Separate where into its own function and split different types completely

This commit is contained in:
Marshall Lochbaum 2022-09-11 08:31:20 -04:00
parent 638121c054
commit 1f40d36712

View File

@ -109,30 +109,37 @@
for (u64 v=X[i]; v; v&=v-1) R[j++] = i*64 + CTZ(v); \ for (u64 v=X[i]; v; v&=v-1) R[j++] = i*64 + CTZ(v); \
} while (0) } while (0)
extern B rt_slash; static B where(B x, usz xia, u64 s) {
B slash_c1(B t, B x) {
if (RARE(isAtm(x)) || RARE(RNK(x)!=1)) thrF("/: Argument must have rank 1 (%H ≡ ≢𝕩)", x);
u64 s = usum(x);
if (s>=USZ_MAX) thrOOM();
if (s==0) { decG(x); return emptyIVec(); }
usz xia = IA(x);
if (RARE(xia>=I32_MAX)) {
usz xia = IA(x);
SGetU(x)
f64* rp; B r = m_f64arrv(&rp, s); usz ri = 0;
for (usz i = 0; i < xia; i++) {
usz c = o2s(GetU(x, i));
for (usz j = 0; j < c; j++) rp[ri++] = i;
}
decG(x);
return r;
}
B r; B r;
u8 xe = TI(x,elType);
if (xe==el_bit) {
u64* xp = bitarr_ptr(x); u64* xp = bitarr_ptr(x);
if (xia > 32768) {
usz q=xia%64; if (q) xp[xia/64] &= ((u64)1<<q) - 1; usz q=xia%64; if (q) xp[xia/64] &= ((u64)1<<q) - 1;
if (xia <= 128) {
#if SINGELI && defined(__BMI2__)
i8* rp = m_tyarrvO(&r, 1, s, t_i8arr, 8);
bmipopc_1slash8(xp, rp, xia);
#else
i8* rp; r=m_i8arrv(&rp,s); WHERE_SPARSE(xp,rp,s);
#endif
} else if (xia <= 32768) {
#if SINGELI && defined(__BMI2__)
if (s >= xia/16) {
i16* rp = m_tyarrvO(&r, 2, s, t_i16arr, 16);
bmipopc_1slash16(xp, rp, xia);
}
#else
if (s >= xia/2) {
i16* rp = m_tyarrvO(&r, 2, s, t_i16arr, 2);
for (usz i=0; i<(xia+7)/8; i++) {
u8 v = ((u8*)xp)[i];
for (usz k=0; k<8; k++) { *rp=8*i+k; rp+=v&1; v>>=1; }
}
}
#endif
else {
i16* rp; r=m_i16arrv(&rp,s); WHERE_SPARSE(xp,rp,s);
}
} else {
assert(xia <= (usz)I32_MAX+1);
#if SINGELI && defined(__BMI2__) #if SINGELI && defined(__BMI2__)
i32* rp; r = m_i32arrv(&rp, s); i32* rp; r = m_i32arrv(&rp, s);
#else #else
@ -153,6 +160,7 @@ B slash_c1(B t, B x) {
if (bs >= b/8+b/16) { if (bs >= b/8+b/16) {
bmipopc_1slash16(xp, buf, b); bmipopc_1slash16(xp, buf, b);
for (usz j=0; j<bs; j++) rq[j] = i+buf[j]; for (usz j=0; j<bs; j++) rq[j] = i+buf[j];
}
#else #else
if (bs >= b/2) { if (bs >= b/2) {
for (usz ii=0; ii<(b+7)/8; ii++) { for (usz ii=0; ii<(b+7)/8; ii++) {
@ -160,8 +168,9 @@ B slash_c1(B t, B x) {
i32* rs=rq; i32* rs=rq;
for (usz k=0; k<8; k++) { *rs=i+8*ii+k; rs+=v&1; v>>=1; } for (usz k=0; k<8; k++) { *rs=i+8*ii+k; rs+=v&1; v>>=1; }
} }
}
#endif #endif
} else if (bs >= b/256) { // Branchless sparse else if (bs >= b/256) { // Branchless sparse
for (usz j=0; j<bs; j++) rq[j]=0; for (usz j=0; j<bs; j++) rq[j]=0;
u32 top = 1<<24; u32 top = 1<<24;
for (usz i=0, j=0; i<(b+63)/64; i++) { for (usz i=0, j=0; i<(b+63)/64; i++) {
@ -185,32 +194,32 @@ B slash_c1(B t, B x) {
xp+= b/64; xp+= b/64;
} }
TFREE(buf); TFREE(buf);
} else
// Sparse method with CTZ
#if SINGELI && defined(__BMI2__)
if (xia>128 && s < xia/16) {
#else
if (xia<=128 || s < xia/2) {
#endif
usz q=xia%64; if (q) xp[xia/64] &= ((u64)1<<q) - 1;
#define WHERE(T) T* rp; r=m_##T##arrv(&rp,s); WHERE_SPARSE(xp,rp,s);
if (0&&xia<=32768) { WHERE(i16) } else { WHERE(i32) }
#undef WHERE
} else
#if SINGELI && defined(__BMI2__)
if (xia<=128) { i8* rp = m_tyarrvO(&r, 1, s, t_i8arr , 8); bmipopc_1slash8 (xp, rp, xia); }
else { i16* rp = m_tyarrvO(&r, 2, s, t_i16arr, 16); bmipopc_1slash16(xp, rp, xia); }
#else
{
i32* rp = m_tyarrvO(&r, 4, s, t_i32arr, 4);
u8* x8 = (u8*)xp;
u8 q=xia%8; if (q) x8[xia/8] &= (1<<q)-1;
for (usz i=0; i<(xia+7)/8; i++) {
u8 v = x8[i];
for (usz k=0; k<8; k++) { *rp=8*i+k; rp+= v&1; v>>=1; }
} }
return r;
} }
#endif
extern B rt_slash;
B slash_c1(B t, B x) {
if (RARE(isAtm(x)) || RARE(RNK(x)!=1)) thrF("/: Argument must have rank 1 (%H ≡ ≢𝕩)", x);
u64 s = usum(x);
if (s>=USZ_MAX) thrOOM();
if (s==0) { decG(x); return emptyIVec(); }
usz xia = IA(x);
if (RARE(xia>=I32_MAX)) {
usz xia = IA(x);
SGetU(x)
f64* rp; B r = m_f64arrv(&rp, s); usz ri = 0;
for (usz i = 0; i < xia; i++) {
usz c = o2s(GetU(x, i));
for (usz j = 0; j < c; j++) rp[ri++] = i;
}
decG(x);
return r;
}
B r;
u8 xe = TI(x,elType);
if (xe==el_bit) {
r = where(x, xia, s);
} else { } else {
i32* rp; r = m_i32arrv(&rp, s); i32* rp; r = m_i32arrv(&rp, s);
if (xe==el_i8) { if (xe==el_i8) {