fix ≍˘ error checking, start of generalizing to join_cells

This commit is contained in:
dzaima 2024-09-13 22:05:31 +03:00
parent 1679d957e8
commit fff96b9475
3 changed files with 62 additions and 52 deletions

View File

@ -23,7 +23,7 @@ B scan_rows_bit(u8, B x, usz m);
B takedrop_highrank(bool take, B w, B x); B takedrop_highrank(bool take, B w, B x);
B rotate_highrank(bool inv, B w, B x); B rotate_highrank(bool inv, B w, B x);
B interleave_cells(B w, B x, ur xr, ur xk, usz* xsh); // from transpose.c B interleave_cells(B w, B x, ur k); // from transpose.c
// from select.c: // from select.c:
B select_rows_B(B x, ux csz, ux cam, B inds); B select_rows_B(B x, ux csz, ux cam, B inds);
@ -694,7 +694,7 @@ NOINLINE B for_cells_AS(B f, B w, B x, ur wcr, ur wr, u32 chr) { // F⟜x⎉wcr
case n_rtack: return const_cells(w, wk, wsh, x, chr); case n_rtack: return const_cells(w, wk, wsh, x, chr);
case n_couple: if (RNK(w)==1) { case n_couple: if (RNK(w)==1) {
x = taga(arr_shVec(reshape_one(IA(w), x))); x = taga(arr_shVec(reshape_one(IA(w), x)));
return interleave_cells(w, x, 1, 1, wsh); return interleave_cells(w, x, 1);
} break; } break;
} }
if (IA(w)!=0 && isPervasiveDy(f)) { if (IA(w)!=0 && isPervasiveDy(f)) {
@ -745,7 +745,7 @@ NOINLINE B for_cells_SA(B f, B w, B x, ur xcr, ur xr, u32 chr) { // w⊸F⎉xcr
break; break;
case n_couple: if (RNK(x)==1) { case n_couple: if (RNK(x)==1) {
w = taga(arr_shVec(reshape_one(IA(x), w))); w = taga(arr_shVec(reshape_one(IA(x), w)));
return interleave_cells(w, x, 1, 1, xsh); return interleave_cells(w, x, 1);
} break; } break;
case n_pick: if (isF64(w) && xcr==1 && TI(x,arrD1)) { case n_pick: if (isF64(w) && xcr==1 && TI(x,arrD1)) {
usz l = xsh[xk]; usz l = xsh[xk];
@ -848,8 +848,8 @@ NOINLINE B for_cells_AA(B f, B w, B x, ur wcr, ur xcr, u32 chr) { // w F⎉wcr
if (rsh) shcpy(rsh, zsh, zk); if (rsh) shcpy(rsh, zsh, zk);
decG(w); decG(x); return taga(r); decG(w); decG(x); return taga(r);
} }
if (rtid==n_couple && wr==xr) { if (rtid==n_couple && wr==xr && eqShPart(wsh+wk, xsh+wk, wcr)) {
return interleave_cells(w, x, xr, xk, xsh); return interleave_cells(w, x, wk);
} }
} }
if (isPervasiveDy(f)) { if (isPervasiveDy(f)) {

