Replaced FMA feature flag for relaxed which also ties in the other pl…

…atform precision variation functions like invsqrt and reciprical. Fixed WASM arch detect
Auburn · Mar 26, 2024 · 185d7ef · 185d7ef
1 parent 9fa0b71
commit 185d7ef
Show file tree

Hide file tree

Showing 20 changed files with 202 additions and 137 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -30,6 +30,10 @@ else()
     target_compile_definitions(FastSIMD PUBLIC FASTSIMD_STATIC_LIB)
 endif()
 
+if(CMAKE_SYSTEM_NAME STREQUAL "Emscripten")
+    target_compile_options(FastSIMD PUBLIC "-msimd128")
+endif()
+
 target_include_directories(FastSIMD PUBLIC 
     $<BUILD_INTERFACE:${FastSIMD_SOURCE_DIR}/include>
     $<INSTALL_INTERFACE:include>

diff --git a/dispatch/cmake/ClassSIMD.cmake b/dispatch/cmake/ClassSIMD.cmake
@@ -1,5 +1,5 @@
 
-function(fastsimd_add_feature_set_source simd_inl feature_set)
+function(fastsimd_add_feature_set_source simd_inl feature_set is_relaxed)
     set(feature_set_source "${simd_library_source_dir}/${simd_library_name}_${feature_set}.cpp")
     set(simd_inl_full "${CMAKE_CURRENT_LIST_DIR}/${simd_inl}")
 
@@ -9,52 +9,64 @@ function(fastsimd_add_feature_set_source simd_inl feature_set)
     if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
         # MSVC 32bit needs SSE2 flag for all SSE levels
         if(${feature_set} MATCHES "SSE[^(0-9)]" AND CMAKE_SIZEOF_VOID_P EQUAL 4)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "/arch:SSE2")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS /arch:SSE2)
 
         elseif(${feature_set} MATCHES "AVX[^(0-9)]")
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "/arch:AVX")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS /arch:AVX)
 
         elseif(${feature_set} MATCHES AVX2)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "/arch:AVX2")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS /arch:AVX2)
 
         elseif(${feature_set} MATCHES AVX512)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "/arch:AVX512")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS /arch:AVX512)
         endif()
     else()
         if(${feature_set} MATCHES SSE2 AND CMAKE_SIZEOF_VOID_P EQUAL 4)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-msse2")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -msse2)
 
         elseif(${feature_set} MATCHES SSE3)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-msse3")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -msse3)
 
         elseif(${feature_set} MATCHES SSSE3)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-mssse3")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mssse3)
 
         elseif(${feature_set} MATCHES SSE41)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-msse4.1")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -msse4.1)
 
         elseif(${feature_set} MATCHES SSE42)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-msse4.2")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -msse4.2)
 
         elseif(${feature_set} MATCHES "AVX[^(0-9)]")
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-mavx")
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mavx)
 
         elseif(${feature_set} MATCHES AVX2)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-mavx2 -mfma")
+            if(is_relaxed)
+                set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mfma)
+            else()
+                set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mno-fma) 
+            endif()
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mavx2)
 
         elseif(${feature_set} MATCHES AVX512)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-mavx512f -mavx512dq -mavx512vl -mavx512bw -mfma")
+            if(is_relaxed)
+                set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mfma)
+            else()
+                set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mno-fma)                
+            endif()
+            set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mavx512f -mavx512dq -mavx512vl -mavx512bw)
 
         elseif(${feature_set} MATCHES WASM)
-            set_source_files_properties(${feature_set_source} PROPERTIES COMPILE_FLAGS "-msimd128 -mrelaxed-simd")
+            if(is_relaxed)
+                set_property(SOURCE ${feature_set_source} APPEND PROPERTY COMPILE_OPTIONS -mrelaxed-simd)
+            endif()
         endif()
     endif()
 
 endfunction()
 
 function(fastsimd_create_dispatch_library simd_library_name)
 
-    cmake_parse_arguments(PARSE_ARGV 0 fastsimd_create_dispatch_library "" "" "SOURCES;FEATURE_SETS")
+    cmake_parse_arguments(PARSE_ARGV 0 fastsimd_create_dispatch_library "RELAXED" "" "SOURCES;FEATURE_SETS")
 
     list(LENGTH fastsimd_create_dispatch_library_FEATURE_SETS FEATURE_SET_COUNT)
     list(LENGTH fastsimd_create_dispatch_library_SOURCES SOURCES_COUNT)
@@ -68,10 +80,10 @@ function(fastsimd_create_dispatch_library simd_library_name)
         set(fastsimd_create_dispatch_library_FEATURE_SETS
             SSE2
             SSE41
-            AVX2_FMA
-            AVX512_FMA
-            NEON_FMA
-            AARCH64_FMA
+            AVX2
+            AVX512
+            NEON
+            AARCH64
             WASM)
     endif()
 
