mirror of
https://github.com/bkaradzic/bx.git
synced 2026-10-23 01:01:06 +00:00
Added SIMD version of BitArray find. (#464)
This commit is contained in:
committed by
GitHub
parent
ad000aeb31
commit
fbadfc3abc
@@ -979,6 +979,38 @@ TEST_CASE("BitArray-ranges-ending-on-word-boundary", "[container]")
|
||||
REQUIRE( (UINT64_MAX >> 28) == words[2]);
|
||||
}
|
||||
|
||||
namespace bx
|
||||
{
|
||||
uint32_t bitSetFindRef(const uint64_t* _ptr, uint32_t _beginBit, uint32_t _endBit, uint32_t _numBits);
|
||||
uint32_t bitSetFindSimd(const uint64_t* _ptr, uint32_t _beginBit, uint32_t _endBit, uint32_t _numBits);
|
||||
uint32_t bitSetFindWordRef(const uint64_t* _ptr, uint32_t _beginWord, uint32_t _endWord, uint32_t _match);
|
||||
uint32_t bitSetFindWordSimd(const uint64_t* _ptr, uint32_t _beginWord, uint32_t _endWord, uint32_t _match);
|
||||
uint32_t bitSetFindLastWordRef(const uint64_t* _ptr, uint32_t _beginWord, uint32_t _endWord, uint32_t _match);
|
||||
uint32_t bitSetFindLastWordSimd(const uint64_t* _ptr, uint32_t _beginWord, uint32_t _endWord, uint32_t _match);
|
||||
}
|
||||
|
||||
static bool getBit(const uint64_t* _bits, uint32_t _idx)
|
||||
{
|
||||
return 0 != ( (_bits[_idx/64] >> (_idx & 63) ) & 1);
|
||||
}
|
||||
|
||||
static uint32_t findFirstFit(const uint64_t* _bits, uint32_t _beginBit, uint32_t _endBit, uint32_t _numBits)
|
||||
{
|
||||
uint32_t run = 0;
|
||||
|
||||
for (uint32_t ii = _beginBit; ii < _endBit; ++ii)
|
||||
{
|
||||
run = getBit(_bits, ii) ? 0 : run + 1;
|
||||
|
||||
if (run == _numBits)
|
||||
{
|
||||
return ii + 1 - _numBits;
|
||||
}
|
||||
}
|
||||
|
||||
return bx::kInvalid;
|
||||
}
|
||||
|
||||
static uint32_t findRef(const bx::MutableBitArrayView& _bs, uint32_t _beginBit, uint32_t _endBit, uint32_t _numBits)
|
||||
{
|
||||
for (uint32_t ii = _beginBit; ii + _numBits <= _endBit; ++ii)
|
||||
@@ -1073,7 +1105,7 @@ TEST_CASE("BitArray-find-bench", "[.][!benchmark]")
|
||||
|
||||
char name[64];
|
||||
|
||||
for (uint32_t numBits : { 4u, 12u, 40u, 200u, 300u })
|
||||
for (uint32_t numBits : { 1u, 2u, 4u, 6u, 7u, 12u, 14u, 15u, 24u, 31u, 40u, 63u, 64u, 127u, 200u, 300u })
|
||||
{
|
||||
uint32_t idx = 0;
|
||||
|
||||
@@ -1084,6 +1116,399 @@ TEST_CASE("BitArray-find-bench", "[.][!benchmark]")
|
||||
return idx;
|
||||
};
|
||||
|
||||
bx::snprintf(name, sizeof(name), "find ref %u", numBits);
|
||||
BENCHMARK(name)
|
||||
{
|
||||
return bx::bitSetFindRef(bits, 0, kNumBits, numBits);
|
||||
};
|
||||
|
||||
REQUIRE( (bx::kInvalid == idx) == (bx::kInvalid == findRef(bs, 0, kNumBits, numBits) ) );
|
||||
}
|
||||
}
|
||||
|
||||
TEST_CASE("BitArray-find-simd-matches-ref", "[container]")
|
||||
{
|
||||
constexpr uint32_t kMaxWords = 80;
|
||||
|
||||
uint64_t bits[kMaxWords];
|
||||
|
||||
bx::RngMwc rng;
|
||||
|
||||
for (uint32_t iter = 0; iter < 50000; ++iter)
|
||||
{
|
||||
const uint32_t numWords = 1 + rng.gen() % kMaxWords;
|
||||
const uint32_t numBits = numWords*64;
|
||||
