Merge pull request #18 from maniatic0/main

Support for WASM with EMSCRIPTEN

CMakeLists.txt

@@ -8,38 +8,126 @@ project(tiny_bvh LANGUAGES CXX)
 
 if (APPLE)
 	find_library(COCOA_LIBRARY Cocoa)
-elseif (UNIX)
+elseif (UNIX AND NOT EMSCRIPTEN)
 	find_package(X11)
+elseif (EMSCRIPTEN)
+	# To auto generate test pages
+	set(CMAKE_EXECUTABLE_SUFFIX ".html")
+
+	# Enable clang's Address Sanitizer for WASM
+	option(enable_asan "Enable ASAN on the build (it'll slow down the app)" OFF)
+
+	# Enable wasm SIMD instructions
+	set(emcc_simd "fixed" CACHE STRING "SIMD type to use in the build (none, fixed, relaxed)")
+	set_property(CACHE emcc_simd PROPERTY STRINGS none fixed relaxed)
+
+	# Enable thread support for WASM (note that you need special http headers now https://emscripten.org/docs/porting/pthreads.html )
+	option(emcc_threads "Enable Threads in WASM (note that means you need to change your http headers)" OFF)
+
+	# Enable C/C++ main thread as a proxy thread to avoid blocking DOM/UI thread https://emscripten.org/docs/porting/pthreads.html
+	option(emcc_proxy_main "Enable C/C++ main thread as a proxy thread to avoid blocking DOM/UI thread" ${emcc_threads})
 endif()
 
 enable_testing()
 
 add_executable(tiny_bvh_test tiny_bvh_test.cpp)
 add_executable(tiny_bvh_renderer tiny_bvh_renderer.cpp)
 add_executable(tiny_bvh_speedtest tiny_bvh_speedtest.cpp)
-add_executable(tiny_bvh_fenster tiny_bvh_fenster.cpp)
+if (NOT EMSCRIPTEN)
+	# EMSCRIPTEN doesn't render anything by default (you would need WebGL/WebGPU)
+	add_executable(tiny_bvh_fenster tiny_bvh_fenster.cpp)
+endif()
 
