diff --git a/src/layer/vulkan/shader/absval_pack8.comp b/src/layer/vulkan/shader/absval_pack8.comp
new file mode 100644
index 000000000..5b3021fbd
--- /dev/null
+++ b/src/layer/vulkan/shader/absval_pack8.comp
@@ -0,0 +1,57 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    v[0] = abs(v[0]);
+    v[1] = abs(v[1]);
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/batchnorm_pack8.comp b/src/layer/vulkan/shader/batchnorm_pack8.comp
new file mode 100644
index 000000000..f1448fafb
--- /dev/null
+++ b/src/layer/vulkan/shader/batchnorm_pack8.comp
@@ -0,0 +1,64 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer a { sfpvec8 a_data[]; };
+layout (binding = 2) readonly buffer b { sfpvec8 b_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    const ivec3 gxyz = ivec3(gx, gy, gz);
+    const int abi = gxyz[p.dims - 1];
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+    afpvec8 b = buffer_ld8(b_data, abi);
+    afpvec8 a = buffer_ld8(a_data, abi);
+
+    v[0] = b[0] * v[0] + a[0];
+    v[1] = b[1] * v[1] + a[1];
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/binaryop_broadcast_a1_pack8.comp b/src/layer/vulkan/shader/binaryop_broadcast_a1_pack8.comp
new file mode 100644
index 000000000..e3627c50b
--- /dev/null
+++ b/src/layer/vulkan/shader/binaryop_broadcast_a1_pack8.comp
@@ -0,0 +1,120 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int op_type = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer a_blob { sfp a_blob_data[]; };
+layout (binding = 1) readonly buffer b_blob { sfpvec8 b_blob_data[]; };
+layout (binding = 2) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int adims;
+    int aw;
+    int ah;
+    int ac;
+    int acstep;
+
+    int bdims;
+    int bw;
+    int bh;
+    int bc;
+    int bcstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    // type 2 3 4
+    afpvec4 v1 = afpvec4(buffer_ld1(a_blob_data, 0));
+    afpvec8 v2 = buffer_ld8(b_blob_data, gi);
+
+    afpvec8 res;
+
+    if (op_type == 0)
+    {
+        res[0] = v1 + v2[0];
+        res[1] = v1 + v2[1];
+    }
+    if (op_type == 1)
+    {
+        res[0] = v1 - v2[0];
+        res[1] = v1 - v2[1];
+    }
+    if (op_type == 2)
+    {
+        res[0] = v1 * v2[0];
+        res[1] = v1 * v2[1];
+    }
+    if (op_type == 3)
+    {
+        res[0] = v1 / v2[0];
+        res[1] = v1 / v2[1];
+    }
+    if (op_type == 4)
+    {
+        res[0] = max(v1, v2[0]);
+        res[1] = max(v1, v2[1]);
+    }
+    if (op_type == 5)
+    {
+        res[0] = min(v1, v2[0]);
+        res[1] = min(v1, v2[1]);
+    }
+    if (op_type == 6)
+    {
+        res[0] = pow(v1, v2[0]);
+        res[1] = pow(v1, v2[1]);
+    }
+    if (op_type == 7)
+    {
+        res[0] = v2[0] - v1;
+        res[1] = v2[1] - v1;
+    }
+    if (op_type == 8)
+    {
+        res[0] = v2[0] / v1;
+        res[1] = v2[1] / v1;
+    }
+
+    buffer_st8(top_blob_data, gi, res);
+}
diff --git a/src/layer/vulkan/shader/binaryop_broadcast_b1_pack8.comp b/src/layer/vulkan/shader/binaryop_broadcast_b1_pack8.comp
new file mode 100644
index 000000000..aa6cf171a
--- /dev/null
+++ b/src/layer/vulkan/shader/binaryop_broadcast_b1_pack8.comp
@@ -0,0 +1,128 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int op_type = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer a_blob { sfpvec8 a_blob_data[]; };
+layout (binding = 1) readonly buffer b_blob { sfp b_blob_data[]; };
+layout (binding = 2) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int adims;
+    int aw;
+    int ah;
+    int ac;
+    int acstep;
+
+    int bdims;
+    int bw;
+    int bh;
+    int bc;
+    int bcstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    int bi = 0;
+
+    if (p.bw == p.aw && p.bh == p.ah && p.bc == 1)
+    {
+        // special type 2
+        bi = gy * p.bw + gx;
+    }
+
+    // type 6 11 16
+    afpvec8 v1 = buffer_ld8(a_blob_data, gi);
+    afpvec4 v2 = afpvec4(buffer_ld1(b_blob_data, bi));
+
+    afpvec8 res;
+
+    if (op_type == 0)
+    {
+        res[0] = v1[0] + v2;
+        res[1] = v1[1] + v2;
+    }
+    if (op_type == 1)
+    {
+        res[0] = v1[0] - v2;
+        res[1] = v1[1] - v2;
+    }
+    if (op_type == 2)
+    {
+        res[0] = v1[0] * v2;
+        res[1] = v1[1] * v2;
+    }
+    if (op_type == 3)
+    {
+        res[0] = v1[0] / v2;
+        res[1] = v1[1] / v2;
+    }
+    if (op_type == 4)
+    {
+        res[0] = max(v1[0], v2);
+        res[1] = max(v1[1], v2);
+    }
+    if (op_type == 5)
+    {
+        res[0] = min(v1[0], v2);
+        res[1] = min(v1[1], v2);
+    }
+    if (op_type == 6)
+    {
+        res[0] = pow(v1[0], v2);
+        res[1] = pow(v1[1], v2);
+    }
+    if (op_type == 7)
+    {
+        res[0] = v2 - v1[0];
+        res[1] = v2 - v1[1];
+    }
+    if (op_type == 8)
+    {
+        res[0] = v2 / v1[0];
+        res[1] = v2 / v1[1];
+    }
+
+    buffer_st8(top_blob_data, gi, res);
+}
diff --git a/src/layer/vulkan/shader/binaryop_broadcast_pack8.comp b/src/layer/vulkan/shader/binaryop_broadcast_pack8.comp
new file mode 100644
index 000000000..9c6f6b864
--- /dev/null
+++ b/src/layer/vulkan/shader/binaryop_broadcast_pack8.comp
@@ -0,0 +1,181 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int op_type = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer a_blob { sfpvec8 a_blob_data[]; };
+layout (binding = 1) readonly buffer b_blob { sfpvec8 b_blob_data[]; };
+layout (binding = 2) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int adims;
+    int aw;
+    int ah;
+    int ac;
+    int acstep;
+
+    int bdims;
+    int bw;
+    int bh;
+    int bc;
+    int bcstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    int ai;
+    int bi;
+
+    if (p.adims == 3)
+    {
+        if (p.bdims == 3)
+        {
+            if (p.bw == 1 && p.bh == 1)
+            {
+                // special type 1
+                ai = gi;
+                bi = gz;
+            }
+        }
+
+        if (p.bdims == 2)
+        {
+            // type 18
+            ai = gi;
+            bi = gz * p.bw + gy;
+        }
+
+        if (p.bdims == 1)
+        {
+            // type 17
+            ai = gi;
+            bi = gz;
+        }
+    }
+    else if (p.adims == 2)
+    {
+        if (p.bdims == 3)
+        {
+            // type 14
+            ai = gz * p.aw + gy;
+            bi = gi;
+        }
+
+        if (p.bdims == 1)
+        {
+            // type 12
+            ai = gi;
+            bi = gy;
+        }
+    }
+    else if (p.adims == 1)
+    {
+        if (p.bdims == 3)
+        {
+            // type 9
+            ai = gz;
+            bi = gi;
+        }
+
+        if (p.bdims == 2)
+        {
+            // type 8
+            ai = gy;
+            bi = gi;
+        }
+    }
+
+    afpvec8 v1 = buffer_ld8(a_blob_data, ai);
+    afpvec8 v2 = buffer_ld8(b_blob_data, bi);
+
+    afpvec8 res;
+
+    if (op_type == 0)
+    {
+        res[0] = v1[0] + v2[0];
+        res[1] = v1[1] + v2[1];
+    }
+    if (op_type == 1)
+    {
+        res[0] = v1[0] - v2[0];
+        res[1] = v1[1] - v2[1];
+    }
+    if (op_type == 2)
+    {
+        res[0] = v1[0] * v2[0];
+        res[1] = v1[1] * v2[1];
+    }
+    if (op_type == 3)
+    {
+        res[0] = v1[0] / v2[0];
+        res[1] = v1[1] / v2[1];
+    }
+    if (op_type == 4)
+    {
+        res[0] = max(v1[0], v2[0]);
+        res[1] = max(v1[1], v2[1]);
+    }
+    if (op_type == 5)
+    {
+        res[0] = min(v1[0], v2[0]);
+        res[1] = min(v1[1], v2[1]);
+    }
+    if (op_type == 6)
+    {
+        res[0] = pow(v1[0], v2[0]);
+        res[1] = pow(v1[1], v2[1]);
+    }
+    if (op_type == 7)
+    {
+        res[0] = v2[0] - v1[0];
+        res[1] = v2[1] - v1[1];
+    }
+    if (op_type == 8)
+    {
+        res[0] = v2[0] / v1[0];
+        res[1] = v2[1] / v1[1];
+    }
+
+    buffer_st8(top_blob_data, gi, res);
+}
diff --git a/src/layer/vulkan/shader/binaryop_pack8.comp b/src/layer/vulkan/shader/binaryop_pack8.comp
new file mode 100644
index 000000000..9b2291c6f
--- /dev/null
+++ b/src/layer/vulkan/shader/binaryop_pack8.comp
@@ -0,0 +1,179 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int op_type = 0;
+layout (constant_id = 1) const int with_scalar = 0;
+layout (constant_id = 2) const float const_b = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer a_blob { sfpvec8 a_blob_data[]; };
+layout (binding = 1) readonly buffer b_blob { sfpvec8 b_blob_data[]; };
+layout (binding = 2) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int adims;
+    int aw;
+    int ah;
+    int ac;
+    int acstep;
+
+    int bdims;
+    int bw;
+    int bh;
+    int bc;
+    int bcstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    afpvec8 v1 = buffer_ld8(a_blob_data, gi);
+
+    afpvec8 res;
+
+    if (with_scalar == 1)
+    {
+        // type 5 10 15
+        afp b = afp(const_b);
+
+        if (op_type == 0)
+        {
+            res[0] = v1[0] + b;
+            res[1] = v1[1] + b;
+        }
+        if (op_type == 1)
+        {
+            res[0] = v1[0] - b;
+            res[1] = v1[1] - b;
+        }
+        if (op_type == 2)
+        {
+            res[0] = v1[0] * b;
+            res[1] = v1[1] * b;
+        }
+        if (op_type == 3)
+        {
+            res[0] = v1[0] / b;
