Widen cells in 𝕨∊𝕩, 𝕨⊐𝕩, 𝕨⊒𝕩

This commit is contained in:
dzaima 2023-02-17 00:27:15 +02:00
parent 089e2a38c6
commit 32142a45d6
2 changed files with 134 additions and 55 deletions

View File

@ -1,3 +1,25 @@
// Dyadic search functions: Member Of (∊), Index of (⊐), Progressive Index of (⊒)
// 𝕨⊐unit or unit∊𝕩: scalar loop with early-exit
// SHOULD use simd
// SHOULD unify implementations
// 𝕩⊒unit or 𝕨⊒𝕩 where 1≥≠𝕩: defer to 𝕨⊐𝕩
// Both arguments with rank≥1:
// High-rank inputs:
// Convert to a (lower-rank) typed integer array if cells are ≤62 bits
// COULD have special hashing for equal type >64 bit cells, skipping squeezing
// COULD try conditionally squeezing ahead-of-time, and not squeezing in bqn_hash
// p⊐n & n∊p with short p & long n: n⊸=¨ p
// bitarr⊐𝕩: more special arithmetic
// SHOULD have impls for long p & short n
// ≤16-bit elements: lookup tables
// Character elements: reinterpret as integer elements
// Otherwise, generic hashtable
// SHOULD handle up to 64 bit cells via proper typed hash tables
// SHOULD have fast path when cell sizes or element types doesn't match
// SHOULD properly handle ¯0
#include "../core.h" #include "../core.h"
#include "../utils/hash.h" #include "../utils/hash.h"
#include "../utils/talloc.h" #include "../utils/talloc.h"
@ -41,20 +63,63 @@ static u64 elRange(u8 eltype) { return 1ull<<(1<<elWidthLogBits(eltype)); }
TFREE(tab0); TFREE(tab0);
typedef struct { B n, p; } B2; typedef struct { B n, p; } B2;
static NOINLINE B toIntCell(B x, ux csz0, ur co) {
assert(TI(x,elType)!=el_B);
usz ria = shProd(SH(x), 0, co);
ShArr* rsh;
if (co>1) { rsh=m_shArr(co); shcpy(rsh->a,SH(x),co); }
B r0 = widenBitArr(x, co);
usz csz = shProd(SH(r0),co,RNK(r0)) << elWidthLogBits(TI(r0,elType));
u8 t;
if (csz==8) t = t_i8slice;
else if (csz==16) t = t_i16slice;
else if (csz==32) t = t_i32slice;
else if (csz==64) t = t_f64slice;
else UD;
TySlice* r = m_arr(sizeof(TySlice), t, ria);
r->p = a(r0);
r->a = tyany_ptr(r0);
if (co>=1) arr_shSetU((Arr*)r, co, rsh);
else arr_shVec((Arr*)r);
return taga(r);
}
static NOINLINE B cpyToElLog(B x, u8 xe, u8 lb) {
switch(lb) { default: UD;
case 0: return taga(cpyBitArr(x));
case 3: return taga(elNum(xe)? cpyI8Arr(x) : cpyC8Arr(x));
case 4: return taga(elNum(xe)? cpyI16Arr(x) : cpyC16Arr(x));
case 5: return taga(elNum(xe)? cpyI32Arr(x) : cpyC32Arr(x));
case 6: return taga(cpyF64Arr(x));
}
}
static NOINLINE B2 splitCells(B n, B p, u8 mode) { // 0:∊ 1:⊐ 2:⊒ static NOINLINE B2 splitCells(B n, B p, u8 mode) { // 0:∊ 1:⊐ 2:⊒
#define SYMB (mode==0? "∊" : mode==1? "⊐" : "⊒") #define SYMB (mode==0? "∊" : mode==1? "⊐" : "⊒")
#define ARG_N (mode? "𝕩" : "𝕨") #define ARG_N (mode? "𝕩" : "𝕨")
#define ARG_P (mode? "𝕨" : "𝕩") #define ARG_P (mode? "𝕨" : "𝕩")
