Files
doris/thirdparty/tests/parallel_hashmap_neon_test.cpp
HappenLee 45d1598da2 [improvement](be) Accelerate phmap control scans with ARM NEON (#69688)
AArch64 currently uses portable scalar phmap control-byte scans. The
subtraction-based H2 mask can include adjacent false-positive
candidates: control bytes `0x12, 0x13` can both be returned when looking
for `0x12`. Key comparison preserves correctness, but the extra
candidates cost work.

Add exact NEON control comparisons on little-endian AArch64 while
keeping eight-byte groups, the portable fallback and the existing ARM
dump/load layout. x86 keeps its existing SSE implementation.
`PHMAP_HAVE_ARM_NEON=0` disables the new implementation at compile time.

The normal third-party downloader applies this patch alongside the
existing AArch64 int128 hash-mixer patch, with separate markers and a
subshell that restores the caller's working directory. The NEON patch
now includes context so a replay cannot silently insert the same code
again. The downloader checks the actual source with a direction-locked
reverse dry-run, recovers a missing marker when all hunks are already
applied, and preflights every hunk before a fresh application. Partial
or duplicated sources fail without further edits and do not retain a
success marker. The standalone tests cover control masks, unaligned
reads, special slots, conversion bounds, map operations, rehash and
dump/load. The standalone C++ benchmark can be compiled directly. For an
incremental comparison, retain the existing int128 patch in both the
baseline and optimized dependency sources, adding NEON only to the
optimized copy.

### Supplied ARM measurements

The following measurements come from the supplied report on original
Doris baseline `70f7dcbe11c8e142b81536c1dfaaf505ec05437b`, not a new ARM
run of this rebased PR. That baseline predates master's AArch64 int128
patch. These numbers therefore do **not** establish the incremental
gains of the final combined patch stack; a new ARM run with the combined
patch stack is still needed.

Environment: Neoverse-N2, 16 cores, nominal 3 GHz; Clang 20.1.8; C++17,
`-O3 -DNDEBUG -march=armv8-a+crc -pthread`; one process pinned to CPU
15. Nine alternating baseline/NEON rounds, each measurement at least 0.2
seconds of thread CPU time. Values are median thread CPU ns/op. Reported
intervals use 10,000 paired bootstrap resamples.

Positive time reduction means faster; negative means slower.

| Scenario | Baseline ns/op | NEON ns/op | Time reduction | Reported 95%
interval |
|---|---:|---:|---:|---:|
| Hit lookup, 4,096 keys | 4.0773 | 3.8350 | +5.94% | [+5.84%, +6.11%] |
| Hit lookup, 100,000 keys | 10.6120 | 9.6412 | +9.15% | [+8.99%,
+12.44%] |
| Hit lookup, 114,000 keys | 11.4606 | 10.5508 | +7.94% | [+7.73%,
+11.42%] |
| Hit lookup, 1,000,000 keys | 16.3554 | 14.4228 | +11.82% | [+10.20%,
+12.74%] |
| Miss lookup, 4,096 keys | 3.4818 | 3.3069 | +5.02% | [+4.81%, +5.08%]
|
| Miss lookup, 100,000 keys | 9.7027 | 9.9144 | -2.18% | [-2.91%,
-1.27%] |
| Miss lookup, 114,000 keys | 16.8860 | 17.5147 | -3.72% | [-3.93%,
-2.89%] |
| Miss lookup, 1,000,000 keys | 8.6927 | 8.1035 | +6.78% | [+4.93%,
+9.01%] |
| Insert, 100,000 keys, including allocation/destruction | 9.8116 |
9.2758 | +5.46% | [+3.94%, +6.25%] |
| Erase/reinsert, 20,000 operations per iteration | 16.3032 | 16.3810 |
-0.48% | [-1.19%, +0.97%] |
| Synthetic adjacent-H2 scan, 32,768 groups | 2.7578 | 1.7641 | +36.03%
| [+35.75%, +36.09%] |

Load factors differ: approximately 50.0%, 76.3%, 87.0% and 47.7% at the
four lookup sizes. Two miss-lookup workloads regress reproducibly. The
erase/reinsert interval crosses zero. The adjacent-H2 case measures
group scans and candidate-key comparisons, not complete table lookups.
These are library microbenchmarks, not end-to-end SQL gains.