+            res[1] = v1[1] / b;
+        }
+        if (op_type == 4)
+        {
+            res[0] = max(v1[0], b);
+            res[1] = max(v1[1], b);
+        }
+        if (op_type == 5)
+        {
+            res[0] = min(v1[0], b);
+            res[1] = min(v1[1], b);
+        }
+        if (op_type == 6)
+        {
+            res[0] = pow(v1[0], afpvec4(b));
+            res[1] = pow(v1[1], afpvec4(b));
+        }
+        if (op_type == 7)
+        {
+            res[0] = b - v1[0];
+            res[1] = b - v1[1];
+        }
+        if (op_type == 8)
+        {
+            res[0] = b / v1[0];
+            res[1] = b / v1[1];
+        }
+
+        buffer_st8(a_blob_data, gi, res);
+    }
+    else
+    {
+        // type 7 13 19
+        afpvec8 v2 = buffer_ld8(b_blob_data, gi);
+
+        if (op_type == 0)
+        {
+            res[0] = v1[0] + v2[0];
+            res[1] = v1[1] + v2[1];
+        }
+        if (op_type == 1)
+        {
+            res[0] = v1[0] - v2[0];
+            res[1] = v1[1] - v2[1];
+        }
+        if (op_type == 2)
+        {
+            res[0] = v1[0] * v2[0];
+            res[1] = v1[1] * v2[1];
+        }
+        if (op_type == 3)
+        {
+            res[0] = v1[0] / v2[0];
+            res[1] = v1[1] / v2[1];
+        }
+        if (op_type == 4)
+        {
+            res[0] = max(v1[0], v2[0]);
+            res[1] = max(v1[1], v2[1]);
+        }
+        if (op_type == 5)
+        {
+            res[0] = min(v1[0], v2[0]);
+            res[1] = min(v1[1], v2[1]);
+        }
+        if (op_type == 6)
+        {
+            res[0] = pow(v1[0], v2[0]);
+            res[1] = pow(v1[1], v2[1]);
+        }
+        if (op_type == 7)
+        {
+            res[0] = v2[0] - v1[0];
+            res[1] = v2[1] - v1[1];
+        }
+        if (op_type == 8)
+        {
+            res[0] = v2[0] / v1[0];
+            res[1] = v2[1] / v1[1];
+        }
+
+        buffer_st8(top_blob_data, gi, res);
+    }
+}
diff --git a/src/layer/vulkan/shader/cast_fp16_to_fp32_pack8.comp b/src/layer/vulkan/shader/cast_fp16_to_fp32_pack8.comp
new file mode 100644
index 000000000..58a358720
--- /dev/null
+++ b/src/layer/vulkan/shader/cast_fp16_to_fp32_pack8.comp
@@ -0,0 +1,61 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { mat2x4 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    const int v_offset = gz * p.cstep + gy * p.w + gx;
+
+    top_blob_data[gi] = mat2x4(buffer_ld8(bottom_blob_data, v_offset));
+}
diff --git a/src/layer/vulkan/shader/cast_fp32_to_fp16_pack8.comp b/src/layer/vulkan/shader/cast_fp32_to_fp16_pack8.comp
new file mode 100644
index 000000000..035b03d1e
--- /dev/null
+++ b/src/layer/vulkan/shader/cast_fp32_to_fp16_pack8.comp
@@ -0,0 +1,61 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { mat2x4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    const int v_offset = gz * p.cstep + gy * p.w + gx;
+
+    buffer_st8(top_blob_data, gi, afpvec8(bottom_blob_data[v_offset]));
+}
diff --git a/src/layer/vulkan/shader/clip_pack8.comp b/src/layer/vulkan/shader/clip_pack8.comp
new file mode 100644
index 000000000..8a3ebb197
--- /dev/null
+++ b/src/layer/vulkan/shader/clip_pack8.comp
@@ -0,0 +1,60 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const float const_min = 0;
+layout (constant_id = 1) const float const_max = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    v[0] = clamp(v[0], afp(const_min), afp(const_max));
+    v[1] = clamp(v[1], afp(const_min), afp(const_max));
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/concat_pack8.comp b/src/layer/vulkan/shader/concat_pack8.comp
new file mode 100644
index 000000000..b308d6dc6
--- /dev/null
+++ b/src/layer/vulkan/shader/concat_pack8.comp
@@ -0,0 +1,69 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int offset;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    ivec3 gxyz = ivec3(gx, gy, gz);
+
+    gxyz[p.dims - 1 - axis] += p.offset;
+
+    int v_offset = gxyz.z * p.outcstep + gxyz.y * p.outw + gxyz.x;
+
+    buffer_cp8(top_blob_data, v_offset, bottom_blob_data, gi);
+}
diff --git a/src/layer/vulkan/shader/concat_pack8to1.comp b/src/layer/vulkan/shader/concat_pack8to1.comp
new file mode 100644
index 000000000..30e93543c
--- /dev/null
+++ b/src/layer/vulkan/shader/concat_pack8to1.comp
@@ -0,0 +1,75 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int offset;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    ivec3 gxyz = ivec3(gx, gy, gz);
+
+    gxyz[p.dims - 1] *= 8;
+    gxyz[p.dims - 1 - axis] += p.offset;
+
+    int v_offset_0 = gxyz.z * p.outcstep + gxyz.y * p.outw + gxyz.x;
+
+    ivec3 gxyz4 = ivec3(1, p.outw, p.outcstep);
+
+    ivec4 v_offset = v_offset_0 + ivec4(0, 1, 2, 3) * gxyz4[p.dims - 1 - axis];
+    ivec4 vv_offset = v_offset + 4 * gxyz4[p.dims - 1 - axis];
+
+    buffer_cp8to1(top_blob_data, v_offset, vv_offset, bottom_blob_data, gi);
+}
diff --git a/src/layer/vulkan/shader/concat_pack8to4.comp b/src/layer/vulkan/shader/concat_pack8to4.comp
new file mode 100644
index 000000000..58d7b0484
--- /dev/null
+++ b/src/layer/vulkan/shader/concat_pack8to4.comp
@@ -0,0 +1,74 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int offset;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    ivec3 gxyz = ivec3(gx, gy, gz);
+
+    gxyz[p.dims - 1] *= 2;
+    gxyz[p.dims - 1 - axis] += p.offset;
+
+    int v_offset_0 = gxyz.z * p.outcstep + gxyz.y * p.outw + gxyz.x;
+
+    ivec3 gxyz4 = ivec3(1, p.outw, p.outcstep);
+
+    ivec2 v_offset = v_offset_0 + ivec2(0, 1) * gxyz4[p.dims - 1 - axis];
+
+    buffer_cp8to4(top_blob_data, v_offset, bottom_blob_data, gi);
+}
diff --git a/src/layer/vulkan/shader/convolution_pack1to8.comp b/src/layer/vulkan/shader/convolution_pack1to8.comp
new file mode 100644
index 000000000..e627600f9
--- /dev/null
+++ b/src/layer/vulkan/shader/convolution_pack1to8.comp
@@ -0,0 +1,131 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    int w_offset = gz * p.c * kernel_w * kernel_h;
+
+    for (int z = 0; z < p.c; z++)
+    {
+        int v_offset = z * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afp v = buffer_ld1(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset + x);
+
+                // sum += v * k;
+                sum[0] += v * k[0];
+                sum[1] += v * k[1];
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolution_pack4to8.comp b/src/layer/vulkan/shader/convolution_pack4to8.comp
new file mode 100644
index 000000000..b9ffb9072
--- /dev/null
+++ b/src/layer/vulkan/shader/convolution_pack4to8.comp
@@ -0,0 +1,144 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    int w_offset = gz * p.c * kernel_w * kernel_h;
+
+    for (int z = 0; z < p.c; z++)
+    {
+        int v_offset = z * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec4 k0 = buffer_ld4(weight_data, (w_offset + x) * 8 + 0);
+                afpvec4 k1 = buffer_ld4(weight_data, (w_offset + x) * 8 + 1);
+                afpvec4 k2 = buffer_ld4(weight_data, (w_offset + x) * 8 + 2);
+                afpvec4 k3 = buffer_ld4(weight_data, (w_offset + x) * 8 + 3);
+                afpvec4 k4 = buffer_ld4(weight_data, (w_offset + x) * 8 + 4);
+                afpvec4 k5 = buffer_ld4(weight_data, (w_offset + x) * 8 + 5);
+                afpvec4 k6 = buffer_ld4(weight_data, (w_offset + x) * 8 + 6);
+                afpvec4 k7 = buffer_ld4(weight_data, (w_offset + x) * 8 + 7);
+
+                // sum += v * k;
+                sum[0].r += dot(v, k0);
+                sum[0].g += dot(v, k1);
+                sum[0].b += dot(v, k2);
+                sum[0].a += dot(v, k3);
+                sum[1].r += dot(v, k4);
+                sum[1].g += dot(v, k5);
+                sum[1].b += dot(v, k6);
+                sum[1].a += dot(v, k7);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolution_pack8.comp b/src/layer/vulkan/shader/convolution_pack8.comp
new file mode 100644
index 000000000..af09f7af1
--- /dev/null
+++ b/src/layer/vulkan/shader/convolution_pack8.comp
@@ -0,0 +1,144 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    int w_offset = gz * p.c * kernel_w * kernel_h;
+
+    for (int z = 0; z < p.c; z++)
+    {
+        int v_offset = z * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k0 = buffer_ld8(weight_data, (w_offset + x) * 8 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, (w_offset + x) * 8 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, (w_offset + x) * 8 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, (w_offset + x) * 8 + 3);
+                afpvec8 k4 = buffer_ld8(weight_data, (w_offset + x) * 8 + 4);
+                afpvec8 k5 = buffer_ld8(weight_data, (w_offset + x) * 8 + 5);
+                afpvec8 k6 = buffer_ld8(weight_data, (w_offset + x) * 8 + 6);
+                afpvec8 k7 = buffer_ld8(weight_data, (w_offset + x) * 8 + 7);
+
+                // sum += v * k
+                sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+                sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
+                sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
+                sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
+                sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolution_pack8to1.comp b/src/layer/vulkan/shader/convolution_pack8to1.comp
