if (c!=0) goto base; } decG(x); return m_i32(i+1 > 0); } if (xe==el_i16) { i16* xp = i16any_ptr(x); usz i=ia; if (!wi) while (i--) { i16 c=xp[i]; if (c==1) break; if (c!=0) goto base; } decG(x); return m_i32(i+1 > 0); } if (xe==el_i32) { i32* xp = i32any_ptr(x); usz i=ia; if (!wi) while (i--) { i32 c=xp[i]; if (c==1) break; if (c!=0) goto base; } decG(x); return m_i32(i+1 > 0); } } } base:; SLOW3("𝕨 F´ 𝕩", w, x, f); B c = w; SGet(x) FC2 fc2 = c2fn(f); for (usz i = ia; i>0; i--) c = fc2(f, Get(x, i-1), c); decG(x); return c; } u64 usum(B x) { // doesn't consume; will error on non-integers, or elements <0, or if sum overflows u64 assert(isArr(x)); u64 r = 0; usz xia = IA(x); u8 xe = TI(x,elType); if (xe==el_bit) return bit_sum(bitarr_ptr(x), xia); else if (xe==el_i8 ) { i8* p = i8any_ptr (x); i8 m=0; for (usz i = 0; i < xia; ) { usz b=1<< 8; i16 s=0; for (usz e = xia-if; if (isAtm(x) || RNK(x)==0) thrM("˝: 𝕩 must have rank at least 1"); usz xia = IA(x); if (xia==0) { SLOW2("!𝕎˝𝕩", f, x); return m1c1(rt_insert, f, x); } if (isFun(f)) { u8 rtid = v(f)->flags-1; ur xr = RNK(x); if (xr==1 && isPervasiveDyExt(f)) return m_unit(fold_c1(d, x)); if (rtid == n_join) { if (xr==1) return x; ShArr* rsh; if (xr>2) { rsh = m_shArr(xr-1); usz* xsh = SH(x); shcpy(rsh->a+1, xsh+2, xr-2); rsh->a[0] = xsh[0] * xsh[1]; } Arr* r = TI(x,slice)(x, 0, IA(x)); if (xr>2) arr_shSetUG(r, xr-1, rsh); else arr_shVec(r); return taga(r); } } return insert_base(f, x, xia, 0, bi_N); } B insert_c2(Md1D* d, B w, B x) { B f = d->f; if (isAtm(x) || RNK(x)==0) thrM("˝: 𝕩 must have rank at least 1"); usz xia = IA(x); B r = w; if (xia==0) { decG(x); return r; } if (isFun(f)) { if (RNK(x)==1 && isPervasiveDyExt(f)) { if (isAtm(w)) { to_fold: return m_unit(fold_c2(d, w, x)); } if (RNK(w)==0) { B w0=w; w = IGet(w,0); decG(w0); goto to_fold; } } } return insert_base(f, x, xia, 1, w); } // Arithmetic fold/insert on rows of flat rank-2 array x B transp_c1(B, B); B join_c2(B, B, B); B fold_rows(Md1D* fd, B x) { assert(isArr(x) && RNK(x)==2); // Target block size trying to avoid power-of-two lengths, from: // {𝕩/˜⌊´⊸= +˝˘ +˝¬∨`2|>⌊∘÷⟜2⍟(↕12) ⌊0.5+32÷˜𝕩÷⌜1+↕64} +⟜↕2⋆16 u64 block = (116053*8) >> arrTypeBitsLog(TY(x)); if (TI(x,elType)==el_bit || IA(x)/2 <= block) { x = C1(transp, x); return insert_c1(fd, x); } else { usz *sh = SH(x); usz n = sh[0]; usz m = sh[1]; usz b = (block + m - 1) / m; // Normal block length usz b_max = b + b/4; // Last block max length MAKE_MUT(r, n); MUT_APPEND_INIT(r); BSS2A slice = TI(x,slice); for (usz i=0, im=0; i b_max) { incG(x); l = b; } usz sia = l * m; Arr* sl = slice(x, im, sia); usz* ssh = arr_shAlloc(sl, 2); ssh[0] = l; ssh[1] = m; B sr = insert_c1(fd, C1(transp, taga(sl))); MUT_APPEND(r, sr, 0, l); decG(sr); i += l; im += sia; } return mut_fv(r); } }