Merge branch 'sortflag'

This commit is contained in:
dzaima 2022-11-17 18:56:15 +02:00
commit 5e94de7983
11 changed files with 408 additions and 268 deletions

View File

@ -212,7 +212,7 @@ ${bd}/%.o: src/jit/%.c
@echo $< | cut -c 5-
@$(CC_INC) $@.d -o $@ -c $<
builtins: ${addprefix ${bd}/, arithm.o arithd.o cmp.o sfns.o squeeze.o select.o slash.o group.o sort.o search.o selfsearch.o md1.o md2.o fns.o sysfn.o internal.o inverse.o}
builtins: ${addprefix ${bd}/, arithm.o arithd.o cmp.o sfns.o squeeze.o select.o slash.o group.o sort.o search.o selfsearch.o scan.o md1.o md2.o fns.o sysfn.o internal.o inverse.o}
${bd}/%.o: src/builtins/%.c
@echo $< | cut -c 5-
@$(CC_INC) $@.d -o $@ -c $<

View File

@ -31,22 +31,22 @@
}
#if SINGELI
extern void (*const avx2_scan_max8)(int8_t* v0,int8_t* v1,uint64_t v2);
extern void (*const avx2_scan_min8)(int8_t* v0,int8_t* v1,uint64_t v2);
extern void (*const avx2_scan_max16)(int16_t* v0,int16_t* v1,uint64_t v2);
extern void (*const avx2_scan_min16)(int16_t* v0,int16_t* v1,uint64_t v2);
extern void (*const avx2_scan_max_i8)(int8_t* v0,int8_t* v1,uint64_t v2);
extern void (*const avx2_scan_min_i8)(int8_t* v0,int8_t* v1,uint64_t v2);
extern void (*const avx2_scan_max_i16)(int16_t* v0,int16_t* v1,uint64_t v2);
extern void (*const avx2_scan_min_i16)(int16_t* v0,int16_t* v1,uint64_t v2);
#define COUNT_THRESHOLD 32
#define WRITE_SPARSE_i8 \
for (usz i=0; i<n; i++) rp[i]=j; \
while (ij<n) { rp[ij]=GRADE_UD(++j,--j); ij+=c0o[j]; } \
GRADE_UD(avx2_scan_max8,avx2_scan_min8)(rp,rp,n);
GRADE_UD(avx2_scan_max_i8,avx2_scan_min_i8)(rp,rp,n);
#define WRITE_SPARSE_i16 \
usz b = 1<<10; \
for (usz k=0; ; ) { \
usz e = b<n-k? k+b : n; \
for (usz i=k; i<e; i++) rp[i]=j; \
while (ij<e) { rp[ij]=GRADE_UD(++j,--j); ij+=c0o[j]; } \
GRADE_UD(avx2_scan_max16,avx2_scan_min16)(rp+k,rp+k,e-k); \
GRADE_UD(avx2_scan_max_i16,avx2_scan_min_i16)(rp+k,rp+k,e-k); \
if (e==n) {break;} k=e; \
}
#define WRITE_SPARSE(T) WRITE_SPARSE_##T

View File

@ -25,7 +25,6 @@
#include "../utils/mut.h"
extern B ud_c1(B, B);
extern B not_c1(B, B);
extern B ne_c2(B, B, B);
extern B slash_c1(B, B);
extern B slash_c2(B, B, B);
@ -100,9 +99,9 @@ static B group_simple(B w, B x, ur xr, usz wia, usz xn, usz* xsh, u8 we) {
}
if (we==el_bit) {
assert(ria == 2);
if (wia>xn) w = take_c2(m_f64(0), m_f64(xn), w);
rp[1] = slash_c2(m_f64(0), inc(w), inc(x));
rp[0] = slash_c2(m_f64(0), not_c1(m_f64(0), w), x);
if (wia>xn) w = C2(take, m_f64(xn), w);
rp[1] = C2(slash, inc(w), inc(x));
rp[0] = C2(slash, bit_negate(w), x);
return taga(r);
}
// Needed to make sure wia>0 for ip[wia-1] below
@ -128,9 +127,6 @@ static B group_simple(B w, B x, ur xr, usz wia, usz xn, usz* xsh, u8 we) {
// Few changes in 𝕨: move in chunks
if (xn>64 && notB && change<(xn*width)/32) {
#define C1(F,X ) F##_c1(m_f64(0),X )
#define C2(F,X,W) F##_c2(m_f64(0),X,W)
u64* mp; B m = m_bitarrv(&mp, xn);
u8* wp0 = tyany_ptr(w);
we = TI(w,elType);
@ -139,8 +135,6 @@ static B group_simple(B w, B x, ur xr, usz wia, usz xn, usz* xsh, u8 we) {
B ind = C1(slash, m);
w = C2(select, inc(ind), w);
#undef C1
#undef C2
if (TI(ind,elType)!=el_i32) ind = taga(cpyI32Arr(ind));
if (TI(w ,elType)!=el_i32) w = taga(cpyI32Arr(w ));
wia = IA(ind);
@ -179,10 +173,10 @@ static B group_simple(B w, B x, ur xr, usz wia, usz xn, usz* xsh, u8 we) {
// Many ¯1s: filter out, then continue
if (xn>32 && neg>(bits?0:xn/4)+xn/8) {
if (wia>xn) w = take_c2(m_f64(0), m_f64(xn), w);
B m = ne_c2(m_f64(0), m_f64(-1), inc(w));
w = slash_c2(m_f64(0), inc(m), w);
x = slash_c2(m_f64(0), m, x); xn = *SH(x);
if (wia>xn) w = C2(take, m_f64(xn), w);
B m = C2(ne, m_f64(-1), inc(w));
w = C2(slash, inc(m), w);
x = C2(slash, m, x); xn = *SH(x);
neg = 0;
}
if (TI(w,elType)!=el_i32) w = taga(cpyI32Arr(w));
@ -331,8 +325,8 @@ B group_c2(B t, B w, B x) {
B group_c1(B t, B x) {
if (isArr(x) && RNK(x)==1 && TI(x,arrD1)) {
usz ia = IA(x);
B range = ud_c1(t, m_f64(ia));
return group_c2(m_f64(0), x, range);
B range = C1(ud, m_f64(ia));
return C2(group, x, range);
}
return c1(rt_group, x);
}