### Validation

**Reported in the supplied ARM bundle:** randomized scalar-oracle
checks, baseline/NEON/forced-portable paths, bidirectional dump/load and
ASAN/UBSAN passed. The archive contains the patch and report; these ARM
executions were not repeated during this submission.

**Executed locally after rebasing onto master
`3d2d9aac54148056529c13ef4b9fbdbcb65fab77` (x86_64, Clang 16.0.6):**

- Standalone correctness tests passed for baseline, patched and
explicitly NEON-disabled builds; x86 selected the existing 16-byte SSE
groups in all three. Tests also exercise the portable group
implementation directly.
- Baseline-to-patched and patched-to-baseline dump/load passed; patched
ASAN/UBSAN passed.
- Benchmark source compiled with `-Wall -Wextra -Werror`; all 11
scenarios passed a short checksum smoke run. No x86 performance
conclusion is drawn from it.
- Standard third-party download/patch script passed twice with `OLDPWD`
unset. The standard downloader produces byte-identical headers to those
used by the standalone tests.
- clang-format 16, repository clang-tidy entry point for both standalone
C++ files, shell syntax and `git diff --check` passed.

**Downloader recovery follow-up:** `bash
thirdparty/test/parallel-hashmap-patch-test.sh` passed using the pinned
cached archive and the real downloader; no dependency downloads were
needed. It covers fresh extraction, normal reruns, an applied patch with
a missing marker, a stale marker over pristine sources, partially
applied sources with/without a marker, and duplicate NEON definitions.
Successful reruns compare both complete headers; failing cases verify
unchanged source and absence of the success marker. The same test
detects the original replay bug. The regenerated patch inserts exactly
the same 67 C++ lines, and its applied header bytes match the original
optimized source. Bash syntax and shell whitespace checks passed;
generated patch context preserves the pinned dependency's existing
whitespace.

Full Doris BE compilation, SQL regression/performance tests and new ARM
measurements against the combined int128+NEON patch stack have not been
run.

The C++ test and benchmark have no Python dependency and can be compiled
directly:

```bash
bash thirdparty/download-thirdparty.sh parallel_hashmap
clang++ -std=c++17 -O2 -pthread -Wall -Wextra -Werror \
  -isystem thirdparty/src/parallel-hashmap-1.3.8 \
  thirdparty/tests/parallel_hashmap_neon_test.cpp -o /tmp/phmap-neon-test
/tmp/phmap-neon-test
clang++ -std=c++17 -O3 -DNDEBUG -pthread -Wall -Wextra -Werror \
  -isystem thirdparty/src/parallel-hashmap-1.3.8 \
  thirdparty/benchmarks/parallel_hashmap_bench.cpp -o /tmp/phmap-neon-bench
taskset -c 15 /tmp/phmap-neon-bench 0.2
```

Select an available isolated CPU instead of CPU 15 when necessary. For
the reported ARM compiler settings, add `-march=armv8-a+crc`. The
benchmark emits one JSON line per scenario with CPU/wall timings and
checksums. To compare against baseline on current master, extract the
verified source archive into another directory, apply only
`parallel-hashmap-1.3.8-aarch64-int128.patch`, compile the same C++
benchmark against that directory with identical flags, and alternate
baseline/optimized execution order.