if (isAtm(p) || RNK(p)==0) thrF("%U: %U cannot have rank 0", SYMB, ARG_P); if (isAtm(p) || RNK(p)==0) thrF("%U: %U cannot have rank 0", SYMB, ARG_P);
ur pr = RNK(p); ur pr = RNK(p);
if (isAtm(n)) n = m_hunit(n); if (isAtm(n)) n = m_atomUnit(n);
ur nr = RNK(n); ur nr = RNK(n);
if (nr < pr-1) thrF("%U: Rank of %U must be at least the cell rank of %U (%H ≡ ≢𝕨, %H ≡ ≢𝕩)", SYMB, ARG_N, ARG_P, mode? p : n, mode? n : p); if (nr < pr-1) thrF("%U: Rank of %U must be at least the cell rank of %U (%H ≡ ≢𝕨, %H ≡ ≢𝕩)", SYMB, ARG_N, ARG_P, mode?p:n, mode?n:p);
ur cr = pr-1; ur pcr = pr-1;
n = toKCells(n, nr-cr); ur nco = nr-pcr;
p = toKCells(p, 1); if (nco>0 && eqShPart(SH(n)+nco, SH(p)+1, pcr)) {
assert(RNK(p)<=1); u8 ne = TI(n,elType);
return (B2){.n=n, .p=p}; u8 pe = TI(p,elType);
if (ne<el_B && pe<el_B && elNum(ne)==elNum(pe)) {
usz csz = arr_csz(p);
u8 neb = elWidthLogBits(ne);
u8 peb = elWidthLogBits(pe);
u8 meb = neb>peb? neb : peb;
ux rb = csz<<meb;
if (rb!=0 && rb<=62) {
if (neb!=meb) n = cpyToElLog(n, ne, meb);
else if (peb!=meb) p = cpyToElLog(p, pe, meb);
return (B2){.n=toIntCell(n,rb,nco), .p=toIntCell(p,rb,1)};
}
}
}
return (B2){.n=toKCells(n,nco), .p=toCells(p)};
#undef ARG_N #undef ARG_N
#undef ARG_P #undef ARG_P
@ -108,27 +173,33 @@ B indexOf_c2(B t, B w, B x) {
u8 we = TI(w,elType); usz wia = IA(w); u8 we = TI(w,elType); usz wia = IA(w);
u8 xe = TI(x,elType); usz xia = IA(x); u8 xe = TI(x,elType); usz xia = IA(x);
if (wia == 0) { B r=taga(arr_shCopy(allZeroes(xia), x)); decG(w); decG(x); return r; } if (wia == 0) { B r=taga(arr_shCopy(allZeroes(xia), x)); decG(w); decG(x); return r; }
if (we==el_bit && xe!=el_B) {
u64* wp = bitarr_ptr(w); if (elNum(we) && elNum(xe)) { tyEls:
u64 w0 = 1 & wp[0]; if (we==el_bit) {
u64 i = bit_find(wp, wia, !w0); decG(w); u64* wp = bitarr_ptr(w);
if (i!=wia) incG(x); u64 w0 = 1 & wp[0];
B r = C2i(mul, wia , C2i(ne, w0, x)) ; u64 i = bit_find(wp, wia, !w0); decG(w);
return i==wia? r : C2(sub, r, C2i(mul, wia-i, C2i(eq, !w0, x))); if (i!=wia) incG(x);
} B r = C2i(mul, wia , C2i(ne, w0, x)) ;
if (wia<=(we<=el_i16?4:16) && xia>16 && we<el_B && xe<el_B) { return i==wia? r : C2(sub, r, C2i(mul, wia-i, C2i(eq, !w0, x)));
SGetU(w); }
#define XEQ(I) C2(ne, GetU(w,I), incG(x))
B r = XEQ(wia-1); if (wia<=(we<=el_i16?4:16) && xia>16) {
for (usz i=wia-1; i--; ) r = C2(mul, XEQ(i), C2i(add, 1, r)); SGetU(w);
#undef XEQ #define XEQ(I) C2(ne, GetU(w,I), incG(x))
decG(w); decG(x); return r; B r = XEQ(wia-1);
} for (usz i=wia-1; i--; ) r = C2(mul, XEQ(i), C2i(add, 1, r));
if (xia+wia>20 && we<=el_i16 && xe<=el_i16) { #undef XEQ
B r; decG(w); decG(x); return r;
TABLE(w, x, i32, wia, i) }
return reduceI32Width(r, wia);
} if (xia+wia>20 && we<=el_i16 && xe<=el_i16) {
B r;
TABLE(w, x, i32, wia, i)