View File

@ -107,29 +107,6 @@ B each_c2(Md1D* d, B w, B x) { B f = d->f;
return homFil2(f, eachd(f, w, x), wf, xf);
}
#if SINGELI && __PCLMUL__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wunused-variable"
#include "../singeli/gen/neq.c"
#pragma GCC diagnostic pop
#endif
B scan_ne(u64 p, u64* xp, u64 ia) {
u64* rp; B r=m_bitarrv(&rp,ia);
#if SINGELI && __PCLMUL__
clmul_scan_ne(p, xp, rp, BIT_N(ia));
#else
for (usz i = 0; i < BIT_N(ia); i++) {
u64 c = xp[i];
u64 r = c ^ (c<<1);
r^= r<< 2; r^= r<< 4; r^= r<<8;
r^= r<<16; r^= r<<32; r^= p;
rp[i] = r;
p = -(r>>63); // repeat sign bit
}
#endif
return r;
}
static bool fold_ne(u64* x, u64 am) {
u64 r = 0;
for (u64 i = 0; i < (am>>6); i++) r^= x[i];
@ -143,204 +120,6 @@ static i64 bit_diff(u64* x, u64 am) {
if (am&63) r+= POPC((x[am>>6]^a)<<(64-am & 63));
return r - (i64)(am/2);
}
#if SINGELI
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wunused-variable"
#include "../singeli/gen/scan.c"
#pragma GCC diagnostic pop
#endif
B slash_c1(B f, B x);
B scan_bit_sum(B x, u64* xp, u64 ia, u64 xs) { // consumes x
u8 re = xs<=I8_MAX? el_i8 : xs<=I16_MAX? el_i16 : el_i32;
if (xs < ia/128) {
B ones = slash_c1(m_f64(0), x);
MAKE_MUT(r0, ia) mut_init(r0, re); MUTG_INIT(r0);
SGetU(ones)
usz ri = 0;
for (usz i = 0; i < xs; i++) {
usz e = o2s(GetU(ones, i));
mut_fillG(r0, ri, m_i32(i), e-ri);
ri = e;
}
if (ri<ia) mut_fillG(r0, ri, m_i32(xs), ia-ri);
decG(ones);
return mut_fv(r0);
}
B r;
void* rp = m_tyarrv(&r, elWidth(re), ia, el2t(re));
#if SINGELI
#define SUM(W,T) avx2_bcs##W(xp, rp, ia);
#else
#define SUM(W,T) { T c=0; for (usz i=0; i<ia; i++) { c+= bitp_get(xp,i); ((T*)rp)[i]=c; } }
#endif
#define CASE(W) case el_i##W: SUM(W, i##W) break;
switch (re) { default:UD; CASE(8) CASE(16) CASE(32) }
#undef CASE
#undef SUM
decG(x); return r;
}
#if !USE_VALGRIND
static u64 vg_rand(u64 x) { return x; }
#endif
B scan_c1(Md1D* d, B x) { B f = d->f;
if (isAtm(x) || RNK(x)==0) thrM("`: Argument cannot have rank 0");
ur xr = RNK(x);
usz ia = IA(x);
if (ia==0) return x;
B xf = getFillQ(x);
u8 xe = TI(x,elType);
if (xr==1 && xe<=el_f64 && isFun(f) && v(f)->flags) {
u8 rtid = v(f)->flags-1;
if (xe==el_bit) {
u64* xp=bitarr_ptr(x);
if (rtid==n_add) {
u64 xs = bit_sum(xp, ia);
if (xs>I32_MAX) goto base;
if (xs<=1) { if (xs==0) return x; goto bit_or; }
return FL_SET(scan_bit_sum(x, xp, ia, xs), fl_asc|fl_squoze);
}
if (rtid==n_or | rtid==n_ceil ) { bit_or:; u64* rp; B r=m_bitarrv(&rp,ia); usz n=BIT_N(ia); u64 xi; usz i=0; while(i<n) if ((xi= vg_rand(xp[i]))!=0) { rp[i] = -(xi&-xi) ; i++; while(i<n) rp[i++] = ~0LL; break; } else rp[i++]= 0 ; decG(x); return r; }
if (rtid==n_and | rtid==n_mul | rtid==n_floor) { u64* rp; B r=m_bitarrv(&rp,ia); usz n=BIT_N(ia); u64 xi; usz i=0; while(i<n) if ((xi=~vg_rand(xp[i]))!=0) { rp[i] = (xi&-xi)-1; i++; while(i<n) rp[i++] = 0 ; break; } else rp[i++]=~0LL; decG(x); return r; }
if (rtid==n_ne) { B r=scan_ne(0, xp, ia); decG(x); return r; }
if (rtid==n_lt) {
u64* rp; B r=m_bitarrv(&rp,ia); usz n=BIT_N(ia);
u64 m10 = 0x5555555555555555;
u64 p = 0;
for (usz i=0; i<n; i++) {
u64 x = xp[i];
u64 c = (m10 & ~(x<<1)) & ~(p>>63);
rp[i] = p = x & (m10 ^ (x + c));
}
decG(x); return r;
}
goto base;
}
if (rtid==n_add) { // +
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=0; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=0; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=0; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
}
if (rtid==n_floor) { // ⌊
#if SINGELI
if (xe==el_i8 ) { i8* rp; B r=m_i8arrv (&rp, ia); avx2_scan_min8 (i8any_ptr (x), rp, ia); decG(x); return r; }
if (xe==el_i16) { i16* rp; B r=m_i16arrv(&rp, ia); avx2_scan_min16(i16any_ptr(x), rp, ia); decG(x); return r; }
if (xe==el_i32) { i32* rp; B r=m_i32arrv(&rp, ia); avx2_scan_min32(i32any_ptr(x), rp, ia); decG(x); return r; }
#else
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); i8* rp; B r=m_i8arrv (&rp, ia); i8 c=I8_MAX ; for (usz i=0; i<ia; i++) { if (xp[i]<c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); i16* rp; B r=m_i16arrv(&rp, ia); i16 c=I16_MAX; for (usz i=0; i<ia; i++) { if (xp[i]<c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=I32_MAX; for (usz i=0; i<ia; i++) { if (xp[i]<c)c=xp[i]; rp[i]=c; } decG(x); return r; }
#endif
}
if (rtid==n_ceil) { // ⌈
#if SINGELI
if (xe==el_i8 ) { i8* rp; B r=m_i8arrv (&rp, ia); avx2_scan_max8 (i8any_ptr (x), rp, ia); decG(x); return r; }
if (xe==el_i16) { i16* rp; B r=m_i16arrv(&rp, ia); avx2_scan_max16(i16any_ptr(x), rp, ia); decG(x); return r; }
if (xe==el_i32) { i32* rp; B r=m_i32arrv(&rp, ia); avx2_scan_max32(i32any_ptr(x), rp, ia); decG(x); return r; }
#else
