diff --git a/.github/workflows/linux-aarch64-cpu-gcc.yml b/.github/workflows/linux-aarch64-cpu-gcc.yml
index d6e387886..f021a1d28 100644
--- a/.github/workflows/linux-aarch64-cpu-gcc.yml
+++ b/.github/workflows/linux-aarch64-cpu-gcc.yml
@@ -39,22 +39,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-aarch64-install-1
+        key: qemu-aarch64-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=aarch64-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=aarch64-linux-user --disable-system
         make -j2
         make install
-        cp -r aarch64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: aarch64-gnu-toolchain
       run: |
@@ -93,22 +97,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-aarch64-install-1
+        key: qemu-aarch64-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=aarch64-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=aarch64-linux-user --disable-system
         make -j2
         make install
-        cp -r aarch64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: aarch64-gnu-toolchain
       run: |
diff --git a/.github/workflows/linux-arm-cpu-gcc.yml b/.github/workflows/linux-arm-cpu-gcc.yml
index 772cdd1f9..6613db561 100644
--- a/.github/workflows/linux-arm-cpu-gcc.yml
+++ b/.github/workflows/linux-arm-cpu-gcc.yml
@@ -41,22 +41,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-arm-install-1
+        key: qemu-arm-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=arm-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-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: |
@@ -95,22 +99,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-arm-install-1
+        key: qemu-arm-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=arm-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-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: |
@@ -149,22 +157,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-arm-install-1
+        key: qemu-arm-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=arm-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-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: |
diff --git a/.github/workflows/linux-mips-cpu-gcc.yml b/.github/workflows/linux-mips-cpu-gcc.yml
index fa668b987..7e79c73f9 100644
--- a/.github/workflows/linux-mips-cpu-gcc.yml
+++ b/.github/workflows/linux-mips-cpu-gcc.yml
@@ -41,22 +41,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-mipsel-install-1
+        key: qemu-mipsel-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=mipsel-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=mipsel-linux-user --disable-system
         make -j2
         make install
-        cp -r mipsel-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: mipsel-gnu-toolchain
       run: |
@@ -84,22 +88,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-mipsel-install-1
+        key: qemu-mipsel-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=mipsel-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=mipsel-linux-user --disable-system
         make -j2
         make install
-        cp -r mipsel-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: mipsisa32r6el-gnu-toolchain
       run: |
diff --git a/.github/workflows/linux-mips64-cpu-gcc.yml b/.github/workflows/linux-mips64-cpu-gcc.yml
index 47396aeb1..2c6a06d9c 100644
--- a/.github/workflows/linux-mips64-cpu-gcc.yml
+++ b/.github/workflows/linux-mips64-cpu-gcc.yml
@@ -41,22 +41,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-mips64el-install-1
+        key: qemu-mips64el-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=mips64el-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=mips64el-linux-user --disable-system
         make -j2
         make install
-        cp -r mips64el-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: mips64el-gnuabi64-toolchain
       run: |
@@ -84,22 +88,38 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-mips64el-install-1
+        key: qemu-mips64el-install-20220502-3
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=mips64el-linux-user --disable-system
