fix armv7 without neon (#3514)

.github/workflows/linux-arm-cpu-gcc.yml

@@ -135,3 +135,57 @@ jobs:
         export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
         cd build-noint8
         TESTS_EXECUTABLE_LOADER=qemu-arm TESTS_EXECUTABLE_LOADER_ARGUMENTS="-L;/usr/arm-linux-gnueabihf" ctest --output-on-failure -j 2
+
+  linux-gcc-armhf-vfpv3-d16:
+    runs-on: ubuntu-20.04
+    steps:
+    - uses: actions/checkout@v2
+
+    - name: cache-qemu
+      id: cache-qemu
+      uses: actions/cache@v2.1.7
+      with:
+        path: qemu-install
+        key: qemu-arm-install-1
+    - name: checkout-qemu
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      uses: actions/checkout@v2
+      with:
+        repository: qemu/qemu
+        path: qemu
+        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+    - name: qemu
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        cd qemu
+        ./configure --prefix=install --target-list=arm-linux-user --disable-system
+        make -j2
+        make install
+        cp -r arm-linux-user/install $GITHUB_WORKSPACE/qemu-install
+
+    - name: arm-gnu-toolchain
+      run: |
+        sudo apt-get update
+        sudo apt-get install g++-arm-linux-gnueabihf
+
+    - name: build
+      run: |
+        mkdir build && cd build
+        cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf-vfpv3-d16.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON ..
+        cmake --build . -j 2
+    - name: test
+      run: |
+        export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
+        cd build
+        TESTS_EXECUTABLE_LOADER=qemu-arm TESTS_EXECUTABLE_LOADER_ARGUMENTS="-L;/usr/arm-linux-gnueabihf" ctest --output-on-failure -j 2
+
+    - name: build-noint8
+      run: |
+        mkdir build-noint8 && cd build-noint8
+        cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/arm-linux-gnueabihf-vfpv3-d16.toolchain.cmake -DNCNN_BUILD_TOOLS=OFF -DNCNN_BUILD_EXAMPLES=OFF -DNCNN_BUILD_TESTS=ON -DNCNN_INT8=OFF ..
+        cmake --build . -j 2
+    - name: test-noint8
+      run: |
+        export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
+        cd build-noint8
+        TESTS_EXECUTABLE_LOADER=qemu-arm TESTS_EXECUTABLE_LOADER_ARGUMENTS="-L;/usr/arm-linux-gnueabihf" ctest --output-on-failure -j 2

build.sh

@@ -19,7 +19,7 @@ popd
 ##### android armv7 without neon
 mkdir -p build-android-armv7-without-neon
 pushd build-android-armv7-without-neon
-cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI="armeabi-v7a" -DANDROID_PLATFORM=android-19 ..
+cmake -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI="armeabi-v7a" -DANDROID_ARM_NEON=OFF -DANDROID_PLATFORM=android-19 ..
 make -j4
 make install
 popd

src/layer/arm/convolution1d_arm.cpp

@@ -147,7 +147,13 @@ int Convolution1D_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
     w = bottom_blob_bordered.w;
     h = bottom_blob_bordered.h;
 
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     const int outw = (w - kernel_extent_w) / stride_w + 1;
@@ -524,7 +530,7 @@ int Convolution1D_arm::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, cons
     w = bottom_blob_bordered.w;
     h = bottom_blob_bordered.h;
 
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = (opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     const int outw = (w - kernel_extent_w) / stride_w + 1;
@@ -1196,7 +1202,6 @@ int Convolution1D_arm::create_pipeline_bf16s(const Option& opt)
 
     int elempack = 1;
     int out_elempack = 1;
-
 #if __ARM_NEON
     if (opt.use_packing_layout)
     {
@@ -1258,7 +1263,13 @@ int Convolution1D_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_blob, cons
     w = bottom_blob_bordered.w;
     h = bottom_blob_bordered.h;
 
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     const int outw = (w - kernel_extent_w) / stride_w + 1;

src/layer/arm/convolution_arm.cpp

@@ -223,8 +223,15 @@ int Convolution_arm::create_pipeline(const Option& opt)
     const int maxk = kernel_w * kernel_h;
     const int num_input = weight_data_size / maxk / num_output;
 
-    int elempack = (support_packing && opt.use_packing_layout && num_input % 4 == 0) ? 4 : 1;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int elempack = 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        elempack = num_input % 4 == 0 ? 4 : 1;
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
 