new file mode 100644
index 000000000..2f40ed7b2
--- /dev/null
+++ b/src/layer/vulkan/shader/convolution_pack8to1.comp
@@ -0,0 +1,126 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfp bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afp sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld1(bias_data, gz);
+    }
+    else
+    {
+        sum = afp(0.f);
+    }
+
+    int w_offset = gz * p.c * kernel_w * kernel_h;
+
+    for (int z = 0; z < p.c; z++)
+    {
+        int v_offset = z * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset + x);
+
+                // sum += dot(v, k);
+                sum += dot(v[0], k[0]) + dot(v[1], k[1]);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = sum < afp(0.f) ? sum * slope : sum;
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st1(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolution_pack8to4.comp b/src/layer/vulkan/shader/convolution_pack8to4.comp
new file mode 100644
index 000000000..7f1a9e95d
--- /dev/null
+++ b/src/layer/vulkan/shader/convolution_pack8to4.comp
@@ -0,0 +1,132 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec4 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld4(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec4(0.f);
+    }
+
+    int w_offset = gz * p.c * kernel_w * kernel_h;
+
+    for (int z = 0; z < p.c; z++)
+    {
+        int v_offset = z * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k0 = buffer_ld8(weight_data, (w_offset + x) * 4 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, (w_offset + x) * 4 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, (w_offset + x) * 4 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, (w_offset + x) * 4 + 3);
+
+                // sum += v * k
+                sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st4(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolutiondepthwise_group_pack1to8.comp b/src/layer/vulkan/shader/convolutiondepthwise_group_pack1to8.comp
new file mode 100644
index 000000000..e78e0ed93
--- /dev/null
+++ b/src/layer/vulkan/shader/convolutiondepthwise_group_pack1to8.comp
@@ -0,0 +1,142 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int z = 0; z < channels_g; z++)
+    {
+        int v_offset = v_offset_0 + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afp v = buffer_ld1(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset + x);
+
+                // sum += v * k;
+                sum[0] += v * k[0];
+                sum[1] += v * k[1];
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+
+        v_offset_0 += p.cstep;
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolutiondepthwise_group_pack4to8.comp b/src/layer/vulkan/shader/convolutiondepthwise_group_pack4to8.comp
new file mode 100644
index 000000000..7df4bde97
--- /dev/null
+++ b/src/layer/vulkan/shader/convolutiondepthwise_group_pack4to8.comp
@@ -0,0 +1,155 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int z = 0; z < channels_g; z++)
+    {
+        int v_offset = v_offset_0 + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec4 k0 = buffer_ld4(weight_data, (w_offset + x) * 8 + 0);
+                afpvec4 k1 = buffer_ld4(weight_data, (w_offset + x) * 8 + 1);
+                afpvec4 k2 = buffer_ld4(weight_data, (w_offset + x) * 8 + 2);
+                afpvec4 k3 = buffer_ld4(weight_data, (w_offset + x) * 8 + 3);
+                afpvec4 k4 = buffer_ld4(weight_data, (w_offset + x) * 8 + 4);
+                afpvec4 k5 = buffer_ld4(weight_data, (w_offset + x) * 8 + 5);
+                afpvec4 k6 = buffer_ld4(weight_data, (w_offset + x) * 8 + 6);
+                afpvec4 k7 = buffer_ld4(weight_data, (w_offset + x) * 8 + 7);
+
+                // sum += v * k;
+                sum[0].r += dot(v, k0);
+                sum[0].g += dot(v, k1);
+                sum[0].b += dot(v, k2);
+                sum[0].a += dot(v, k3);
+                sum[1].r += dot(v, k4);
+                sum[1].g += dot(v, k5);
+                sum[1].b += dot(v, k6);
+                sum[1].a += dot(v, k7);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+
+        v_offset_0 += p.cstep;
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolutiondepthwise_group_pack8.comp b/src/layer/vulkan/shader/convolutiondepthwise_group_pack8.comp
new file mode 100644
index 000000000..197f67061
--- /dev/null
+++ b/src/layer/vulkan/shader/convolutiondepthwise_group_pack8.comp
@@ -0,0 +1,155 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int z = 0; z < channels_g; z++)
+    {
+        int v_offset = v_offset_0 + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k0 = buffer_ld8(weight_data, (w_offset + x) * 8 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, (w_offset + x) * 8 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, (w_offset + x) * 8 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, (w_offset + x) * 8 + 3);
+                afpvec8 k4 = buffer_ld8(weight_data, (w_offset + x) * 8 + 4);
+                afpvec8 k5 = buffer_ld8(weight_data, (w_offset + x) * 8 + 5);
+                afpvec8 k6 = buffer_ld8(weight_data, (w_offset + x) * 8 + 6);
+                afpvec8 k7 = buffer_ld8(weight_data, (w_offset + x) * 8 + 7);
+
+                // sum += v * k
+                sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+                sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
+                sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
+                sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
+                sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+
+        v_offset_0 += p.cstep;
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolutiondepthwise_group_pack8to1.comp b/src/layer/vulkan/shader/convolutiondepthwise_group_pack8to1.comp
new file mode 100644
index 000000000..42f809bcd
--- /dev/null
+++ b/src/layer/vulkan/shader/convolutiondepthwise_group_pack8to1.comp
@@ -0,0 +1,137 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfp bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afp sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld1(bias_data, gz);
+    }
+    else
+    {
+        sum = afp(0.f);
+    }
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int z = 0; z < channels_g; z++)
+    {
+        int v_offset = v_offset_0 + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset + x);
+
+                // sum += dot(v, k);
+                sum += dot(v[0], k[0]) + dot(v[1], k[1]);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+
+        v_offset_0 += p.cstep;
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = sum < afp(0.f) ? sum * slope : sum;
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st1(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolutiondepthwise_group_pack8to4.comp b/src/layer/vulkan/shader/convolutiondepthwise_group_pack8to4.comp
new file mode 100644
index 000000000..521339101
--- /dev/null
+++ b/src/layer/vulkan/shader/convolutiondepthwise_group_pack8to4.comp
@@ -0,0 +1,143 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec4 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld4(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec4(0.f);
+    }
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int z = 0; z < channels_g; z++)
+    {
+        int v_offset = v_offset_0 + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
+
+                afpvec8 k0 = buffer_ld8(weight_data, (w_offset + x) * 4 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, (w_offset + x) * 4 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, (w_offset + x) * 4 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, (w_offset + x) * 4 + 3);
+
+                // sum += v * k
+                sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+            }
+
+            v_offset += dilation_h * p.w;
+            w_offset += kernel_w;
+        }
+
+        v_offset_0 += p.cstep;
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st4(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/convolutiondepthwise_pack8.comp b/src/layer/vulkan/shader/convolutiondepthwise_pack8.comp
new file mode 100644
index 000000000..4985e2a54
--- /dev/null
+++ b/src/layer/vulkan/shader/convolutiondepthwise_pack8.comp
@@ -0,0 +1,129 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    // depth-wise convolution
+    int w_offset = gz * kernel_w * kernel_h;
+    int v_offset = gz * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        for (int x = 0; x < kernel_w; x++)
+        {
+            afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
+
+            afpvec8 k = buffer_ld8(weight_data, w_offset + x);
+
+            // sum += v * k;
+            sum[0] += v[0] * k[0];
+            sum[1] += v[1] * k[1];
+        }
+
+        v_offset += dilation_h * p.w;
+        w_offset += kernel_w;
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/crop_pack1to8.comp b/src/layer/vulkan/shader/crop_pack1to8.comp
new file mode 100644
index 000000000..ebd86e781
--- /dev/null
+++ b/src/layer/vulkan/shader/crop_pack1to8.comp
@@ -0,0 +1,69 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int woffset;
+    int hoffset;
+    int coffset;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    int x = gx + p.woffset;