bx::MutableBitArrayView bs(bits, numWords*sizeof(uint64_t) );
|
||||
|
||||
const uint32_t density = rng.gen() % 6;
|
||||
|
||||
bs.clear();
|
||||
|
||||
for (uint32_t ii = 0, num = rng.gen() % (4 + numWords*4); ii < num; ++ii)
|
||||
{
|
||||
const uint32_t begin = rng.gen() % numBits;
|
||||
const uint32_t end = bx::min(numBits, begin + 1 + rng.gen() % (4 << density) );
|
||||
bs.set(begin, end, true);
|
||||
}
|
||||
|
||||
if (0 == (rng.gen() & 3) )
|
||||
{
|
||||
bs.set(0, numBits, true);
|
||||
|
||||
for (uint32_t ii = 0, num = rng.gen() % 4; ii < num; ++ii)
|
||||
{
|
||||
const uint32_t begin = rng.gen() % numBits;
|
||||
const uint32_t end = bx::min(numBits, begin + 1 + rng.gen() % 320);
|
||||
bs.set(begin, end, false);
|
||||
}
|
||||
}
|
||||
|
||||
const uint32_t runBits = 0 == (rng.gen() & 1) ? 1 + rng.gen() % 8 : 1 + rng.gen() % 300;
|
||||
const uint32_t beginBit = rng.gen() % numBits;
|
||||
const uint32_t endBit = beginBit + 1 + rng.gen() % (numBits - beginBit);
|
||||
|
||||
const uint32_t ref = bx::bitSetFindRef (bits, beginBit, endBit, runBits);
|
||||
const uint32_t simd = bx::bitSetFindSimd(bits, beginBit, endBit, runBits);
|
||||
|
||||
INFO("Iteration " << iter << ", words " << numWords << ", range [" << beginBit << ", " << endBit << "), num bits " << runBits << ", simd " << simd << ", ref " << ref);
|
||||
|
||||
REQUIRE(ref == simd);
|
||||
REQUIRE(findFirstFit(bits, beginBit, endBit, runBits) == simd);
|
||||
}
|
||||
}
|
||||
|
||||
TEST_CASE("BitArray-ops-bench", "[.][!benchmark]")
|
||||
{
|
||||
constexpr uint32_t kNumBits = 64<<10;
|
||||
|
||||
alignas(64) static uint64_t bits[kNumBits/64];
|
||||
bx::MutableBitArrayView bs(bits, sizeof(bits) );
|
||||
|
||||
char name[64];
|
||||
|
||||
for (uint32_t numBits : { 7u, 64u, 300u, 4096u })
|
||||
{
|
||||
bs.clear();
|
||||
|
||||
bx::snprintf(name, sizeof(name), "set %u", numBits);
|
||||
BENCHMARK(name)
|
||||
{
|
||||
bs.set(13, 13 + numBits, true);
|
||||
bs.set(13, 13 + numBits, false);
|
||||
return bits[0];
|
||||
};
|
||||
}
|
||||
|
||||
bs.clear();
|
||||
|
||||
BENCHMARK("countBits 64K")
|
||||
{
|
||||
return bs.countBits(1, kNumBits - 1);
|
||||
};
|
||||
|
||||
BENCHMARK("testNone 64K")
|
||||
{
|
||||
return bs.testNone(1, kNumBits - 1);
|
||||
};
|
||||
|
||||
BENCHMARK("testAny 64K")
|
||||
{
|
||||
return bs.testAny(1, kNumBits - 1);
|
||||
};
|
||||
|
||||
BENCHMARK("findSet first 64K")
|
||||
{
|
||||
return bs.findSet(bx::kFirst);
|
||||
};
|
||||
|
||||
BENCHMARK("findSet range 64K")
|
||||
{
|
||||
return bs.findSet(1, kNumBits - 1);
|
||||
};
|
||||
|
||||
bs.set(0, kNumBits, true);
|
||||
|
||||
BENCHMARK("testAll 64K")
|
||||
{
|
||||
return bs.testAll(1, kNumBits - 1);
|
||||
};
|
||||
|
||||
BENCHMARK("findClear range 64K")
|
||||
{
|
||||
return bs.findClear(1, kNumBits - 1);
|
||||
};
|
||||
|
||||
BENCHMARK("find 1 64K")
|
||||
{
|
||||
return bs.find(1);
|
||||
};
|
||||
|
||||
BENCHMARK("find 12 64K")
|
||||
{
|
||||
return bs.find(12);
|
||||
};
|
||||
|
||||
