flexible clearN_ that can start at any index

This commit is contained in:
daanx 2024-12-15 13:21:13 -08:00
parent b5dfd233e9
commit 4aeb2e1005

View file

@ -26,6 +26,10 @@ static inline size_t mi_bfield_ctz(mi_bfield_t x) {
return mi_ctz(x);
}
static inline size_t mi_bfield_clz(mi_bfield_t x) {
return mi_clz(x);
}
static inline size_t mi_bfield_popcount(mi_bfield_t x) {
return mi_popcount(x);
}
@ -41,6 +45,15 @@ static inline bool mi_bfield_find_least_bit(mi_bfield_t x, size_t* idx) {
return mi_bsf(x,idx);
}
// find the most significant bit that is set.
// return false if `x==0` (with `*idx` undefined) and true otherwise,
// with the `idx` is set to the bit index (`0 <= *idx < MI_BFIELD_BITS`).
static inline bool mi_bfield_find_highest_bit(mi_bfield_t x, size_t* idx) {
return mi_bsr(x, idx);
}
// find each set bit in a bit field `x` and clear it, until it becomes zero.
static inline bool mi_bfield_foreach_bit(mi_bfield_t* x, size_t* idx) {
const bool found = mi_bfield_find_least_bit(*x, idx);
@ -598,9 +611,9 @@ static mi_decl_noinline bool mi_bchunk_try_find_and_clear8(mi_bchunk_t* chunk, s
#if MI_OPT_SIMD && defined(__AVX2__) && (MI_BCHUNK_BITS==512)
while (true) {
// since a cache-line is 64b, load all at once
const __m256i vec1 = _mm256_load_si256((const __m256i*)chunk->bfields);
const __m256i vec2 = _mm256_load_si256((const __m256i*)chunk->bfields+1);
const __m256i cmpv = mi_mm256_ones();
const __m256i vec1 = _mm256_load_si256((const __m256i*)chunk->bfields);
const __m256i vec2 = _mm256_load_si256((const __m256i*)chunk->bfields+1);
const __m256i cmpv = mi_mm256_ones();
const __m256i vcmp1 = _mm256_cmpeq_epi8(vec1, cmpv); // (byte == ~0 ? 0xFF : 0)
const __m256i vcmp2 = _mm256_cmpeq_epi8(vec2, cmpv); // (byte == ~0 ? 0xFF : 0)
const uint32_t mask1 = _mm256_movemask_epi8(vcmp1); // mask of most significant bit of each byte
@ -610,7 +623,7 @@ static mi_decl_noinline bool mi_bchunk_try_find_and_clear8(mi_bchunk_t* chunk, s
if (mask==0) return false;
const size_t bidx = _tzcnt_u64(mask); // byte-idx of the byte in the chunk
const size_t chunk_idx = bidx / 8;
const size_t idx = (bidx % 8)*8;
const size_t idx = (bidx % 8)*8;
mi_assert_internal(chunk_idx < MI_BCHUNK_FIELDS);
if mi_likely(mi_bfield_atomic_try_clear8(&chunk->bfields[chunk_idx], idx, NULL)) { // clear it atomically
*pidx = (chunk_idx*MI_BFIELD_BITS) + idx;
@ -618,6 +631,7 @@ static mi_decl_noinline bool mi_bchunk_try_find_and_clear8(mi_bchunk_t* chunk, s
return true;
}
// try again
// note: there must be an atomic release/acquire in between or otherwise the registers may not be reloaded }
}
#else
// first skip allset fields to reduce fragmentation
@ -664,6 +678,7 @@ static mi_decl_noinline bool mi_bchunk_try_find_and_clearX(mi_bchunk_t* chunk,
return true;
}
// try again
// note: there must be an atomic release/acquire in between or otherwise the registers may not be reloaded
}
#else
for (int i = 0; i < MI_BCHUNK_FIELDS; i++) {
@ -684,7 +699,8 @@ static inline bool mi_bchunk_try_find_and_clear_X(mi_bchunk_t* chunk, size_t n,