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0001-target-mips-Fix-SAT_S-trans-helper.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0002-target-mips-Fix-df_extract_val-and-df_extract_df-dfe.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0003-target-mips-Fix-msa-checking-condition-in-trans_msa_.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0004-target-mips-Do-not-treat-msa-INSERT-as-NOP-when-wd-i.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0005-target-mips-Fix-FTRUNC_S-and-FTRUNC_U-trans-helper.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0006-target-mips-Fix-store-adress-of-high-64bit-in-helper.patch
+        patch -p1 -i 0001-target-mips-Fix-SAT_S-trans-helper.patch
+        patch -p1 -i 0002-target-mips-Fix-df_extract_val-and-df_extract_df-dfe.patch
+        patch -p1 -i 0003-target-mips-Fix-msa-checking-condition-in-trans_msa_.patch
+        patch -p1 -i 0004-target-mips-Do-not-treat-msa-INSERT-as-NOP-when-wd-i.patch
+        patch -p1 -i 0005-target-mips-Fix-FTRUNC_S-and-FTRUNC_U-trans-helper.patch
+        patch -p1 -i 0006-target-mips-Fix-store-adress-of-high-64bit-in-helper.patch
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=mips64el-linux-user --disable-system
         make -j2
         make install
-        cp -r mips64el-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: mipsisa64r6el-gnuabi64-toolchain
       run: |
diff --git a/.github/workflows/linux-ppc64-cpu-gcc.yml b/.github/workflows/linux-ppc64-cpu-gcc.yml
index acaca922f..398f09bcd 100644
--- a/.github/workflows/linux-ppc64-cpu-gcc.yml
+++ b/.github/workflows/linux-ppc64-cpu-gcc.yml
@@ -37,22 +37,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-ppc64le-install-1
+        key: qemu-ppc64le-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=ppc64le-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=ppc64le-linux-user --disable-system
         make -j2
         make install
-        cp -r ppc64le-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: powerpc64le-gnu-toolchain
       run: |
diff --git a/.github/workflows/linux-riscv64-cpu-gcc.yml b/.github/workflows/linux-riscv64-cpu-gcc.yml
index a00b50d6f..902d63f1a 100644
--- a/.github/workflows/linux-riscv64-cpu-gcc.yml
+++ b/.github/workflows/linux-riscv64-cpu-gcc.yml
@@ -41,22 +41,28 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-riscv64-install-1
+        key: qemu-riscv64-install-20220502-3
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=riscv64-linux-user --disable-system
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        patch -p1 -i 0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system
         make -j2
         make install
-        cp -r riscv64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: riscv64-gnu-toolchain
       run: |
@@ -84,37 +90,28 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-riscv64-rvv-install-20210610
-
+        key: qemu-riscv64-install-20220502-3
     - name: install-qemu-build-deps
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         sudo apt-get update
         sudo apt-get install autoconf automake autotools-dev ninja-build
-
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
-        repository: sifive/qemu
+        repository: qemu/qemu
         path: qemu
-        ref: 7a3e8e23b4cf1422ec48e9d4b4009337a05a635d
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
-        wget https://download.qemu.org/qemu-6.0.0.tar.xz
-        tar -xf qemu-6.0.0.tar.xz
-        cp -r qemu/fpu qemu-6.0.0/
-        cp -r qemu/include/fpu qemu-6.0.0/include/
-        cp -r qemu/target/riscv qemu-6.0.0/target/
-        cp -r qemu/linux-user/elfload.c qemu-6.0.0/linux-user/
-        cd qemu-6.0.0
-        mkdir build
-        cd build
-        ../configure --prefix=`pwd`/install --target-list=riscv64-linux-user --disable-system
+        cd qemu
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        patch -p1 -i 0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system
         make -j2
         make install
-        cp -r install $GITHUB_WORKSPACE/qemu-install
 
