diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7e8b319a2..e09e14fff 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -219,6 +219,7 @@ else()
     include(CheckCXXCompilerFlag)
     if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
         check_cxx_compiler_flag("/arch:AVX" NCNN_COMPILER_SUPPORT_X86_AVX)
+        check_cxx_compiler_flag("/arch:AVX" NCNN_COMPILER_SUPPORT_X86_FMA)
         check_cxx_compiler_flag("/arch:AVX" NCNN_COMPILER_SUPPORT_X86_XOP)
         check_cxx_compiler_flag("/arch:AVX2" NCNN_COMPILER_SUPPORT_X86_AVX2)
         check_cxx_compiler_flag("/arch:AVX2" NCNN_COMPILER_SUPPORT_X86_AVX_VNNI)
@@ -232,6 +233,7 @@ else()
         endif()
     else()
         check_cxx_compiler_flag("-mavx" NCNN_COMPILER_SUPPORT_X86_AVX)
+        check_cxx_compiler_flag("-mfma" NCNN_COMPILER_SUPPORT_X86_FMA)
         check_cxx_compiler_flag("-mxop" NCNN_COMPILER_SUPPORT_X86_XOP)
         check_cxx_compiler_flag("-mfma -mf16c -mavx2" NCNN_COMPILER_SUPPORT_X86_AVX2)
         check_cxx_compiler_flag("-mfma -mf16c -mavx512f -mavx512vl" NCNN_COMPILER_SUPPORT_X86_AVX512)
@@ -247,6 +249,13 @@ else()
 
     if(NOT CMAKE_SYSTEM_NAME STREQUAL "Emscripten" AND NCNN_COMPILER_SUPPORT_X86_AVX)
         option(NCNN_AVX "optimize x86 platform with avx extension" ON)
+        if(NCNN_COMPILER_SUPPORT_X86_FMA)
+            if(NCNN_AVX)
+                option(NCNN_FMA "optimize x86 platform with fma extension" ON)
+            endif()
+        else()
+            message(WARNING "The compiler does not support fma extension. NCNN_FMA will be OFF.")
+        endif()
         if(NCNN_COMPILER_SUPPORT_X86_XOP)
             if(NCNN_AVX)
                 option(NCNN_XOP "optimize x86 platform with xop extension" ON)
diff --git a/cmake/ncnn_add_layer.cmake b/cmake/ncnn_add_layer.cmake
index f08ec8844..0c1903c3a 100644
--- a/cmake/ncnn_add_layer.cmake
+++ b/cmake/ncnn_add_layer.cmake
@@ -159,82 +159,69 @@ macro(ncnn_add_layer class)
         set(layer_registry "${layer_registry}#if NCNN_STRING\n{\"${class}\", 0},\n#else\n{0},\n#endif\n")
     endif()
 