### Release note

Accelerate phmap control-byte scanning on little-endian AArch64 while
preserving eight-byte ARM groups and the portable fallback.
2026-10-30 16:56:52 +08:00

273 lines
9.4 KiB
C++

// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements. See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership. The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License. You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied. See the License for the
// specific language governing permissions and limitations
// under the License.
#include <parallel_hashmap/phmap.h>
#include <parallel_hashmap/phmap_dump.h>
#include <algorithm>
#include <array>
#include <cstdint>
#include <fstream>
#include <iostream>
#include <random>
#include <sstream>
#include <stdexcept>
#include <string>
#include <type_traits>
#include <unordered_map>
#include <vector>
namespace {
using phmap::priv::ctrl_t;
using phmap::priv::Group;
using phmap::priv::GroupPortableImpl;
using phmap::priv::h2_t;
using phmap::priv::kDeleted;
using phmap::priv::kEmpty;
using phmap::priv::kSentinel;
void require(bool condition, const char* message) {
if (!condition) {
throw std::runtime_error(message);
}
}
template <typename Mask>
std::vector<uint32_t> positions(Mask mask) {
std::vector<uint32_t> result;
for (uint32_t position : mask) {
result.push_back(position);
}
require(static_cast<bool>(mask) == !result.empty(), "mask bool");
if (!result.empty()) {
require(mask.TrailingZeros() == result.front(), "trailing zeros");
}
return result;
}
template <typename Impl>
void check_group(const ctrl_t* bytes, h2_t hash) {
constexpr int width = Impl::kWidth;
Impl group(bytes);
std::vector<uint32_t> matches;
std::vector<uint32_t> empty;
std::vector<uint32_t> available;
for (int i = 0; i < width; ++i) {
if (bytes[i] == static_cast<ctrl_t>(hash)) {
matches.push_back(i);
}
if (bytes[i] == kEmpty) {
empty.push_back(i);
}
if (bytes[i] == kEmpty || bytes[i] == kDeleted) {
available.push_back(i);
}
}
auto actual = positions(group.Match(hash));
if constexpr (std::is_same_v<Impl, GroupPortableImpl>) {
// The scalar subtraction can borrow into an adjacent byte: 0x12,0x13
// can both match 0x12. Only full slots may be false positives.
require(matches.empty() == actual.empty(), "portable match emptiness");
for (auto expected : matches) {
require(std::find(actual.begin(), actual.end(), expected) != actual.end(),
"portable missed an exact match");
}
for (auto position : actual) {
require(position < width && bytes[position] >= 0, "portable matched special slot");
}
} else {
require(actual == matches, "SIMD match differs from scalar oracle");
}
require(positions(group.MatchEmpty()) == empty, "empty mask");
require(positions(group.MatchEmptyOrDeleted()) == available, "empty/deleted mask");
if (!actual.empty()) {
require(group.Match(hash).LeadingZeros() == width - 1 - actual.back(), "leading zeros");
}
uint32_t leading = 0;
while (leading < width && (bytes[leading] == kEmpty || bytes[leading] == kDeleted)) {
++leading;
}
require(group.CountLeadingEmptyOrDeleted() == leading, "leading empty/deleted count");
std::array<ctrl_t, width + 16> converted;
converted.fill(42);
group.ConvertSpecialToEmptyAndFullToDeleted(converted.data() + 7);
for (int i = 0; i < width; ++i) {
require(converted[i + 7] == (bytes[i] < 0 ? kEmpty : kDeleted), "special conversion");
}
for (int i = 0; i < 7; ++i) {
require(converted[i] == 42, "conversion underrun");
}
for (size_t i = width + 7; i < converted.size(); ++i) {
require(converted[i] == 42, "conversion overrun");
}
}
void check_groups() {
std::mt19937 random(81384);
std::array<ctrl_t, 32> bytes;
constexpr std::array<ctrl_t, 3> special_values {kEmpty, kDeleted, kSentinel};
for (int iteration = 0; iteration < 100000; ++iteration) {
bytes.fill(kEmpty);
auto* data = bytes.data() + (iteration & 7);
for (int i = 0; i < Group::kWidth; ++i) {
uint32_t value = random() % 131;
