[libcxx-commits] [libcxx] [libc++] Optimize std::min_element (PR #100616)
via libcxx-commits
libcxx-commits at lists.llvm.org
Thu Jul 25 15:18:00 PDT 2024
https://github.com/mrdaybird updated https://github.com/llvm/llvm-project/pull/100616
>From 01c71b48fa1b4a9dfe11e14238f7a8fb4f0ab286 Mon Sep 17 00:00:00 2001
From: Vaibhav Pathak <pathakvaibhav at protonmail.com>
Date: Thu, 25 Jul 2024 22:01:54 +0530
Subject: [PATCH 1/3] [libc++] Optimize std::min_element
[libc++] Fix formatting and move min_element.bench.cpp
[libc++] Use __invoke instead of invoke
[libc++] Fix build issues with find
---
libcxx/benchmarks/CMakeLists.txt | 1 +
.../algorithms/min_element.bench.cpp | 80 +++++++++++++++++++
libcxx/include/__algorithm/find.h | 2 +-
libcxx/include/__algorithm/min_element.h | 56 +++++++++++--
4 files changed, 133 insertions(+), 6 deletions(-)
create mode 100644 libcxx/benchmarks/algorithms/min_element.bench.cpp
diff --git a/libcxx/benchmarks/CMakeLists.txt b/libcxx/benchmarks/CMakeLists.txt
index d96ccc1e49f66..952777d5346f7 100644
--- a/libcxx/benchmarks/CMakeLists.txt
+++ b/libcxx/benchmarks/CMakeLists.txt
@@ -121,6 +121,7 @@ set(BENCHMARK_TESTS
algorithms/make_heap_then_sort_heap.bench.cpp
algorithms/min.bench.cpp
algorithms/minmax.bench.cpp
+ algorithms/min_element.bench.cpp
algorithms/min_max_element.bench.cpp
algorithms/mismatch.bench.cpp
algorithms/pop_heap.bench.cpp
diff --git a/libcxx/benchmarks/algorithms/min_element.bench.cpp b/libcxx/benchmarks/algorithms/min_element.bench.cpp
new file mode 100644
index 0000000000000..7edb8ddbb15f2
--- /dev/null
+++ b/libcxx/benchmarks/algorithms/min_element.bench.cpp
@@ -0,0 +1,80 @@
+#include <algorithm>
+#include <limits>
+#include <vector>
+
+#include <benchmark/benchmark.h>
+#include <random>
+
+template <typename T>
+static void BM_stdmin_element_decreasing(benchmark::State& state) {
+ std::vector<T> v(state.range(0));
+ T start = std::numeric_limits<T>::max();
+ T end = std::numeric_limits<T>::min();
+
+ for (size_t i = 0; i < v.size(); i++)
+ v[i] = ((start != end) ? start-- : end);
+
+ for (auto _ : state) {
+ benchmark::DoNotOptimize(v);
+ benchmark::DoNotOptimize(std::min_element(v.begin(), v.end()));
+ }
+}
+
+BENCHMARK(BM_stdmin_element_decreasing<char>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+BENCHMARK(BM_stdmin_element_decreasing<short>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+BENCHMARK(BM_stdmin_element_decreasing<int>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+BENCHMARK(BM_stdmin_element_decreasing<long long>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+BENCHMARK(BM_stdmin_element_decreasing<unsigned char>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+BENCHMARK(BM_stdmin_element_decreasing<unsigned short>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+BENCHMARK(BM_stdmin_element_decreasing<unsigned int>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+BENCHMARK(BM_stdmin_element_decreasing<unsigned long long>)
+ ->DenseRange(1, 8)
+ ->Range(32, 128)
+ ->Range(256, 4096)
+ ->DenseRange(5000, 10000, 1000)
+ ->Range(1 << 14, 1 << 16)
+ ->Arg(70000);
+
+BENCHMARK_MAIN();
diff --git a/libcxx/include/__algorithm/find.h b/libcxx/include/__algorithm/find.h
index 7f58dbb13a577..8ecd53e63af21 100644
--- a/libcxx/include/__algorithm/find.h
+++ b/libcxx/include/__algorithm/find.h