-    if(NCNN_RUNTIME_CPU AND NCNN_AVX512 AND NCNN_TARGET_ARCH STREQUAL "x86")
+    if(NCNN_RUNTIME_CPU AND NCNN_TARGET_ARCH STREQUAL "x86")
         if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
-            ncnn_add_arch_opt_layer(${class} avx512 "/arch:AVX512")
-        else()
-            ncnn_add_arch_opt_layer(${class} avx512 "-mfma -mf16c -mavx512f -mavx512vl")
-        endif()
-    endif()
-
-    if(NCNN_RUNTIME_CPU AND NCNN_AVX512VNNI AND NCNN_TARGET_ARCH STREQUAL "x86")
-        if(NCNN_COMPILER_SUPPORT_X86_AVX512_VNNI)
-            if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
-                ncnn_add_arch_opt_source(${class} avx512vnni "/arch:AVX512 /D__AVX512VNNI__")
-            else()
-                ncnn_add_arch_opt_source(${class} avx512vnni "-mfma -mf16c -mavx512f -mavx512vl -mavx512vnni")
+            if(NCNN_AVX512)
+                ncnn_add_arch_opt_layer(${class} avx512 "/arch:AVX512 /D__FMA__")
             endif()
-        endif()
-    endif()
-
-    if(NCNN_RUNTIME_CPU AND NCNN_AVX2 AND NCNN_TARGET_ARCH STREQUAL "x86")
-        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
-            ncnn_add_arch_opt_layer(${class} avx2 "/arch:AVX2")
-        else()
-            ncnn_add_arch_opt_layer(${class} avx2 "-mfma -mf16c -mavx2")
-        endif()
-    endif()
-
-    if(NCNN_RUNTIME_CPU AND NCNN_AVXVNNI AND NCNN_TARGET_ARCH STREQUAL "x86")
-        if(NCNN_COMPILER_SUPPORT_X86_AVX_VNNI)
-            if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
-                ncnn_add_arch_opt_source(${class} avxvnni "/arch:AVX2 /D__AVXVNNI__")
-            else()
-                ncnn_add_arch_opt_source(${class} avxvnni "-mfma -mf16c -mavx2 -mavxvnni")
+            if(NCNN_FMA)
+                ncnn_add_arch_opt_layer(${class} fma "/arch:AVX /D__FMA__")
             endif()
-        endif()
-    endif()
-
-    if(NCNN_RUNTIME_CPU AND NCNN_AVX AND NCNN_TARGET_ARCH STREQUAL "x86")
-        if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
-            ncnn_add_arch_opt_layer(${class} avx "/arch:AVX")
-        else()
-            ncnn_add_arch_opt_layer(${class} avx "-mavx")
-        endif()
-    endif()
-
-    if(NCNN_RUNTIME_CPU AND NCNN_XOP AND NCNN_TARGET_ARCH STREQUAL "x86")
-        if(NCNN_COMPILER_SUPPORT_X86_XOP)
-            if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
+            if(NCNN_AVX)
+                ncnn_add_arch_opt_layer(${class} avx "/arch:AVX")
+            endif()
+            if(NCNN_AVX512VNNI)
+                ncnn_add_arch_opt_source(${class} avx512vnni "/arch:AVX512 /D__FMA__ /D__AVX512VNNI__")
+            endif()
+            if(NCNN_AVXVNNI)
+                ncnn_add_arch_opt_source(${class} avxvnni "/arch:AVX2 /D__FMA__ /D__AVXVNNI__")
+            endif()
+            if(NCNN_AVX2)
+                ncnn_add_arch_opt_source(${class} avx2 "/arch:AVX2 /D__FMA__")
+            endif()
+            if(NCNN_XOP)
                 ncnn_add_arch_opt_source(${class} xop "/arch:AVX /D__XOP__")
-            else()
+            endif()
+        else()
+            if(NCNN_AVX512)
+                ncnn_add_arch_opt_layer(${class} avx512 "-mavx512f -mavx512vl -mfma -mf16c")
+            endif()
+            if(NCNN_FMA)
+                ncnn_add_arch_opt_layer(${class} fma "-mavx -mfma")
+            endif()
+            if(NCNN_AVX)
+                ncnn_add_arch_opt_layer(${class} avx "-mavx")
+            endif()
+            if(NCNN_AVX512VNNI)
+                ncnn_add_arch_opt_source(${class} avx512vnni "-mavx512f -mavx512vl -mfma -mf16c -mavx512vnni")
+            endif()
+            if(NCNN_AVXVNNI)
+                ncnn_add_arch_opt_source(${class} avxvnni "-mavx2 -mfma -mf16c -mavxvnni")
+            endif()
+            if(NCNN_AVX2)
+                ncnn_add_arch_opt_source(${class} avx2 "-mavx2 -mfma -mf16c")
+            endif()
+            if(NCNN_XOP)
                 ncnn_add_arch_opt_source(${class} xop "-mavx -mxop")
             endif()
         endif()
     endif()
 