 #if __ARM_NEON
     // pack4
@@ -501,7 +508,13 @@ int Convolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option
 
     int outw = (w - kernel_extent_w) / stride_w + 1;
     int outh = (h - kernel_extent_h) / stride_h + 1;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
@@ -1384,7 +1397,7 @@ int Convolution_arm::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const
 
     int outw = (w - kernel_extent_w) / stride_w + 1;
     int outh = (h - kernel_extent_h) / stride_h + 1;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = (opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
@@ -1908,8 +1921,15 @@ int Convolution_arm::create_pipeline_bf16s(const Option& opt)
     const int maxk = kernel_w * kernel_h;
     const int num_input = weight_data_size / maxk / num_output;
 
-    int elempack = (support_packing && opt.use_packing_layout && num_input % 4 == 0) ? 4 : 1;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int elempack = 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        elempack = num_input % 4 == 0 ? 4 : 1;
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
 
 #if __ARM_NEON
     // pack4
@@ -2055,7 +2075,13 @@ int Convolution_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const
 
     int outw = (w - kernel_extent_w) / stride_w + 1;
     int outh = (h - kernel_extent_h) / stride_h + 1;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);

src/layer/arm/convolutiondepthwise_arm.cpp

@@ -535,8 +535,15 @@ int ConvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, con
     const int channels_g = channels * elempack / group;
     const int num_output_g = num_output / group;
 
-    int g_elempack = (support_packing && opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
-    int out_g_elempack = (support_packing && opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
+    int g_elempack = 1;
+    int out_g_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        g_elempack = channels_g % 4 == 0 ? 4 : 1;
+        out_g_elempack = num_output_g % 4 == 0 ? 4 : 1;
+    }
+#endif
 
     // unpacking
     Mat bottom_blob_bordered_unpacked = bottom_blob_bordered;
@@ -856,8 +863,8 @@ int ConvolutionDepthWise_arm::forward_fp16s(const Mat& bottom_blob, Mat& top_blo
     const int channels_g = channels * elempack / group;
     const int num_output_g = num_output / group;
 
-    int g_elempack = (support_packing && opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
-    int out_g_elempack = (support_packing && opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
+    int g_elempack = (opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
+    int out_g_elempack = (opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
 
     // unpacking
     Mat bottom_blob_bordered_unpacked = bottom_blob_bordered;
@@ -1491,8 +1498,15 @@ int ConvolutionDepthWise_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_blo
     const int channels_g = channels * elempack / group;
     const int num_output_g = num_output / group;
 
-    int g_elempack = (support_packing && opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
-    int out_g_elempack = (support_packing && opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
+    int g_elempack = 1;
+    int out_g_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        g_elempack = channels_g % 4 == 0 ? 4 : 1;
+        out_g_elempack = num_output_g % 4 == 0 ? 4 : 1;
+    }
+#endif
 
     // unpacking
     Mat bottom_blob_bordered_unpacked = bottom_blob_bordered;

src/layer/arm/deconvolution_arm.cpp

@@ -120,8 +120,15 @@ int Deconvolution_arm::create_pipeline(const Option& opt)
         }
     }
 
-    int elempack = (support_packing && opt.use_packing_layout && num_input % 4 == 0) ? 4 : 1;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int elempack = 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        elempack = num_input % 4 == 0 ? 4 : 1;
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
 
 #if __ARM_NEON
     // pack4
@@ -333,7 +340,13 @@ int Deconvolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Opti
 
     int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
     int outh = (h - 1) * stride_h + kernel_extent_h + output_pad_bottom;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     Mat top_blob_bordered;
@@ -1911,8 +1924,15 @@ int Deconvolution_arm::create_pipeline_bf16s(const Option& opt)
     const int maxk = kernel_w * kernel_h;
     const int num_input = weight_data_size / maxk / num_output;
 
-    int elempack = opt.use_packing_layout && num_input % 4 == 0 ? 4 : 1;
-    int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
+    int elempack = 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        elempack = num_input % 4 == 0 ? 4 : 1;
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
 
     Mat weight_data_transposed(weight_data.w);
     {
@@ -1985,7 +2005,13 @@ int Deconvolution_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_blob, cons
 
     int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
     int outh = (h - 1) * stride_h + kernel_extent_h + output_pad_bottom;
-    int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     Mat top_blob_bordered;

src/layer/arm/deconvolutiondepthwise_arm.cpp

@@ -47,7 +47,13 @@ int DeconvolutionDepthWise_arm::create_pipeline(const Option& opt)
     // depth-wise
     if (channels == group && group == num_output)
     {
-        int elempack = (support_packing && opt.use_packing_layout && channels % 4 == 0) ? 4 : 1;
+        int elempack = 1;
+#if __ARM_NEON
+        if (opt.use_packing_layout)
+        {
+            elempack = channels % 4 == 0 ? 4 : 1;
+        }
+#endif
 