}
// find a sequence of `n` bits in a chunk with `n < MI_BFIELD_BITS` with all bits set,
// and try to clear them atomically.
// and try to clear them atomically.
// Currently does not cross bfield boundaries.
// set `*pidx` to its bit index (0 <= *pidx <= MI_BCHUNK_BITS - n) on success.
// (We do not cross bfield boundaries)
mi_decl_noinline static bool mi_bchunk_try_find_and_clearNX(mi_bchunk_t* chunk, size_t n, size_t* pidx) {
@ -732,35 +748,51 @@ mi_decl_noinline static bool mi_bchunk_try_find_and_clearNX(mi_bchunk_t* chunk,
static mi_decl_noinline bool mi_bchunk_try_find_and_clearN_(mi_bchunk_t* chunk, size_t n, size_t* pidx) {
if (n == 0 || n > MI_BCHUNK_BITS) return false; // cannot be more than a chunk
// we align at a bfield, and scan `field_count` fields
// n >= MI_BFIELD_BITS; find a first field that is 0
const size_t field_count = _mi_divide_up(n, MI_BFIELD_BITS); // we need this many fields
for (size_t i = 0; i <= MI_BCHUNK_FIELDS - field_count; i++)
const size_t skip_count = n/MI_BFIELD_BITS;
size_t cidx;
for (size_t i = 0; i <= MI_BCHUNK_FIELDS - skip_count; i++)
{
// first pre-scan for a range of fields that are all set (up to the last one)
bool allset = true;
size_t j = 0;
size_t m = n;
do {
mi_assert_internal(i + j < MI_BCHUNK_FIELDS);
mi_bfield_t b = mi_atomic_load_relaxed(&chunk->bfields[i+j]);
size_t idx;
if (mi_bfield_find_least_bit(~b,&idx)) {
if (m > idx) {
allset = false;
i += j; // no need to look again at the previous fields
break;
}
size_t j = 1; // field count from i
size_t m = n; // bits to go
// first field
mi_bfield_t b = mi_atomic_load_relaxed(&chunk->bfields[i]);
size_t ones = mi_bfield_clz(~b);
cidx = i*MI_BFIELD_BITS + (MI_BFIELD_BITS - ones); // start index
if (ones >= m) {
// we found enough bits!
m = 0;
}
else {
m -= ones;
mi_assert_internal(m>0);
}
// keep scanning further fields?
while (i+j < MI_BCHUNK_FIELDS) {
mi_assert_internal(m > 0);
b = mi_atomic_load_relaxed(&chunk->bfields[i+j]);
ones = mi_bfield_ctz(~b);
if (ones >= m) {
// we found enough bits
m = 0;
break;
}
else if (ones == MI_BFIELD_BITS) {
// not enough yet, proceed to the next field
j++;
m -= MI_BFIELD_BITS;
}
else {
// all bits in b were set
m -= MI_BFIELD_BITS; // note: can underflow
// the range was not enough, start from scratch
i = i + j - 1; // no need to re-scan previous fields, except the last one (with clz this time)
mi_assert_internal(m>0);
break;
}
} while (++j < field_count);
// if all set, we can try to atomically clear them
if (allset) {
const size_t cidx = i*MI_BFIELD_BITS;
}
// did we find a range?
if (m==0) {
if (mi_bchunk_try_clearN(chunk, cidx, n, NULL)) {
// we cleared all atomically
*pidx = cidx;
@ -768,8 +800,9 @@ static mi_decl_noinline bool mi_bchunk_try_find_and_clearN_(mi_bchunk_t* chunk,
mi_assert_internal(*pidx + n <= MI_BCHUNK_BITS);
return true;
}
// note: if we fail for a small `n` on the first field, we don't rescan that field (as `i` is incremented)
}
// continue
// otherwise continue searching
}
return false;
}