-    if(NCNN_RUNTIME_CPU AND NCNN_ARM82 AND ((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm64|aarch64)")))
-        if(NCNN_COMPILER_SUPPORT_ARM82_FP16)
+    if(NCNN_RUNTIME_CPU AND ((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm64|aarch64)")))
+        if(NCNN_ARM82)
             ncnn_add_arch_opt_layer(${class} arm82 "-march=armv8.2-a+fp16")
         endif()
-    endif()
-
-    if(NCNN_RUNTIME_CPU AND NCNN_ARM82DOT AND ((IOS AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (APPLE AND CMAKE_OSX_ARCHITECTURES MATCHES "arm64") OR (CMAKE_SYSTEM_PROCESSOR MATCHES "^(arm64|aarch64)")))
-        if(NCNN_COMPILER_SUPPORT_ARM82_FP16_DOTPROD)
+        if(NCNN_ARM82DOT)
             ncnn_add_arch_opt_source(${class} arm82dot "-march=armv8.2-a+fp16+dotprod")
         endif()
     endif()
 
     if(NCNN_RUNTIME_CPU AND NCNN_MSA AND NCNN_TARGET_ARCH STREQUAL "mips")
-        if(NCNN_COMPILER_SUPPORT_MIPS_MSA)
-            ncnn_add_arch_opt_layer(${class} msa "-mmsa")
-        endif()
+        ncnn_add_arch_opt_layer(${class} msa "-mmsa")
     endif()
 
     if(NCNN_RUNTIME_CPU AND NCNN_MMI AND NCNN_TARGET_ARCH STREQUAL "mips")
-        if(NCNN_COMPILER_SUPPORT_LOONGSON_MMI)
-            ncnn_add_arch_opt_layer(${class} mmi "-mmsa -mloongson-mmi")
-        endif()
+        ncnn_add_arch_opt_layer(${class} mmi "-mmsa -mloongson-mmi")
     endif()
 
     if(NCNN_RUNTIME_CPU AND NCNN_RVV AND NCNN_TARGET_ARCH STREQUAL "riscv")
diff --git a/cmake/ncnn_generate_avx2_source.cmake b/cmake/ncnn_generate_fma_source.cmake
similarity index 60%
rename from cmake/ncnn_generate_avx2_source.cmake
rename to cmake/ncnn_generate_fma_source.cmake
index c00792545..d15d73a0b 100644
--- a/cmake/ncnn_generate_avx2_source.cmake
+++ b/cmake/ncnn_generate_fma_source.cmake
@@ -7,8 +7,8 @@ file(READ ${SRC} source_data)
 string(TOUPPER ${CLASS} CLASS_UPPER)
 string(TOLOWER ${CLASS} CLASS_LOWER)
 
-string(REGEX REPLACE "LAYER_${CLASS_UPPER}_X86_H" "LAYER_${CLASS_UPPER}_X86_AVX2_H" source_data "${source_data}")
-string(REGEX REPLACE "${CLASS}_x86" "${CLASS}_x86_avx2" source_data "${source_data}")
-string(REGEX REPLACE "#include \"${CLASS_LOWER}_x86.h\"" "#include \"${CLASS_LOWER}_x86_avx2.h\"" source_data "${source_data}")
+string(REGEX REPLACE "LAYER_${CLASS_UPPER}_X86_H" "LAYER_${CLASS_UPPER}_X86_FMA_H" source_data "${source_data}")
+string(REGEX REPLACE "${CLASS}_x86" "${CLASS}_x86_fma" source_data "${source_data}")
+string(REGEX REPLACE "#include \"${CLASS_LOWER}_x86.h\"" "#include \"${CLASS_LOWER}_x86_fma.h\"" source_data "${source_data}")
 