return reduceI32Width(r, wia);
}
} else if (elChr(we) && elChr(xe)) { we-= el_c8-el_i8; xe-= el_c8-el_i8; goto tyEls; }
i32* rp; B r = m_i32arrc(&rp, x); i32* rp; B r = m_i32arrc(&rp, x);
H_b2i* map = m_b2i(64); H_b2i* map = m_b2i(64);
SGetU(x) SGetU(x)
@ -143,8 +214,7 @@ B indexOf_c2(B t, B w, B x) {
} }
} }
B enclosed_0; B enclosed_0, enclosed_1;
B enclosed_1;
B memberOf_c2(B t, B w, B x) { B memberOf_c2(B t, B w, B x) {
if (isAtm(x) || RNK(x)!=1) { if (isAtm(x) || RNK(x)!=1) {
B2 t = splitCells(w, x, false); B2 t = splitCells(w, x, false);
@ -176,27 +246,32 @@ B memberOf_c2(B t, B w, B x) {
u8 we = TI(w,elType); usz wia = IA(w); u8 we = TI(w,elType); usz wia = IA(w);
u8 xe = TI(x,elType); usz xia = IA(x); u8 xe = TI(x,elType); usz xia = IA(x);
if (xia == 0) { r=taga(arr_shCopy(allZeroes(wia), w)); decG(w); goto dec_x; } if (xia == 0) { r=taga(arr_shCopy(allZeroes(wia), w)); decG(w); goto dec_x; }
#define WEQ(V) C2(eq, incG(w), V)
if (xe==el_bit && we!=el_B) { if (elNum(we) && elNum(xe)) { tyEls:
u64* xp = bitarr_ptr(x); #define WEQ(V) C2(eq, incG(w), V)
u64 x0 = 1 & xp[0]; if (xe==el_bit) {
r = WEQ(m_usz(x0)); u64* xp = bitarr_ptr(x);
if (bit_has(xp, xia, !x0)) r = C2(or, r, WEQ(m_usz(!x0))); u64 x0 = 1 & xp[0];
decG(w); goto dec_x; r = WEQ(m_usz(x0));
} if (bit_has(xp, xia, !x0)) r = C2(or, r, WEQ(m_usz(!x0)));
if (xia<=(xe==el_i16?8:16) && wia>16 && we<el_B && xe<el_B) { decG(w); goto dec_x;
SGetU(x); }
r = WEQ(GetU(x,0));
for (usz i=1; i<xia; i++) r = C2(or, r, WEQ(GetU(x,i))); if (xia<=(xe==el_i16?8:16) && wia>16) {
decG(w); goto dec_x; SGetU(x);
} r = WEQ(GetU(x,0));
#undef WEQ for (usz i=1; i<xia; i++) r = C2(or, r, WEQ(GetU(x,i)));
// TODO O(wia×xia) for small wia decG(w); goto dec_x;
if (xia+wia>20 && we<=el_i16 && xe<=el_i16) { }
B r; #undef WEQ
TABLE(x, w, i8, 0, 1)
return taga(cpyBitArr(r)); if (xia+wia>20 && we<=el_i16 && xe<=el_i16) {
} B r;
TABLE(x, w, i8, 0, 1)
return taga(cpyBitArr(r));
}
} else if (elChr(we) && elChr(xe)) { we-= el_c8-el_i8; xe-= el_c8-el_i8; goto tyEls; }
H_Sb* set = m_Sb(64); H_Sb* set = m_Sb(64);
SGetU(x) SGetU(w) SGetU(x) SGetU(w)
bool had; bool had;
@ -228,6 +303,7 @@ B count_c2(B t, B w, B x) {
if (we<=el_i16 && xe<=el_i16) { if (we<=el_i16 && xe<=el_i16) {
if (we==el_bit) { w = toI8Any(w); we = TI(w,elType); } if (we==el_bit) { w = toI8Any(w); we = TI(w,elType); }
if (xe==el_bit) { x = toI8Any(x); xe = TI(x,elType); } if (xe==el_bit) { x = toI8Any(x); xe = TI(x,elType); }
el8or16:;
usz it = elRange(we); // Range of writes usz it = elRange(we); // Range of writes
usz ft = elRange(xe); // Range of lookups usz ft = elRange(xe); // Range of lookups
usz t = it>ft? it : ft; // Table allocation width usz t = it>ft? it : ft; // Table allocation width
@ -248,6 +324,9 @@ B count_c2(B t, B w, B x) {
