Update minimum CPU arch requirement on Windows to AVX2 with SSE4.2

94.48% support on PC according to Steam Hardware & Software Survey: May 2025 (https://store.steampowered.com/hwsurvey/)
This commit is contained in:
Wojtek Figat
2025-06-12 17:03:19 +02:00
parent 4240646ec7
commit 7606c9ac12
6 changed files with 188 additions and 29 deletions

View File

@@ -207,28 +207,22 @@ API_ENUM() enum class ArchitectureType
#define PLATFORM_UNIX_FAMILY (PLATFORM_LINUX || PLATFORM_ANDROID || PLATFORM_PS4 || PLATFORM_PS5 || PLATFORM_APPLE_FAMILY)
// SIMD defines
#if defined(__i386__) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64) || defined(__SSE2__)
#if !defined(PLATFORM_SIMD_SSE2) && (defined(__i386__) || defined(_M_IX86) || defined(__x86_64__) || defined(_M_X64) || defined(__SSE2__))
#define PLATFORM_SIMD_SSE2 1
#if defined(__SSE3__)
#if !defined(PLATFORM_SIMD_SSE3) && (defined(__SSE3__))
#define PLATFORM_SIMD_SSE3 1
#endif
#if defined(__SSE4__)
#define PLATFORM_SIMD_SSE4 1
#endif
#if defined(__SSE4_1__)
#if !defined(PLATFORM_SIMD_SSE4_1) && (defined(__SSE4_1__))
#define PLATFORM_SIMD_SSE4_1 1
#endif
#if defined(__SSE4_2__)
#if !defined(PLATFORM_SIMD_SSE4_2) && (defined(__SSE4_2__))
#define PLATFORM_SIMD_SSE4_2 1
#endif
#endif
#if defined(_M_ARM) || defined(__ARM_NEON__) || defined(__ARM_NEON)
#if !defined(PLATFORM_SIMD_NEON) && (defined(_M_ARM) || defined(__ARM_NEON__) || defined(__ARM_NEON))
#define PLATFORM_SIMD_NEON 1
#endif
#if defined(_M_PPC) || defined(__CELLOS_LV2__)
#define PLATFORM_SIMD_VMX 1
#endif
#define PLATFORM_SIMD (PLATFORM_SIMD_SSE2 || PLATFORM_SIMD_SSE3 || PLATFORM_SIMD_SSE4 || PLATFORM_SIMD_NEON || PLATFORM_SIMD_VMX)
#define PLATFORM_SIMD (PLATFORM_SIMD_SSE2 || PLATFORM_SIMD_SSE3 || PLATFORM_SIMD_SSE4_1 || PLATFORM_SIMD_SSE4_2 || PLATFORM_SIMD_NEON)
// Unicode text macro
#if !defined(TEXT)