@@ -93,6 +105,10 @@ function(fastsimd_create_dispatch_library simd_library_name)
     if(CMAKE_COMPILER_IS_GNUCC)
         set_target_properties(${simd_library_name} PROPERTIES COMPILE_FLAGS "-Wno-ignored-attributes")
     endif()
+
+    if(fastsimd_create_dispatch_library_RELAXED)
+        target_compile_definitions(${simd_library_name} PUBLIC FASTSIMD_IS_RELAXED=1)
+    endif()
 
     set(feature_set_list "")
     set(feature_set_list_debug "")
@@ -112,7 +128,7 @@ function(fastsimd_create_dispatch_library simd_library_name)
             if ("${COMPILE_OUTPUT}" MATCHES "TEST_FEATURE_SET_ACTIVE_SUCCESS")
                 list(APPEND feature_set_list "FastSIMD::FeatureSet::${feature_set}")
                 list(APPEND feature_set_list_debug "${feature_set}")
-                fastsimd_add_feature_set_source(${simd_inl} ${feature_set})
+                fastsimd_add_feature_set_source(${simd_inl} ${feature_set} ${fastsimd_create_dispatch_library_RELAXED})
             endif()
         endforeach()
     endforeach()

diff --git a/dispatch/impl/DispatchClassImpl.h b/dispatch/impl/DispatchClassImpl.h
@@ -58,8 +58,8 @@ namespace FastSIMD
     };
 
 
-    template<typename T, FastSIMD::FeatureSet SIMD>
-    FS_FORCEINLINE static T* DispatchClassFactoryIterator( FastSIMD::FeatureSet maxFeatureSet, MemoryAllocator allocator )
+    template<typename T, FeatureSet SIMD>
+    FS_FORCEINLINE static T* DispatchClassFactoryIterator( FeatureSet maxFeatureSet, MemoryAllocator allocator )
     {
         if( maxFeatureSet < SIMD )
         {
@@ -68,7 +68,7 @@ namespace FastSIMD
 
         constexpr auto NextCompiled = FastSIMD::FASTSIMD_LIBRARY_NAME::CompiledFeatureSets::NextAfter<SIMD>;
 
-        if constexpr( NextCompiled != FastSIMD::FeatureSet::Max )
+        if constexpr( NextCompiled != FeatureSet::Max )
         {
             if( maxFeatureSet >= NextCompiled )
             {
@@ -80,7 +80,7 @@ namespace FastSIMD
     }
 
     template<typename T>
-    FASTSIMD_API T* NewDispatchClass( FastSIMD::FeatureSet maxFeatureSet, MemoryAllocator allocator )
+    FASTSIMD_API T* NewDispatchClass( FeatureSet maxFeatureSet, MemoryAllocator allocator )
     {
         if( maxFeatureSet == FeatureSet::Max )
         {

diff --git a/examples/header_only/main.cpp b/examples/header_only/main.cpp
@@ -10,5 +10,5 @@ int main()
 
     auto out = FS::Masked( invSqrt != FS::f32<4>( INFINITY ), dist );
 
-    return 0;
+    return FS::Extract0( FS::Convert<int>( out ) );
 }
diff --git a/include/FastSIMD/ToolSet.h b/include/FastSIMD/ToolSet.h
@@ -12,6 +12,12 @@ namespace FastSIMD
     {
         return SIMD;
     }
+
+    template<auto = 0, bool RELAXED = FASTSIMD_IS_RELAXED>
+    static constexpr bool IsRelaxed()
+    {
+        return RELAXED;
+    }
 } // namespace FastSIMD
 
