Merge pull request #137 from mlochbaum/scalreduce

Scalar-based arithmetic insert and stride-1 scan-cells for short rows
This commit is contained in:
dzaima 2025-03-11 19:23:57 +02:00 committed by GitHub
commit dfce405111
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
4 changed files with 126 additions and 23 deletions

View File

@ -102,6 +102,21 @@ B select_cells_base(B inds, B x0, ux csz, ux cam) { // consumes inds,x0; Used by
return taga(APD_SH_GET(r, '\0')); return taga(APD_SH_GET(r, '\0'));
} }
static B scan_cells_stride1(B f, B x, usz m) {
// m is cell size == scan axis length
B xf = getFillR(x);
HArr_p r = m_harr0c(x);
SGet(x)
FC2 fc2 = c2fn(f);
for (usz i=0, ia=IA(x); i<ia; ) {
usz e = i + m;
r.a[i] = Get(x, i); i++;
for (; i<e; i++) r.a[i] = fc2(f, inc(r.a[i-1]), Get(x, i));
}
decG(x);
return withFill(r.b, xf);
}
B scan_arith(B f, B w, B x, usz* xsh) { // Used by scan.c B scan_arith(B f, B w, B x, usz* xsh) { // Used by scan.c
bool has_w = w.u != m_f64(0).u; bool has_w = w.u != m_f64(0).u;
assert(isArr(x) && (!has_w || isArr(w))); assert(isArr(x) && (!has_w || isArr(w)));
@ -537,12 +552,15 @@ B for_cells_c1(B f, u32 xr, u32 cr, u32 k, B x, u32 chr) { // F⎉cr x; array x,
if (cr==0) break; if (cr==0) break;
usz m = xsh[k]; usz m = xsh[k];
if (m<=1 || IA(x)==0) return x; if (m<=1 || IA(x)==0) return x;
if (!isFun(fd->f)) break; B f = fd->f;
u8 frtid = RTID(fd->f); if (!isFun(f)) break;
u8 frtid = RTID(f);
if (frtid==n_rtack) return x; if (frtid==n_rtack) return x;
if (TI(x,elType)==el_bit && (isPervasiveDyExt(fd->f)||frtid==n_ltack) if (cr==1 || cam*m == IA(x)) {
&& 1==shProd(xsh, k+1, xr)) { if (TI(x,elType)==el_bit && (isPervasiveDy(f) || frtid==n_ltack)) {
B r = scan_rows_bit(frtid, x, m); if (!q_N(r)) return r; B r = scan_rows_bit(frtid, x, m); if (!q_N(r)) return r;
}
if (m <= 6) return scan_cells_stride1(f, x, m);
} }
break; break;
} }