@@ -104,7 +104,7 @@ __find_bool(__bit_iterator<_Cp, _IsConst> __first, typename _Cp::size_type __n)
// do first partial word
if (__first.__ctz_ != 0) {
__storage_type __clz_f = static_cast<__storage_type>(__bits_per_word - __first.__ctz_);
- __storage_type __dn = std::min(__clz_f, __n);
+ __storage_type __dn = (__clz_f < __n) ? __clz_f : __n;
__storage_type __m = (~__storage_type(0) << __first.__ctz_) & (~__storage_type(0) >> (__clz_f - __dn));
__storage_type __b = std::__invert_if<!_ToFind>(*__first.__seg_) & __m;
if (__b)
diff --git a/libcxx/include/__algorithm/min_element.h b/libcxx/include/__algorithm/min_element.h
index 65f3594d630ce..25964a9e3a52f 100644
--- a/libcxx/include/__algorithm/min_element.h
+++ b/libcxx/include/__algorithm/min_element.h
@@ -11,6 +11,8 @@
#include <__algorithm/comp.h>
#include <__algorithm/comp_ref_type.h>
+#include <__algorithm/find.h>
+#include <__algorithm/iterator_operations.h>
#include <__config>
#include <__functional/identity.h>
#include <__functional/invoke.h>
@@ -33,12 +35,56 @@ __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj) {
if (__first == __last)
return __first;
- _Iter __i = __first;
- while (++__i != __last)
- if (std::__invoke(__comp, std::__invoke(__proj, *__i), std::__invoke(__proj, *__first)))
- __first = __i;
+ const size_t __n = static_cast<size_t>(std::distance(__first, __last));
- return __first;
+ if (__n <= 64) {
+ _Iter __i = __first;
+ while (++__i != __last)
+ if (std::__invoke(__comp, std::__invoke(__proj, *__i), std::__invoke(__proj, *__first)))
+ __first = __i;
+ return __first;
+ }
+
+ size_t __block_size = 256;
+
+ size_t __n_blocked = __n - (__n % __block_size);
+ _Iter __block_start = __first, __block_end = __first;
+
+ typedef typename std::iterator_traits<_Iter>::value_type value_type;
+ value_type __min_val = std::__invoke(__proj, *__first);
+
+ _Iter __curr = __first;
+ for (size_t __i = 0; __i < __n_blocked; __i += __block_size) {
+ _Iter __start = __curr;
+ value_type __block_min = __min_val;
+ for (size_t j = 0; j < __block_size; j++) {
+ if (std::__invoke(__comp, std::__invoke(__proj, *__curr), __block_min)) {
+ __block_min = *__curr;
+ }
+ __curr++;
+ }
+ if (std::__invoke(__comp, __block_min, __min_val)) {
+ __min_val = __block_min;
+ __block_start = __start;
+ __block_end = __curr;
+ }
+ }
+
+ value_type __epilogue_min = __min_val;
+ _Iter __epilogue_start = __curr;
+ while (__curr != __last) {
+ if (std::__invoke(__comp, std::__invoke(__proj, *__curr), __epilogue_min)) {
+ __epilogue_min = *__curr;
+ }
+ __curr++;
+ }
+ if (std::__invoke(__comp, __epilogue_min, __min_val)) {
+ __min_val = __epilogue_min;
+ __block_start = __epilogue_start;
+ __block_end = __last;
+ }
+
+ return std::__find(__block_start, __block_end, __min_val, __proj);
}
template <class _Comp, class _Iter, class _Sent>
>From 7ad9834c6764a57df8bc21b72d440a187641e75c Mon Sep 17 00:00:00 2001
From: Vaibhav Pathak <pathakvaibhav at protonmail.com>
Date: Fri, 26 Jul 2024 03:22:53 +0530
Subject: [PATCH 2/3] [libc++] fix build issues
---
libcxx/include/__algorithm/min_element.h | 10 ++++++----
1 file changed, 6 insertions(+), 4 deletions(-)
diff --git a/libcxx/include/__algorithm/min_element.h b/libcxx/include/__algorithm/min_element.h
index 25964a9e3a52f..d0b609c6405bc 100644
--- a/libcxx/include/__algorithm/min_element.h
+++ b/libcxx/include/__algorithm/min_element.h
