add expand128() implementation for NEON

This commit is contained in:
Konstantinos Margaritis 2021-01-15 17:33:41 +02:00
parent 0372a8120a
commit 6a11c83630
5 changed files with 48 additions and 7 deletions

View File

@ -106,7 +106,6 @@ u64a compress64_impl(u64a x, u64a m) {
static really_inline static really_inline
m128 compress128_impl(m128 x, m128 m) { m128 compress128_impl(m128 x, m128 m) {
m128 one = set1_2x64(1); m128 one = set1_2x64(1);
m128 bitset = one; m128 bitset = one;
m128 vres = zeroes128(); m128 vres = zeroes128();
@ -118,7 +117,7 @@ m128 compress128_impl(m128 x, m128 m) {
m128 mask = not128(eq64_m128(tv, zeroes128())); m128 mask = not128(eq64_m128(tv, zeroes128()));
mask = vandq_s64(bitset, mask); mask = vandq_s64(bitset, mask);
vres = or128(vres, mask); vres = or128(vres, mask);
m = and128(m, sub_2x64(m, set1_2x64(1))); m = and128(m, sub_2x64(m, one));
bitset = lshift64_m128(bitset, 1); bitset = lshift64_m128(bitset, 1);
} }
return vres; return vres;
@ -134,6 +133,25 @@ u64a expand64_impl(u64a x, u64a m) {
return expand64_impl_c(x, m); return expand64_impl_c(x, m);
} }
static really_inline
m128 expand128_impl(m128 x, m128 m) {
m128 one = set1_2x64(1);
m128 bitset = one;
m128 vres = zeroes128();
while (isnonzero128(m)) {
m128 tv = and128(x, m);
m128 mm = sub_2x64(zeroes128(), m);
m128 mask = not128(eq64_m128(tv, zeroes128()));
mask = vandq_s64(bitset, mask);
mask = and128(mask, mm);
vres = or128(vres, mask);
m = and128(m, sub_2x64(m, one));
bitset = lshift64_m128(bitset, 1);
}
return vres;
}
/* returns the first set bit after begin (if not ~0U). If no bit is set after /* returns the first set bit after begin (if not ~0U). If no bit is set after
* begin returns ~0U * begin returns ~0U
*/ */

View File

@ -301,6 +301,18 @@ u64a expand64_impl_c(u64a x, u64a m) {
return x & m0; // clear out extraneous bits*/ return x & m0; // clear out extraneous bits*/
} }
static really_inline
m128 expand128_impl_c(m128 xvec, m128 mvec) {
u64a ALIGN_ATTR(16) x[2];
u64a ALIGN_ATTR(16) m[2];
store128(x, xvec);
store128(m, mvec);
expand64_impl_c(x[0], m[0]);
expand64_impl_c(x[1], m[1]);
return xvec;
}
/* returns the first set bit after begin (if not ~0U). If no bit is set after /* returns the first set bit after begin (if not ~0U). If no bit is set after
* begin returns ~0U * begin returns ~0U

View File

@ -239,6 +239,11 @@ u64a expand64_impl(u64a x, u64a m) {
#endif #endif
} }
static really_inline
m128 expand128_impl(m128 x, m128 m) {
return expand128_impl_c(x, m);
}
/* returns the first set bit after begin (if not ~0U). If no bit is set after /* returns the first set bit after begin (if not ~0U). If no bit is set after
* begin returns ~0U * begin returns ~0U
*/ */

View File

@ -135,6 +135,10 @@ u64a expand64(u64a x, u64a m) {
return expand64_impl(x, m); return expand64_impl(x, m);
} }
static really_inline
m128 expand128(m128 x, m128 m) {
return expand128_impl(x, m);
}
/* returns the first set bit after begin (if not ~0U). If no bit is set after /* returns the first set bit after begin (if not ~0U). If no bit is set after
* begin returns ~0U * begin returns ~0U

View File

@ -162,14 +162,16 @@ m128 loadcompressed128_64bit(const void *ptr, m128 mvec) {
u64a ALIGN_ATTR(16) m[2]; u64a ALIGN_ATTR(16) m[2];
store128(m, mvec); store128(m, mvec);
u32 bits[2] = { popcount64(m[0]), popcount64(m[1]) }; // Count the number of bits of compressed state we're writing out per
// chunk.
u32 ALIGN_ATTR(16) bits[2] = { popcount64(m[0]), popcount64(m[1]) };
u64a ALIGN_ATTR(16) v[2]; u64a ALIGN_ATTR(16) v[2];
unpack_bits_64(v, (const u8 *)ptr, bits, 2); unpack_bits_64(v, (const u8 *)ptr, bits, 2);
m128 xvec = load128(v);
u64a x[2] = { expand64(v[0], m[0]), expand64(v[1], m[1]) }; // Expand vector
return expand128(xvec, mvec);
return set2x64(x[1], x[0]);
} }
#endif #endif