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); i8* rp; B r=m_i8arrv (&rp, ia); i8 c=I8_MIN ; for (usz i=0; i<ia; i++) { if (xp[i]>c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); i16* rp; B r=m_i16arrv(&rp, ia); i16 c=I16_MIN; for (usz i=0; i<ia; i++) { if (xp[i]>c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=I32_MIN; for (usz i=0; i<ia; i++) { if (xp[i]>c)c=xp[i]; rp[i]=c; } decG(x); return r; }
#endif
}
if (rtid==n_ne) { // ≠
f64 x0 = IGetU(x,0).f; if (x0!=0 && x0!=1) goto base;
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=x0; rp[0]=c; for (usz i=1; i<ia; i++) { c = c!=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=x0; rp[0]=c; for (usz i=1; i<ia; i++) { c = c!=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=x0; rp[0]=c; for (usz i=1; i<ia; i++) { c = c!=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
}
if (rtid==n_or) { //
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=0; for (usz i=0; i<ia; i++) { if ((xp[i]&1)!=xp[i])goto base; c|=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=0; for (usz i=0; i<ia; i++) { if ((xp[i]&1)!=xp[i])goto base; c|=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=0; for (usz i=0; i<ia; i++) { if ((xp[i]&1)!=xp[i])goto base; c|=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
}
}
base:;
SLOW2("𝕎` 𝕩", f, x);
bool reuse = TY(x)==t_harr && reusable(x);
HArr_p r = reuse? harr_parts(REUSE(x)) : m_harr0c(x);
AS2B xget = reuse? TI(x,getU) : TI(x,get); Arr* xa = a(x);
BBB2B fc2 = c2fn(f);
if (xr==1) {
r.a[0] = xget(xa,0);
for (usz i=1; i<ia; i++) r.a[i] = fc2(f, inc(r.a[i-1]), xget(xa,i));
} else {
usz csz = arr_csz(x);
usz i = 0;
for (; i<csz; i++) r.a[i] = xget(xa,i);
for (; i<ia; i++) r.a[i] = fc2(f, inc(r.a[i-csz]), xget(xa,i));
}
if (!reuse) decG(x);
return withFill(r.b, xf);
}
B scan_c2(Md1D* d, B w, B x) { B f = d->f;
if (isAtm(x) || RNK(x)==0) thrM("`: 𝕩 cannot have rank 0");
ur xr = RNK(x); usz* xsh = SH(x); usz ia = IA(x);
B wf = getFillQ(w);
u8 xe = TI(x,elType);
if (xr==1 && q_i32(w) && elInt(xe) && isFun(f) && v(f)->flags) {
u8 rtid = v(f)->flags-1;
i32 wv = o2iG(w);
if (xe==el_bit) {
u64* xp=bitarr_ptr(x);
if (rtid==n_add) { i32* rp; B r=m_i32arrv(&rp, ia); i64 c=wv; for (usz i=0; i<ia; i++) { c+= bitp_get(xp,i); rp[i]=c; } decG(x); return r; }
if (rtid==n_ne) { B r=scan_ne(-(u64)(q_ibit(wv)?wv:1&~*xp), xp, ia); decG(x); return r; }
goto base;
}
if (rtid==n_add) { // +
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
}
if (rtid==n_floor) { // ⌊
if (xe==el_i8 && wv==(i8 )wv) { i8* xp=i8any_ptr (x); i8* rp; B r=m_i8arrv (&rp, ia); i8 c=wv; for (usz i=0; i<ia; i++) { if (xp[i]<c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i16 && wv==(i16)wv) { i16* xp=i16any_ptr(x); i16* rp; B r=m_i16arrv(&rp, ia); i16 c=wv; for (usz i=0; i<ia; i++) { if (xp[i]<c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i32 && wv==(i32)wv) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if (xp[i]<c)c=xp[i]; rp[i]=c; } decG(x); return r; }
}
if (rtid==n_ceil) { // ⌈
if (xe==el_i8 && wv==(i8 )wv) { i8* xp=i8any_ptr (x); i8* rp; B r=m_i8arrv (&rp, ia); i8 c=wv; for (usz i=0; i<ia; i++) { if (xp[i]>c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i16 && wv==(i16)wv) { i16* xp=i16any_ptr(x); i16* rp; B r=m_i16arrv(&rp, ia); i16 c=wv; for (usz i=0; i<ia; i++) { if (xp[i]>c)c=xp[i]; rp[i]=c; } decG(x); return r; }
if (xe==el_i32 && wv==(i32)wv) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if (xp[i]>c)c=xp[i]; rp[i]=c; } decG(x); return r; }
}
if (rtid==n_ne) { // ≠
if (!q_ibit(wv)) { goto base; } bool c=wv;
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); u64* rp; B r=m_bitarrv(&rp, ia); for (usz i=0; i<ia; i++) { c^= xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); u64* rp; B r=m_bitarrv(&rp, ia); for (usz i=0; i<ia; i++) { c^= xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); u64* rp; B r=m_bitarrv(&rp, ia); for (usz i=0; i<ia; i++) { c^= xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
}
}
base:;
SLOW3("𝕨 F` 𝕩", w, x, f);
bool reuse = (TY(x)==t_harr && reusable(x)) | !ia;
usz i = 0;
HArr_p r = reuse? harr_parts(REUSE(x)) : m_harr0c(x);
AS2B xget = reuse? TI(x,getU) : TI(x,get); Arr* xa = a(x);
BBB2B fc2 = c2fn(f);
if (isArr(w)) {
ur wr = RNK(w); usz* wsh = SH(w); SGet(w)
if (wr+1!=xr || !eqShPart(wsh, xsh+1, wr)) thrF("`: Shape of 𝕨 must match the cell of 𝕩 (%H ≡ ≢𝕨, %H ≡ ≢𝕩)", w, x);
if (ia==0) return x;
usz csz = arr_csz(x);
for (; i < csz; i++) r.a[i] = fc2(f, Get(w,i), xget(xa,i));
for (; i < ia; i++) r.a[i] = fc2(f, inc(r.a[i-csz]), xget(xa,i));
decG(w);
} else {
if (xr!=1) thrF("`: Shape of 𝕨 must match the cell of 𝕩 (%H ≡ ≢𝕨, %H ≡ ≢𝕩)", w, x);
if (ia==0) return x;
B pr = r.a[0] = fc2(f, w, xget(xa,0)); i++;
for (; i < ia; i++) r.a[i] = pr = fc2(f, inc(pr), xget(xa,i));
}
if (!reuse) decG(x);
return withFill(r.b, wf);
}
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);