     - name: cache-riscv
       id: cache-riscv
@@ -167,10 +164,10 @@ jobs:
       run: |
         export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
         cd build
-        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,x-v=true,x-Zfh=true,vlen=256,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
+        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,v=true,Zfh=true,vlen=256,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
 
     - name: test-vlen128
       run: |
         export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
         cd build
-        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,x-v=true,x-Zfh=true,vlen=128,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
+        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,v=true,Zfh=true,vlen=128,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
diff --git a/.github/workflows/test-coverage.yml b/.github/workflows/test-coverage.yml
index 3306939e6..e394bcb1a 100644
--- a/.github/workflows/test-coverage.yml
+++ b/.github/workflows/test-coverage.yml
@@ -278,22 +278,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-arm-install-1
+        key: qemu-arm-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=arm-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-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: |
@@ -336,22 +340,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-aarch64-install-1
+        key: qemu-aarch64-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=aarch64-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=aarch64-linux-user --disable-system
         make -j2
         make install
-        cp -r aarch64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: aarch64-gnu-toolchain
       run: |
@@ -394,22 +402,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-aarch64-install-1
+        key: qemu-aarch64-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=aarch64-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=aarch64-linux-user --disable-system
         make -j2
         make install
-        cp -r aarch64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: aarch64-gnu-toolchain
       run: |
@@ -451,22 +463,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-aarch64-install-1
+        key: qemu-aarch64-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=aarch64-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=aarch64-linux-user --disable-system
         make -j2
         make install
-        cp -r aarch64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: aarch64-gnu-toolchain
       run: |
@@ -508,22 +524,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-aarch64-install-1
+        key: qemu-aarch64-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=aarch64-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=aarch64-linux-user --disable-system
         make -j2
         make install
-        cp -r aarch64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: aarch64-gnu-toolchain
       run: |
@@ -565,22 +585,26 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-mipsel-install-1
+        key: qemu-mipsel-install-20220502
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=mipsel-linux-user --disable-system
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=mipsel-linux-user --disable-system
         make -j2
         make install
-        cp -r mipsel-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: mipsisa32r6el-gnu-toolchain
       run: |
@@ -623,22 +647,38 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-mips64el-install-1
+        key: qemu-mips64el-install-20220502-3
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=mips64el-linux-user --disable-system
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0001-target-mips-Fix-SAT_S-trans-helper.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0002-target-mips-Fix-df_extract_val-and-df_extract_df-dfe.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0003-target-mips-Fix-msa-checking-condition-in-trans_msa_.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0004-target-mips-Do-not-treat-msa-INSERT-as-NOP-when-wd-i.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0005-target-mips-Fix-FTRUNC_S-and-FTRUNC_U-trans-helper.patch
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0006-target-mips-Fix-store-adress-of-high-64bit-in-helper.patch
+        patch -p1 -i 0001-target-mips-Fix-SAT_S-trans-helper.patch
+        patch -p1 -i 0002-target-mips-Fix-df_extract_val-and-df_extract_df-dfe.patch
+        patch -p1 -i 0003-target-mips-Fix-msa-checking-condition-in-trans_msa_.patch
+        patch -p1 -i 0004-target-mips-Do-not-treat-msa-INSERT-as-NOP-when-wd-i.patch
+        patch -p1 -i 0005-target-mips-Fix-FTRUNC_S-and-FTRUNC_U-trans-helper.patch
+        patch -p1 -i 0006-target-mips-Fix-store-adress-of-high-64bit-in-helper.patch
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=mips64el-linux-user --disable-system
         make -j2
         make install
-        cp -r mips64el-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: mipsisa64r6el-gnuabi64-toolchain
       run: |
@@ -682,22 +722,28 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-riscv64-install-1
+        key: qemu-riscv64-install-20220502-3
+    - name: install-qemu-build-deps
+      if: steps.cache-qemu.outputs.cache-hit != 'true'
+      run: |
+        sudo apt-get update
+        sudo apt-get install autoconf automake autotools-dev ninja-build
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
         repository: qemu/qemu
         path: qemu
-        ref: 8746309137ba470d1b2e8f5ce86ac228625db940
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         cd qemu
-        ./configure --prefix=install --target-list=riscv64-linux-user --disable-system
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        patch -p1 -i 0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system
         make -j2
         make install
-        cp -r riscv64-linux-user/install $GITHUB_WORKSPACE/qemu-install
 