+    int y = gy + p.hoffset;
+    ivec4 z4 = gz * 8 + ivec4(0, 1, 2, 3) + p.coffset;
+    ivec4 v_offset = z4 * p.cstep + y * p.w + x;
+    ivec4 vv_offset = (z4 + 4) * p.cstep + y * p.w + x;
+
+    buffer_cp1to8(top_blob_data, gi, bottom_blob_data, v_offset, vv_offset);
+}
diff --git a/src/layer/vulkan/shader/crop_pack8to1.comp b/src/layer/vulkan/shader/crop_pack8to1.comp
new file mode 100644
index 000000000..949e2d804
--- /dev/null
+++ b/src/layer/vulkan/shader/crop_pack8to1.comp
@@ -0,0 +1,74 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int woffset;
+    int hoffset;
+    int coffset;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    int x = gx + p.woffset;
+    int y = gy + p.hoffset;
+    int z = gz + p.coffset;
+
+    int sz = z / 8;
+    int sz4 = (z % 8) / 4;
+    int lane = z % 4;
+    int v_offset = sz * p.cstep + y * p.w + x;
+
+    afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+    buffer_st1(top_blob_data, gi, v[sz4][lane]);
+}
diff --git a/src/layer/vulkan/shader/deconvolution_pack1to8.comp b/src/layer/vulkan/shader/deconvolution_pack1to8.comp
new file mode 100644
index 000000000..b56f6e582
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolution_pack1to8.comp
@@ -0,0 +1,151 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    int w_offset_0 = gz * p.c * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < p.c; z++)
+            {
+                afp v = buffer_ld1(bottom_blob_data, v_offset);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset);
+
+                // sum += v * k;
+                sum[0] += v * k[0];
+                sum[1] += v * k[1];
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolution_pack4to8.comp b/src/layer/vulkan/shader/deconvolution_pack4to8.comp
new file mode 100644
index 000000000..904fe4d1d
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolution_pack4to8.comp
@@ -0,0 +1,164 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    int w_offset_0 = gz * p.c * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < p.c; z++)
+            {
+                afpvec4 v = buffer_ld4(bottom_blob_data, v_offset);
+
+                afpvec4 k0 = buffer_ld4(weight_data, w_offset * 8 + 0);
+                afpvec4 k1 = buffer_ld4(weight_data, w_offset * 8 + 1);
+                afpvec4 k2 = buffer_ld4(weight_data, w_offset * 8 + 2);
+                afpvec4 k3 = buffer_ld4(weight_data, w_offset * 8 + 3);
+                afpvec4 k4 = buffer_ld4(weight_data, w_offset * 8 + 4);
+                afpvec4 k5 = buffer_ld4(weight_data, w_offset * 8 + 5);
+                afpvec4 k6 = buffer_ld4(weight_data, w_offset * 8 + 6);
+                afpvec4 k7 = buffer_ld4(weight_data, w_offset * 8 + 7);
+
+                // sum += v * k;
+                sum[0].r += dot(v, k0);
+                sum[0].g += dot(v, k1);
+                sum[0].b += dot(v, k2);
+                sum[0].a += dot(v, k3);
+                sum[1].r += dot(v, k4);
+                sum[1].g += dot(v, k5);
+                sum[1].b += dot(v, k6);
+                sum[1].a += dot(v, k7);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolution_pack8.comp b/src/layer/vulkan/shader/deconvolution_pack8.comp
new file mode 100644
index 000000000..88b14694d
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolution_pack8.comp
@@ -0,0 +1,164 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    int w_offset_0 = gz * p.c * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < p.c; z++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+                afpvec8 k0 = buffer_ld8(weight_data, w_offset * 8 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, w_offset * 8 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, w_offset * 8 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, w_offset * 8 + 3);
+                afpvec8 k4 = buffer_ld8(weight_data, w_offset * 8 + 4);
+                afpvec8 k5 = buffer_ld8(weight_data, w_offset * 8 + 5);
+                afpvec8 k6 = buffer_ld8(weight_data, w_offset * 8 + 6);
+                afpvec8 k7 = buffer_ld8(weight_data, w_offset * 8 + 7);
+
+                // sum += v * k
+                sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+                sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
+                sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
+                sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
+                sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolution_pack8to1.comp b/src/layer/vulkan/shader/deconvolution_pack8to1.comp
new file mode 100644
index 000000000..00cd87371
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolution_pack8to1.comp
@@ -0,0 +1,146 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfp bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afp sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld1(bias_data, gz);
+    }
+    else
+    {
+        sum = afp(0.f);
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    int w_offset_0 = gz * p.c * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < p.c; z++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset);
+
+                // sum += dot(v, k);
+                sum += dot(v[0], k[0]) + dot(v[1], k[1]);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = sum < afp(0.f) ? sum * slope : sum;
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st1(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolution_pack8to4.comp b/src/layer/vulkan/shader/deconvolution_pack8to4.comp
new file mode 100644
index 000000000..af6f4e7cd
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolution_pack8to4.comp
@@ -0,0 +1,152 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int activation_type = 0;
+layout (constant_id = 8) const float activation_param_0 = 0;
+layout (constant_id = 9) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec4 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld4(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec4(0.f);
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    int w_offset_0 = gz * p.c * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < p.c; z++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+                afpvec8 k0 = buffer_ld8(weight_data, w_offset * 4 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, w_offset * 4 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, w_offset * 4 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, w_offset * 4 + 3);
+
+                // sum += v * k
+                sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st4(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack1to8.comp b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack1to8.comp
new file mode 100644
index 000000000..2e01d9555
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack1to8.comp
@@ -0,0 +1,160 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < channels_g; z++)
+            {
+                afp v = buffer_ld1(bottom_blob_data, v_offset);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset);
+
+                // sum += v * k;
+                sum[0] += v * k[0];
+                sum[1] += v * k[1];
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack4to8.comp b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack4to8.comp
new file mode 100644
index 000000000..80bf9a273
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack4to8.comp
@@ -0,0 +1,173 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < channels_g; z++)
+            {
+                afpvec4 v = buffer_ld4(bottom_blob_data, v_offset);
+
+                afpvec4 k0 = buffer_ld4(weight_data, w_offset * 8 + 0);
+                afpvec4 k1 = buffer_ld4(weight_data, w_offset * 8 + 1);
+                afpvec4 k2 = buffer_ld4(weight_data, w_offset * 8 + 2);
+                afpvec4 k3 = buffer_ld4(weight_data, w_offset * 8 + 3);
+                afpvec4 k4 = buffer_ld4(weight_data, w_offset * 8 + 4);
+                afpvec4 k5 = buffer_ld4(weight_data, w_offset * 8 + 5);
+                afpvec4 k6 = buffer_ld4(weight_data, w_offset * 8 + 6);
+                afpvec4 k7 = buffer_ld4(weight_data, w_offset * 8 + 7);
+
+                // sum += v * k;
+                sum[0].r += dot(v, k0);
+                sum[0].g += dot(v, k1);
+                sum[0].b += dot(v, k2);
+                sum[0].a += dot(v, k3);
+                sum[1].r += dot(v, k4);
+                sum[1].g += dot(v, k5);
+                sum[1].b += dot(v, k6);
+                sum[1].a += dot(v, k7);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8.comp b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8.comp
new file mode 100644
index 000000000..630b5d071
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8.comp
@@ -0,0 +1,173 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < channels_g; z++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+                afpvec8 k0 = buffer_ld8(weight_data, w_offset * 8 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, w_offset * 8 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, w_offset * 8 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, w_offset * 8 + 3);
+                afpvec8 k4 = buffer_ld8(weight_data, w_offset * 8 + 4);