261
src/builtins/scan.c Normal file
View File

@ -0,0 +1,261 @@
#include "../core.h"
#include "../utils/mut.h"
#include "../builtins.h"
#include <math.h>
#define F64_MIN -INFINITY
#define F64_MAX INFINITY
#if !USE_VALGRIND
static u64 vg_rand(u64 x) { return x; }
#endif
#if SINGELI
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wunused-variable"
#include "../singeli/gen/scan.c"
#pragma GCC diagnostic pop
#endif
#if SINGELI && __PCLMUL__
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wunused-variable"
#include "../singeli/gen/neq.c"
#pragma GCC diagnostic pop
#endif
B scan_ne(B x, u64 p, u64 ia) { // consumes x
u64* xp = bitarr_ptr(x);
u64* rp; B r=m_bitarrv(&rp,ia);
#if SINGELI && __PCLMUL__
clmul_scan_ne(p, xp, rp, BIT_N(ia));
#else
for (usz i = 0; i < BIT_N(ia); i++) {
u64 c = xp[i];
u64 r = c ^ (c<<1);
r^= r<< 2; r^= r<< 4; r^= r<<8;
r^= r<<16; r^= r<<32; r^= p;
rp[i] = r;
p = -(r>>63); // repeat sign bit
}
#endif
decG(x); return r;
}
static B scan_or(B x, u64 ia) { // consumes x
u64* xp = bitarr_ptr(x);
u64* rp; B r=m_bitarrv(&rp,ia);
usz n=BIT_N(ia); u64 xi; usz i=0;
while (i<n) if ((xi= vg_rand(xp[i]))!=0) { rp[i] = -(xi&-xi) ; i++; while(i<n) rp[i++] = ~0LL; break; } else rp[i++]=0;
decG(x); return FL_SET(r, fl_asc|fl_squoze);
}
static B scan_and(B x, u64 ia) { // consumes x
u64* xp = bitarr_ptr(x);
u64* rp; B r=m_bitarrv(&rp,ia);
usz n=BIT_N(ia); u64 xi; usz i=0;
while (i<n) if ((xi=~vg_rand(xp[i]))!=0) { rp[i] = (xi&-xi)-1; i++; while(i<n) rp[i++] = 0 ; break; } else rp[i++]=~0LL;
decG(x); return FL_SET(r, fl_dsc|fl_squoze);
}
B slash_c1(B f, B x);
B scan_add_bool(B x, u64 ia) { // consumes x
u64* xp = bitarr_ptr(x);
u64 xs = bit_sum(xp, ia);
if (xs<=1) return xs==0? x : scan_or(x, ia);
B r;
u8 re = xs<=I8_MAX? el_i8 : xs<=I16_MAX? el_i16 : xs<=I32_MAX? el_i32 : el_f64;
if (xs < ia/128) {
B ones = C1(slash, x);
MAKE_MUT(r0, ia) mut_init(r0, re); MUTG_INIT(r0);
SGetU(ones)
usz ri = 0;
for (usz i = 0; i < xs; i++) {
usz e = o2s(GetU(ones, i));
mut_fillG(r0, ri, m_usz(i), e-ri);
ri = e;
}
if (ri<ia) mut_fillG(r0, ri, m_usz(xs), ia-ri);
decG(ones);
r = mut_fv(r0);
} else {
void* rp = m_tyarrv(&r, elWidth(re), ia, el2t(re));
#define SUM_BITWISE(T) { T c=0; for (usz i=0; i<ia; i++) { c+= bitp_get(xp,i); ((T*)rp)[i]=c; } }
#if SINGELI
#define SUM(W,T) avx2_bcs##W(xp, rp, ia);
#else
#define SUM(W,T) SUM_BITWISE(T)
#endif
#define CASE(W) case el_i##W: SUM(W, i##W) break;
switch (re) { default:UD;
CASE(8) CASE(16) CASE(32) case el_f64: SUM_BITWISE(f64) break;
}
#undef CASE
#undef SUM
#undef SUM_BITWISE
decG(x);
}
return FL_SET(r, fl_asc|fl_squoze);
}
// min/max-scan
#if SINGELI
#define MINMAX_SCAN(T,NAME,C,I) avx2_scan_##NAME##_init_##T(xp, rp, ia, I);
#else
#define MINMAX_SCAN(T,NAME,C,I) T c=I; for (usz i=0; i<ia; i++) { if (xp[i] C c)c=xp[i]; rp[i]=c; }
#endif
#define MM_CASE(T,N,C,I) \
case el_##T : { T* xp=T##any_ptr(x); T* rp; r=m_##T##arrv(&rp, ia); MINMAX_SCAN(T,N,C,I); break; }
#define MINMAX(NAME,C,INIT,BIT,ORD) \
B r; switch (xe) { default:UD; \
case el_bit: return scan_##BIT(x, ia); \
MM_CASE(i8 ,NAME,C,I8_##INIT ) \
MM_CASE(i16,NAME,C,I16_##INIT) \
MM_CASE(i32,NAME,C,I32_##INIT) \
MM_CASE(f64,NAME,C,F64_##INIT) \
} \
decG(x); return FL_SET(r, fl_##ORD);
B scan_min_num(B x, u8 xe, u64 ia) { MINMAX(min,<,MAX,and,dsc) }
B scan_max_num(B x, u8 xe, u64 ia) { MINMAX(max,>,MIN,or ,asc) }
#undef MINMAX
// Initialized: try to convert 𝕨 to type of 𝕩
// (could do better for out-of-range floats)
B shape_c2(B, B, B);
#define MM2_ICASE(T,N,C,I) \
case el_##T : { \
if (wv!=(T)wv) { if (wv C 0) { r=C2(shape,m_f64(ia),w); break; } else wv=I; } \
T* xp=T##any_ptr(x); T* rp; r=m_##T##arrv(&rp, ia); MINMAX_SCAN(T,N,C,wv); \
break; }
#define MINMAX2(NAME,C,INIT,BIT,BI,ORD) \
i32 wv=0; if (q_i32(w)) { wv=w.f; } else { x=taga(cpyF64Arr(x)); xe=el_f64; } \
B r; switch (xe) { default:UD; \
case el_bit: if (wv C BI) r=C2(shape,m_f64(ia),w); else return scan_##BIT(x, ia); break; \
MM2_ICASE(i8 ,NAME,C,I8_##INIT ) \
MM2_ICASE(i16,NAME,C,I16_##INIT) \
MM_CASE(i32,NAME,C,wv) \
MM_CASE(f64,NAME,C,w.f) \
} \
decG(x); return FL_SET(r, fl_##ORD);
static B scan2_min_num(B w, B x, u8 xe, usz ia) { MINMAX2(min,<,MAX,and,1,dsc) }
static B scan2_max_num(B w, B x, u8 xe, usz ia) { MINMAX2(max,>,MIN,or ,0,asc) }
#undef MINMAX2
#undef MM_CASE
#undef MM_CASE2
#undef MINMAX_SCAN
B scan_c1(Md1D* d, B x) { B f = d->f;
if (isAtm(x) || RNK(x)==0) thrM("`: Argument cannot have rank 0");