     - name: riscv64-gnu-toolchain
       run: |
@@ -741,37 +787,28 @@ jobs:
       uses: actions/cache@v3
       with:
         path: qemu-install
-        key: qemu-riscv64-rvv-install-20210610
-
+        key: qemu-riscv64-install-20220502-3
     - name: install-qemu-build-deps
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
         sudo apt-get update
         sudo apt-get install autoconf automake autotools-dev ninja-build
-
     - name: checkout-qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       uses: actions/checkout@v3
       with:
-        repository: sifive/qemu
+        repository: qemu/qemu
         path: qemu
-        ref: 7a3e8e23b4cf1422ec48e9d4b4009337a05a635d
+        ref: f5643914a9e8f79c606a76e6a9d7ea82a3fc3e65
     - name: qemu
       if: steps.cache-qemu.outputs.cache-hit != 'true'
       run: |
-        wget https://download.qemu.org/qemu-6.0.0.tar.xz
-        tar -xf qemu-6.0.0.tar.xz
-        cp -r qemu/fpu qemu-6.0.0/
-        cp -r qemu/include/fpu qemu-6.0.0/include/
-        cp -r qemu/target/riscv qemu-6.0.0/target/
-        cp -r qemu/linux-user/elfload.c qemu-6.0.0/linux-user/
-        cd qemu-6.0.0
-        mkdir build
-        cd build
-        ../configure --prefix=`pwd`/install --target-list=riscv64-linux-user --disable-system
+        cd qemu
+        wget https://raw.githubusercontent.com/nihui/ncnn-assets/master/qemu-patches/0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        patch -p1 -i 0007-linux-user-Expose-risc-v-V-isa-bit-in-get_elf_hwcap.patch
+        ./configure --prefix=$GITHUB_WORKSPACE/qemu-install --target-list=riscv64-linux-user --disable-system
         make -j2
         make install
-        cp -r install $GITHUB_WORKSPACE/qemu-install
 
     - name: cache-riscv
       id: cache-riscv
@@ -824,7 +861,7 @@ jobs:
       run: |
         export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
         cd build
-        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,x-v=true,x-Zfh=true,vlen=128,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
+        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,v=true,Zfh=true,vlen=128,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
 
     - name: lcov-collect-vlen128
       run: |
@@ -844,7 +881,7 @@ jobs:
       run: |
         export PATH=$GITHUB_WORKSPACE/qemu-install/bin:$PATH
         cd build
-        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,x-v=true,x-Zfh=true,vlen=256,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
+        TESTS_EXECUTABLE_LOADER=qemu-riscv64 TESTS_EXECUTABLE_LOADER_ARGUMENTS="-cpu;rv64,v=true,Zfh=true,vlen=256,elen=64,vext_spec=v1.0;-L;$GITHUB_WORKSPACE/rv64gcv-install/sysroot" ctest --output-on-failure -j 2
 
     - name: lcov-collect-vlen256
       run: |
diff --git a/src/layer/arm/convolution_sgemm_bf16s.h b/src/layer/arm/convolution_sgemm_bf16s.h
index 036bf8b15..c304b3969 100644
--- a/src/layer/arm/convolution_sgemm_bf16s.h
+++ b/src/layer/arm/convolution_sgemm_bf16s.h
@@ -56,7 +56,7 @@ static void im2col_sgemm_bf16s_neon(const Mat& bottom_im2col, Mat& top_blob, con
 #else
                     asm volatile(
                         "pld        [%0, #128]          \n"
-                        "vld1.u16   {d0-d1}, [%0 :64]   \n"
+                        "vld1.u16   {d0-d1}, [%0]       \n"
                         "vst1.u16   {d0-d1}, [%1 :64]!  \n"
                         : "=r"(img0),  // %0
                         "=r"(tmpptr) // %1
@@ -104,7 +104,7 @@ static void im2col_sgemm_bf16s_neon(const Mat& bottom_im2col, Mat& top_blob, con
 #else
                     asm volatile(
                         "pld        [%0, #64]       \n"
-                        "vld1.u16   {d0}, [%0 :64]  \n"
+                        "vld1.u16   {d0}, [%0]      \n"
                         "vst1.u16   {d0}, [%1 :64]! \n"
                         : "=r"(img0),  // %0
                         "=r"(tmpptr) // %1
@@ -931,18 +931,18 @@ static void im2col_sgemm_bf16s_neon(const Mat& bottom_im2col, Mat& top_blob, con
                 "pld        [%4, #256]          \n"
                 "vld1.u16   {d12-d15}, [%4 :64]! \n"
 