View File

@ -93,13 +93,7 @@ static void interleave_bits(u64* rp, void* x0v, void* x1v, usz n) {
} }
} }
B toBPtrAny(B x) { NOINLINE B toElTypeArr(u8 re, B x) { // consumes; returns an array with the given element type (re==el_B guarantees TO_BPTR working)
if (arr_bptr(x)!=NULL) return x;
return taga(cpyHArr(x));
}
NOINLINE
B toElTypeArr(u8 re, B x) { // consumes; returns an array with the given element type (re==el_B guarantees TO_BPTR working)
switch (re) { default: UD; switch (re) { default: UD;
case el_bit: return toBitAny(x); case el_bit: return toBitAny(x);
case el_i8: return toI8Any(x); case el_i8: return toI8Any(x);
@ -109,14 +103,11 @@ B toElTypeArr(u8 re, B x) { // consumes; returns an array with the given element
case el_c8: return toC8Any(x); case el_c8: return toC8Any(x);
case el_c16: return toC16Any(x); case el_c16: return toC16Any(x);
case el_c32: return toC32Any(x); case el_c32: return toC32Any(x);
case el_B: return toBPtrAny(x); case el_B: TO_BPTR(x); return x;
} }
} }
Arr* join_cells(B w, B x, ur k) { // consumes w,x; join k-cells, 𝕨 ∾○⥊⎉(-k) 𝕩; result has unset shape
// interleave arrays, 𝕨≍⎉(-xk)𝕩
B interleave_cells(B w, B x, ur xr, ur xk, usz* xsh) { // consumes w,x
assert(RNK(w)==xr && xr>=1);
u8 we = TI(w,elType); u8 we = TI(w,elType);
u8 xe = TI(x,elType); u8 xe = TI(x,elType);
@ -125,44 +116,47 @@ B interleave_cells(B w, B x, ur xr, ur xk, usz* xsh) { // consumes w,x
// delay doing this until it's known that there will be code that can utilize it // delay doing this until it's known that there will be code that can utilize it
if (re!=we) w = toElTypeArr(re, w); if (re!=we) w = toElTypeArr(re, w);
if (re!=xe) x = toElTypeArr(re, x); if (re!=xe) x = toElTypeArr(re, x);
return interleave_cells(w, x, xr, xk, SH(x)); return join_cells(w, x, k);
} }
Arr *r; Arr *r;
u8 xlw = elwBitLog(re); u8 xlw = elwBitLog(re);
usz n = shProd(xsh, 0, xk); usz n = shProd(SH(w), 0, k);
usz csz = shProd(xsh, xk, xr); usz wcsz = shProd(SH(w), k, RNK(w));
usz ia = 2*n*csz; usz xcsz = shProd(SH(x), k, RNK(x));
usz ia = IA(w)+IA(x);
if (csz & (csz-1)) { if (wcsz == xcsz) {
goto generic; usz csz = wcsz;
} else if (csz==1 && xlw==0) { // we & xe are trivially el_bit if (csz & (csz-1)) {
u64* rp; r=m_bitarrp(&rp, ia); goto generic;
interleave_bits(rp, bitany_ptr(w), bitany_ptr(x), ia); } else if (csz==1 && xlw==0) { // we & xe are trivially el_bit
#if SINGELI u64* rp; r=m_bitarrp(&rp, ia);
} else if (csz==1 && re==el_B) { interleave_bits(rp, bitany_ptr(w), bitany_ptr(x), ia);
if (we!=xe) goto to_equal_types; #if SINGELI
B* wp = TO_BPTR(w); B* xp = TO_BPTR_RUN(x, xsh = SH(x)); } else if (csz==1 && re==el_B) {
if (we!=xe) goto to_equal_types;
HArr_p p = m_harrUv(ia); // Debug build complains with harrUp B* wp = TO_BPTR(w); B* xp = TO_BPTR(x);
interleave_fns[3](p.a, wp, xp, n);
for (usz i=0; i<ia; i++) inc(p.a[i]); HArr_p p = m_harrUv(ia); // Debug build complains with harrUp
NOGC_E; interleave_fns[3](p.a, wp, xp, n);
r = (Arr*) p.c; for (usz i=0; i<ia; i++) inc(p.a[i]);
goto add_fill; NOGC_E;
} else if (csz<=64>>xlw && csz<<xlw>=8) { // Require CPU-sized cells r = (Arr*) p.c;
if (we!=xe) goto to_equal_types; goto add_fill;
assert(re!=el_B); } else if (csz<=64>>xlw && csz<<xlw>=8) { // Require CPU-sized cells
void* rv; if (we!=xe) goto to_equal_types;
if (xlw==0) { u64* rp; r = m_bitarrp(&rp, ia); rv=rp; } assert(re!=el_B);
else rv = m_tyarrp(&r,elWidth(re),ia,el2t(re)); void* rv;
interleave_fns[CTZ(csz<<xlw)-3](rv, tyany_ptr(w), tyany_ptr(x), n); if (xlw==0) { u64* rp; r = m_bitarrp(&rp, ia); rv=rp; }
#endif else rv = m_tyarrp(&r,elWidth(re),ia,el2t(re));
interleave_fns[CTZ(csz<<xlw)-3](rv, tyany_ptr(w), tyany_ptr(x), n);
#endif
} else goto generic;
} else { generic:; } else { generic:;
MAKE_MUT_INIT(rm, ia, re); MUTG_INIT(rm); MAKE_MUT_INIT(rm, ia, re); MUTG_INIT(rm);
for (ux o = 0; o < n*csz; o+= csz) { for (ux ow=0, ox=0, or=0; or < ia; or+= wcsz+xcsz) {
mut_copyG(rm, o*2, w, o, csz); mut_copyG(rm, or, w, ow, wcsz); ow+= wcsz;
mut_copyG(rm, o*2+csz, x, o, csz); mut_copyG(rm, or+wcsz, x, ox, xcsz); ox+= xcsz;
} }
r = a(mut_fv(rm)); r = a(mut_fv(rm));
goto add_fill; goto add_fill;
@ -172,10 +166,21 @@ B interleave_cells(B w, B x, ur xr, ur xk, usz* xsh) { // consumes w,x
if (SFNS_FILLS) r = a(qWithFill(taga(r), fill_both(w, x))); if (SFNS_FILLS) r = a(qWithFill(taga(r), fill_both(w, x)));
} }
usz* sh = arr_shAlloc(r, xr+1);
shcpy(sh, xsh, xk); sh[xk]=2; shcpy(sh+xk+1, xsh+xk, xr-xk);
decG(w); decG(x); decG(w); decG(x);
return taga(r); return r;
}
B join_c2(B, B, B);
B interleave_cells(B w, B x, ur k) { // consumes w,x; interleave arrays, 𝕨 ≍⎉(-xk) 𝕩; assumes equal-shape args
ux xr = RNK(x);
if (xr==0) return C2(join, w, x);
ShArr* rsh = m_shArr(xr+1); // TODO handle leak if join_cells fails
usz* xsh = SH(x);
shcpy(rsh->a, xsh, k);
rsh->a[k] = 2;
shcpy(rsh->a+k+1, xsh+k, xr-k);
return taga(arr_shSetUG(join_cells(w, x, k), xr+1, rsh));
} }
static void transpose_move(void* rv, void* xv, u8 xe, usz w, usz h) { static void transpose_move(void* rv, void* xv, u8 xe, usz w, usz h) {

