[OMPT-svn-commit] OpenMPT r25333 - trunk/OpenMPT/src/openmpt/soundbase

[<sv...@op...>]

Author: manx
Date: Wed May 20 17:45:31 2026
New Revision: 25333
URL: https://source.openmpt.org/browse/openmpt/?op=revision&rev=25333

Log:
[Imp] openmpt/soundbase/SampleConvert.hpp: Reduce scope of fastround() implementation. Originally, this had been motivated by inefficient code generation with MSVC. However, this decision was made when we were using /fp:fast or equivalent settings in the past, where in particular it can avoid a call to a library function for SSE2 targets. With /fp:precise, this turns out to be a pessimization when targeting SSE2, which is the most common target for 32bit. Only use round(x)=(floor(x)+0.5) when MSVC actually generates simpler code, otherwise prefer the canonical std::round(x). For modern amd64 (SSE4.2 or AVX2), this probably needs further benchmarking in context in order to determine whether the inlined additional code folds away in context. For now, the previous behaviour is kept for these targets.

Modified:
   trunk/OpenMPT/src/openmpt/soundbase/SampleConvert.hpp

Modified: trunk/OpenMPT/src/openmpt/soundbase/SampleConvert.hpp
==============================================================================
--- trunk/OpenMPT/src/openmpt/soundbase/SampleConvert.hpp	Wed May 20 17:28:03 2026	(r25332)
+++ trunk/OpenMPT/src/openmpt/soundbase/SampleConvert.hpp	Wed May 20 17:45:31 2026	(r25333)
@@ -7,6 +7,7 @@
 #include "openmpt/all/BuildSettings.hpp"
 
 #include "mpt/base/arithmetic_shift.hpp"
+#include "mpt/base/detect.hpp"
 #include "mpt/base/macros.hpp"
 #include "mpt/base/math.hpp"
 #include "mpt/base/saturate_cast.hpp"
@@ -31,7 +32,66 @@
 
 
 
+// fastround:
+// Improves code generation on certain MSVC targets.
+//
+// arch  SSE    /fp:    std::round   fastround    better
+//
+// i386  x87    fast    call         fixup+call   -
+// i386  SSE2   fast    call         inline+fixup +
+// amd64 SSE2   fast    call         inline       +
+// amd64 SSE4.2 fast    inline+fixup inline       +
+// amd64 AVX2   fast    inline+fixup inline       +
+// 
+// i386  x87    precise call         fixup+call   -
+// i386  SSE2   precise call         fixup+call   -
+// amd64 SSE2   precise call         call+fixup   -
+// amd64 SSE4.2 precise inline+fixup inline       +
+// amd64 AVX2   precise inline+fixup inline       +
+//
+// See <https://godbolt.org/z/dYYGenjKT> for a microbenchmark.
+//
+// TODO: Real-world benchmark of amd64 SSE4.2 and amd64 AVX2.
+// The inlined additional code might fold away, or nearly away,
+// in which case the canonical semantically correct std::round should be preferred.
+
 #if MPT_COMPILER_MSVC
+#if (defined(_M_FP_FAST) && (_M_FP_FAST == 1))
+// /fp:fast
+#if 0
+#elif MPT_ARCH_X86 && (defined(MPT_ARCH_X86_FPU) && !defined(MPT_ARCH_X86_SSE2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 0
+#elif MPT_ARCH_X86 && (defined(MPT_ARCH_X86_SSE2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 1
+#elif MPT_ARCH_AMD64 && (defined(MPT_ARCH_X86_SSE2) && !defined(MPT_ARCH_X86_SSE4_2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 1
+#elif MPT_ARCH_AMD64 && (defined(MPT_ARCH_X86_SSE4_2) && !defined(MPT_ARCH_X86_AVX2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 1
+#elif MPT_ARCH_AMD64 && (defined(MPT_ARCH_X86_AVX2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 1
+#endif
+#else
+// /fp:precise / ...
+#if 0
+#elif MPT_ARCH_X86 && (defined(MPT_ARCH_X86_FPU) && !defined(MPT_ARCH_X86_SSE2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 0
+#elif MPT_ARCH_X86 && (defined(MPT_ARCH_X86_SSE2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 0
+#elif MPT_ARCH_AMD64 && (defined(MPT_ARCH_X86_SSE2) && !defined(MPT_ARCH_X86_SSE4_2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 0
+#elif MPT_ARCH_AMD64 && (defined(MPT_ARCH_X86_SSE4_2) && !defined(MPT_ARCH_X86_AVX2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 1
+#elif MPT_ARCH_AMD64 && (defined(MPT_ARCH_X86_AVX2))
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 1
+#endif
+#endif
+#endif // MPT_COMPILER_MSVC
+
+#ifndef MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND
+#define MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND 0
+#endif
+
+#if MPT_SOUNDBASE_SAMPLECONVERT_HPP_FASTROUND
 template <typename Tfloat>
 MPT_ATTR_ALWAYSINLINE MPT_INLINE_FORCE Tfloat fastround(Tfloat x)
 {