ur xr = RNK(x);
usz ia = IA(x);
if (ia==0) return x;
B xf = getFillQ(x);
u8 xe = TI(x,elType);
if (xr==1 && xe<=el_f64 && isFun(f) && v(f)->flags) {
u8 rtid = v(f)->flags-1;
if (xe==el_bit) {
u64* xp=bitarr_ptr(x);
if (rtid==n_add ) return scan_add_bool(x, ia);
if (rtid==n_or | rtid==n_ceil ) return scan_or(x, ia);
if (rtid==n_and | rtid==n_mul | rtid==n_floor) return scan_and(x, ia);
if (rtid==n_ne ) return scan_ne(x, 0, ia);
if (rtid==n_lt) {
u64* rp; B r=m_bitarrv(&rp,ia); usz n=BIT_N(ia);
u64 m10 = 0x5555555555555555;
u64 p = 0;
for (usz i=0; i<n; i++) {
u64 x = xp[i];
u64 c = (m10 & ~(x<<1)) & ~(p>>63);
rp[i] = p = x & (m10 ^ (x + c));
}
decG(x); return r;
}
goto base;
}
if (rtid==n_add) { // +
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=0; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=0; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=0; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
}
if (rtid==n_floor) return scan_min_num(x, xe, ia); // ⌊
if (rtid==n_ceil ) return scan_max_num(x, xe, ia); // ⌈
if (rtid==n_ne) { // ≠
f64 x0 = IGetU(x,0).f; if (x0!=0 && x0!=1) goto base;
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=x0; rp[0]=c; for (usz i=1; i<ia; i++) { c = c!=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=x0; rp[0]=c; for (usz i=1; i<ia; i++) { c = c!=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); u64* rp; B r=m_bitarrv(&rp,ia); bool c=x0; rp[0]=c; for (usz i=1; i<ia; i++) { c = c!=xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
}
if (rtid==n_or) { x=num_squeezeChk(x); xe=TI(x,elType); if (xe==el_bit) return scan_or(x, ia); }
}
base:;
SLOW2("𝕎` 𝕩", f, x);
bool reuse = TY(x)==t_harr && reusable(x);
HArr_p r = reuse? harr_parts(REUSE(x)) : m_harr0c(x);
AS2B xget = reuse? TI(x,getU) : TI(x,get); Arr* xa = a(x);
BBB2B fc2 = c2fn(f);
if (xr==1) {
r.a[0] = xget(xa,0);
for (usz i=1; i<ia; i++) r.a[i] = fc2(f, inc(r.a[i-1]), xget(xa,i));
} else {
usz csz = arr_csz(x);
usz i = 0;
for (; i<csz; i++) r.a[i] = xget(xa,i);
for (; i<ia; i++) r.a[i] = fc2(f, inc(r.a[i-csz]), xget(xa,i));
}
if (!reuse) decG(x);
return withFill(r.b, xf);
}
B scan_c2(Md1D* d, B w, B x) { B f = d->f;
if (isAtm(x) || RNK(x)==0) thrM("`: 𝕩 cannot have rank 0");
ur xr = RNK(x); usz* xsh = SH(x); usz ia = IA(x);
B wf = getFillQ(w);
u8 xe = TI(x,elType);
if (xr==1 && xe<=el_f64 && isFun(f) && v(f)->flags) {
u8 rtid = v(f)->flags-1;
if (rtid==n_floor) return scan2_min_num(w, x, xe, ia); // ⌊
if (rtid==n_ceil ) return scan2_max_num(w, x, xe, ia); // ⌈
if (!q_i32(w)) goto base;
i32 wv = o2iG(w);
if (xe==el_bit) {
u64* xp=bitarr_ptr(x);
if (rtid==n_add) { i32* rp; B r=m_i32arrv(&rp, ia); i64 c=wv; for (usz i=0; i<ia; i++) { c+= bitp_get(xp,i); rp[i]=c; } decG(x); return r; }
if (rtid==n_ne) return scan_ne(x, -(u64)(q_ibit(wv)?wv:1&~*xp), ia);
goto base;
}
if (rtid==n_add) { // +
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); i32* rp; B r=m_i32arrv(&rp, ia); i32 c=wv; for (usz i=0; i<ia; i++) { if(addOn(c,xp[i]))goto base; rp[i]=c; } decG(x); return r; }
}
if (rtid==n_ne) { // ≠
if (!q_ibit(wv)) { goto base; } bool c=wv;
if (xe==el_i8 ) { i8* xp=i8any_ptr (x); u64* rp; B r=m_bitarrv(&rp, ia); for (usz i=0; i<ia; i++) { c^= xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i16) { i16* xp=i16any_ptr(x); u64* rp; B r=m_bitarrv(&rp, ia); for (usz i=0; i<ia; i++) { c^= xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
if (xe==el_i32) { i32* xp=i32any_ptr(x); u64* rp; B r=m_bitarrv(&rp, ia); for (usz i=0; i<ia; i++) { c^= xp[i]; bitp_set(rp,i,c); } decG(x); return r; }
}
}
base:;
SLOW3("𝕨 F` 𝕩", w, x, f);
bool reuse = (TY(x)==t_harr && reusable(x)) | !ia;
usz i = 0;
HArr_p r = reuse? harr_parts(REUSE(x)) : m_harr0c(x);
AS2B xget = reuse? TI(x,getU) : TI(x,get); Arr* xa = a(x);
BBB2B fc2 = c2fn(f);
if (isArr(w)) {
ur wr = RNK(w); usz* wsh = SH(w); SGet(w)
if (wr+1!=xr || !eqShPart(wsh, xsh+1, wr)) thrF("`: Shape of 𝕨 must match the cell of 𝕩 (%H ≡ ≢𝕨, %H ≡ ≢𝕩)", w, x);
if (ia==0) return x;
usz csz = arr_csz(x);
for (; i < csz; i++) r.a[i] = fc2(f, Get(w,i), xget(xa,i));
for (; i < ia; i++) r.a[i] = fc2(f, inc(r.a[i-csz]), xget(xa,i));
decG(w);
} else {
if (xr!=1) thrF("`: Shape of 𝕨 must match the cell of 𝕩 (%H ≡ ≢𝕨, %H ≡ ≢𝕩)", w, x);
if (ia==0) return x;
B pr = r.a[0] = fc2(f, w, xget(xa,0)); i++;
for (; i < ia; i++) r.a[i] = pr = fc2(f, inc(pr), xget(xa,i));
}
if (!reuse) decG(x);
return withFill(r.b, wf);
}