bx::RngMwc rng;
|
||||
|
||||
for (uint32_t ii = 0; ii < kNumBits; ++ii)
|
||||
{
|
||||
bits[ii/64] = (uint64_t(rng.gen() ) << 32) | rng.gen();
|
||||
}
|
||||
|
||||
BENCHMARK("countBits random 64K")
|
||||
{
|
||||
return bs.countBits(1, kNumBits - 1);
|
||||
};
|
||||
|
||||
for (uint32_t numBits : { 300u, 4096u })
|
||||
{
|
||||
bx::snprintf(name, sizeof(name), "blit %u", numBits);
|
||||
BENCHMARK(name)
|
||||
{
|
||||
bs.blit(1000, 3, numBits);
|
||||
return bits[20];
|
||||
};
|
||||
|
||||
bx::snprintf(name, sizeof(name), "blit aligned %u", numBits);
|
||||
BENCHMARK(name)
|
||||
{
|
||||
bs.blit(1024, 128, numBits);
|
||||
return bits[20];
|
||||
};
|
||||
}
|
||||
}
|
||||
|
||||
static void fillRandom(bx::RngMwc& _rng, uint64_t* _bits, uint32_t _numWords)
|
||||
{
|
||||
const uint32_t mode = _rng.gen() % 4;
|
||||
|
||||
for (uint32_t ii = 0; ii < _numWords; ++ii)
|
||||
{
|
||||
const uint64_t word = (uint64_t(_rng.gen() ) << 32) | _rng.gen();
|
||||
const uint32_t pick = _rng.gen() % 8;
|
||||
|
||||
switch (mode)
|
||||
{
|
||||
case 0: _bits[ii] = word; break;
|
||||
case 1: _bits[ii] = 0 == pick ? word : 0; break;
|
||||
case 2: _bits[ii] = 0 == pick ? word : UINT64_MAX; break;
|
||||
default: _bits[ii] = 0 == pick ? word : 1 == pick ? 0 : UINT64_MAX; break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
TEST_CASE("BitArray-find-word-simd-matches-ref", "[container]")
|
||||
{
|
||||
constexpr uint32_t kMaxWords = 80;
|
||||
|
||||
uint64_t bits[kMaxWords];
|
||||
|
||||
bx::RngMwc rng;
|
||||
|
||||
for (uint32_t iter = 0; iter < 50000; ++iter)
|
||||
{
|
||||
const uint32_t numWords = 1 + rng.gen() % kMaxWords;
|
||||
fillRandom(rng, bits, numWords);
|
||||
|
||||
const uint32_t begin = rng.gen() % (numWords + 1);
|
||||
const uint32_t end = begin + rng.gen() % (numWords - begin + 1);
|
||||
const uint32_t match = rng.gen() % 3;
|
||||
|
||||
INFO("Iteration " << iter << ", words " << numWords << ", range [" << begin << ", " << end << "), match " << match);
|
||||
|
||||
REQUIRE(bx::bitSetFindWordRef(bits, begin, end, match) == bx::bitSetFindWordSimd(bits, begin, end, match) );
|
||||
REQUIRE(bx::bitSetFindLastWordRef(bits, begin, end, match) == bx::bitSetFindLastWordSimd(bits, begin, end, match) );
|
||||
}
|
||||
}
|
||||
|
||||
TEST_CASE("BitArray-range-ops-random", "[container]")
|
||||
{
|
||||
constexpr uint32_t kMaxWords = 80;
|
||||
|
||||
uint64_t bits[kMaxWords];
|
||||
|
||||
bx::RngMwc rng;
|
||||
|
||||
for (uint32_t iter = 0; iter < 50000; ++iter)
|
||||
{
|
||||
const uint32_t numWords = 1 + rng.gen() % kMaxWords;
|
||||
const uint32_t numBits = numWords*64;
|
||||
fillRandom(rng, bits, numWords);
|
||||
|
||||
bx::MutableBitArrayView bs(bits, numWords*sizeof(uint64_t) );
|
||||
|
||||
const uint32_t beginBit = rng.gen() % (numBits + 1);
|
||||
const uint32_t endBit = beginBit + rng.gen() % (numBits - beginBit + 1);
|
||||
|
||||
uint32_t count = 0;
|
||||
uint32_t firstSet = bx::kInvalid;
|
||||
uint32_t firstClr = bx::kInvalid;