 #include "ToolSet/Generic/Functions.h"

diff --git a/include/FastSIMD/ToolSet/Generic/Functions.h b/include/FastSIMD/ToolSet/Generic/Functions.h
@@ -464,8 +464,22 @@ namespace FS
     template<typename T, std::size_t N, FastSIMD::FeatureSet SIMD>
     FS_FORCEINLINE Register<T, N, SIMD> InvSqrt( const Register<T, N, SIMD>& a )
     {
-        static_assert( !IsNativeV<Register<T, N, SIMD>>, "FastSIMD: FS::InvSqrt not supported with provided types" ); 
-        return Register<T, N, SIMD>{ InvSqrt( a.v0 ), InvSqrt( a.v1 ) };        
+        if constexpr( IsNativeV<Register<T, N, SIMD>> )
+        {
+            return Register<T, N, SIMD>( 1 ) / Sqrt( a );
+        }
+        else
+        {
+            return Register<T, N, SIMD>{ InvSqrt( a.v0 ), InvSqrt( a.v1 ) };
+        }   
+    }
+
+    // Sqrt
+    template<typename T, std::size_t N, FastSIMD::FeatureSet SIMD>
+    FS_FORCEINLINE Register<T, N, SIMD> Sqrt( const Register<T, N, SIMD>& a )
+    {
+        static_assert( !IsNativeV<Register<T, N, SIMD>>, "FastSIMD: FS::Sqrt not supported with provided types" );
+        return Register<T, N, SIMD>{ Sqrt( a.v0 ), Sqrt( a.v1 ) };
     }
 
     template<std::size_t N, FastSIMD::FeatureSet SIMD>

diff --git a/include/FastSIMD/ToolSet/Generic/Register.h b/include/FastSIMD/ToolSet/Generic/Register.h
@@ -287,6 +287,9 @@ namespace FS
     template<typename T>
     constexpr bool IsNativeV = IsNative<T>::value;
 
+    template<auto T = 0>
+    using EnableIfRelaxed = std::enable_if_t<FastSIMD::IsRelaxed<T>()>;
+
 
     template<std::size_t N, FastSIMD::FeatureSet SIMD = FastSIMD::FeatureSetDefault()>
     using i32 = Register<std::int32_t, N, SIMD>;

diff --git a/include/FastSIMD/ToolSet/Generic/Scalar/f32x1.h b/include/FastSIMD/ToolSet/Generic/Scalar/f32x1.h
@@ -123,8 +123,8 @@ namespace FS
 
 
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<1, SIMD>>>
-    FS_FORCEINLINE f32<1, SIMD> InvSqrt( const f32<1, SIMD>& a )
+    FS_FORCEINLINE f32<1, SIMD> Sqrt( const f32<1, SIMD>& a )
     {
-        return 1 / std::sqrt( a.native.f );
+        return std::sqrt( a.native.f );
     }
 }
diff --git a/include/FastSIMD/ToolSet/WASM/128/f32x4.h b/include/FastSIMD/ToolSet/WASM/128/f32x4.h
@@ -156,13 +156,27 @@ namespace FS
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
     FS_FORCEINLINE f32<4, SIMD> Min( const f32<4, SIMD>& a, const f32<4, SIMD>& b )
     {
-        return wasm_f32x4_pmin( a.native, b.native );
+        if constexpr( FastSIMD::IsRelaxed() )
+        {
+            return wasm_f32x4_relaxed_min( a.native, b.native );        
+        }
+        else
+        {
+            return wasm_f32x4_min( a.native, b.native );
+        }
     }
 
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
     FS_FORCEINLINE f32<4, SIMD> Max( const f32<4, SIMD>& a, const f32<4, SIMD>& b )
     {
-        return wasm_f32x4_pmax( a.native, b.native );
+        if constexpr( FastSIMD::IsRelaxed() )
+        {
+            return wasm_f32x4_relaxed_max( a.native, b.native );
+        }
+        else
+        {
+            return wasm_f32x4_max( a.native, b.native );
+        }
     }
 
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
@@ -182,50 +196,33 @@ namespace FS
     FS_FORCEINLINE f32<4, SIMD> MaskedIncrement( const typename f32<4, SIMD>::MaskTypeArg& mask, const f32<4, SIMD>& a )
     {
         return wasm_f32x4_sub( a.native,
-            wasm_f32x4_convert_i32x4(static_cast<v128_t>(mask.native)) );
+            wasm_f32x4_convert_i32x4( static_cast<v128_t>( mask.native ) ) );
     }
 