View File

@ -1,8 +1,8 @@
#include "../core.h"
#include "../utils/hash.h"
#include "../utils/talloc.h"
#include "../utils/calls.h"
#define C2(F,X,W) F##_c2(m_f64(0),X,W)
#define C2i(F, W, X) C2(F, m_i32(W), X)
extern B eq_c2(B,B,B);
extern B ne_c2(B,B,B);

View File

@ -1,11 +1,49 @@
// Self-search: Mark Firsts (∊), Occurrence Count (⊒), Classify (⊐), Deduplicate (⍷)
// Except for trivial cases, ⍷ is implemented as ∊⊸/
// Other functions use adaptations of the same set of methods
// Boolean cases all use special code, including ⍷
// COULD vectorize boolean ⊒ with +`
// Sorted flags: start with r0⌾⊑»⊸≠𝕩 (r0←0 for ⊐, 1 otherwise)
// ∊: ⊢; ⊐: +`; ⊒: ↕∘≠⊸(⊣-⌈`∘×)
// COULD determine ⊒ result type by direct comparisons on 𝕩
// Brute force or all-pairs comparison for small lengths
// Branchless, not vectorized (+´∧` structure for ⊐)
// COULD use direct all-pairs filter, not ∊⊸/, for short ⍷
// Full-size table lookups for 1- and 2-byte 𝕩
// 2-byte table can be "sparse" initialized with an extra pass over 𝕩
// 4-byte ⊐ can use a small-range lookup table
// COULD add small-range 4-byte tables for ∊ and ⊒
// Radix-assisted lookups are fallbacks for 4-byte ∊ and ⊒
// COULD do radix-assisted ⊐ as ⍷⊸⊐ or similar
// Specializes on constant top 1/2 bytes, but hashes make this rare
// Specialized 4-byte and 8-byte hash tables
// In-place resizing by factor of 4 based on measured collisions
// Max collisions ensures bounded time spent here before giving up
// First element used as sentinel (not good for ⊒)
// COULD prefetch when table gets larger
// Generic hash table for other cases
// Resizing is pretty expensive here
// SHOULD widen small odd sizes
#include "../core.h"
#include "../utils/hash.h"
#include "../utils/talloc.h"
#include "../utils/calls.h"
#include "../builtins.h"
B not_c1(B t, B x);
B shape_c1(B t, B x);
B slash_c2(B t, B w, B x);
extern B shape_c1(B, B);
extern B slash_c2(B, B, B);
extern B ud_c1(B, B);
extern B sub_c2(B, B, B);
extern B mul_c2(B, B, B);
extern B scan_add_bool(B x, u64 ia);
extern B scan_max_num(B x, u8 xe, u64 ia);
// These hashes are stored in tables and must be invertible!
#if defined(__SSE4_2__)
static inline u32 hash32(u32 x) { return _mm_crc32_u32(0x973afb51, x); }
#else
@ -24,6 +62,20 @@ static inline u64 hash64(u64 x) {
return x;
}
static inline bool use_sorted(B x, u8 logw) {
if (!FL_HAS(x, fl_asc|fl_dsc)) return 0;
if (logw==6) return TI(x, elType) == el_f64;
return 3<=logw & logw<=5;
}
static inline B shift_ne(B x, usz n, u8 lw, bool r0) { // consumes x
u64* rp; B r = m_bitarrv(&rp, n);
u8* xp = tyany_ptr(x);
u8 lb = lw - 3;
CMP_AA_IMM(ne, el_i8+lb, rp, xp-(1<<lb), xp, n);
bitp_set(rp, 0, r0);
decG(x); return r;
}
static bool canCompare64_norm(B x, usz n) {
u8 e = TI(x,elType);
if (e == el_B) return 0;
@ -169,6 +221,9 @@ B memberOf_c1(B t, B x) {
rp[0]=1; if (i<n) bitp_set(rp, i, 1);
return r;
}
if (use_sorted(x, lw)) {
return shift_ne(x, n, lw, 1);
}
#define BRUTE(T) \
i##T* xp = xv; \
u64 rv = 1; \
@ -247,7 +302,32 @@ B count_c1(B t, B x) {
if (n>(usz)I32_MAX+1) thrM("⊒: Argument length >2⋆31 not supported");
u8 lw = cellWidthLog(x);