 file(WRITE ${DST} "${source_data}")
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index 65c2064ce..7cde1e194 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -331,29 +331,37 @@ if(NCNN_TARGET_ARCH STREQUAL "x86")
     if(NOT NCNN_RUNTIME_CPU AND NCNN_AVX512)
         if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
             if(NCNN_AVX512VNNI)
-                target_compile_options(ncnn PRIVATE /arch:AVX512 /D__AVX512VNNI__)
+                target_compile_options(ncnn PRIVATE /arch:AVX512 /D__FMA__ /D__AVX512VNNI__)
             else()
-                target_compile_options(ncnn PRIVATE /arch:AVX512)
+                target_compile_options(ncnn PRIVATE /arch:AVX512 /D__FMA__)
             endif()
         else()
             if(NCNN_AVX512VNNI)
-                target_compile_options(ncnn PRIVATE -mfma -mf16c -mavx512f -mavx512vl -mavx512vnni)
+                target_compile_options(ncnn PRIVATE -mavx512f -mavx512vl -mfma -mf16c -mavx512vnni)
             else()
-                target_compile_options(ncnn PRIVATE -mfma -mf16c -mavx512f -mavx512vl)
+                target_compile_options(ncnn PRIVATE -mavx512f -mavx512vl -mfma -mf16c)
             endif()
         endif()
-    elseif(NOT NCNN_RUNTIME_CPU AND NCNN_AVX2)
+    elseif(NOT NCNN_RUNTIME_CPU AND NCNN_FMA)
         if(CMAKE_CXX_COMPILER_ID MATCHES "MSVC" OR (CMAKE_CXX_COMPILER_ID MATCHES "Clang" AND CMAKE_CXX_SIMULATE_ID MATCHES "MSVC" AND CMAKE_CXX_COMPILER_FRONTEND_VARIANT MATCHES "MSVC"))
             if(NCNN_AVXVNNI)
-                target_compile_options(ncnn PRIVATE /arch:AVX2 /D__AVXVNNI__)
+                target_compile_options(ncnn PRIVATE /arch:AVX2 /D__FMA__ /D__AVXVNNI__)
+            elseif(NCNN_AVX2)
+                target_compile_options(ncnn PRIVATE /arch:AVX2 /D__FMA__)
+            elseif(NCNN_XOP)
+                target_compile_options(ncnn PRIVATE /arch:AVX /D__FMA__ /D__XOP__)
             else()
-                target_compile_options(ncnn PRIVATE /arch:AVX2)
+                target_compile_options(ncnn PRIVATE /arch:AVX /D__FMA__)
             endif()
         else()
             if(NCNN_AVXVNNI)
-                target_compile_options(ncnn PRIVATE -mfma -mf16c -mavx2 -mavxvnni)
+                target_compile_options(ncnn PRIVATE -mavx2 -mfma -mf16c -mavxvnni)
+            elseif(NCNN_AVX2)
+                target_compile_options(ncnn PRIVATE -mavx2 -mfma -mf16c)
+            elseif(NCNN_XOP)
+                target_compile_options(ncnn PRIVATE -mavx -mfma -mxop)
             else()
-                target_compile_options(ncnn PRIVATE -mfma -mf16c -mavx2)
+                target_compile_options(ncnn PRIVATE -mavx -mfma)
             endif()
         endif()
     elseif(NOT NCNN_RUNTIME_CPU AND NCNN_AVX)
diff --git a/src/cpu.cpp b/src/cpu.cpp
index f5bbfe0e4..bd71177fd 100644
--- a/src/cpu.cpp
+++ b/src/cpu.cpp
@@ -512,6 +512,30 @@ static int get_cpu_support_x86_avx()
     return 1;
 }
 
+static int get_cpu_support_x86_fma()
+{
+#if !NCNN_FMA
+    return 0;
+#endif
+    unsigned int cpu_info[4] = {0};
+    x86_cpuid(0, cpu_info);
+
+    int nIds = cpu_info[0];
+    if (nIds < 7)
+        return 0;
+
+    x86_cpuid(1, cpu_info);
+    // check AVX XSAVE OSXSAVE
+    if (!(cpu_info[2] & (1u << 28)) || !(cpu_info[2] & (1u << 26)) || !(cpu_info[2] & (1u << 27)))
+        return 0;
+
+    // check XSAVE enabled by kernel
+    if ((x86_get_xcr0() & 6) != 6)
+        return 0;
+
+    return cpu_info[2] & (1u << 12);
+}
+
 static int get_cpu_support_x86_xop()
 {
 #if !NCNN_XOP
@@ -637,6 +661,7 @@ static int get_cpu_support_x86_avx512_vnni()
 }
 
 static int g_cpu_support_x86_avx = get_cpu_support_x86_avx();
+static int g_cpu_support_x86_fma = get_cpu_support_x86_fma();
 static int g_cpu_support_x86_xop = get_cpu_support_x86_xop();
 static int g_cpu_support_x86_avx2 = get_cpu_support_x86_avx2();
 static int g_cpu_support_x86_avx_vnni = get_cpu_support_x86_avx_vnni();
@@ -644,6 +669,7 @@ static int g_cpu_support_x86_avx512 = get_cpu_support_x86_avx512();
 static int g_cpu_support_x86_avx512_vnni = get_cpu_support_x86_avx512_vnni();
 #else  // defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
 static const int g_cpu_support_x86_avx = 0;
+static const int g_cpu_support_x86_fma = 0;
 static const int g_cpu_support_x86_xop = 0;
 static const int g_cpu_support_x86_avx2 = 0;
 static const int g_cpu_support_x86_avx_vnni = 0;
@@ -656,6 +682,11 @@ int cpu_support_x86_avx()
     return g_cpu_support_x86_avx;
 }
 