|
||||
|
||||
for (uint32_t ii = beginBit; ii < endBit; ++ii)
|
||||
{
|
||||
const bool bit = getBit(bits, ii);
|
||||
count += bit;
|
||||
firstSet = bit && bx::kInvalid == firstSet ? ii : firstSet;
|
||||
firstClr = !bit && bx::kInvalid == firstClr ? ii : firstClr;
|
||||
}
|
||||
|
||||
const uint32_t numRange = endBit - beginBit;
|
||||
|
||||
INFO("Iteration " << iter << ", words " << numWords << ", range [" << beginBit << ", " << endBit << ")");
|
||||
|
||||
REQUIRE(count == bs.countBits(beginBit, endBit) );
|
||||
REQUIRE( (0 != count) == bs.testAny(beginBit, endBit) );
|
||||
REQUIRE( (0 == count) == bs.testNone(beginBit, endBit) );
|
||||
REQUIRE( (numRange == count) == bs.testAll(beginBit, endBit) );
|
||||
REQUIRE(firstSet == bs.findSet(beginBit, endBit) );
|
||||
REQUIRE(firstClr == bs.findClear(beginBit, endBit) );
|
||||
|
||||
uint32_t first = bx::kInvalid;
|
||||
uint32_t last = bx::kInvalid;
|
||||
|
||||
for (uint32_t ii = 0; ii < numBits; ++ii)
|
||||
{
|
||||
if (getBit(bits, ii) )
|
||||
{
|
||||
first = bx::kInvalid == first ? ii : first;
|
||||
last = ii;
|
||||
}
|
||||
}
|
||||
|
||||
REQUIRE(first == bs.findSet(bx::kFirst) );
|
||||
REQUIRE(last == bs.findSet(bx::kLast) );
|
||||
}
|
||||
}
|
||||
|
||||
TEST_CASE("BitArray-blit-random", "[container]")
|
||||
{
|
||||
constexpr uint32_t kNumWords = 40;
|
||||
constexpr uint32_t kNumBits = kNumWords*64;
|
||||
|
||||
uint64_t bits[kNumWords];
|
||||
uint64_t expected[kNumWords];
|
||||
|
||||
bx::RngMwc rng;
|
||||
|
||||
for (uint32_t iter = 0; iter < 50000; ++iter)
|
||||
{
|
||||
fillRandom(rng, bits, kNumWords);
|
||||
bx::memCopy(expected, bits, sizeof(bits) );
|
||||
|
||||
const uint32_t numBits = 1 + rng.gen() % (0 == (rng.gen() & 1) ? 160 : kNumBits/2);
|
||||
const uint32_t srcBit = rng.gen() % (kNumBits - numBits + 1);
|
||||
const uint32_t dstBit = 0 == (rng.gen() & 3)
|
||||
? bx::min(kNumBits - numBits, (srcBit & ~7u) + (rng.gen() % 64)*8)
|
||||
: rng.gen() % (kNumBits - numBits + 1)
|
||||
;
|
||||
|
||||
bool tmp[kNumBits];
|
||||
|
||||
for (uint32_t ii = 0; ii < numBits; ++ii)
|
||||
{
|
||||
tmp[ii] = getBit(expected, srcBit + ii);
|
||||
}
|
||||
|
||||
for (uint32_t ii = 0; ii < numBits; ++ii)
|
||||
{
|
||||
const uint32_t idx = dstBit + ii;
|
||||
expected[idx/64] = (expected[idx/64] & ~(UINT64_C(1) << (idx & 63) ) ) | (uint64_t(tmp[ii]) << (idx & 63) );
|
||||
}
|
||||
|
||||
bx::MutableBitArrayView bs(bits, sizeof(bits) );
|
||||
bs.blit(dstBit, srcBit, numBits);
|
||||
|
||||
INFO("Iteration " << iter << ", dst " << dstBit << ", src " << srcBit << ", num bits " << numBits);
|
||||
|
||||
REQUIRE(0 == bx::memCmp(bits, expected, sizeof(bits) ) );
|
||||
}
|
||||
}
|
||||
|
||||
TEST_CASE("BitArray-set-random", "[container]")
|
||||
{
|
||||
constexpr uint32_t kNumWords = 40;
|
||||
constexpr uint32_t kNumBits = kNumWords*64;
|
||||
|
||||
uint64_t bits[kNumWords];
|
||||
uint64_t expected[kNumWords];
|
||||
|
||||
bx::RngMwc rng;
|
||||
|
||||
