From: Matthias Kretz <m.kretz@gsi.de>
Date: Mon, 20 Feb 2023 15:33:31 +0000 (+0100)
Subject: libstdc++: More efficient masked inc-/decrement implementation
X-Git-Tag: basepoints/gcc-14~874
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=6ce55180d494b616e2e3e68ffedfe9007e42ca06;p=thirdparty%2Fgcc.git

libstdc++: More efficient masked inc-/decrement implementation

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>

libstdc++-v3/ChangeLog:

	PR libstdc++/108856
	* include/experimental/bits/simd_builtin.h
	(_SimdImplBuiltin::_S_masked_unary): More efficient
	implementation of masked inc-/decrement for integers and floats
	without AVX2.
	* include/experimental/bits/simd_x86.h
	(_SimdImplX86::_S_masked_unary): New. Use AVX512 masked subtract
	builtins for masked inc-/decrement.
---

diff --git a/libstdc++-v3/include/experimental/bits/simd_builtin.h b/libstdc++-v3/include/experimental/bits/simd_builtin.h
index 9736db65e016..0cf5c9897cd1 100644
--- a/libstdc++-v3/include/experimental/bits/simd_builtin.h
+++ b/libstdc++-v3/include/experimental/bits/simd_builtin.h
@@ -2546,8 +2546,31 @@ template <typename _Abi, typename>
 	_Op<decltype(__vv)> __op;
 	if (__k._M_is_constprop_all_of())
 	  return __data(__op(__vv));
-	else
-	  return _CommonImpl::_S_blend(__k, __v, __data(__op(__vv)));
+	else if constexpr (is_same_v<_Op<void>, __increment<void>>)
+	  {
+	    static_assert(not std::is_same_v<_K, bool>);
+	    if constexpr (is_integral_v<_Tp>)
+	      // Take a shortcut knowing that __k is an integer vector with values -1 or 0.
+	      return __v._M_data - __vector_bitcast<_Tp>(__k._M_data);
+	    else if constexpr (not __have_avx2)
+	      return __v._M_data
+		       + __vector_bitcast<_Tp>(__k._M_data & __builtin_bit_cast(
+							       _K, _Tp(1)));
+	    // starting with AVX2 it is more efficient to blend after add
+	  }
+	else if constexpr (is_same_v<_Op<void>, __decrement<void>>)
+	  {
+	    static_assert(not std::is_same_v<_K, bool>);
+	    if constexpr (is_integral_v<_Tp>)
+	      // Take a shortcut knowing that __k is an integer vector with values -1 or 0.
+	      return __v._M_data + __vector_bitcast<_Tp>(__k._M_data);
+	    else if constexpr (not __have_avx2)
+	      return __v._M_data
+		       - __vector_bitcast<_Tp>(__k._M_data & __builtin_bit_cast(
+							       _K, _Tp(1)));
+	    // starting with AVX2 it is more efficient to blend after sub
+	  }
+	return _CommonImpl::_S_blend(__k, __v, __data(__op(__vv)));
       }
 
     //}}}2
diff --git a/libstdc++-v3/include/experimental/bits/simd_x86.h b/libstdc++-v3/include/experimental/bits/simd_x86.h
index cd642cb3290f..8872ca301b9e 100644
--- a/libstdc++-v3/include/experimental/bits/simd_x86.h
+++ b/libstdc++-v3/include/experimental/bits/simd_x86.h
@@ -3462,6 +3462,74 @@ template <typename _Abi, typename>
       }
 