 if (NOT MSVC)
+	# Produce debug symbols and set optimization level
 	set(common_cxx_flags
-		-g -march=native
+		$<$<OR:$<CONFIG:RelWithDebInfo>,$<CONFIG:Debug>>:-g> $<$<OR:$<CONFIG:RelWithDebInfo>,$<CONFIG:Release>>:-O3>
 	)
+
+	set(common_link_flags
+		$<$<OR:$<CONFIG:RelWithDebInfo>,$<CONFIG:Debug>>:-g>
+	)
+
+	if (NOT EMSCRIPTEN)
+		# EMSCRIPTEN doesn't know about archs
+		set(common_cxx_flags
+			${common_cxx_flags} -march=native
+		)
+	else()
+		# EMSCRIPTEN flags:
+		# -sALLOW_MEMORY_GROWTH (allow runtime allocations) (for some reason cmake passes this as a linker flag when in theory it's a compiler one)
+		set(common_link_flags ${common_link_flags} -sALLOW_MEMORY_GROWTH=1)
+
+		# -fsanitize=address (enable ASAN)
+		# More at https://emscripten.org/docs/debugging/Sanitizers.html#address-sanitizer 
+		if (enable_asan) 
+			set(common_cxx_flags ${common_cxx_flags} -fsanitize=address)
+			set(common_link_flags ${common_link_flags} -fsanitize=address)
+		endif()
+
+		# Enable SIMD instructions and AVX intrinsics
+		string(TOLOWER "${emcc_simd}" emcc_simd_lower)
+		if (emcc_simd_lower STREQUAL fixed)
+			# Supported by all browsers https://webassembly.org/features/
+			set(common_cxx_flags ${common_cxx_flags} -msimd128 -mavx)
+		elseif (emcc_simd_lower STREQUAL relaxed)
+			# Supported some browsers https://webassembly.org/features/
+			set(common_cxx_flags ${common_cxx_flags} -mrelaxed-simd -mavx)
+		endif()
+
+	endif()
+
 	target_compile_options(tiny_bvh_test PRIVATE ${common_cxx_flags})
+	target_link_options(tiny_bvh_test PRIVATE ${common_link_flags})
+
 	target_compile_options(tiny_bvh_renderer PRIVATE ${common_cxx_flags})
-	if (NOT APPLE)
-		target_compile_options(tiny_bvh_speedtest PRIVATE ${common_cxx_flags} -fopenmp)
-		target_link_options(tiny_bvh_speedtest PRIVATE -fopenmp)
-	else()
-		# No openmp support in default compiler
-		target_compile_options(tiny_bvh_speedtest PRIVATE ${common_cxx_flags})
-		target_link_options(tiny_bvh_speedtest PRIVATE)
+	target_link_options(tiny_bvh_renderer PRIVATE ${common_link_flags})
+
+	# No openmp support in default compiler
+	set(tiny_bvh_speedtest_cxx_flags ${common_cxx_flags})
+	set(tiny_bvh_speedtest_link_flags ${common_link_flags})
+
+	if (NOT APPLE AND NOT EMSCRIPTEN)
+		set(tiny_bvh_speedtest_cxx_flags ${tiny_bvh_speedtest_cxx_flags} -fopenmp)
+		set(tiny_bvh_speedtest_link_flags ${tiny_bvh_speedtest_link_flags} -fopenmp)
+	elseif (EMSCRIPTEN)
+		if (emcc_threads)
+			set(tiny_bvh_speedtest_cxx_flags ${tiny_bvh_speedtest_cxx_flags} -pthread)
+			set(tiny_bvh_speedtest_link_flags ${tiny_bvh_speedtest_link_flags} -pthread -sUSE_PTHREADS=1 -sPTHREAD_POOL_SIZE=navigator.hardwareConcurrency)
+
+			if (FALSE)
+				# From:
+				#	- https://github.com/llvm/llvm-project/pull/95169 
+				#	- https://github.com/llvm/llvm-project/pull/71297 
+				#	- https://reviews.llvm.org/D142593?id=492403
+				# llvm's and clang's OpenMP has been ported to WASM around August 2024
+				# Sadly emcc's clang doesn't include a copy of the lib
+				# So you would need to do compile it yourself https://github.com/abrown/wasm-openmp-examples/blob/main/emscripten.sh
+				# Then link it here
+				set(tiny_bvh_speedtest_cxx_flags ${tiny_bvh_speedtest_cxx_flags} -fopenmp)
+				set(tiny_bvh_speedtest_link_flags ${tiny_bvh_speedtest_link_flags} -fopenmp)
+			endif()
+
+			if (emcc_proxy_main)
+				set(tiny_bvh_speedtest_link_flags ${tiny_bvh_speedtest_link_flags} -sPROXY_TO_PTHREAD=1)
+			endif()
+		endif()
 	endif()
-	target_compile_options(tiny_bvh_fenster PRIVATE ${common_cxx_flags})
-	if (WIN32)
-		target_link_libraries(tiny_bvh_fenster -mwindows)
-	elseif (APPLE)
-		target_link_libraries(tiny_bvh_fenster ${COCOA_LIBRARY})
-	elseif (X11_FOUND)
-		target_link_libraries(tiny_bvh_fenster ${X11_LIBRARIES})
+	target_compile_options(tiny_bvh_speedtest PRIVATE ${tiny_bvh_speedtest_cxx_flags})
+	target_link_options(tiny_bvh_speedtest PRIVATE ${tiny_bvh_speedtest_link_flags})
+
+	if (TARGET tiny_bvh_fenster)
+		# Not all platforms support all targets. E. g.:
+		#	- EMSCRIPTEN doesn't render anything by default (you would need WebGL/WebGPU)
+		target_compile_options(tiny_bvh_fenster PRIVATE ${common_cxx_flags})
+		target_link_options(tiny_bvh_fenster PRIVATE ${common_link_flags})
+		if (WIN32)
+			target_link_libraries(tiny_bvh_fenster -mwindows)
+		elseif (APPLE)
+			target_link_libraries(tiny_bvh_fenster ${COCOA_LIBRARY})
+		elseif (X11_FOUND)
+			target_link_libraries(tiny_bvh_fenster ${X11_LIBRARIES})
+		endif()
 	endif()
 else()
 	if (WIN32)