+int cpu_support_x86_fma()
+{
+    return g_cpu_support_x86_fma;
+}
+
 int cpu_support_x86_xop()
 {
     return g_cpu_support_x86_xop;
diff --git a/src/cpu.h b/src/cpu.h
index 979ce8522..9a0f64646 100644
--- a/src/cpu.h
+++ b/src/cpu.h
@@ -54,17 +54,19 @@ NCNN_EXPORT int cpu_support_arm_asimdhp();
 // asimddp = aarch64 asimd dot product
 NCNN_EXPORT int cpu_support_arm_asimddp();
 
-// avx = x86_64 avx
+// avx = x86 avx
 NCNN_EXPORT int cpu_support_x86_avx();
-// xop = x86_64 xop
+// xop = x86 xop
 NCNN_EXPORT int cpu_support_x86_xop();
-// avx2 = x86_64 avx2 + fma + f16c
+// fma = x86 fma
+NCNN_EXPORT int cpu_support_x86_fma();
+// avx2 = x86 avx2 + fma + f16c
 NCNN_EXPORT int cpu_support_x86_avx2();
-// avx_vnni = x86_64 avx vnni
+// avx_vnni = x86 avx vnni
 NCNN_EXPORT int cpu_support_x86_avx_vnni();
-// avx512 = x86_64 avx512f + avx512vl
+// avx512 = x86 avx512f + avx512vl
 NCNN_EXPORT int cpu_support_x86_avx512();
-// avx512_vnni = x86_64 avx512 vnni
+// avx512_vnni = x86 avx512 vnni
 NCNN_EXPORT int cpu_support_x86_avx512_vnni();
 
