Disables AVX and fp16 intrinsics by default

Should fix #24

README.md

@@ -31,7 +31,7 @@ will fix the links to the lib and asset folders, so you can run e.g. `love demos
 
 ## Building TÖVE Yourself
 
-On Linux and macOS:
+### On Linux and macOS
 
 ```
 git clone --recurse-submodules https://github.com/poke1024/tove2d
@@ -40,16 +40,31 @@ scons
 love demos/hearts
 ```
 
-On Windows:
+### On Windows
 
 ```
+# in git bash:
 git clone --recurse-submodules https://github.com/poke1024/tove2d
 cd tove2d
+
+# in regular cmd shell:
+cd path/to/tove2d
 setup.bat /C
 ```
 
 Take a look inside `setup.bat` to learn more about installing a compiler environment.
 
+### Experimental
+
+Use `scons --arch=sandybridge` or `scons --arch=haswell` to compile for a custom CPU architecture. On Windows, this enables AVX extensions,
+which mainly speeds up the rasterizer dithering code.
+
+On POSIX, in addition, you can enable hardware intrinsics for 16-bit floating point operations by using `scons --f16c`. These intrinsics
+might give small performance benefits in the `gpux` mode.
+
+Since enabling custom options causes crashes on some CPUs (https://github.com/poke1024/tove2d/issues/24), I strongly recommend sticking to
+the defaults, as they ensure that your binary will run on many CPUs.
+
 ## Roadmap
 
 I keep my ideas of what might eventually end up in TÖVE in [TÖVE's Public Trello Board](https://trello.com/b/p5nWCZVC/t%C3%B6ve).

SConstruct

@@ -13,6 +13,21 @@ AddOption(
 	help='build static library',
 	default=False)
 
+AddOption(
+	'--arch',
+	dest='arch',
+	type='string',
+	nargs=1,
+	action='store',
+	help='CPU architecture to use (e.g. sandybridge, haswell)',
+	default=None)
+
+AddOption(
+	'--f16c',
+	action='store_true',
+	help='assume hardware support for 16-bit floating point numbers',
+	default=False)
+
 sources = [
 	"src/cpp/version.cpp",
 	"src/cpp/interface/api.cpp",
@@ -38,32 +53,55 @@ sources = [
 
 if env["PLATFORM"] == 'win32':
 	env["CCFLAGS"] = ' /EHsc /std:c++17 '
+
 	if not GetOption('tovedebug'):
-		env["CCFLAGS"] += ' /Oi /Ot /Oy /Ob2 /GF /Gy /fp:fast /arch:AVX '
-
-		# enabling /O2 (or /Og above) will crash demos/retro for some reason.
-		# env["CCFLAGS"] += ' /O2 /fp:fast '
-
+		env["CCFLAGS"] += ' /Oi /Ot /Oy /Ob2 /GF /Gy /fp:fast '
+		# enabling /O2 (or /Og) will crash demos/retro for some reason.
+
+
+		# concering CPU architectures and AVX:
+
+		# by default, we do not specify a value for /arch - even though in theory, /arch:AVX should be
+		# fine for processors not older than 2011 (https://en.wikipedia.org/wiki/Advanced_Vector_Extensions),
+		# in practice setting AVX causes issues (see https://github.com/poke1024/tove2d/issues/24).
+		# for /arch:AVX2, we even need Haswell architectures.
+
+		# the main benefit for enabling AVX2 would probably be src/thirdparty/nanosvg/tove/svgrast.cpp,
+		# where various dithering operations are optimizable for SIMD (AVX can then be seen in asm code).
+
+		arch = GetOption('arch')
+		if arch == 'sandybridge':
+			env["CCFLAGS"] += ' /arch:AVX '
+		elif arch == 'haswell':
+			env["CCFLAGS"] += ' /arch:AVX2 '
+		elif arch is not None:
+			raise RuntimeError("please specify sandybridge or haswell for arch")
+
+
 		env["LINKFLAGS"] = ' /OPT:REF '
 	else:
 		env["CCFLAGS"] += ' /DEBUG:FULL '
 		env["LINKFLAGS"] = ' /DEBUG:FULL '
 
 else:
-	# might want to trigger these with additional options:
-	# -march=haswell
 	# -Wreorder -Wunused-variable
 
 	CCFLAGS = ' -std=c++17 -fvisibility=hidden -funsafe-math-optimizations '
 
+	if GetOption('arch'):
+		CCFLAGS += ' -march=%s ' % GetOption('arch')
+
 	if GetOption('tovedebug'):
 		CCFLAGS += '-g '
 	else:
 		CCFLAGS += '-O3 '
 
-	if env["PLATFORM"] == 'posix':
+	if env["PLATFORM"] == 'posix' and GetOption('f16c'):
 		CCFLAGS += ' -mf16c '
 
+	if GetOption('f16c'):
+		CCFLAGS += ' -DTOVE_F16C=1 '
+
 	env["CCFLAGS"] = CCFLAGS
 
 env["CPPPATH"] = "src/thirdparty/fp16/include"

src/cpp/common.h

@@ -63,11 +63,11 @@ inline void store_gpu_float(float &p, float x) {
 	p = x;
 }
 
-#if defined(__clang__)
+#if defined(TOVE_F16C) && defined(__clang__)
 inline void store_gpu_float(uint16_t &p, float x) {
 	*reinterpret_cast<__fp16*>(&p) = x;
 }
-#elif defined(__GNUC__)
+#elif defined(TOVE_F16C) && defined(__GNUC__)
 #include <x86intrin.h>
 inline void store_gpu_float(uint16_t &p, float x) {
 	p = _cvtss_sh(x, 0);