+                afpvec8 k5 = buffer_ld8(weight_data, w_offset * 8 + 5);
+                afpvec8 k6 = buffer_ld8(weight_data, w_offset * 8 + 6);
+                afpvec8 k7 = buffer_ld8(weight_data, w_offset * 8 + 7);
+
+                // sum += v * k
+                sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+                sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
+                sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
+                sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
+                sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to1.comp b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to1.comp
new file mode 100644
index 000000000..99b65c922
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to1.comp
@@ -0,0 +1,155 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfp bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afp sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld1(bias_data, gz);
+    }
+    else
+    {
+        sum = afp(0.f);
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < channels_g; z++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+                afpvec8 k = buffer_ld8(weight_data, w_offset);
+
+                // sum += dot(v, k);
+                sum += dot(v[0], k[0]) + dot(v[1], k[1]);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = sum < afp(0.f) ? sum * slope : sum;
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st1(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to4.comp b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to4.comp
new file mode 100644
index 000000000..ff889e6ab
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolutiondepthwise_group_pack8to4.comp
@@ -0,0 +1,161 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec4 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld4(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec4(0.f);
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // group convolution
+    const int channels_g = p.c / group;
+    const int num_output_g = p.outc / group;
+
+    // group id
+    const int gg = gz / num_output_g;
+
+    int w_offset_0 = gz * channels_g * kernel_w * kernel_h;
+    int v_offset_0 = gg * channels_g * p.cstep;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            for (int z = 0; z < channels_g; z++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+                afpvec8 k0 = buffer_ld8(weight_data, w_offset * 4 + 0);
+                afpvec8 k1 = buffer_ld8(weight_data, w_offset * 4 + 1);
+                afpvec8 k2 = buffer_ld8(weight_data, w_offset * 4 + 2);
+                afpvec8 k3 = buffer_ld8(weight_data, w_offset * 4 + 3);
+
+                // sum += v * k
+                sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+                sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+                sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+                sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+
+                v_offset += p.cstep;
+                w_offset += kernel_w * kernel_h;
+            }
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st4(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/deconvolutiondepthwise_pack8.comp b/src/layer/vulkan/shader/deconvolutiondepthwise_pack8.comp
new file mode 100644
index 000000000..fc2db1e21
--- /dev/null
+++ b/src/layer/vulkan/shader/deconvolutiondepthwise_pack8.comp
@@ -0,0 +1,148 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int kernel_w = 1;
+layout (constant_id = 1) const int kernel_h = 1;
+layout (constant_id = 2) const int dilation_w = 1;
+layout (constant_id = 3) const int dilation_h = 1;
+layout (constant_id = 4) const int stride_w = 1;
+layout (constant_id = 5) const int stride_h = 1;
+layout (constant_id = 6) const int bias_term = 0;
+layout (constant_id = 7) const int group = 1;
+layout (constant_id = 8) const int activation_type = 0;
+layout (constant_id = 9) const float activation_param_0 = 0;
+layout (constant_id = 10) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
+    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
+
+    // depth-wise deconvolution
+    int v_offset_0 = gz * p.cstep;
+    int w_offset_0 = gz * kernel_w * kernel_h;
+
+    for (int y = 0; y < kernel_h; y++)
+    {
+        int sys = (gy + y * dilation_h - (kernel_extent_h - 1));
+        if (sys < 0 || sys % stride_h != 0)
+            continue;
+
+        int sy = sys / stride_h;
+        if (sy >= p.h)
+            continue;
+
+        for (int x = 0; x < kernel_w; x++)
+        {
+            int sxs = (gx + x * dilation_w - (kernel_extent_w - 1));
+            if (sxs < 0 || sxs % stride_w != 0)
+                continue;
+
+            int sx = sxs / stride_w;
+            if (sx >= p.w)
+                continue;
+
+            int v_offset = v_offset_0 + sy * p.w + sx;
+            int w_offset = w_offset_0 + y * kernel_w + x;
+
+            afpvec8 v = buffer_ld8(bottom_blob_data, v_offset);
+
+            afpvec8 k = buffer_ld8(weight_data, w_offset);
+
+            // sum += v * k;
+            sum[0] += v[0] * k[0];
+            sum[1] += v[1] * k[1];
+        }
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, sum);
+}
diff --git a/src/layer/vulkan/shader/dropout_pack8.comp b/src/layer/vulkan/shader/dropout_pack8.comp
new file mode 100644
index 000000000..03c6b4fde
--- /dev/null
+++ b/src/layer/vulkan/shader/dropout_pack8.comp
@@ -0,0 +1,59 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const float scale = 1;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    v[0] = v[0] * afp(scale);
+    v[1] = v[1] * afp(scale);
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/eltwise_pack8.comp b/src/layer/vulkan/shader/eltwise_pack8.comp
new file mode 100644
index 000000000..7c85d0ee1
--- /dev/null
+++ b/src/layer/vulkan/shader/eltwise_pack8.comp
@@ -0,0 +1,106 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int op_type = 0;
+layout (constant_id = 1) const int coeff_term = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob1 { sfpvec8 bottom_blob1_data[]; };
+layout (binding = 1) readonly buffer bottom_blob2 { sfpvec8 bottom_blob2_data[]; };
+layout (binding = 2) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    float coeff0;
+    float coeff1;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v1 = buffer_ld8(bottom_blob1_data, gi);
+    afpvec8 v2 = buffer_ld8(bottom_blob2_data, gi);
+
+    afpvec8 res;
+
+    if (coeff_term == 0)
+    {
+        if (op_type == 0)
+        {
+            res[0] = v1[0] * v2[0];
+            res[1] = v1[1] * v2[1];
+        }
+
+        if (op_type == 1)
+        {
+            res[0] = v1[0] + v2[0];
+            res[1] = v1[1] + v2[1];
+        }
+
+        if (op_type == 2)
+        {
+            res[0] = max(v1[0], v2[0]);
+            res[1] = max(v1[1], v2[1]);
+        }
+    }
+    else
+    {
+        if (op_type == 0)
+        {
+            res[0] = v1[0] * afp(p.coeff0) * v2[0] * afp(p.coeff1);
+            res[1] = v1[1] * afp(p.coeff0) * v2[1] * afp(p.coeff1);
+        }
+
+        if (op_type == 1)
+        {
+            res[0] = v1[0] * afp(p.coeff0) + v2[0] * afp(p.coeff1);
+            res[1] = v1[1] * afp(p.coeff0) + v2[1] * afp(p.coeff1);
+        }
+
+        if (op_type == 2)
+        {
+            res[0] = max(v1[0] * afp(p.coeff0), v2[0] * afp(p.coeff1));
+            res[1] = max(v1[1] * afp(p.coeff0), v2[1] * afp(p.coeff1));
+        }
+    }
+
+    buffer_st8(top_blob_data, gi, res);
+}
diff --git a/src/layer/vulkan/shader/flatten_pack1to8.comp b/src/layer/vulkan/shader/flatten_pack1to8.comp
new file mode 100644
index 000000000..b87c1ece1
--- /dev/null
+++ b/src/layer/vulkan/shader/flatten_pack1to8.comp
@@ -0,0 +1,73 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= 1 || gz >= 1)
+        return;
+
+    ivec4 i4 = gx * 8 + ivec4(0, 1, 2, 3);
+    ivec4 ii4 = i4 + 4;
+
+    int size = p.w * p.h;
+
+    ivec4 z4 = i4 / size;
+    ivec4 y4 = i4 % size / p.w;
+    ivec4 x4 = i4 % size % p.w;
+
+    ivec4 zz4 = ii4 / size;
+    ivec4 yy4 = ii4 % size / p.w;
+    ivec4 xx4 = ii4 % size % p.w;
+
+    ivec4 v_offset = z4 * p.cstep + y4 * p.w + x4;
+    ivec4 vv_offset = zz4 * p.cstep + yy4 * p.w + xx4;
+
+    buffer_cp1to8(top_blob_data, gx, bottom_blob_data, v_offset, vv_offset);
+}
diff --git a/src/layer/vulkan/shader/flatten_pack4.comp b/src/layer/vulkan/shader/flatten_pack4.comp
index 7a79bb36b..49215a6eb 100644
--- a/src/layer/vulkan/shader/flatten_pack4.comp
+++ b/src/layer/vulkan/shader/flatten_pack4.comp
@@ -26,7 +26,7 @@ layout (local_size_y_id = 234) in;
 layout (local_size_z_id = 235) in;
 
 #if NCNN_fp16_packed
-layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 0) readonly buffer bottom_blob { sfpvec2 bottom_blob_data[]; };
 #else
 layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
 #endif
@@ -60,15 +60,15 @@ void main()
     ivec4 i4 = gx * 4 + ivec4(0, 1, 2, 3);
 
     ivec4 v_offset;
-    ivec4 lane4;
+    ivec4 lane2;
 
     if (p.dims == 2)
     {
         ivec4 y4 = i4 / p.w;
         ivec4 x4 = i4 % p.w;
 
-        v_offset = (y4 / 4) * p.w + x4;
-        lane4 = y4 % 4;