-                "vshll.u16  q4, d12, #16    \n"
-                "vshll.u16  q5, d13, #16    \n"
-                "vshll.u16  q6, d14, #16    \n"
-                "vshll.u16  q7, d15, #16    \n"
+                "vshll.u16  q4, d12, #16        \n"
+                "vshll.u16  q5, d13, #16        \n"
+                "vshll.u16  q6, d14, #16        \n"
+                "vshll.u16  q7, d15, #16        \n"
 
                 "pld        [%5, #256]          \n"
                 "vld1.u16   {d4-d7}, [%5 :64]!  \n"
 
-                "vshll.u16  q0, d4, #16     \n"
-                "vshll.u16  q1, d5, #16     \n"
-                "vshll.u16  q2, d6, #16     \n"
-                "vshll.u16  q3, d7, #16     \n"
+                "vshll.u16  q0, d4, #16         \n"
+                "vshll.u16  q1, d5, #16         \n"
+                "vshll.u16  q2, d6, #16         \n"
+                "vshll.u16  q3, d7, #16         \n"
 
                 "vmla.f32   q8, q4, d0[0]       \n"
                 "vmla.f32   q10, q4, d0[1]      \n"
@@ -967,10 +967,10 @@ static void im2col_sgemm_bf16s_neon(const Mat& bottom_im2col, Mat& top_blob, con
                 "pld        [%4, #256]          \n"
                 "vld1.u16   {d12-d15}, [%4 :64]! \n"
 
-                "vshll.u16  q4, d12, #16    \n"
-                "vshll.u16  q5, d13, #16    \n"
-                "vshll.u16  q6, d14, #16    \n"
-                "vshll.u16  q7, d15, #16    \n"
+                "vshll.u16  q4, d12, #16        \n"
+                "vshll.u16  q5, d13, #16        \n"
+                "vshll.u16  q6, d14, #16        \n"
+                "vshll.u16  q7, d15, #16        \n"
 
                 "vmla.f32   q8, q4, d4[0]       \n"
                 "vmla.f32   q10, q4, d4[1]      \n"
@@ -1324,18 +1324,18 @@ static void im2col_sgemm_bf16s_neon(const Mat& bottom_im2col, Mat& top_blob, con
                 "pld        [%4, #256]          \n"
                 "vld1.u16   {d12-d15}, [%4 :64]! \n"
 
-                "vshll.u16  q4, d12, #16    \n"
-                "vshll.u16  q5, d13, #16    \n"
-                "vshll.u16  q6, d14, #16    \n"
-                "vshll.u16  q7, d15, #16    \n"
+                "vshll.u16  q4, d12, #16        \n"
+                "vshll.u16  q5, d13, #16        \n"
+                "vshll.u16  q6, d14, #16        \n"
+                "vshll.u16  q7, d15, #16        \n"
 
                 "pld        [%5, #256]          \n"
                 "vld1.u16   {d4-d7}, [%5 :64]!  \n"
 
-                "vshll.u16  q0, d4, #16     \n"
-                "vshll.u16  q1, d5, #16     \n"
-                "vshll.u16  q2, d6, #16     \n"
-                "vshll.u16  q3, d7, #16     \n"
+                "vshll.u16  q0, d4, #16         \n"
+                "vshll.u16  q1, d5, #16         \n"
+                "vshll.u16  q2, d6, #16         \n"
+                "vshll.u16  q3, d7, #16         \n"
 