View File

@@ -257,6 +257,37 @@ void GetWindowsVersion(String& windowsName, int32& versionMajor, int32& versionM
RegCloseKey(hKey);
}
#if PLATFORM_ARCH_X86 || PLATFORM_ARCH_X64
struct CPUBrand
{
char Buffer[0x40];
CPUBrand()
{
Buffer[0] = 0;
int32 cpuInfo[4];
__cpuid(cpuInfo, 0x80000000);
if (cpuInfo[0] >= 0x80000004)
{
// Get name
for (uint32 i = 0; i < 3; i++)
{
__cpuid(cpuInfo, 0x80000002 + i);
memcpy(Buffer + i * sizeof(cpuInfo), cpuInfo, sizeof(cpuInfo));
}
// Trim ending whitespaces
int32 size = StringUtils::Length(Buffer);
while (size > 1 && Buffer[size - 1] == ' ')
size--;
Buffer[size] = 0;
}
}
};
#endif
LRESULT CALLBACK WndProc(HWND hwnd, UINT msg, WPARAM wParam, LPARAM lParam)
{
// Find window to process that message
@@ -517,6 +548,60 @@ void WindowsPlatform::PreInit(void* hInstance)
ASSERT(hInstance);
Instance = hInstance;
#if PLATFORM_ARCH_X86 || PLATFORM_ARCH_X64
// Check the minimum vector instruction set support
int32 cpuInfo[4] = { -1 };
__cpuid(cpuInfo, 0);
int32 cpuInfoSize = cpuInfo[0];
__cpuid(cpuInfo, 1);
bool SSE2 = cpuInfo[3] & (1u << 26);
bool SSE3 = cpuInfo[2] & (1u << 0);
bool SSE41 = cpuInfo[2] & (1u << 19);
bool SSE42 = cpuInfo[2] & (1u << 20);
bool AVX = cpuInfo[2] & (1u << 28);
bool POPCNT = cpuInfo[2] & (1u << 23);
bool AVX2 = false;
if (cpuInfoSize >= 7)
{
__cpuid(cpuInfo, 7);
AVX2 = cpuInfo[1] & (1u << 5) && (_xgetbv(0) & 6) == 6;
}
const Char* missingFeature = nullptr;
#if defined(__AVX__)
if (!AVX)
missingFeature = TEXT("AVX");
#endif
#if defined(__AVX2__)
if (!AVX2)
missingFeature = TEXT("AVX2");
#endif
#if PLATFORM_SIMD_SSE2
if (!SSE2)
missingFeature = TEXT("SSE2");
#endif
#if PLATFORM_SIMD_SSE3
if (!SSE3)
missingFeature = TEXT("SSE3");
#endif
#if PLATFORM_SIMD_SSE4_1
if (!SSE41)
missingFeature = TEXT("SSE4.1");
#endif
#if PLATFORM_SIMD_SSE4_2
if (!SSE42)
missingFeature = TEXT("SSE4.2");
if (!POPCNT)
missingFeature = TEXT("POPCNT");
#endif
if (missingFeature)
{
// Not supported CPU
CPUBrand cpu;
Error(String::Format(TEXT("Cannot start program due to lack of CPU feature {}.\n\n{}"), missingFeature, String(cpu.Buffer)));
exit(-1);
}
#endif
// Disable the process from being showing "ghosted" while not responding messages during slow tasks
DisableProcessWindowsGhosting();
@@ -707,20 +792,8 @@ void WindowsPlatform::LogInfo()
#if PLATFORM_ARCH_X86 || PLATFORM_ARCH_X64
// Log CPU brand
{
char brandBuffer[0x40] = {};
int32 cpuInfo[4] = { -1 };
__cpuid(cpuInfo, 0x80000000);
if (cpuInfo[0] >= 0x80000004)
{
for (uint32 i = 0; i < 3; i++)
{
__cpuid(cpuInfo, 0x80000002 + i);
memcpy(brandBuffer + i * sizeof(cpuInfo), cpuInfo, sizeof(cpuInfo));
}
}
LOG(Info, "CPU: {0}", String(brandBuffer));
}
CPUBrand cpu;
LOG(Info, "CPU: {0}", String(cpu.Buffer));
#endif
LOG(Info, "Microsoft {0} {1}-bit ({2}.{3}.{4})", WindowsName, Platform::Is64BitPlatform() ? TEXT("64") : TEXT("32"), VersionMajor, VersionMinor, VersionBuild);

View File

@@ -84,6 +84,47 @@ namespace Flax.Build.NativeCpp
Latest,
}
/// <summary>
/// The SIMD architecture to use for code generation.
/// </summary>
public enum CpuArchitecture
{
/// <summary>
/// No specific architecture set.
/// </summary>
None,
/// <summary>
/// Intel Advanced Vector Extensions.
/// </summary>
AVX,
/// <summary>
/// Enables Intel Advanced Vector Extensions 2.
/// </summary>
AVX2,
/// <summary>
/// Intel Advanced Vector Extensions 512.
/// </summary>
AVX512,
/// <summary>
/// Intel Streaming SIMD Extensions 2.
/// </summary>
SSE2,
/// <summary>
/// Intel Streaming SIMD Extensions 4.2.
/// </summary>
SSE4_2,
/// <summary>
/// ARM Neon.
/// </summary>
NEON,
}
/// <summary>
/// The C++ compilation environment required to build source files in the native modules.
/// </summary>
@@ -104,6 +145,11 @@ namespace Flax.Build.NativeCpp
/// </summary>
public Sanitizer Sanitizers = Sanitizer.None;
/// <summary>
/// SIMD architecture to use.
/// </summary>
public CpuArchitecture CpuArchitecture = CpuArchitecture.None;
/// <summary>
/// Enables exceptions support.
/// </summary>
@@ -222,6 +268,7 @@ namespace Flax.Build.NativeCpp
CppVersion = CppVersion,
FavorSizeOrSpeed = FavorSizeOrSpeed,
Sanitizers = Sanitizers,
CpuArchitecture = CpuArchitecture,
EnableExceptions = EnableExceptions,
RuntimeTypeInfo = RuntimeTypeInfo,
Inlining = Inlining,