+        v_offset = ((y4 / 4) * p.w + x4) * 2 + (y4 % 4) / 2;
+        lane2 = y4 % 2;
     }
     else // if (p.dims == 3)
     {
@@ -78,16 +78,16 @@ void main()
         ivec4 y4 = i4 % size / p.w;
         ivec4 x4 = i4 % size % p.w;
 
-        v_offset = (z4 / 4) * p.cstep + y4 * p.w + x4;
-        lane4 = z4 % 4;
+        v_offset = ((z4 / 4) * p.cstep + y4 * p.w + x4) * 2 + (z4 % 4) / 2;
+        lane2 = z4 % 2;
     }
 
-    vec4 vr = buffer_ld4(bottom_blob_data, v_offset.r);
-    vec4 vg = buffer_ld4(bottom_blob_data, v_offset.g);
-    vec4 vb = buffer_ld4(bottom_blob_data, v_offset.b);
-    vec4 va = buffer_ld4(bottom_blob_data, v_offset.a);
+    afpvec2 vr = buffer_ld2(bottom_blob_data, v_offset.r);
+    afpvec2 vg = buffer_ld2(bottom_blob_data, v_offset.g);
+    afpvec2 vb = buffer_ld2(bottom_blob_data, v_offset.b);
+    afpvec2 va = buffer_ld2(bottom_blob_data, v_offset.a);
 
-    vec4 v = vec4(vr[lane4.r], vg[lane4.g], vb[lane4.b], va[lane4.a]);
+    afpvec4 v = afpvec4(vr[lane2.r], vg[lane2.g], vb[lane2.b], va[lane2.a]);
 
     buffer_st4(top_blob_data, gx, v);
 #else
diff --git a/src/layer/vulkan/shader/flatten_pack8.comp b/src/layer/vulkan/shader/flatten_pack8.comp
new file mode 100644
index 000000000..5e8ea4e21
--- /dev/null
+++ b/src/layer/vulkan/shader/flatten_pack8.comp
@@ -0,0 +1,152 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+#if NCNN_fp16_packed
+layout (binding = 0) readonly buffer bottom_blob { sfpvec2 bottom_blob_data[]; };
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+#endif
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= 1 || gz >= 1)
+        return;
+
+#if NCNN_fp16_packed
+    ivec4 i4 = gx * 8 + ivec4(0, 1, 2, 3);
+    ivec4 ii4 = i4 + 4;
+
+    ivec4 v_offset;
+    ivec4 lane4;
+    ivec4 vv_offset;
+    ivec4 lane8;
+
+    if (p.dims == 2)
+    {
+        ivec4 y4 = i4 / p.w;
+        ivec4 x4 = i4 % p.w;
+
+        ivec4 yy4 = ii4 / p.w;
+        ivec4 xx4 = ii4 % p.w;
+
+        v_offset = ((y4 / 8) * p.w + x4) * 2 + (y4 % 8) / 2;
+        lane4 = y4 % 2;
+
+        vv_offset = ((yy4 / 8) * p.w + xx4) * 2 + (yy4 % 8) / 2;
+        lane8 = yy4 % 2;
+    }
+    else // if (p.dims == 3)
+    {
+        int size = p.w * p.h;
+
+        ivec4 z4 = i4 / size;
+        ivec4 y4 = i4 % size / p.w;
+        ivec4 x4 = i4 % size % p.w;
+
+        ivec4 zz4 = ii4 / size;
+        ivec4 yy4 = ii4 % size / p.w;
+        ivec4 xx4 = ii4 % size % p.w;
+
+        v_offset = ((z4 / 8) * p.cstep + y4 * p.w + x4) * 2 + (z4 % 4) / 2;
+        lane4 = z4 % 2;
+
+        vv_offset = ((zz4 / 8) * p.cstep + yy4 * p.w + xx4) * 2 + (zz4 % 4) / 2;
+        lane8 = zz4 % 2;
+    }
+
+    afpvec2 vr = buffer_ld2(bottom_blob_data, v_offset.r);
+    afpvec2 vg = buffer_ld2(bottom_blob_data, v_offset.g);
+    afpvec2 vb = buffer_ld2(bottom_blob_data, v_offset.b);
+    afpvec2 va = buffer_ld2(bottom_blob_data, v_offset.a);
+
+    afpvec2 vvr = buffer_ld2(bottom_blob_data, vv_offset.r);
+    afpvec2 vvg = buffer_ld2(bottom_blob_data, vv_offset.g);
+    afpvec2 vvb = buffer_ld2(bottom_blob_data, vv_offset.b);
+    afpvec2 vva = buffer_ld2(bottom_blob_data, vv_offset.a);
+
+    afpvec8 v = afpvec8(vr[lane4.r], vg[lane4.g], vb[lane4.b], va[lane4.a], vvr[lane8.r], vvg[lane8.g], vvb[lane8.b], vva[lane8.a]);
+
+    buffer_st8(top_blob_data, gx, v);
+#else
+    ivec4 v_offset;
+    ivec4 vv_offset;
+
+    if (p.dims == 2)
+    {
+        int y = gx / p.w;
+        ivec4 i4 = (gx % p.w) * 8 + ivec4(0, 1, 2, 3);
+        ivec4 ii4 = i4 + 4;
+
+        ivec4 k4 = i4 / p.w;
+        ivec4 si4 = i4 % p.w;
+        ivec4 kk4 = ii4 / p.w;
+        ivec4 sii4 = ii4 % p.w;
+
+        v_offset = ivec4(y * 8 * p.w) + si4 * 8 + k4;
+        vv_offset = ivec4(y * 8 * p.w) + sii4 * 8 + kk4;
+    }
+    else // if (p.dims == 3)
+    {
+        int size = p.w * p.h;
+
+        int z = gx / size;
+        ivec4 i4 = (gx % size) * 8 + ivec4(0, 1, 2, 3);
+        ivec4 ii4 = i4 + 4;
+
+        ivec4 k4 = i4 / size;
+        ivec4 si4 = i4 % size;
+        ivec4 kk4 = ii4 / size;
+        ivec4 sii4 = ii4 % size;
+
+        v_offset = ivec4(z * 8 * p.cstep) + si4 * 8 + k4;
+        vv_offset = ivec4(z * 8 * p.cstep) + sii4 * 8 + kk4;
+    }
+
+    buffer_cp1to8(top_blob_data, gx, bottom_blob_data, v_offset, vv_offset);
+#endif
+}
diff --git a/src/layer/vulkan/shader/hardsigmoid_pack8.comp b/src/layer/vulkan/shader/hardsigmoid_pack8.comp
new file mode 100644
index 000000000..84e4a3fd4
--- /dev/null
+++ b/src/layer/vulkan/shader/hardsigmoid_pack8.comp
@@ -0,0 +1,60 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const float alpha = 0;
+layout (constant_id = 1) const float beta = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    v[0] = clamp(v[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
+    v[1] = clamp(v[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/hardswish_pack8.comp b/src/layer/vulkan/shader/hardswish_pack8.comp
new file mode 100644
index 000000000..fe5d7882f
--- /dev/null
+++ b/src/layer/vulkan/shader/hardswish_pack8.comp
@@ -0,0 +1,60 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const float alpha = 0;
+layout (constant_id = 1) const float beta = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    v[0] = v[0] * clamp(v[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
+    v[1] = v[1] * clamp(v[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/innerproduct_pack1to8.comp b/src/layer/vulkan/shader/innerproduct_pack1to8.comp
new file mode 100644
index 000000000..2b06e3f6e
--- /dev/null
+++ b/src/layer/vulkan/shader/innerproduct_pack1to8.comp
@@ -0,0 +1,112 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int bias_term = 0;
+layout (constant_id = 1) const int activation_type = 0;
+layout (constant_id = 2) const float activation_param_0 = 0;
+layout (constant_id = 3) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= 1 || gz >= 1)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    int w_offset = gx * p.w;
+
+    for (int i = 0; i < p.w; i++)
+    {
+        afp v = buffer_ld1(bottom_blob_data, i);
+
+        afpvec8 k = buffer_ld8(weight_data, w_offset + i);
+
+        // sum += v * k;
+        sum[0] += v * k[0];
+        sum[1] += v * k[1];
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    buffer_st8(top_blob_data, gx, sum);
+}
diff --git a/src/layer/vulkan/shader/innerproduct_pack4to8.comp b/src/layer/vulkan/shader/innerproduct_pack4to8.comp
new file mode 100644
index 000000000..0b66b5174
--- /dev/null
+++ b/src/layer/vulkan/shader/innerproduct_pack4to8.comp
@@ -0,0 +1,125 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int bias_term = 0;
+layout (constant_id = 1) const int activation_type = 0;
+layout (constant_id = 2) const float activation_param_0 = 0;
+layout (constant_id = 3) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= 1 || gz >= 1)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    int w_offset = gx * p.w;
+
+    for (int i = 0; i < p.w; i++)
+    {
+        afpvec4 v = buffer_ld4(bottom_blob_data, i);
+
+        afpvec4 k0 = buffer_ld4(weight_data, (w_offset + i) * 8 + 0);
+        afpvec4 k1 = buffer_ld4(weight_data, (w_offset + i) * 8 + 1);
+        afpvec4 k2 = buffer_ld4(weight_data, (w_offset + i) * 8 + 2);
+        afpvec4 k3 = buffer_ld4(weight_data, (w_offset + i) * 8 + 3);
+        afpvec4 k4 = buffer_ld4(weight_data, (w_offset + i) * 8 + 4);
+        afpvec4 k5 = buffer_ld4(weight_data, (w_offset + i) * 8 + 5);
+        afpvec4 k6 = buffer_ld4(weight_data, (w_offset + i) * 8 + 6);
+        afpvec4 k7 = buffer_ld4(weight_data, (w_offset + i) * 8 + 7);
+
+        // sum += v * k;
+        sum[0].r += dot(v, k0);
+        sum[0].g += dot(v, k1);
+        sum[0].b += dot(v, k2);
+        sum[0].a += dot(v, k3);
+        sum[1].r += dot(v, k4);
+        sum[1].g += dot(v, k5);
+        sum[1].b += dot(v, k6);
+        sum[1].a += dot(v, k7);
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    buffer_st8(top_blob_data, gx, sum);
+}
diff --git a/src/layer/vulkan/shader/innerproduct_pack8.comp b/src/layer/vulkan/shader/innerproduct_pack8.comp
new file mode 100644
index 000000000..f7bb49eeb
--- /dev/null
+++ b/src/layer/vulkan/shader/innerproduct_pack8.comp
@@ -0,0 +1,125 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int bias_term = 0;
+layout (constant_id = 1) const int activation_type = 0;
+layout (constant_id = 2) const float activation_param_0 = 0;
+layout (constant_id = 3) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= 1 || gz >= 1)
+        return;
+
+    afpvec8 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld8(bias_data, gx);
+    }
+    else
+    {
+        sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+    }
+
+    int w_offset = gx * p.w;
+
+    for (int i = 0; i < p.w; i++)
+    {
+        afpvec8 v = buffer_ld8(bottom_blob_data, i);
+
+        afpvec8 k0 = buffer_ld8(weight_data, (w_offset + i) * 8 + 0);
+        afpvec8 k1 = buffer_ld8(weight_data, (w_offset + i) * 8 + 1);
+        afpvec8 k2 = buffer_ld8(weight_data, (w_offset + i) * 8 + 2);
+        afpvec8 k3 = buffer_ld8(weight_data, (w_offset + i) * 8 + 3);
+        afpvec8 k4 = buffer_ld8(weight_data, (w_offset + i) * 8 + 4);
+        afpvec8 k5 = buffer_ld8(weight_data, (w_offset + i) * 8 + 5);