                 "vmla.f32   q8, q4, d0[0]       \n"
                 "vmla.f32   q9, q4, d0[1]       \n"
diff --git a/src/layer/arm/convolution_sgemm_pack4_bf16s.h b/src/layer/arm/convolution_sgemm_pack4_bf16s.h
index f9c495872..07cb9576b 100644
--- a/src/layer/arm/convolution_sgemm_pack4_bf16s.h
+++ b/src/layer/arm/convolution_sgemm_pack4_bf16s.h
@@ -181,7 +181,7 @@ static void im2col_sgemm_pack4_bf16s_neon(const Mat& bottom_im2col, Mat& top_blo
 #else
                     asm volatile(
                         "pld        [%0, #256]          \n"
-                        "vld1.u16   {d0-d3}, [%0 :128]  \n"
+                        "vld1.u16   {d0-d3}, [%0 :64]   \n"
                         "vst1.u16   {d0-d3}, [%1 :128]! \n"
                         : "=r"(img0),  // %0
                         "=r"(tmpptr) // %1
@@ -227,7 +227,7 @@ static void im2col_sgemm_pack4_bf16s_neon(const Mat& bottom_im2col, Mat& top_blo
 #else
                     asm volatile(
                         "pld        [%0, #128]          \n"
-                        "vld1.u16   {d0-d1}, [%0 :128]  \n"
+                        "vld1.u16   {d0-d1}, [%0 :64]   \n"
                         "vst1.u16   {d0-d1}, [%1 :128]! \n"
                         : "=r"(img0),  // %0
                         "=r"(tmpptr) // %1
@@ -1574,7 +1574,7 @@ static void im2col_sgemm_pack4_bf16s_neon(const Mat& bottom_im2col, Mat& top_blo
                 "0:                         \n"
 
                 "pld        [%2, #128]      \n"
-                "vld1.u16   {d4-d5}, [%2 :128]! \n"
+                "vld1.u16   {d4-d5}, [%2 :64]! \n"
 
                 "vshll.u16  q0, d4, #16     \n"
                 "vshll.u16  q1, d5, #16     \n"
@@ -1606,7 +1606,7 @@ static void im2col_sgemm_pack4_bf16s_neon(const Mat& bottom_im2col, Mat& top_blo
                 "vshrn.u32  d16, q8, #16    \n"
                 "vshrn.u32  d17, q9, #16    \n"
 
-                "vst1.u16   {d16-d17}, [%1 :128]! \n"
+                "vst1.u16   {d16-d17}, [%1 :64]! \n"
 