         Mat weight_data_transposed(weight_data.w);
         {
@@ -253,7 +259,13 @@ int DeconvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, c
 
     int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
     int outh = (h - 1) * stride_h + kernel_extent_h + output_pad_bottom;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     Mat top_blob_bordered;
@@ -427,8 +439,15 @@ int DeconvolutionDepthWise_arm::forward(const Mat& bottom_blob, Mat& top_blob, c
         const int channels_g = channels * elempack / group;
         const int num_output_g = num_output / group;
 
-        int g_elempack = (support_packing && opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
-        int out_g_elempack = (support_packing && opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
+        int g_elempack = 1;
+        int out_g_elempack = 1;
+#if __ARM_NEON
+        if (opt.use_packing_layout)
+        {
+            g_elempack = channels_g % 4 == 0 ? 4 : 1;
+            out_g_elempack = num_output_g % 4 == 0 ? 4 : 1;
+        }
+#endif
 
         // unpacking
         Mat bottom_blob_unpacked = bottom_blob;
@@ -493,7 +512,7 @@ int DeconvolutionDepthWise_arm::forward_fp16s(const Mat& bottom_blob, Mat& top_b
 
     int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
     int outh = (h - 1) * stride_h + kernel_extent_h + output_pad_bottom;
-    int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
+    int out_elempack = (opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     Mat top_blob_bordered;
@@ -648,8 +667,8 @@ int DeconvolutionDepthWise_arm::forward_fp16s(const Mat& bottom_blob, Mat& top_b
         const int channels_g = channels * elempack / group;
         const int num_output_g = num_output / group;
 
-        int g_elempack = (support_packing && opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
-        int out_g_elempack = (support_packing && opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
+        int g_elempack = (opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
+        int out_g_elempack = (opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
 
         // unpacking
         Mat bottom_blob_unpacked = bottom_blob;
@@ -1008,7 +1027,13 @@ int DeconvolutionDepthWise_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_b
 
     int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
     int outh = (h - 1) * stride_h + kernel_extent_h + output_pad_bottom;
-    int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     Mat top_blob_bordered;
@@ -1182,8 +1207,15 @@ int DeconvolutionDepthWise_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_b
         const int channels_g = channels * elempack / group;
         const int num_output_g = num_output / group;
 
-        int g_elempack = (support_packing && opt.use_packing_layout && channels_g % 4 == 0) ? 4 : 1;
-        int out_g_elempack = (support_packing && opt.use_packing_layout && num_output_g % 4 == 0) ? 4 : 1;
+        int g_elempack = 1;
+        int out_g_elempack = 1;
+#if __ARM_NEON
+        if (opt.use_packing_layout)
+        {
+            g_elempack = channels_g % 4 == 0 ? 4 : 1;
+            out_g_elempack = num_output_g % 4 == 0 ? 4 : 1;
+        }
+#endif
 
         // unpacking
         Mat bottom_blob_unpacked = bottom_blob;

src/layer/arm/innerproduct_arm.cpp

@@ -1543,7 +1543,13 @@ int InnerProduct_arm::create_pipeline_bf16s(const Option& opt)
 {
     const int num_input = weight_data_size / num_output;
 
-    int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif // __ARM_NEON
 
     // src = inch-outch
     // dst = pb-inch-outch/pb
@@ -1763,7 +1769,13 @@ int InnerProduct_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const
     size_t elemsize = bottom_blob_flattened.elemsize;
     int elempack = bottom_blob_flattened.elempack;
 
-    int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
+    int out_elempack = 1;
+#if __ARM_NEON
+    if (opt.use_packing_layout)
+    {
+        out_elempack = num_output % 4 == 0 ? 4 : 1;
+    }
+#endif // __ARM_NEON
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     top_blob.create(num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);

src/layer/arm/pixelshuffle_arm.cpp

@@ -61,10 +61,12 @@ int PixelShuffle_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Optio
     int outc = channels * elempack / (upscale_factor * upscale_factor);
 
     int out_elempack = 1;
+#if __ARM_NEON
     if (opt.use_packing_layout)
     {
         out_elempack = outc % 4 == 0 ? 4 : 1;
     }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     if (upscale_factor != 2 || mode != 0)
@@ -234,6 +236,7 @@ int PixelShuffle_arm::forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob,
     int outc = channels * elempack / (upscale_factor * upscale_factor);
 
     int out_elempack = 1;
+#if __ARM_NEON
     if (opt.use_packing_layout)
     {
 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
@@ -242,6 +245,7 @@ int PixelShuffle_arm::forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob,
         out_elempack = outc % 4 == 0 ? 4 : 1;
 #endif
     }
+#endif
     size_t out_elemsize = elemsize / elempack * out_elempack;
 
     if (upscale_factor != 2 || mode != 0)

src/layer/arm/slice_arm.cpp

@@ -64,7 +64,13 @@ int Slice_arm::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& t
                 slice = (w - q) / (top_blobs.size() - i);
             }
 