if (xe==el_i8) { GET(i8) } else { GET(i16) } if (xe==el_i8) { GET(i8) } else { GET(i16) }
#undef GET #undef GET
TFREE(tab0); TFREE(tab0);
} else if (we>=el_c8 && we<=el_c16 && xe>=el_c8 && xe<=el_c16) {
we-= el_c8-el_i8; xe-= el_c8-el_i8;
goto el8or16;
} else { } else {
H_b2i* map = m_b2i(64); H_b2i* map = m_b2i(64);
SGetU(x) SGetU(x)

View File

@ -3,7 +3,7 @@
// Except for trivial cases, ⍷ is implemented as ∊⊸/ // Except for trivial cases, ⍷ is implemented as ∊⊸/
// Other functions use adaptations of the same set of methods // Other functions use adaptations of the same set of methods
// Boolean cases all use special code, including ⍷ // Boolean cases (1-bit cells) all use special code, including ⍷
// COULD vectorize boolean ⊒ with +` // COULD vectorize boolean ⊒ with +`
// Sorted flags: start with r0⌾⊑»⊸≠𝕩 (r0←0 for ⊐, 1 otherwise) // Sorted flags: start with r0⌾⊑»⊸≠𝕩 (r0←0 for ⊐, 1 otherwise)
// ∊: ⊢; ⊐: +`; ⊒: ↕∘≠⊸(⊣-⌈`∘×) // ∊: ⊢; ⊐: +`; ⊒: ↕∘≠⊸(⊣-⌈`∘×)
@ -18,6 +18,8 @@
// Radix-assisted lookups are fallbacks for 4-byte ∊ and ⊒ // Radix-assisted lookups are fallbacks for 4-byte ∊ and ⊒
// COULD do radix-assisted ⊐ as ⍷⊸⊐ or similar // COULD do radix-assisted ⊐ as ⍷⊸⊐ or similar
// Specializes on constant top 1/2 bytes, but hashes make this rare // Specializes on constant top 1/2 bytes, but hashes make this rare
// Cells with 2≤n≤63 bits, ¬n∊8‿16‿32:
// Pad to the nearest nice size (8, 16, 32, 64 bits)
// Specialized 4-byte and 8-byte hash tables // Specialized 4-byte and 8-byte hash tables
// In-place resizing by factor of 4 based on measured collisions // In-place resizing by factor of 4 based on measured collisions
@ -27,8 +29,6 @@
// Generic hash table for other cases // Generic hash table for other cases
// Resizing is pretty expensive here // Resizing is pretty expensive here
// SHOULD widen small odd sizes
#include "../core.h" #include "../core.h"
#include "../utils/hash.h" #include "../utils/hash.h"
#include "../utils/talloc.h" #include "../utils/talloc.h"
@ -94,7 +94,7 @@ static bool canCompare64_norm(B x, usz n) {
static bool shouldWidenBitarr(B x, usz csz) { // assumes cells won't anymore have sizes of 0, 8, or 16 bits static bool shouldWidenBitarr(B x, usz csz) { // assumes cells won't anymore have sizes of 0, 8, or 16 bits
u8 xe = TI(x,elType); u8 xe = TI(x,elType);
ux bcsz = ((ux)csz)<<elWidthLogBits(xe); ux bcsz = ((ux)csz)<<elWidthLogBits(xe);
assert(csz!=0 && bcsz!=8 && bcsz!=16); assert(csz!=0 && bcsz!=1 && bcsz!=8 && bcsz!=16);
if (bcsz<64 && bcsz!=32) { assert(xe!=el_B); return true; } // not el_B because csz>0 → csz*sizeof(B) >= 64 if (bcsz<64 && bcsz!=32) { assert(xe!=el_B); return true; } // not el_B because csz>0 → csz*sizeof(B) >= 64
return false; return false;
} }