for (uint32_t iter = 0; iter < 50000; ++iter)
|
||||
{
|
||||
fillRandom(rng, bits, kNumWords);
|
||||
bx::memCopy(expected, bits, sizeof(bits) );
|
||||
|
||||
const uint32_t beginBit = rng.gen() % (kNumBits + 1);
|
||||
const uint32_t endBit = beginBit + rng.gen() % (kNumBits - beginBit + 1);
|
||||
const bool value = 0 != (rng.gen() & 1);
|
||||
|
||||
for (uint32_t idx = beginBit; idx < endBit; ++idx)
|
||||
{
|
||||
expected[idx/64] = (expected[idx/64] & ~(UINT64_C(1) << (idx & 63) ) ) | (uint64_t(value) << (idx & 63) );
|
||||
}
|
||||
|
||||
bx::MutableBitArrayView bs(bits, sizeof(bits) );
|
||||
bs.set(beginBit, endBit, value);
|
||||
|
||||
INFO("Iteration " << iter << ", range [" << beginBit << ", " << endBit << "), value " << value);
|
||||
|
||||
REQUIRE(0 == bx::memCmp(bits, expected, sizeof(bits) ) );
|
||||
}
|
||||
}
|
||||
|
||||
TEST_CASE("MippedBitArray-random", "[container]")
|
||||
{
|
||||
constexpr uint32_t kNumBits = 64*130;
|
||||
constexpr uint32_t kNumWords = kNumBits/64;
|
||||
|
||||
bx::FixedMippedBitArrayT<kNumBits> array;
|
||||
array.setCount(kNumBits);
|
||||
|
||||
bx::RngMwc rng;
|
||||
|
||||
for (uint32_t iter = 0; iter < 20000; ++iter)
|
||||
{
|
||||
const uint32_t op = rng.gen() % 6;
|
||||
const uint32_t beginBit = rng.gen() % kNumBits;
|
||||
const uint32_t numBits = 1 + rng.gen() % (0 == (rng.gen() & 3) ? kNumBits - beginBit : bx::min<uint32_t>(kNumBits - beginBit, 300) );
|
||||
const uint32_t endBit = beginBit + numBits;
|
||||
|
||||
switch (op)
|
||||
{
|
||||
case 0: array.set(beginBit, endBit, true); break;
|
||||
case 1: array.set(beginBit, endBit, false); break;
|
||||
case 2: array.set(beginBit, 0 != (rng.gen() & 1) ); break;
|
||||
case 3: array.set(beginBit, bx::min(endBit, beginBit + 64), (uint64_t(rng.gen() ) << 32) | rng.gen() ); break;
|
||||
case 4: array.blit(rng.gen() % (kNumBits - numBits + 1), beginBit, numBits); break;
|
||||
default: if (0 == rng.gen() % 64) { array.clear(); } break;
|
||||
}
|
||||
|
||||
const uint64_t* words = (const uint64_t*)array.getPtr();
|
||||
const uint64_t* mip = array.getMip();
|
||||
|
||||
INFO("Iteration " << iter << ", op " << op << ", range [" << beginBit << ", " << endBit << ")");
|
||||
|
||||
for (uint32_t ii = 0; ii < kNumWords; ++ii)
|
||||
{
|
||||
const bool full = UINT64_MAX == words[ii];
|
||||
const bool mipFull = 0 != ( (mip[ii/64] >> (ii & 63) ) & 1);
|
||||
REQUIRE(full == mipFull);
|
||||
}
|
||||
|
||||
for (uint32_t ii = kNumWords; ii < bx::alignUp(kNumWords, 64u); ++ii)
|
||||
{
|
||||
REQUIRE(0 != ( (mip[ii/64] >> (ii & 63) ) & 1) );
|
||||
}
|
||||
|
||||
const uint32_t findBegin = rng.gen() % kNumBits;
|
||||
uint32_t expected = bx::kInvalid;
|
||||
|
||||
for (uint32_t ii = findBegin; ii < kNumBits && bx::kInvalid == expected; ++ii)
|
||||
{
|
||||
expected = getBit(words, ii) ? bx::kInvalid : ii;
|
||||
}
|
||||
|
||||
REQUIRE(expected == array.findClear(findBegin, kNumBits) );
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user