 // msa = mips mas
diff --git a/src/layer.cpp b/src/layer.cpp
index 199882239..21ae073a9 100644
--- a/src/layer.cpp
+++ b/src/layer.cpp
@@ -239,13 +239,13 @@ Layer* create_layer(int index)
     }
     else
 #endif// NCNN_RUNTIME_CPU && NCNN_AVX512
-#if NCNN_RUNTIME_CPU && NCNN_AVX2
-    if (ncnn::cpu_support_x86_avx2())
+#if NCNN_RUNTIME_CPU && NCNN_FMA
+    if (ncnn::cpu_support_x86_fma())
     {
-        layer_creator = layer_registry_avx2[index].creator;
+        layer_creator = layer_registry_fma[index].creator;
     }
     else
-#endif// NCNN_RUNTIME_CPU && NCNN_AVX2
+#endif// NCNN_RUNTIME_CPU && NCNN_FMA
 #if NCNN_RUNTIME_CPU && NCNN_AVX
     if (ncnn::cpu_support_x86_avx())
     {
diff --git a/src/layer/x86/convolution_3x3_pack8to1_int8.h b/src/layer/x86/convolution_3x3_pack8to1_int8.h
index f7b8b42f1..e9de884c7 100644
--- a/src/layer/x86/convolution_3x3_pack8to1_int8.h
+++ b/src/layer/x86/convolution_3x3_pack8to1_int8.h
@@ -32,6 +32,15 @@ static void conv3x3s1_winograd42_transform_kernel_pack8to1_int8_sse(const Mat& k
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void conv3x3s1_winograd42_transform_kernel_pack8to1_int8_sse_avx2(const Mat& kernel, Mat& kernel_tm_pack8to1, int inch, int outch, const Option& opt);
+        conv3x3s1_winograd42_transform_kernel_pack8to1_int8_sse_avx2(kernel, kernel_tm_pack8to1, inch, outch, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
@@ -190,6 +199,15 @@ static void conv3x3s1_winograd42_pack8to1_int8_sse(const Mat& bottom_blob, Mat&
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void conv3x3s1_winograd42_pack8to1_int8_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt);
+        conv3x3s1_winograd42_pack8to1_int8_sse_avx2(bottom_blob, top_blob, kernel_tm, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
diff --git a/src/layer/x86/convolution_3x3_pack8to4_int8.h b/src/layer/x86/convolution_3x3_pack8to4_int8.h
index 6a9647d32..01106f9a7 100644
--- a/src/layer/x86/convolution_3x3_pack8to4_int8.h
+++ b/src/layer/x86/convolution_3x3_pack8to4_int8.h
@@ -32,6 +32,15 @@ static void conv3x3s1_winograd42_transform_kernel_pack8to4_int8_sse(const Mat& k
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void conv3x3s1_winograd42_transform_kernel_pack8to4_int8_sse_avx2(const Mat& kernel, Mat& kernel_tm_pack8, int inch, int outch, const Option& opt);
+        conv3x3s1_winograd42_transform_kernel_pack8to4_int8_sse_avx2(kernel, kernel_tm_pack8, inch, outch, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
@@ -174,6 +183,15 @@ static void conv3x3s1_winograd42_pack8to4_int8_sse(const Mat& bottom_blob, Mat&
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void conv3x3s1_winograd42_pack8to4_int8_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt);
+        conv3x3s1_winograd42_pack8to4_int8_sse_avx2(bottom_blob, top_blob, kernel_tm, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
diff --git a/src/layer/x86/convolution_sgemm_int8.h b/src/layer/x86/convolution_sgemm_int8.h
index a3c63f119..322ee0993 100644
--- a/src/layer/x86/convolution_sgemm_int8.h
+++ b/src/layer/x86/convolution_sgemm_int8.h
@@ -32,6 +32,15 @@ static void im2col_sgemm_int8_sse(const Mat& bottom_im2col, Mat& top_blob, const
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void im2col_sgemm_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+        im2col_sgemm_int8_sse_avx2(bottom_im2col, top_blob, kernel, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
diff --git a/src/layer/x86/convolution_sgemm_pack1to4_int8.h b/src/layer/x86/convolution_sgemm_pack1to4_int8.h
index 8e7399398..a9dd0b160 100644
--- a/src/layer/x86/convolution_sgemm_pack1to4_int8.h
+++ b/src/layer/x86/convolution_sgemm_pack1to4_int8.h
@@ -32,6 +32,15 @@ static void im2col_sgemm_pack1to4_int8_sse(const Mat& bottom_im2col, Mat& top_bl
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void im2col_sgemm_pack1to4_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+        im2col_sgemm_pack1to4_int8_sse_avx2(bottom_im2col, top_blob, kernel, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
diff --git a/src/layer/x86/convolution_sgemm_pack8to1_int8.h b/src/layer/x86/convolution_sgemm_pack8to1_int8.h
index d01e72def..5662266ad 100644
--- a/src/layer/x86/convolution_sgemm_pack8to1_int8.h
+++ b/src/layer/x86/convolution_sgemm_pack8to1_int8.h
@@ -32,6 +32,15 @@ static void im2col_sgemm_pack8to1_int8_sse(const Mat& bottom_im2col, Mat& top_bl
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void im2col_sgemm_pack8to1_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+        im2col_sgemm_pack8to1_int8_sse_avx2(bottom_im2col, top_blob, kernel, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
diff --git a/src/layer/x86/convolution_sgemm_pack8to4_int8.h b/src/layer/x86/convolution_sgemm_pack8to4_int8.h
index 42e4b952e..e6304743e 100644
--- a/src/layer/x86/convolution_sgemm_pack8to4_int8.h
+++ b/src/layer/x86/convolution_sgemm_pack8to4_int8.h
@@ -32,6 +32,15 @@ static void im2col_sgemm_pack8to4_int8_sse(const Mat& bottom_im2col, Mat& top_bl
     }
 #endif
 