+        afpvec8 k6 = buffer_ld8(weight_data, (w_offset + i) * 8 + 6);
+        afpvec8 k7 = buffer_ld8(weight_data, (w_offset + i) * 8 + 7);
+
+        // sum += v * k
+        sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+        sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+        sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+        sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+        sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
+        sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
+        sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
+        sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
+    }
+
+    if (activation_type == 1)
+    {
+        sum[0] = max(sum[0], afp(0.f));
+        sum[1] = max(sum[1], afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum[0] = mix(sum[0], sum[0] * afp(slope), lessThan(sum[0], afpvec4(0.f)));
+        sum[1] = mix(sum[1], sum[1] * afp(slope), lessThan(sum[1], afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum[0] = clamp(sum[0], const_min, const_max);
+        sum[1] = clamp(sum[1], const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum[0] = afp(1.f) / (afp(1.f) + exp(-sum[0]));
+        sum[1] = afp(1.f) / (afp(1.f) + exp(-sum[1]));
+    }
+
+    buffer_st8(top_blob_data, gx, sum);
+}
diff --git a/src/layer/vulkan/shader/innerproduct_pack8to1.comp b/src/layer/vulkan/shader/innerproduct_pack8to1.comp
new file mode 100644
index 000000000..e7e3912cb
--- /dev/null
+++ b/src/layer/vulkan/shader/innerproduct_pack8to1.comp
@@ -0,0 +1,107 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int bias_term = 0;
+layout (constant_id = 1) const int activation_type = 0;
+layout (constant_id = 2) const float activation_param_0 = 0;
+layout (constant_id = 3) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfp bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= 1 || gz >= 1)
+        return;
+
+    afp sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld1(bias_data, gx);
+    }
+    else
+    {
+        sum = afp(0.f);
+    }
+
+    int w_offset = gx * p.w;
+
+    for (int i = 0; i < p.w; i++)
+    {
+        afpvec8 v = buffer_ld8(bottom_blob_data, i);
+
+        afpvec8 k = buffer_ld8(weight_data, w_offset + i);
+
+        // sum += dot(v, k);
+        sum += dot(v[0], k[0]) + dot(v[1], k[1]);
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = sum < afp(0.f) ? sum * slope : sum;
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    buffer_st1(top_blob_data, gx, sum);
+}
diff --git a/src/layer/vulkan/shader/innerproduct_pack8to4.comp b/src/layer/vulkan/shader/innerproduct_pack8to4.comp
new file mode 100644
index 000000000..d5f5edca4
--- /dev/null
+++ b/src/layer/vulkan/shader/innerproduct_pack8to4.comp
@@ -0,0 +1,113 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int bias_term = 0;
+layout (constant_id = 1) const int activation_type = 0;
+layout (constant_id = 2) const float activation_param_0 = 0;
+layout (constant_id = 3) const float activation_param_1 = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
+layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= 1 || gz >= 1)
+        return;
+
+    afpvec4 sum;
+
+    if (bias_term == 1)
+    {
+        sum = buffer_ld4(bias_data, gz);
+    }
+    else
+    {
+        sum = afpvec4(0.f);
+    }
+
+    int w_offset = gx * p.w;
+
+    for (int i = 0; i < p.w; i++)
+    {
+        afpvec8 v = buffer_ld8(bottom_blob_data, i);
+
+        afpvec8 k0 = buffer_ld8(weight_data, (w_offset + i) * 4 + 0);
+        afpvec8 k1 = buffer_ld8(weight_data, (w_offset + i) * 4 + 1);
+        afpvec8 k2 = buffer_ld8(weight_data, (w_offset + i) * 4 + 2);
+        afpvec8 k3 = buffer_ld8(weight_data, (w_offset + i) * 4 + 3);
+
+        // sum += v * k
+        sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
+        sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
+        sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
+        sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
+    }
+
+    if (activation_type == 1)
+    {
+        sum = max(sum, afp(0.f));
+    }
+    if (activation_type == 2)
+    {
+        const afp slope = afp(activation_param_0);
+        sum = mix(sum, sum * afp(slope), lessThan(sum, afpvec4(0.f)));
+    }
+    if (activation_type == 3)
+    {
+        const afp const_min = afp(activation_param_0);
+        const afp const_max = afp(activation_param_1);
+        sum = clamp(sum, const_min, const_max);
+    }
+    if (activation_type == 4)
+    {
+        sum = afp(1.f) / (afp(1.f) + exp(-sum));
+    }
+
+    buffer_st4(top_blob_data, gx, sum);
+}
diff --git a/src/layer/vulkan/shader/packing_1to8.comp b/src/layer/vulkan/shader/packing_1to8.comp
new file mode 100644
index 000000000..622281dd4
--- /dev/null
+++ b/src/layer/vulkan/shader/packing_1to8.comp
@@ -0,0 +1,84 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    ivec4 v_offset;
+    ivec4 vv_offset;
+
+    if (p.dims == 1)
+    {
+        ivec4 x4 = ivec4(gx * 8) + ivec4(0, 1, 2, 3);
+
+        v_offset = x4;
+        vv_offset = x4 + 4;
+    }
+    else if (p.dims == 2)
+    {
+        ivec4 y4 = ivec4(gy * 8) + ivec4(0, 1, 2, 3);
+
+        v_offset = y4 * p.w + gx;
+        vv_offset = (y4 + 4) * p.w + gx;
+    }
+    else // if (p.dims == 3)
+    {
+        ivec4 z4 = ivec4(gz * 8) + ivec4(0, 1, 2, 3);
+
+        v_offset = z4 * p.cstep + ivec4(gy * p.w + gx);
+        vv_offset = (z4 + 4) * p.cstep + ivec4(gy * p.w + gx);
+    }
+
+    int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_cp1to8(top_blob_data, gi, bottom_blob_data, v_offset, vv_offset);
+}
diff --git a/src/layer/vulkan/shader/packing_4to8.comp b/src/layer/vulkan/shader/packing_4to8.comp
new file mode 100644
index 000000000..851600293
--- /dev/null
+++ b/src/layer/vulkan/shader/packing_4to8.comp
@@ -0,0 +1,80 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    ivec2 v_offset;
+
+    if (p.dims == 1)
+    {
+        ivec2 x2 = ivec2(gx * 2) + ivec2(0, 1);
+
+        v_offset = x2;
+    }
+    else if (p.dims == 2)
+    {
+        ivec2 y2 = ivec2(gy * 2) + ivec2(0, 1);
+
+        v_offset = y2 * p.w + gx;
+    }
+    else // if (p.dims == 3)
+    {
+        ivec2 z2 = ivec2(gz * 2) + ivec2(0, 1);
+
+        v_offset = z2 * p.cstep + ivec2(gy * p.w + gx);
+    }
+
+    int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_cp4to8(top_blob_data, gi, bottom_blob_data, v_offset);
+}
diff --git a/src/layer/vulkan/shader/packing_8to1.comp b/src/layer/vulkan/shader/packing_8to1.comp
new file mode 100644
index 000000000..6aa321c6a
--- /dev/null
+++ b/src/layer/vulkan/shader/packing_8to1.comp
@@ -0,0 +1,84 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    ivec4 v_offset;
+    ivec4 vv_offset;
+
+    if (p.dims == 1)
+    {
+        ivec4 x4 = ivec4(gx * 8) + ivec4(0, 1, 2, 3);
+
+        v_offset = x4;
+        vv_offset = x4 + 4;
+    }
+    else if (p.dims == 2)
+    {
+        ivec4 y4 = ivec4(gy * 8) + ivec4(0, 1, 2, 3);
+
+        v_offset = y4 * p.outw + gx;
+        vv_offset = (y4 + 4) * p.outw + gx;
+    }
+    else // if (p.dims == 3)
+    {
+        ivec4 z4 = ivec4(gz * 8) + ivec4(0, 1, 2, 3);
+
+        v_offset = z4 * p.outcstep + ivec4(gy * p.outw + gx);
+        vv_offset = (z4 + 4) * p.outcstep + ivec4(gy * p.outw + gx);
+    }
+
+    int gi = gz * p.cstep + gy * p.w + gx;
+
+    buffer_cp8to1(top_blob_data, v_offset, vv_offset, bottom_blob_data, gi);
+}
diff --git a/src/layer/vulkan/shader/packing_8to4.comp b/src/layer/vulkan/shader/packing_8to4.comp
new file mode 100644
index 000000000..474545a1e
--- /dev/null
+++ b/src/layer/vulkan/shader/packing_8to4.comp
@@ -0,0 +1,80 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    ivec2 v_offset;
+
+    if (p.dims == 1)
+    {
+        ivec2 x2 = ivec2(gx * 2) + ivec2(0, 1);
+
+        v_offset = x2;
+    }
+    else if (p.dims == 2)
+    {
+        ivec2 y2 = ivec2(gy * 2) + ivec2(0, 1);
+
+        v_offset = y2 * p.outw + gx;
+    }
+    else // if (p.dims == 3)
+    {
+        ivec2 z2 = ivec2(gz * 2) + ivec2(0, 1);
+
+        v_offset = z2 * p.outcstep + ivec2(gy * p.outw + gx);
+    }
+
+    int gi = gz * p.cstep + gy * p.w + gx;
+
+    buffer_cp8to4(top_blob_data, v_offset, bottom_blob_data, gi);
+}
diff --git a/src/layer/vulkan/shader/padding_pack8.comp b/src/layer/vulkan/shader/padding_pack8.comp
new file mode 100644
index 000000000..6798a11ee
--- /dev/null
+++ b/src/layer/vulkan/shader/padding_pack8.comp
@@ -0,0 +1,104 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int type = 1;
+layout (constant_id = 1) const float value = 0;
+layout (constant_id = 2) const int per_channel_pad = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+layout (binding = 2) readonly buffer per_channel_pad_blob { sfpvec8 per_channel_pad_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int left;
+    int top;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    int x = gx - p.left;
+    int y = gy - p.top;
+
+    if (type == 0)
+    {
+        if (x >= 0 && x < p.w && y >= 0 && y < p.h)
+        {
+            int v_offset = gz * p.cstep + y * p.w + x;
+            buffer_cp8(top_blob_data, gi, bottom_blob_data, v_offset);
+        }
+        else if (per_channel_pad == 1)