     //}}} }}}
+    template <template <typename> class _Op, typename _Tp, typename _K,
+	      size_t _Np>
+      _GLIBCXX_SIMD_INTRINSIC static _SimdWrapper<_Tp, _Np>
+      _S_masked_unary(const _SimdWrapper<_K, _Np> __k,
+		      const _SimdWrapper<_Tp, _Np> __v)
+      {
+	if (__k._M_is_constprop_none_of())
+	  return __v;
+	else if (__k._M_is_constprop_all_of())
+	  {
+	    auto __vv = _Base::_M_make_simd(__v);
+	    _Op<decltype(__vv)> __op;
+	    return __data(__op(__vv));
+	  }
+	else if constexpr (__is_bitmask_v<decltype(__k)>
+			     && (is_same_v<_Op<void>, __increment<void>>
+				   || is_same_v<_Op<void>, __decrement<void>>))
+	  {
+	    // optimize masked unary increment and decrement as masked sub +/-1
+	    constexpr int __pm_one
+	      = is_same_v<_Op<void>, __increment<void>> ? -1 : 1;
+	    if constexpr (is_integral_v<_Tp>)
+	      {
+		constexpr bool __lp64 = sizeof(long) == sizeof(long long);
+		using _Ip = std::make_signed_t<_Tp>;
+		using _Up = std::conditional_t<
+			      std::is_same_v<_Ip, long>,
+			      std::conditional_t<__lp64, long long, int>,
+			      std::conditional_t<
+				std::is_same_v<_Ip, signed char>, char, _Ip>>;
+		const auto __value = __vector_bitcast<_Up>(__v._M_data);
+#define _GLIBCXX_SIMD_MASK_SUB(_Sizeof, _Width, _Instr)                        \
+  if constexpr (sizeof(_Tp) == _Sizeof && sizeof(__v) == _Width)               \
+    return __vector_bitcast<_Tp>(__builtin_ia32_##_Instr##_mask(__value,       \
+	     __vector_broadcast<_Np>(_Up(__pm_one)), __value, __k._M_data))
+		_GLIBCXX_SIMD_MASK_SUB(1, 64, psubb512);
+		_GLIBCXX_SIMD_MASK_SUB(1, 32, psubb256);
+		_GLIBCXX_SIMD_MASK_SUB(1, 16, psubb128);
+		_GLIBCXX_SIMD_MASK_SUB(2, 64, psubw512);
+		_GLIBCXX_SIMD_MASK_SUB(2, 32, psubw256);
+		_GLIBCXX_SIMD_MASK_SUB(2, 16, psubw128);
+		_GLIBCXX_SIMD_MASK_SUB(4, 64, psubd512);
+		_GLIBCXX_SIMD_MASK_SUB(4, 32, psubd256);
+		_GLIBCXX_SIMD_MASK_SUB(4, 16, psubd128);
+		_GLIBCXX_SIMD_MASK_SUB(8, 64, psubq512);
+		_GLIBCXX_SIMD_MASK_SUB(8, 32, psubq256);
+		_GLIBCXX_SIMD_MASK_SUB(8, 16, psubq128);
+#undef _GLIBCXX_SIMD_MASK_SUB
+	      }
+	    else
+	      {
+#define _GLIBCXX_SIMD_MASK_SUB(_Sizeof, _Width, _Instr)                        \
+  if constexpr (sizeof(_Tp) == _Sizeof && sizeof(__v) == _Width)               \
+    return __builtin_ia32_##_Instr##_mask(                                     \
+	     __v._M_data, __vector_broadcast<_Np>(_Tp(__pm_one)), __v._M_data, \
+	     __k._M_data, _MM_FROUND_CUR_DIRECTION)
+		_GLIBCXX_SIMD_MASK_SUB(4, 64, subps512);
+		_GLIBCXX_SIMD_MASK_SUB(4, 32, subps256);
+		_GLIBCXX_SIMD_MASK_SUB(4, 16, subps128);
+		_GLIBCXX_SIMD_MASK_SUB(8, 64, subpd512);
+		_GLIBCXX_SIMD_MASK_SUB(8, 32, subpd256);
+		_GLIBCXX_SIMD_MASK_SUB(8, 16, subpd128);
+#undef _GLIBCXX_SIMD_MASK_SUB
+	      }
+	  }
+	else
+	  return _Base::template _S_masked_unary<_Op>(__k, __v);
+      }
   };
 
 // }}}