tiny_bvh.h

@@ -71,7 +71,7 @@ THE SOFTWARE.
 #define BVHBINS 8
 
 // include fast AVX BVH builder
-#if defined(__x86_64__) || defined(_M_X64)
+#if defined(__x86_64__) || defined(_M_X64) || defined(__wasm_simd128__) || defined(__wasm_relaxed_simd__)
 #define BVH_USEAVX
 #endif
 
@@ -95,6 +95,24 @@ THE SOFTWARE.
 #define ALIGNED( x ) __declspec( align( x ) )
 #define ALIGNED_MALLOC( x ) ( ( x ) == 0 ? 0 : _aligned_malloc( ( x ), 64 ) )
 #define ALIGNED_FREE( x ) _aligned_free( x )
+#elif defined(__EMSCRIPTEN__)
+// EMSCRIPTEN (needs to be before gcc and clang to avoid misdetection) 
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#define ALIGNED( x ) __attribute__( ( aligned( x ) ) )
+#if defined(__wasm_simd128__) || defined(__wasm_relaxed_simd__)
+// https://emscripten.org/docs/porting/simd.html
+#include <xmmintrin.h>
+#define ALIGNED_MALLOC( x ) ( ( x ) == 0 ? 0 : _mm_malloc( ( x ), 64 ) )
+#define ALIGNED_FREE( x ) _mm_free( x )
+#else
+// Size needs to be a multiple of Alignment in the standard (in EMSCRIPTEN this is enforced)
+// See https://en.cppreference.com/w/c/memory/aligned_alloc
+#define MAKE_MULIPLE_64( x ) ( ( ( x ) + 63 ) & ( ~0x3f ) )
+#define ALIGNED_MALLOC( x ) ( ( x ) == 0 ? 0 : aligned_alloc( 64, MAKE_MULIPLE_64( x ) ) )
+#define ALIGNED_FREE( x ) free( x )
+#endif
 #else
 // gcc / clang
 #include <cstdlib>
@@ -2578,6 +2596,8 @@ static unsigned __bfind( unsigned x ) // https://github.com/mackron/refcode/blob
 {
 #if defined(_MSC_VER) && !defined(__clang__)
 	return 31 - __lzcnt( x );
+#elif defined(__EMSCRIPTEN__)
+	return 31 - __builtin_clz(x);
 #elif defined(__GNUC__) || defined(__clang__)
 	unsigned r;
 	__asm__ __volatile__( "lzcnt{l %1, %0| %0, %1}" : "=r"(r) : "r"(x) : "cc" );

tiny_bvh_speedtest.cpp

@@ -9,6 +9,9 @@
 #ifdef _WIN32
 #include <intrin.h>		// for __cpuidex
 #endif
+#ifdef __EMSCRIPTEN__ 
+#include <emscripten/version.h> // for __EMSCRIPTEN_major__, __EMSCRIPTEN_minor__
+#endif
 
 // 'screen resolution': see tiny_bvh_fenster.cpp; this program traces the
 // same rays, but without visualization - just performance statistics.
@@ -122,6 +125,9 @@ int main()
 	// determine compiler
 #ifdef _MSC_VER
 	printf( "(MSVC %i build)\n", _MSC_VER );
+#elif defined __EMSCRIPTEN__
+	// EMSCRIPTEN needs to be before clang or gcc
+	printf("(emcc %i.%i build)\n", __EMSCRIPTEN_major__, __EMSCRIPTEN_minor__);
 #elif defined __clang__
 	printf( "(clang %i.%i build)\n", __clang_major__, __clang_minor__ );
 #elif defined __GNUC__
@@ -405,6 +411,6 @@ int main()
 
 #endif
 
-	// all done.
+	printf( "all done." );
 	return 0;
 }