                 : "=r"(nn),      // %0
                 "=r"(outptr0), // %1
diff --git a/src/layer/arm/convolution_sgemm_pack4to1_bf16s.h b/src/layer/arm/convolution_sgemm_pack4to1_bf16s.h
index 613248a40..d43215859 100644
--- a/src/layer/arm/convolution_sgemm_pack4to1_bf16s.h
+++ b/src/layer/arm/convolution_sgemm_pack4to1_bf16s.h
@@ -180,8 +180,8 @@ static void im2col_sgemm_pack4to1_bf16s_neon(const Mat& bottom_im2col, Mat& top_
 #else
                     asm volatile(
                         "pld        [%0, #256]          \n"
-                        "vld4.u16   {d0-d3}, [%0 :128]  \n"
-                        "vst1.u16   {d0-d3}, [%1 :128]! \n"
+                        "vld4.u16   {d0-d3}, [%0 :64]   \n"
+                        "vst1.u16   {d0-d3}, [%1 :64]!  \n"
                         : "=r"(img0),  // %0
                         "=r"(tmpptr) // %1
                         : "0"(img0),
diff --git a/src/layer/arm/packing_arm.cpp b/src/layer/arm/packing_arm.cpp
index 2a724fd39..e5211d158 100644
--- a/src/layer/arm/packing_arm.cpp
+++ b/src/layer/arm/packing_arm.cpp
@@ -522,14 +522,14 @@ int Packing_arm::forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, cons
                         : "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23");
 #else
                     asm volatile(
-                        "vld1.u16   {d16-d17}, [%0 : 128]! \n"
-                        "vld1.u16   {d18-d19}, [%1 : 128]! \n"
-                        "vld1.u16   {d20-d21}, [%2 : 128]! \n"
-                        "vld1.u16   {d22-d23}, [%3 : 128]! \n"
-                        "vld1.u16   {d24-d25}, [%4 : 128]! \n"
-                        "vld1.u16   {d26-d27}, [%5 : 128]! \n"
-                        "vld1.u16   {d28-d29}, [%6 : 128]! \n"
-                        "vld1.u16   {d30-d31}, [%7 : 128]! \n"
+                        "vld1.u16   {d16-d17}, [%0]!    \n"
+                        "vld1.u16   {d18-d19}, [%1]!    \n"
+                        "vld1.u16   {d20-d21}, [%2]!    \n"
+                        "vld1.u16   {d22-d23}, [%3]!    \n"
+                        "vld1.u16   {d24-d25}, [%4]!    \n"
+                        "vld1.u16   {d26-d27}, [%5]!    \n"
+                        "vld1.u16   {d28-d29}, [%6]!    \n"
+                        "vld1.u16   {d30-d31}, [%7]!    \n"
 
                         "vtrn.u16   q8, q9              \n"
                         "vtrn.u16   q10, q11            \n"
@@ -667,14 +667,14 @@ int Packing_arm::forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, cons
                         "vswp       d21, d28            \n"
                         "vswp       d23, d30            \n"
 
-                        "vst1.u16   {d16-d17}, [%1 : 128]! \n"
-                        "vst1.u16   {d18-d19}, [%2 : 128]! \n"
-                        "vst1.u16   {d20-d21}, [%3 : 128]! \n"
-                        "vst1.u16   {d22-d23}, [%4 : 128]! \n"
-                        "vst1.u16   {d24-d25}, [%5 : 128]! \n"
-                        "vst1.u16   {d26-d27}, [%6 : 128]! \n"
-                        "vst1.u16   {d28-d29}, [%7 : 128]! \n"
-                        "vst1.u16   {d30-d31}, [%8 : 128]! \n"
+                        "vst1.u16   {d16-d17}, [%1]!    \n"
+                        "vst1.u16   {d18-d19}, [%2]!    \n"
+                        "vst1.u16   {d20-d21}, [%3]!    \n"
+                        "vst1.u16   {d22-d23}, [%4]!    \n"
+                        "vst1.u16   {d24-d25}, [%5]!    \n"
+                        "vst1.u16   {d26-d27}, [%6]!    \n"
+                        "vst1.u16   {d28-d29}, [%7]!    \n"
+                        "vst1.u16   {d30-d31}, [%8]!    \n"
                         : "=r"(r0),      // %0
                         "=r"(outptr0), // %1
                         "=r"(outptr1), // %2
@@ -744,8 +744,8 @@ int Packing_arm::forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, cons
                         : "memory", "v0", "v1", "v2", "v3");
 #else
                     asm volatile(
-                        "vld1.u16   {d0-d1}, [%0 :128]! \n"
-                        "vld1.u16   {d2-d3}, [%1 :128]! \n"
+                        "vld1.u16   {d0-d1}, [%0 :64]!  \n"
+                        "vld1.u16   {d2-d3}, [%1 :64]!  \n"
 
                         "vswp       d1, d2              \n"
 
@@ -813,8 +813,8 @@ int Packing_arm::forward_bf16s_fp16s(const Mat& bottom_blob, Mat& top_blob, cons
 
                         "vswp       d1, d2              \n"
 