if (lw==0) { x = toI8Any(x); lw = cellWidthLog(x); }
if (lw==0) {
u64* xp = bitarr_ptr(x);
B r;
#define COUNT_BOOL(T) \
T* rp; r = m_##T##arrv(&rp, n); \
usz n1 = 0; \
for (usz i=0; i<n; ) { \
u64 bb = xp[i/64]; \
for (usz e=n-i<64?n:i+64; i<e; i++) { \
bool b = bb&1; bb>>=1; \
rp[i] = b? n1 : i-n1; \
n1 += b; \
} \
}
if (n <= 128) { COUNT_BOOL(i8) }
else if (n <= 1<<15) { COUNT_BOOL(i16) }
else { COUNT_BOOL(i32) }
decG(x); return r;
#undef COUNT_BOOL
}
if (use_sorted(x, lw) && n>16 && (lw>4 || n<1<<16)) { // ↕∘≠(⊣-⌈`∘×)∊
B c = shift_ne(x, n, lw, 1);
B i = C1(ud, m_f64(n));
B m = C2(mul, c, inc(i));
return C2(sub, i, scan_max_num(m, TI(m,elType), n));
}
void* xv = tyany_ptr(x);
#define BRUTE(T) \
i##T* xp = xv; \
@ -280,7 +360,7 @@ B count_c1(B t, B x) {
/*THRESHMUL*/1, THRESH, \
/*INIT*/u32 ctr0 = 1;)
if (lw==5) {
if (n<12) { BRUTE(32); }
if (n<20) { BRUTE(32); }
i32* rp; B r = m_i32arrv(&rp, n);
HASHTAB(u32, 32, 1, n/2, sz==msz? 1 : sz>=(1<<14)? 3 : 5)
// Radix-assisted lookup
@ -334,8 +414,11 @@ B indexOf_c1(B t, B x) {
u8 lw = cellWidthLog(x);
void* xv = tyany_ptr(x);
if (lw == 0) {
B r = 1&*(u64*)xv ? not_c1(m_f64(0), x) : x;
return shape_c1(m_f64(0), r);
B r = 1&*(u64*)xv ? bit_negate(x) : x;
return C1(shape, r);
}
if (use_sorted(x, lw) && n>8) {
return scan_add_bool(shift_ne(x, n, lw, 0), n);
}
#define BRUTE(T) \
i##T* xp = xv; \
@ -426,5 +509,13 @@ B find_c1(B t, B x) {
if (isAtm(x) || RNK(x)==0) thrM("⍷: Argument cannot have rank 0");
usz n = *SH(x);
if (n<=1) return x;
return slash_c2(m_f64(0), memberOf_c1(m_f64(0), inc(x)), x);
if (TI(x,elType)==el_bit && RNK(x)==1) {
u64* xp = bitarr_ptr(x);
u64 x0 = 1 & *xp;
usz i = bit_find(xp, n, !x0); decG(x);
u64* rp; B r = m_bitarrv(&rp, 1 + (i<n));
rp[0] = 2 ^ -x0;
return r;
}
return C2(slash, C1(memberOf, inc(x)), x);
}

View File

@ -168,7 +168,7 @@
extern void (*const avx2_scan_pluswrap_u32)(uint32_t* v0,uint32_t* v1,uint64_t v2,uint32_t v3);
#define avx2_scan_pluswrap_u64(V0,V1,V2,V3) for (usz i=k; i<e; i++) js=rp[i]+=js;
#define PLUS_SCAN(T) avx2_scan_pluswrap_##T(rp+k,rp+k,e-k,js); js=rp[e-1];
extern void (*const avx2_scan_max32)(int32_t* v0,int32_t* v1,uint64_t v2);
extern void (*const avx2_scan_max_i32)(int32_t* v0,int32_t* v1,uint64_t v2);
#else
#define PLUS_SCAN(T) for (usz i=k; i<e; i++) js=rp[i]+=js;
#endif
@ -413,7 +413,6 @@ static NOINLINE B zeroCells(B x) { // doesn't consume
return r;
}
B not_c1(B t, B x);
B grade_bool(B x, usz xia, bool up) {
#define BRANCHLESS_GRADE(T) \
T* rp; r = m_##T##arrv(&rp, xia); \
@ -433,7 +432,7 @@ B grade_bool(B x, usz xia, bool up) {
#if SINGELI && defined(__BMI2__)
if (xia < 16) { BRANCHLESS_GRADE(i8) }
else if (xia <= 1<<15) {
B notx = not_c1(m_f64(0), inc(x));
B notx = bit_negate(inc(x));
u64* xp0 = bitarr_ptr(notx);
u64* xp1 = xp;
if (!up) { u64* t=xp1; xp1=xp0; xp0=t; }
@ -607,7 +606,7 @@ static B compress(B w, B x, usz wia, u8 xl, u8 xt) {
#if SINGELI
#define IND_BY_SCAN \
SCAN_CORE(xp[j], rp[ij]=j, rp[k]=j, avx2_scan_max32(rp+k,rp+k,e-k))
SCAN_CORE(xp[j], rp[ij]=j, rp[k]=j, avx2_scan_max_i32(rp+k,rp+k,e-k))
#else
#define IND_BY_SCAN usz js=0; SUM_CORE(i32, xp[j], , 1)
#endif