+#if NCNN_AVX2 && __AVX__ && !__AVX2__
+    if (ncnn::cpu_support_x86_avx2())
+    {
+        extern void im2col_sgemm_pack8to4_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+        im2col_sgemm_pack8to4_int8_sse_avx2(bottom_im2col, top_blob, kernel, opt);
+        return;
+    }
+#endif
+
 #if NCNN_XOP && __SSE2__ && !__XOP__
     if (ncnn::cpu_support_x86_xop())
     {
diff --git a/src/layer/x86/convolution_x86_avx2.cpp b/src/layer/x86/convolution_x86_avx2.cpp
new file mode 100644
index 000000000..d2972a1ac
--- /dev/null
+++ b/src/layer/x86/convolution_x86_avx2.cpp
@@ -0,0 +1,74 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "convolution_x86_avx2.h"
+
+#include "cpu.h"
+#include "mat.h"
+#include "x86_usability.h"
+
+namespace ncnn {
+
+#include "convolution_sgemm_int8.h"
+#include "convolution_sgemm_pack1to4_int8.h"
+#include "convolution_sgemm_pack8to1_int8.h"
+#include "convolution_sgemm_pack8to4_int8.h"
+#include "convolution_3x3_pack8to1_int8.h"
+#include "convolution_3x3_pack8to4_int8.h"
+
+// pack1
+void im2col_sgemm_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_int8_sse(bottom_im2col, top_blob, kernel, opt);
+}
+
+// pack1to4
+void im2col_sgemm_pack1to4_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_pack1to4_int8_sse(bottom_im2col, top_blob, kernel, opt);
+}
+
+// pack8to1
+void im2col_sgemm_pack8to1_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_pack8to1_int8_sse(bottom_im2col, top_blob, kernel, opt);
+}
+
+void conv3x3s1_winograd42_transform_kernel_pack8to1_int8_sse_avx2(const Mat& kernel, Mat& kernel_tm, int inch, int outch, const Option& opt)
+{
+    conv3x3s1_winograd42_transform_kernel_pack8to1_int8_sse(kernel, kernel_tm, inch, outch, opt);
+}
+
+void conv3x3s1_winograd42_pack8to1_int8_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    conv3x3s1_winograd42_pack8to1_int8_sse(bottom_blob, top_blob, kernel, opt);
+}
+
+// pack8to4
+void im2col_sgemm_pack8to4_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    im2col_sgemm_pack8to4_int8_sse(bottom_im2col, top_blob, kernel, opt);
+}
+
+void conv3x3s1_winograd42_transform_kernel_pack8to4_int8_sse_avx2(const Mat& kernel, Mat& kernel_tm, int inch, int outch, const Option& opt)
+{
+    conv3x3s1_winograd42_transform_kernel_pack8to4_int8_sse(kernel, kernel_tm, inch, outch, opt);
+}
+
+void conv3x3s1_winograd42_pack8to4_int8_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt)
+{
+    conv3x3s1_winograd42_pack8to4_int8_sse(bottom_blob, top_blob, kernel, opt);
+}
+
+} // namespace ncnn
diff --git a/src/layer/x86/convolution_x86_avx2.h b/src/layer/x86/convolution_x86_avx2.h
new file mode 100644
index 000000000..32bae2170
--- /dev/null
+++ b/src/layer/x86/convolution_x86_avx2.h
@@ -0,0 +1,37 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
+//
+// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
+// in compliance with the License. You may obtain a copy of the License at
+//
+// https://opensource.org/licenses/BSD-3-Clause
+//
+// Unless required by applicable law or agreed to in writing, software distributed
+// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
+// CONDITIONS OF ANY KIND, either express or implied. See the License for the
+// specific language governing permissions and limitations under the License.
+
+#include "mat.h"
+
+namespace ncnn {
+
+// pack1
+void im2col_sgemm_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+
+// pack1to4
+void im2col_sgemm_pack1to4_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+
+// pack8to1
+void im2col_sgemm_pack8to1_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+
+void conv3x3s1_winograd42_transform_kernel_pack8to1_int8_sse_avx2(const Mat& kernel, Mat& kernel_tm, int inch, int outch, const Option& opt);
+void conv3x3s1_winograd42_pack8to1_int8_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt);
+
+// pack8to4
+void im2col_sgemm_pack8to4_int8_sse_avx2(const Mat& bottom_im2col, Mat& top_blob, const Mat& kernel, const Option& opt);
+
+void conv3x3s1_winograd42_transform_kernel_pack8to4_int8_sse_avx2(const Mat& kernel, Mat& kernel_tm, int inch, int outch, const Option& opt);
+void conv3x3s1_winograd42_pack8to4_int8_sse_avx2(const Mat& bottom_blob, Mat& top_blob, const Mat& kernel, const Option& opt);
+
+} // namespace ncnn
diff --git a/src/layer/x86/x86_activation.h b/src/layer/x86/x86_activation.h
index 6c8dce8b0..eece18451 100644
--- a/src/layer/x86/x86_activation.h
+++ b/src/layer/x86/x86_activation.h
@@ -131,7 +131,7 @@ static NCNN_FORCEINLINE __m256 tanh_avx(__m256 inputs)
 {
     const __m256 one = _mm256_set1_ps(1.0f);
     const __m256 two = _mm256_set1_ps(2.0f);
-#if __AVX2__
+#if __FMA__
     return _mm256_fmsub_ps(sigmoid_avx(_mm256_mul_ps(inputs, two)), two, one);
 #else
     return _mm256_sub_ps(_mm256_mul_ps(sigmoid_avx(_mm256_mul_ps(inputs, two)), two), one);
diff --git a/src/layer/x86/x86_usability.h b/src/layer/x86/x86_usability.h
index 635a0843f..dceebc67d 100644
--- a/src/layer/x86/x86_usability.h
+++ b/src/layer/x86/x86_usability.h
@@ -147,7 +147,7 @@ static NCNN_FORCEINLINE __m128i float2int8_sse(const __m128& _v0, const __m128&
     return _v8;
 }
 