-                        "vst1.u16   {d0-d1}, [%1 :128]! \n"
-                        "vst1.u16   {d2-d3}, [%2 :128]! \n"
+                        "vst1.u16   {d0-d1}, [%1 :64]!  \n"
+                        "vst1.u16   {d2-d3}, [%2 :64]!  \n"
                         : "=r"(r0),      // %0
                         "=r"(outptr0), // %1
                         "=r"(outptr1)  // %2
diff --git a/src/mat.h b/src/mat.h
index d5585dd99..8fdbe9f28 100644
--- a/src/mat.h
+++ b/src/mat.h
@@ -913,48 +913,16 @@ NCNN_FORCEINLINE void Mat::fill(float _v)
     int size = (int)total();
     float* ptr = (float*)data;
 
-#if __ARM_NEON
-    int nn = size >> 2;
-    int remain = size - (nn << 2);
-#else
-    int remain = size;
-#endif // __ARM_NEON
-
+    int i = 0;
 #if __ARM_NEON
     float32x4_t _c = vdupq_n_f32(_v);
-#if __aarch64__
-    if (nn > 0)
+    for (; i + 3 < size; i += 4)
     {
-        asm volatile(
-            "0:                             \n"
-            "subs       %w0, %w0, #1        \n"
-            "st1        {%4.4s}, [%1], #16  \n"
-            "bne        0b                  \n"
-            : "=r"(nn), // %0
-            "=r"(ptr) // %1
-            : "0"(nn),
-            "1"(ptr),
-            "w"(_c) // %4
-            : "cc", "memory");
-    }
-#else
-    if (nn > 0)
-    {
-        asm volatile(
-            "0:                             \n"
-            "subs       %0, #1              \n"
-            "vst1.f32   {%e4-%f4}, [%1 :128]!\n"
-            "bne        0b                  \n"
-            : "=r"(nn), // %0
-            "=r"(ptr) // %1
-            : "0"(nn),
-            "1"(ptr),
-            "w"(_c) // %4
-            : "cc", "memory");
+        vst1q_f32(ptr, _c);
+        ptr += 4;
     }
-#endif // __aarch64__
 #endif // __ARM_NEON
-    for (; remain > 0; remain--)
+    for (; i < size; i++)
     {
         *ptr++ = _v;
     }
@@ -965,48 +933,16 @@ NCNN_FORCEINLINE void Mat::fill(int _v)
     int size = (int)total();
     int* ptr = (int*)data;
 
-#if __ARM_NEON
-    int nn = size >> 2;
-    int remain = size - (nn << 2);
-#else
-    int remain = size;
-#endif // __ARM_NEON
-
+    int i = 0;
 #if __ARM_NEON
     int32x4_t _c = vdupq_n_s32(_v);
-#if __aarch64__
-    if (nn > 0)
+    for (; i + 3 < size; i += 4)
     {
-        asm volatile(
-            "0:                             \n"
-            "subs       %w0, %w0, #1        \n"
-            "st1        {%4.4s}, [%1], #16  \n"
-            "bne        0b                  \n"
-            : "=r"(nn), // %0
-            "=r"(ptr) // %1
-            : "0"(nn),
-            "1"(ptr),
-            "w"(_c) // %4
-            : "cc", "memory");
-    }
-#else
-    if (nn > 0)
-    {
-        asm volatile(
-            "0:                             \n"
-            "subs       %0, #1              \n"
-            "vst1.s32   {%e4-%f4}, [%1 :128]!\n"
-            "bne        0b                  \n"
-            : "=r"(nn), // %0
-            "=r"(ptr) // %1
-            : "0"(nn),
-            "1"(ptr),
-            "w"(_c) // %4
-            : "cc", "memory");
+        vst1q_s32(ptr, _c);
+        ptr += 4;
     }
-#endif // __aarch64__
 #endif // __ARM_NEON
-    for (; remain > 0; remain--)
+    for (; i < size; i++)
     {
         *ptr++ = _v;
     }