@@ -19,6 +19,7 @@
#include <__iterator/iterator_traits.h>
#include <__type_traits/is_callable.h>
#include <__utility/move.h>
+#include <cmath>
#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
# pragma GCC system_header
@@ -37,7 +38,7 @@ __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj) {
const size_t __n = static_cast<size_t>(std::distance(__first, __last));
- if (__n <= 64) {
+ if (__n <= 16) {
_Iter __i = __first;
while (++__i != __last)
if (std::__invoke(__comp, std::__invoke(__proj, *__i), std::__invoke(__proj, *__first)))
@@ -45,12 +46,13 @@ __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj) {
return __first;
}
- size_t __block_size = 256;
+ typedef typename std::iterator_traits<_Iter>::value_type value_type;
+ size_t __block_size = sizeof(value_type) * std::ceil(3.0 * (double)__n / 64.0);
+ __block_size = __block_size < 8 ? 8 : __block_size > 8192 ? 8192 : __block_size;
size_t __n_blocked = __n - (__n % __block_size);
_Iter __block_start = __first, __block_end = __first;
- typedef typename std::iterator_traits<_Iter>::value_type value_type;
value_type __min_val = std::__invoke(__proj, *__first);
_Iter __curr = __first;
@@ -84,7 +86,7 @@ __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj) {
__block_end = __last;
}
- return std::__find(__block_start, __block_end, __min_val, __proj);
+ return std::find(__block_start, __block_end, __min_val);
}
template <class _Comp, class _Iter, class _Sent>
>From 74bad7e8fce68bd7f690690e26d54fd7c7f636a5 Mon Sep 17 00:00:00 2001
From: Vaibhav Pathak <pathakvaibhav at protonmail.com>
Date: Fri, 26 Jul 2024 03:47:38 +0530
Subject: [PATCH 3/3] [libcxx] revert block_size logic
---
libcxx/include/__algorithm/min_element.h | 9 +++------
1 file changed, 3 insertions(+), 6 deletions(-)
diff --git a/libcxx/include/__algorithm/min_element.h b/libcxx/include/__algorithm/min_element.h
index d0b609c6405bc..606330cb0925b 100644
--- a/libcxx/include/__algorithm/min_element.h
+++ b/libcxx/include/__algorithm/min_element.h
@@ -19,7 +19,6 @@
#include <__iterator/iterator_traits.h>
#include <__type_traits/is_callable.h>
#include <__utility/move.h>
-#include <cmath>
#if !defined(_LIBCPP_HAS_NO_PRAGMA_SYSTEM_HEADER)
# pragma GCC system_header
@@ -38,7 +37,7 @@ __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj) {
const size_t __n = static_cast<size_t>(std::distance(__first, __last));
- if (__n <= 16) {
+ if (__n <= 64) {
_Iter __i = __first;
while (++__i != __last)
if (std::__invoke(__comp, std::__invoke(__proj, *__i), std::__invoke(__proj, *__first)))
@@ -46,9 +45,7 @@ __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj) {
return __first;
}
- typedef typename std::iterator_traits<_Iter>::value_type value_type;
- size_t __block_size = sizeof(value_type) * std::ceil(3.0 * (double)__n / 64.0);
- __block_size = __block_size < 8 ? 8 : __block_size > 8192 ? 8192 : __block_size;
+ size_t __block_size = 256;
size_t __n_blocked = __n - (__n % __block_size);
_Iter __block_start = __first, __block_end = __first;
@@ -86,7 +83,7 @@ __min_element(_Iter __first, _Sent __last, _Comp __comp, _Proj& __proj) {
__block_end = __last;
}
- return std::find(__block_start, __block_end, __min_val);
+ return std::__find(__block_start, __block_end, __min_val, __proj);
}
template <class _Comp, class _Iter, class _Sent>
More information about the libcxx-commits
mailing list