-#ifndef __AVX2__
+#ifndef __FMA__
 
 static NCNN_FORCEINLINE __m128 _mm_comp_fmadd_ps(__m128 _a, const __m128 _b, const __m128 _c)
 {
@@ -156,7 +156,7 @@ static NCNN_FORCEINLINE __m128 _mm_comp_fmadd_ps(__m128 _a, const __m128 _b, con
 #endif
 
 #if __AVX__
-#ifndef __AVX2__
+#ifndef __FMA__
 static NCNN_FORCEINLINE __m256 _mm256_comp_fmadd_ps(__m256 _a, const __m256 _b, const __m256 _c)
 {
     return _mm256_add_ps(_mm256_mul_ps(_a, _b), _c);
diff --git a/src/layer_registry.h.in b/src/layer_registry.h.in
index e3add4f93..e7d85ad2b 100644
--- a/src/layer_registry.h.in
+++ b/src/layer_registry.h.in
@@ -11,11 +11,11 @@ static const layer_registry_entry layer_registry_avx512[] = {
 @layer_registry_avx512@
 };
 #endif // NCNN_RUNTIME_CPU && NCNN_AVX512
-#if NCNN_RUNTIME_CPU && NCNN_AVX2
-static const layer_registry_entry layer_registry_avx2[] = {
-@layer_registry_avx2@
+#if NCNN_RUNTIME_CPU && NCNN_FMA
+static const layer_registry_entry layer_registry_fma[] = {
+@layer_registry_fma@
 };
-#endif // NCNN_RUNTIME_CPU && NCNN_AVX2
+#endif // NCNN_RUNTIME_CPU && NCNN_FMA
 #if NCNN_RUNTIME_CPU && NCNN_AVX
 static const layer_registry_entry layer_registry_avx[] = {
 @layer_registry_avx@
diff --git a/src/platform.h.in b/src/platform.h.in
index fc3091966..3f62069b7 100644
--- a/src/platform.h.in
+++ b/src/platform.h.in
@@ -32,6 +32,7 @@
 #cmakedefine01 NCNN_RUNTIME_CPU
 #cmakedefine01 NCNN_AVX
 #cmakedefine01 NCNN_XOP
+#cmakedefine01 NCNN_FMA
 #cmakedefine01 NCNN_AVX2
 #cmakedefine01 NCNN_AVXVNNI
 #cmakedefine01 NCNN_AVX512