View File

@@ -248,6 +248,25 @@ namespace Flax.Build
Modules.Add("Main");
}
switch (options.CompileEnv.CpuArchitecture)
{
case CpuArchitecture.AVX:
case CpuArchitecture.SSE2:
// Basic SEE2
options.CompileEnv.PreprocessorDefinitions.Add("PLATFORM_SIMD_SSE2=1"); break;
case CpuArchitecture.AVX2:
case CpuArchitecture.SSE4_2:
// Assume full support of SEE4.2 and older
options.CompileEnv.PreprocessorDefinitions.Add("PLATFORM_SIMD_SSE2=1");
options.CompileEnv.PreprocessorDefinitions.Add("PLATFORM_SIMD_SSE3=1");
options.CompileEnv.PreprocessorDefinitions.Add("PLATFORM_SIMD_SSE4_1=1");
options.CompileEnv.PreprocessorDefinitions.Add("PLATFORM_SIMD_SSE4_2=1");
break;
case CpuArchitecture.NEON:
options.CompileEnv.PreprocessorDefinitions.Add("PLATFORM_SIMD_NEON=1");
break;
}
options.CompileEnv.EnableExceptions = true; // TODO: try to disable this!
options.CompileEnv.Sanitizers = Configuration.Sanitizers;
switch (options.Configuration)

View File

@@ -102,6 +102,11 @@ namespace Flax.Build
{
options.CompileEnv.IncludePaths.AddRange(SystemIncludePaths);
options.LinkEnv.LibraryPaths.AddRange(SystemLibraryPaths);
if (options.Architecture == TargetArchitecture.x64 || options.Architecture == TargetArchitecture.x86)
options.CompileEnv.CpuArchitecture = CpuArchitecture.AVX;
else if (options.Architecture == TargetArchitecture.ARM64 || options.Architecture == TargetArchitecture.ARM)
options.CompileEnv.CpuArchitecture = CpuArchitecture.NEON;
}
/// <summary>

View File

@@ -15,6 +15,12 @@ namespace Flax.Build
/// </summary>
[CommandLine("winMinVer", "<version>", "Specifies the minimum Windows version to use (eg. 10).")]
public static string WindowsMinVer = "10";
/// <summary>
/// Specifies the minimum CPU architecture type to support (on x86/x64).
/// </summary>
[CommandLine("winCpuArch", "<arch>", "Specifies the minimum CPU architecture type to support (om x86/x64).")]
public static CpuArchitecture WindowsCpuArch = CpuArchitecture.AVX2; // 94.48% support on PC according to Steam Hardware & Software Survey: May 2025 (https://store.steampowered.com/hwsurvey/)
}
}
@@ -80,6 +86,18 @@ namespace Flax.Build.Platforms
options.CompileEnv.PreprocessorDefinitions.Add("USE_SOFT_INTRINSICS");
options.LinkEnv.InputLibraries.Add("softintrin.lib");
}
options.CompileEnv.CpuArchitecture = Configuration.WindowsCpuArch;
if (_minVersion.Major <= 7 && options.CompileEnv.CpuArchitecture == CpuArchitecture.AVX2)
{
// Old Windows had lower support ratio for latest CPU features
options.CompileEnv.CpuArchitecture = CpuArchitecture.AVX;
}
if (_minVersion.Major >= 11 && options.CompileEnv.CpuArchitecture == CpuArchitecture.AVX)
{
// Windows 11 has hard requirement on SSE4.2
options.CompileEnv.CpuArchitecture = CpuArchitecture.SSE4_2;
}
}
/// <inheritdoc />
@@ -87,10 +105,13 @@ namespace Flax.Build.Platforms
{
base.SetupCompileCppFilesArgs(graph, options, args);
if (Toolset >= WindowsPlatformToolset.v142 && _minVersion.Major >= 11)
switch (options.CompileEnv.CpuArchitecture)
{
// Windows 11 requires SSE4.2
args.Add("/d2archSSE42");
case CpuArchitecture.AVX: args.Add("/arch:AVX"); break;
case CpuArchitecture.AVX2: args.Add("/arch:AVX2"); break;
case CpuArchitecture.AVX512: args.Add("/arch:AVX512"); break;
case CpuArchitecture.SSE2: args.Add("/arch:SSE2"); break;
case CpuArchitecture.SSE4_2: args.Add("/arch:SSE4.2"); break;
}
}