-            int out_elempack = opt.use_packing_layout && slice % 4 == 0 ? 4 : 1;
+            int out_elempack = 1;
+#if __ARM_NEON
+            if (opt.use_packing_layout)
+            {
+                out_elempack = slice % 4 == 0 ? 4 : 1;
+            }
+#endif
             size_t out_elemsize = elemsize / elempack * out_elempack;
 
             Mat& top_blob = top_blobs[i];
@@ -95,7 +101,13 @@ int Slice_arm::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& t
                 slice = (h - q) / (top_blobs.size() - i);
             }
 
-            int out_elempack = opt.use_packing_layout && slice % 4 == 0 ? 4 : 1;
+            int out_elempack = 1;
+#if __ARM_NEON
+            if (opt.use_packing_layout)
+            {
+                out_elempack = slice % 4 == 0 ? 4 : 1;
+            }
+#endif
             size_t out_elemsize = elemsize / elempack * out_elempack;
 
             Mat& top_blob = top_blobs[i];
@@ -216,7 +228,13 @@ int Slice_arm::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& t
                 slice = (channels - q) / (top_blobs.size() - i);
             }
 
-            int out_elempack = opt.use_packing_layout && slice % 4 == 0 ? 4 : 1;
+            int out_elempack = 1;
+#if __ARM_NEON
+            if (opt.use_packing_layout)
+            {
+                out_elempack = slice % 4 == 0 ? 4 : 1;
+            }
+#endif
             size_t out_elemsize = elemsize / elempack * out_elempack;
 
             Mat& top_blob = top_blobs[i];
@@ -398,6 +416,7 @@ int Slice_arm::forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::ve
             }
 
             int out_elempack = 1;
+#if __ARM_NEON
             if (opt.use_packing_layout)
             {
 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
@@ -406,6 +425,7 @@ int Slice_arm::forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::ve
                 out_elempack = slice % 4 == 0 ? 4 : 1;
 #endif
             }
+#endif
             size_t out_elemsize = elemsize / elempack * out_elempack;
 
             Mat& top_blob = top_blobs[i];
@@ -437,6 +457,7 @@ int Slice_arm::forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::ve
             }
 
             int out_elempack = 1;
+#if __ARM_NEON
             if (opt.use_packing_layout)
             {
 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
@@ -445,6 +466,7 @@ int Slice_arm::forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::ve
                 out_elempack = slice % 4 == 0 ? 4 : 1;
 #endif
             }
+#endif
             size_t out_elemsize = elemsize / elempack * out_elempack;
 
             Mat& top_blob = top_blobs[i];
@@ -628,6 +650,7 @@ int Slice_arm::forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::ve
             }
 
             int out_elempack = 1;
+#if __ARM_NEON
             if (opt.use_packing_layout)
             {
 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
@@ -636,6 +659,7 @@ int Slice_arm::forward_bf16s_fp16s(const std::vector<Mat>& bottom_blobs, std::ve
                 out_elempack = slice % 4 == 0 ? 4 : 1;
 #endif
             }
+#endif
             size_t out_elemsize = elemsize / elempack * out_elempack;
 
             Mat& top_blob = top_blobs[i];

src/layer/convolutiondepthwise.cpp

@@ -243,11 +243,9 @@ static int convolutiondepthwise(const Mat& bottom_blob, Mat& top_blob, const Mat
                 float* outptr = top_blob.channel(g * outch_g + p);
                 const float* weight_data_ptr = (const float*)weight_data + maxk * inch_g * outch_g * g;
 
-#if NCNN_SIMPLEOMP
                 // shadowed variable for less openmp task args
                 const int outw = top_blob.w;
                 const int outh = top_blob.h;
-#endif
 
                 for (int i = 0; i < outh; i++)
                 {

toolchains/arm-linux-gnueabihf-vfpv3-d16.toolchain.cmake

@@ -0,0 +1,16 @@
+set(CMAKE_SYSTEM_NAME Linux)
+set(CMAKE_SYSTEM_PROCESSOR arm)
+
+set(CMAKE_C_COMPILER "arm-linux-gnueabihf-gcc")
+set(CMAKE_CXX_COMPILER "arm-linux-gnueabihf-g++")
+
+set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
+set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY ONLY)
+set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE ONLY)
+
+set(CMAKE_C_FLAGS "-march=armv7-a -mfloat-abi=hard -mfpu=vfpv3-d16")
+set(CMAKE_CXX_FLAGS "-march=armv7-a -mfloat-abi=hard -mfpu=vfpv3-d16")
+
+# cache flags
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}" CACHE STRING "c flags")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS}" CACHE STRING "c++ flags")