View File

@ -383,6 +383,68 @@ static B m1c1(B t, B f, B x) { // consumes x
} }
extern B insert_base(B f, B x, bool has_w, B w); // from cells.c extern B insert_base(B f, B x, bool has_w, B w); // from cells.c
static bool fillNumeric(B x) {
if (x.u==0) return true;
if (isAtm(x)) return false;
u8 xe = TI(x,elType);
usz ia = IA(x);
if (xe!=el_B) return ia==0 || elNum(xe);
if (!fillNumeric(getFillN(x))) return false;
if (ia==0) return true;
SGetU(x)
for (usz i=0; i<ia; i++) if (!fillNumeric(GetU(x,i))) return false;
return true;
}
// Do arithmetic 𝔽˝ with short rows like 𝔽¨˝ to cut per-row overhead
static B insert_scal(B f, FC2 fc2, B x, bool has_w, B fxw, usz xia, ur rr) {
usz csz = arr_csz(x);
HArr_p r = m_harr0p(csz);
usz* rsh = arr_shAlloc((Arr*)r.c, rr);
if (rr>1) shcpy(rsh, SH(x)+1, rr);
usz xi = xia - csz;
B xf = getFillN(x);
B rf;
if (has_w) {
// fxw is (⊢˝𝕩)𝔽𝕨 so shape errors have been caught
rf = getFillR(fxw);
COPY_TO(r.a, el_B, 0, fxw, 0, csz);
decG(fxw);
} else {
rf = inc(xf);
COPY_TO(r.a, el_B, 0, x, xi, csz);
}
SGet(x)
while (xi) {
xi -= csz;
for (usz i=0; i<csz; i++) r.a[i] = fc2(f, Get(x, xi+i), r.a[i]);
}
// Fill might oscillate between 0 and ' ' or error if either argument
// has characters in its fill
if (!noFill(rf) && !(fillNumeric(rf) && (!has_w || fillNumeric(xf)))) {
usz n = *SH(x);
assert(n > 2);
#if SEMANTIC_CATCH
if (CATCH) {
freeThrown();
rf = bi_noFill;
} else {
if (!has_w) rf = fc2(f, inc(xf), rf);
if (n%2 == !has_w) rf = fc2(f, inc(xf), rf);
else fc2(f, inc(xf), inc(rf)); // could error, -˜˝"abc"
popCatch();
}
#else
if (n%2 == 0) { dec(rf); rf = bi_noFill; }
#endif
}
decG(x);
return withFill(r.b, rf);
}
B insert_c1(Md1D* d, B x) { B f = d->f; B insert_c1(Md1D* d, B x) { B f = d->f;
ur xr; ur xr;
if (isAtm(x) || (xr=RNK(x))==0) thrM("𝔽˝𝕩: 𝕩 must have rank at least 1"); if (isAtm(x) || (xr=RNK(x))==0) thrM("𝔽˝𝕩: 𝕩 must have rank at least 1");
@ -410,7 +472,8 @@ B insert_c1(Md1D* d, B x) { B f = d->f;
if (RARE(!isFun(f))) { decG(x); if (isMd(f)) thrM("Calling a modifier"); return inc(f); } if (RARE(!isFun(f))) { decG(x); if (isMd(f)) thrM("Calling a modifier"); return inc(f); }
if (isPervasiveDyExt(f)) { if (isPervasiveDyExt(f)) {
if (xr==1) return m_unit(fold_c1(d, x)); if (xr==1) return m_unit(fold_c1(d, x));
if (len==IA(x)) { usz xia = IA(x);
if (len==xia) {
B r = m_vec1(fold_c1(d, C1(shape, x))); B r = m_vec1(fold_c1(d, C1(shape, x)));
ur rr = xr - 1; ur rr = xr - 1;
if (rr > 1) { if (rr > 1) {
@ -420,6 +483,9 @@ B insert_c1(Md1D* d, B x) { B f = d->f;
} }
return r; return r;
} }
if (len>2 && xia<6*(u64)len) {
return insert_scal(f, c2fn(f), x, 0, m_f64(0), xia, xr-1);
}
} }
if (RTID(f) != RTID_NONE) { if (RTID(f) != RTID_NONE) {
u8 rtid = RTID(f); u8 rtid = RTID(f);
@ -448,26 +514,34 @@ B insert_c2(Md1D* d, B w, B x) { B f = d->f;
usz len = *SH(x); usz len = *SH(x);
if (len==0) { decG(x); return w; } if (len==0) { decG(x); return w; }
if (RARE(!isFun(f))) { dec(w); decG(x); if (isMd(f)) thrM("Calling a modifier"); return inc(f); } if (RARE(!isFun(f))) { dec(w); decG(x); if (isMd(f)) thrM("Calling a modifier"); return inc(f); }
if (isPervasiveDyExt(f) && len==IA(x)) { if (isPervasiveDyExt(f)) {
// 1-element arrays are always conformable usz xia = IA(x);
// final rank is higher of w, cell rank of x
ur rr = xr - 1; ur rr = xr - 1;
if (isArr(w)) { if (len==xia) {
if (IA(w) != 1) goto skip; // 1-element arrays are always conformable
ur wr = RNK(w); if (wr>rr) rr = wr; // final rank is higher of w, cell rank of x
w = TO_GET(w, 0); if (isArr(w)) {
} if (IA(w) != 1) goto skip;
if (xr > 1) x = C1(shape, x); ur wr = RNK(w); if (wr>rr) rr = wr;
B r = m_unit(fold_c2(d, w, x)); w = TO_GET(w, 0);
if (rr > 0) {
if (rr == 1) arr_shVec(a(r));
else {
ShArr* rsh = m_shArr(rr);
PLAINLOOP for (ur i=0; i<rr; i++) rsh->a[i] = 1;
arr_shReplace(a(r), rr, rsh);
} }
if (xr > 1) x = C1(shape, x);
B r = m_unit(fold_c2(d, w, x));
if (rr > 0) {
if (rr == 1) arr_shVec(a(r));
else {
ShArr* rsh = m_shArr(rr);
PLAINLOOP for (ur i=0; i<rr; i++) rsh->a[i] = 1;
arr_shReplace(a(r), rr, rsh);
}
}
return r;
}
if (len>2 && xia<6*(u64)len && !(isArr(w) && RNK(w)>rr)) {
FC2 fc2 = c2fn(f);
w = fc2(f, C2(select, m_f64(-1), incG(x)), w);
return insert_scal(f, fc2, x, 1, w, xia, rr);
} }
return r;
skip:; skip:;
} }
if (RTID(f) != RTID_NONE) { if (RTID(f) != RTID_NONE) {