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
     FS_FORCEINLINE f32<4, SIMD> MaskedDecrement( const typename f32<4, SIMD>::MaskTypeArg& mask, const f32<4, SIMD>& a )
     {
         return wasm_f32x4_add( a.native,
-            wasm_f32x4_convert_i32x4(static_cast<v128_t>(mask.native)) );
+            wasm_f32x4_convert_i32x4( static_cast<v128_t>( mask.native ) ) );
     }
 
 
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
-    FS_FORCEINLINE f32<4, SIMD> Reciprocal( const f32<4, SIMD>& a )
+    FS_FORCEINLINE f32<4, SIMD> Sqrt( const f32<4, SIMD>& a )
     {
-        return wasm_f32x4_div( f32<4, SIMD>{1.0f}.native, a.native );
+        return wasm_f32x4_sqrt( a.native );
     }
 
-    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
-    FS_FORCEINLINE f32<4, SIMD> InvSqrt( const f32<4, SIMD>& a )
-    {
-        return wasm_f32x4_div( f32<4, SIMD>{1.0f}.native, wasm_f32x4_sqrt( a.native ) );
-    }
 
-    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = std::enable_if_t<SIMD & FastSIMD::FeatureFlag::FMA>>
+    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = EnableIfRelaxed<SIMD>()>
     FS_FORCEINLINE f32<4, SIMD> FMulAdd( const f32<4, SIMD>& a, const f32<4, SIMD>& b, const f32<4, SIMD>& c )
     {
         return wasm_f32x4_relaxed_madd( c.native, a.native, b.native );
     }
 
-    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = std::enable_if_t<SIMD & FastSIMD::FeatureFlag::FMA>>
-    FS_FORCEINLINE f32<4, SIMD> FMulSub( const f32<4, SIMD>& a, const f32<4, SIMD>& b, const f32<4, SIMD>& c )
-    {
-        return wasm_f32x4_relaxed_msub( c.native, a.native, b.native );
-    }
-
-    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = std::enable_if_t<SIMD & FastSIMD::FeatureFlag::FMA>>
+    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = EnableIfRelaxed<SIMD>()>
     FS_FORCEINLINE f32<4, SIMD> FNMulAdd( const f32<4, SIMD>& a, const f32<4, SIMD>& b, const f32<4, SIMD>& c )
     {
         return wasm_f32x4_relaxed_nmadd( c.native, a.native, b.native );
     }
-
-    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = std::enable_if_t<SIMD & FastSIMD::FeatureFlag::FMA>>
-    FS_FORCEINLINE f32<4, SIMD> FNMulSub( const f32<4, SIMD>& a, const f32<4, SIMD>& b, const f32<4, SIMD>& c )
-    {
-        return wasm_f32x4_relaxed_nmsub( c.native, a.native, b.native );
-    }
 }
diff --git a/include/FastSIMD/ToolSet/WASM/WASM.h b/include/FastSIMD/ToolSet/WASM/WASM.h
@@ -13,7 +13,14 @@ namespace FS
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
     FS_FORCEINLINE i32<4, SIMD> Convert( const f32<4, SIMD>& a, TypeDummy<int32_t> )
     {
-        return wasm_i32x4_trunc_sat_f32x4( Round( a ).native );
+        if constexpr( FastSIMD::IsRelaxed() )
+        {
+            return wasm_i32x4_relaxed_trunc_f32x4( Round( a ).native );
+        }
+        else
+        {
+            return wasm_i32x4_trunc_sat_f32x4( Round( a ).native );
+        }
     }
 
     template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<i32<4, SIMD>>>

diff --git a/include/FastSIMD/ToolSet/x86/128/f32x4.h b/include/FastSIMD/ToolSet/x86/128/f32x4.h
@@ -247,15 +247,21 @@ namespace FS
     }
 
 
-    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
+    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = EnableIfRelaxed<SIMD>()>
     FS_FORCEINLINE f32<4, SIMD> Reciprocal( const f32<4, SIMD>& a )
     {            
         return _mm_rcp_ps( a.native );
     }
 
-    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
+    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>, typename = EnableIfRelaxed<SIMD>()>
     FS_FORCEINLINE f32<4, SIMD> InvSqrt( const f32<4, SIMD>& a )
     {            
         return _mm_rsqrt_ps( a.native );
     }
+
+    template<FastSIMD::FeatureSet SIMD, typename = EnableIfNative<f32<4, SIMD>>>
+    FS_FORCEINLINE f32<4, SIMD> Sqrt( const f32<4, SIMD>& a )
+    {            
+        return _mm_sqrt_ps( a.native );
+    }
 }