View File

@ -124,6 +124,11 @@
%USE eqvar ˘1e9 _eqvar 10071 %% 1002 (1e9× 2|200) 2/0=7|100 %USE eqvar ˘1e9 _eqvar 10071 %% 1002 (1e9× 2|200) 2/0=7|100
%USE eqvar ˘'!' _eqvar 10071 %% '!'•internal.Keep˘ 0=7|100 %USE eqvar ˘'!' _eqvar 10071 %% '!'•internal.Keep˘ 0=7|100
%USE eqvar ˘_eqvar˜ 339 %% 323012012345345678678 %USE eqvar ˘_eqvar˜ 339 %% 323012012345345678678
%USE eqvar (2345111) 2 _eqvar 234517 %% (2345111) •internal.PureKeep2 234517
!"˘: Argument frames don't agree (3‿4 ≡ ≢𝕨, 5‿4 ≡ ≢𝕩, common frame of 1 axes)" % (342) ˘ 542
!"≍: 𝕨 and 𝕩 must have equal shapes (⟨4⟩ ≡ ≢𝕨, ⟨5⟩ ≡ ≢𝕩)" % (342) ˘ 352
!"≍: 𝕨 and 𝕩 must have equal shapes (2‿2‿2 ≡ ≢𝕨, 2‿3‿2 ≡ ≢𝕩)" % (32222) ˘ 32322
!"≍: 𝕨 and 𝕩 must have equal shapes (4‿1 ≡ ≢𝕨, ⟨4⟩ ≡ ≢𝕩)" % (3410) ˘ 340
( (
%USE IS_HEAPVERIFY %USE IS_HEAPVERIFY