View File

@ -30,6 +30,10 @@
%USE fill Fill ¨"" %EACH_FILLS %% ' ' %USE fill Fill ¨"" %EACH_FILLS %% ' '
%USE fill Fill ¨"" %EACH_FILLS %% ' ' %USE fill Fill ¨"" %EACH_FILLS %% ' '
{3 -˝𝕩2<<'a'}¨ 56 %!HEAPVERIFY %% <¨¨ "aa ", 30
{3 (22)-˝𝕩2<<'x'}¨ 56 %!HEAPVERIFY %% <¨¨ "vv ", 220
{3 "ee"-˝𝕩2<<'a'}¨ 56 %!HEAPVERIFY %% <¨¨ ¯4¯40, "ee "
# TODO: # TODO:
# ⟨↑‿3⥊↕10 ⋄ ↑‿3⥊<¨↕10 ⋄ ↑‿3⥊<˘↕10⟩ # ⟨↑‿3⥊↕10 ⋄ ↑‿3⥊<¨↕10 ⋄ ↑‿3⥊<˘↕10⟩
# 5‿¯5↑⌜⟨↕2 ⋄ "ab" ⋄ <¨↕2 ⋄ <˘↕2⟩ # 5‿¯5↑⌜⟨↕2 ⋄ "ab" ⋄ <¨↕2 ⋄ <˘↕2⟩

View File

@ -515,6 +515,13 @@ a←↕2 ⋄ ! "e" ≡ (↕10){b←a‿a‿a‿a‿a‿a‿a‿a‿a‿a ⋄
# ˝ # ˝
!"𝔽˝𝕩: 𝕩 must have rank at least 1" % ˝0 !"𝔽˝𝕩: 𝕩 must have rank at least 1" % ˝0
!"𝔽˝𝕩: Identity not found" % •internal.Keep˝ !"𝔽˝𝕩: Identity not found" % •internal.Keep˝
(4)+˝7 %% 21222324
(224)+˝72228 %% 228492100108
(224)+˝2228 %% 222367
+˝ 104 %% 45 ¨ 10×4
( 4) +˝ 104 %% (4) + 45 ¨ 10×4
(<¨ 4) +˝ 104 %% (4) + 45 ¨ 10×4
(05+¨4) +˝ 104 %% ((4) + 45) ¨ 5 + 11×4
# ⌜ # ⌜
!"𝕨𝔽⌜𝕩: Result rank too large (200≡=𝕨, 200≡=𝕩)" % +˜(2001)1 !"𝕨𝔽⌜𝕩: Result rank too large (200≡=𝕨, 200≡=𝕩)" % +˜(2001)1