data[i] = value < 128 ? static_cast<ctrl_t>(value) : special_values[value - 128];
}
h2_t hash = random() % 128;
check_group<Group>(data, hash);
check_group<GroupPortableImpl>(data, hash);
}
for (int offset = 0; offset < 8; ++offset) {
for (auto special : {kEmpty, kDeleted, kSentinel}) {
bytes.fill(special);
check_group<Group>(bytes.data() + offset, 0);
check_group<GroupPortableImpl>(bytes.data() + offset, 0);
}
for (int lane = 0; lane < Group::kWidth; ++lane) {
for (int value = 0; value < 128; ++value) {
bytes.fill(kEmpty);
bytes[offset + lane] = static_cast<ctrl_t>(value);
check_group<Group>(bytes.data() + offset, value);
}
}
}
bytes.fill(kEmpty);
bytes[2] = 0x12;
bytes[3] = 0x13;
check_group<Group>(bytes.data(), 0x12);
require(positions(GroupPortableImpl(bytes.data()).Match(0x12)) == std::vector<uint32_t>({2, 3}),
"adjacent-H2 scalar baseline");
}
template <typename Map>
void check_map() {
Map map;
std::unordered_map<uint64_t, uint64_t> reference;
std::mt19937_64 random(81384);
map.reserve(65536);
for (int i = 0; i < 200000; ++i) {
uint64_t key = random() % 60000;
uint64_t value = random();
switch (i % 4) {
case 0:
case 1:
map[key] = value;
reference[key] = value;
break;
case 2:
require(map.erase(key) == reference.erase(key), "erase result");
break;
default: {
auto found = map.find(key);
auto expected = reference.find(key);
require((found == map.end()) == (expected == reference.end()), "find result");
if (found != map.end()) {
require(found->second == expected->second, "find value");
}
}
}
if (i % 16384 == 0) {
map.rehash(map.size() * 2);
}
}
require(map.size() == reference.size(), "map size");
for (const auto& [key, value] : reference) {
auto found = map.find(key);
require(found != map.end() && found->second == value, "map final content");
}
}
struct Archive {
std::iostream& stream;
void saveBinary(const void* data, size_t bytes) {
stream.write(static_cast<const char*>(data), bytes);
require(static_cast<bool>(stream), "dump write");
}
void loadBinary(void* data, size_t bytes) {
stream.read(static_cast<char*>(data), bytes);
require(static_cast<bool>(stream), "dump read");
}
};
using DumpMap = phmap::flat_hash_map<uint64_t, uint64_t>;
DumpMap dump_fixture() {
DumpMap map;
for (uint64_t i = 0; i < 10000; ++i) {
map[i * 7 + 3] = i * 11 + 5;
}
for (uint64_t i = 0; i < 10000; i += 3) {
map.erase(i * 7 + 3);
}
return map;
}
void check_loaded(const DumpMap& map) {
auto expected = dump_fixture();
require(map.size() == expected.size(), "dump/load size");
for (const auto& [key, value] : expected) {
auto found = map.find(key);
require(found != map.end() && found->second == value, "dump/load content");
}
}
} // namespace
int main(int argc, char** argv) {
try {
if (argc == 3) {
const bool dumping = std::string(argv[1]) == "--dump";
auto mode = std::ios::in | std::ios::out | std::ios::binary;
if (dumping) {
mode |= std::ios::trunc;
}
std::fstream stream(argv[2], mode);
require(static_cast<bool>(stream), "open archive");
Archive archive {stream};
if (dumping) {
auto map = dump_fixture();
require(map.phmap_dump(archive), "dump");
} else {
require(std::string(argv[1]) == "--load", "unknown archive operation");
DumpMap map;
require(map.phmap_load(archive), "load");
check_loaded(map);
}
} else {
require(argc == 1, "unexpected arguments");
check_groups();
check_map<phmap::flat_hash_map<uint64_t, uint64_t>>();
check_map<phmap::parallel_flat_hash_map<uint64_t, uint64_t>>();
std::stringstream stream;
Archive archive {stream};
auto source = dump_fixture();
require(source.phmap_dump(archive), "roundtrip dump");
DumpMap loaded;
require(loaded.phmap_load(archive), "roundtrip load");
check_loaded(loaded);
}
std::cout << "PASS group_width=" << Group::kWidth
<< " portable=" << std::is_same_v<Group, GroupPortableImpl> << '\n';
return 0;
} catch (const std::exception& error) {
std::cerr << "FAIL: " << error.what() << '\n';
return 1;
}
}