View File

@ -22,6 +22,7 @@
#include "../builtins/sort.c"
#include "../builtins/search.c"
#include "../builtins/selfsearch.c"
#include "../builtins/scan.c"
#include "../builtins/arithm.c"
#include "../builtins/arithd.c"
#include "../builtins/cmp.c"

View File

@ -32,14 +32,15 @@ def scan_post{T, init, x:*T, r:*T, len:u64, op, pre} = {
def last{v, p} = op{pre{v}, p}
def scan{v, p} = {
n:= last{v, p}
p = sel{[8]i32, spread{n}, broadcast{[8]i32, 7}}
p = (if (width{T}<=32) sel{[8]i32, spread{n}, broadcast{[8]i32, 7}};
else shuf{[4]u64, n, 4b3333})
n
}
scan_loop{T, init, x, r, len, scan, last}
}
# Associative scan ?` if a?b?a = a?b = b?a, used for ⌊⌈
avx2_scan_idem{T, op, id}(x:*T, r:*T, len:u64) : void = {
avx2_scan_idem{T, op}(x:*T, r:*T, len:u64, init:T) : void = {
# Within each lane, scan using shifts by powers of 2. First k elements
# when shifting by k don't need to change, so leave them alone.
def w = width{T}
@ -53,18 +54,25 @@ avx2_scan_idem{T, op, id}(x:*T, r:*T, len:u64) : void = {
op{b, sel{[8]i32, spread{b}, make{[8]i32, 3*(3<iota{8})}}}
}
scan_post{T, id, x, r, len, op, pre}
scan_post{T, init, x, r, len, op, pre}
}
def avx2_scan_idem{T, op} = {
avx2_scan_idem{T==f64, op}(x:*T, r:*T, len:u64, init:T) : void = {
def sh{s, a} = op{a, shuf{[4]u64, a, s}}
scan_post{T, init, x, r, len, op, {a}=>sh{4b1110,sh{4b2200,a}}}
}
'avx2_scan_min_init_i8' = avx2_scan_idem{i8 , min}; 'avx2_scan_max_init_i8' = avx2_scan_idem{i8 , max}
'avx2_scan_min_init_i16' = avx2_scan_idem{i16, min}; 'avx2_scan_max_init_i16' = avx2_scan_idem{i16, max}
'avx2_scan_min_init_i32' = avx2_scan_idem{i32, min}; 'avx2_scan_max_init_i32' = avx2_scan_idem{i32, max}
'avx2_scan_min_init_f64' = avx2_scan_idem{f64, min}; 'avx2_scan_max_init_f64' = avx2_scan_idem{f64, max}
avx2_scan_idem_id{T, op}(x:*T, r:*T, len:u64) : void = {
def m = 1 << (width{T}-1)
avx2_scan_idem{T, op, (if (match{op,min}) m-1; else -m)}
def id = (if (match{op,min}) m-1; else -m)
avx2_scan_idem{T, op}(x, r, len, id)
}
'avx2_scan_min8' = avx2_scan_idem{i8 , min}
'avx2_scan_max8' = avx2_scan_idem{i8 , max}
'avx2_scan_min16' = avx2_scan_idem{i16, min}
'avx2_scan_max16' = avx2_scan_idem{i16, max}
'avx2_scan_min32' = avx2_scan_idem{i32, min}
'avx2_scan_max32' = avx2_scan_idem{i32, max}
'avx2_scan_min_i8' = avx2_scan_idem_id{i8 , min}; 'avx2_scan_max_i8' = avx2_scan_idem_id{i8 , max}
'avx2_scan_min_i16' = avx2_scan_idem_id{i16, min}; 'avx2_scan_max_i16' = avx2_scan_idem_id{i16, max}
'avx2_scan_min_i32' = avx2_scan_idem_id{i32, min}; 'avx2_scan_max_i32' = avx2_scan_idem_id{i32, max}
# Associative scan
avx2_scan_assoc_0{T, op}(x:*T, r:*T, len:u64, init:T) : void = {

View File

@ -1,5 +1,12 @@
#pragma once
#define C1(F, X) F##_c1(m_f64(0), X)
#define C2(F,W,X) F##_c2(m_f64(0),W,X)
#define M1C1(M,F, X) m1c1_unsafe(M##_c1, bi_##F, X)
#define M1C2(M,F,W,X) m1c2_unsafe(M##_c2, bi_##F, W, X)
static inline B m1c1_unsafe(D1C1 m, B f, B x) { Md1D d; d.f=f; return m(&d, x); }
static inline B m1c2_unsafe(D1C2 m, B f, B w, B x) { Md1D d; d.f=f; return m(&d, w, x); }
typedef void (*M_CopyF)(void*, usz, B, usz, usz);
typedef void (*M_FillF)(void*, usz, B, usz);
extern M_CopyF copyFns[el_MAX];
@ -28,4 +35,4 @@ CMP_DEF(le, AS);
#define CMP_AA_IMM(FN, ELT, WHERE, WP, XP, LEN) CMP_AA_CALL(CMP_AA_FN(FN, ELT), WHERE, WP, XP, LEN)
#define CMP_AS_IMM(FN, ELT, WHERE, WP, X, LEN) CMP_AS_CALL(CMP_AS_FN(FN, ELT), WHERE, WP, X, LEN)
void bit_negatePtr(u64* rp, u64* xp, usz count); // count is number of u64-s
void bit_negatePtr(u64* rp, u64* xp, usz count); // count is number of u64-s