+        {
+            buffer_cp8(top_blob_data, gi, per_channel_pad_blob_data, gz);
+        }
+        else
+        {
+            afpvec8 v = afpvec8(afpvec4(value), afpvec4(value));
+            buffer_st8(top_blob_data, gi, v);
+        }
+    }
+    else if (type == 1)
+    {
+        x = clamp(x, 0, p.w - 1);
+        y = clamp(y, 0, p.h - 1);
+
+        int v_offset = gz * p.cstep + y * p.w + x;
+        buffer_cp8(top_blob_data, gi, bottom_blob_data, v_offset);
+    }
+    else if (type == 2)
+    {
+        x = abs(x);
+        y = abs(y);
+        x = (p.w - 1) - abs(x - (p.w - 1));
+        y = (p.h - 1) - abs(y - (p.h - 1));
+
+        int v_offset = gz * p.cstep + y * p.w + x;
+        buffer_cp8(top_blob_data, gi, bottom_blob_data, v_offset);
+    }
+}
diff --git a/src/layer/vulkan/shader/pooling_global_pack8.comp b/src/layer/vulkan/shader/pooling_global_pack8.comp
new file mode 100644
index 000000000..63f747445
--- /dev/null
+++ b/src/layer/vulkan/shader/pooling_global_pack8.comp
@@ -0,0 +1,93 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+layout (constant_id = 0) const int pooling_type = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    int size = p.w * p.h;
+    int v_offset = gx * p.cstep;
+
+    afpvec8 res;
+
+    if (pooling_type == 0)
+    {
+        res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
+
+        for (int i = 0; i < size; i++)
+        {
+            afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + i);
+            res[0] = max(res[0], v[0]);
+            res[1] = max(res[1], v[1]);
+        }
+    }
+    else if (pooling_type == 1)
+    {
+        res = afpvec8(afpvec4(0.f), afpvec4(0.f));
+
+        for (int i = 0; i < size; i++)
+        {
+            afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + i);
+            res[0] += v[0];
+            res[1] += v[1];
+        }
+
+        afp area = afp(size);
+        res[0] /= area;
+        res[1] /= area;
+    }
+
+    buffer_st8(top_blob_data, gx, res);
+}
diff --git a/src/layer/vulkan/shader/pooling_pack8.comp b/src/layer/vulkan/shader/pooling_pack8.comp
new file mode 100644
index 000000000..195a29c24
--- /dev/null
+++ b/src/layer/vulkan/shader/pooling_pack8.comp
@@ -0,0 +1,163 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+layout (constant_id = 0) const int pooling_type = 0;
+layout (constant_id = 1) const int kernel_w = 1;
+layout (constant_id = 2) const int kernel_h = 1;
+layout (constant_id = 3) const int stride_w = 1;
+layout (constant_id = 4) const int stride_h = 1;
+layout (constant_id = 5) const int pad_left = 0;
+layout (constant_id = 6) const int pad_right = 0;
+layout (constant_id = 7) const int pad_top = 0;
+layout (constant_id = 8) const int pad_bottom = 0;
+layout (constant_id = 9) const int global_pooling = 0;
+layout (constant_id = 10) const int pad_mode = 0;
+layout (constant_id = 11) const int avgpool_count_include_pad = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int wtailpad;
+    int htailpad;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 res;
+
+    if (pooling_type == 0)
+    {
+        res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
+
+        int v_offset = gz * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
+                res[0] = max(res[0], v[0]);
+                res[1] = max(res[1], v[1]);
+            }
+
+            v_offset += p.w;
+        }
+    }
+    else if (pooling_type == 1 && avgpool_count_include_pad == 0)
+    {
+        res = afpvec8(afpvec4(0.f), afpvec4(0.f));
+
+        int sx = gx * stride_w;
+        int sy = gy * stride_h;
+
+        int v_offset = gz * p.cstep + sy * p.w + sx;
+        int area = 0;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            if (sy + y < pad_top)
+            {
+                v_offset += p.w;
+                continue;
+            }
+
+            if (sy + y >= p.h - pad_bottom - p.htailpad)
+                break;
+
+            for (int x = 0; x < kernel_w; x++)
+            {
+                if (sx + x < pad_left)
+                {
+                    continue;
+                }
+
+                if (sx + x >= p.w - pad_right - p.wtailpad)
+                    break;
+
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
+                res[0] += v[0];
+                res[1] += v[1];
+                area += 1;
+            }
+
+            v_offset += p.w;
+        }
+
+        res[0] /= afp(area);
+        res[1] /= afp(area);
+    }
+    else if (pooling_type == 1 && avgpool_count_include_pad == 1)
+    {
+        res = afpvec8(afpvec4(0.f), afpvec4(0.f));
+
+        int v_offset = gz * p.cstep + gy * stride_h * p.w + gx * stride_w;
+
+        for (int y = 0; y < kernel_h; y++)
+        {
+            for (int x = 0; x < kernel_w; x++)
+            {
+                afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
+                res[0] += v[0];
+                res[1] += v[1];
+            }
+
+            v_offset += p.w;
+        }
+
+        afp area = afp(kernel_w * kernel_h);
+        res[0] /= area;
+        res[1] /= area;
+    }
+
+    const int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    buffer_st8(top_blob_data, gi, res);
+}
diff --git a/src/layer/vulkan/shader/prelu_pack8.comp b/src/layer/vulkan/shader/prelu_pack8.comp
new file mode 100644
index 000000000..27cc60d17
--- /dev/null
+++ b/src/layer/vulkan/shader/prelu_pack8.comp
@@ -0,0 +1,64 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int num_slope = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer slope_blob { sfpvec8 slope_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    const ivec3 gxyz = ivec3(gx, gy, gz);
+    const int abi = num_slope > 1 ? gxyz[p.dims - 1] : 0;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+    afpvec8 slope = buffer_ld8(slope_blob_data, abi);
+
+    v[0] = mix(v[0], v[0] * slope[0], lessThan(v[0], afpvec4(0.f)));
+    v[1] = mix(v[1], v[1] * slope[1], lessThan(v[1], afpvec4(0.f)));
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/relu_pack8.comp b/src/layer/vulkan/shader/relu_pack8.comp
new file mode 100644
index 000000000..5d16f4e43
--- /dev/null
+++ b/src/layer/vulkan/shader/relu_pack8.comp
@@ -0,0 +1,67 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const float slope = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    if (slope == 0)
+    {
+        v[0] = max(v[0], afp(0.f));
+        v[1] = max(v[1], afp(0.f));
+    }
+    else
+    {
+        v[0] = mix(v[0], v[0] * afp(slope), lessThan(v[0], afpvec4(0.f)));
+        v[1] = mix(v[1], v[1] * afp(slope), lessThan(v[1], afpvec4(0.f)));
+    }
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/scale_pack8.comp b/src/layer/vulkan/shader/scale_pack8.comp
new file mode 100644
index 000000000..f069e225f
--- /dev/null
+++ b/src/layer/vulkan/shader/scale_pack8.comp
@@ -0,0 +1,74 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int bias_term = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer scale_blob { sfpvec8 scale_blob_data[]; };
+layout (binding = 2) readonly buffer bias_blob { sfpvec8 bias_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    const ivec3 gxyz = ivec3(gx, gy, gz);
+    const int abi = gxyz[p.dims - 1];
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+    afpvec8 scale = buffer_ld8(scale_blob_data, abi);
+
+    if (bias_term == 1)
+    {
+        afpvec8 bias = buffer_ld8(bias_blob_data, abi);
+        v[0] = scale[0] * v[0] + bias[0];
+        v[1] = scale[1] * v[1] + bias[1];
+    }
+    else
+    {
+        v[0] = scale[0] * v[0];
+        v[1] = scale[1] * v[1];
+    }
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/shufflechannel_pack8.comp b/src/layer/vulkan/shader/shufflechannel_pack8.comp
new file mode 100644
index 000000000..4836b8d09
--- /dev/null
+++ b/src/layer/vulkan/shader/shufflechannel_pack8.comp
@@ -0,0 +1,91 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int group = 1;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int channels_g = p.c * 8 / group;
+
+    ivec4 gz4 = ivec4(gz * 8) + ivec4(0, 1, 2, 3);
+    ivec4 gzz4 = gz4 + 4;
+
+    ivec4 z4 = (gz4 % group) * channels_g + (gz4 / group);
+    ivec4 zz4 = (gzz4 % group) * channels_g + (gzz4 / group);
+
+    ivec4 v_offset = (z4 / 8) * p.cstep + gy * p.w + gx;
+    ivec4 vv_offset = (zz4 / 8) * p.cstep + gy * p.w + gx;
+
+    ivec4 sz4 = (z4 % 8) / 4;
+    ivec4 szz4 = (zz4 % 8) / 4;
+
+    ivec4 lane4 = z4 % 4;
+    ivec4 lane8 = zz4 % 4;
+
+    // v = v4[lane]
+    int gi = gz * p.outcstep + gy * p.outw + gx;
+
+    afpvec8 vr = buffer_ld8(bottom_blob_data, v_offset.r);
+    afpvec8 vg = buffer_ld8(bottom_blob_data, v_offset.g);
+    afpvec8 vb = buffer_ld8(bottom_blob_data, v_offset.b);
+    afpvec8 va = buffer_ld8(bottom_blob_data, v_offset.a);
+
+    afpvec8 vvr = buffer_ld8(bottom_blob_data, vv_offset.r);
+    afpvec8 vvg = buffer_ld8(bottom_blob_data, vv_offset.g);
+    afpvec8 vvb = buffer_ld8(bottom_blob_data, vv_offset.b);
+    afpvec8 vva = buffer_ld8(bottom_blob_data, vv_offset.a);
+
+    afpvec8 v = afpvec8(vr[sz4.r][lane4.r], vg[sz4.g][lane4.g], vb[sz4.b][lane4.b], va[sz4.a][lane4.a], vvr[szz4.r][lane8.r], vvg[szz4.g][lane8.g], vvb[szz4.b][lane8.b], vva[szz4.a][lane8.a]);
+
+    buffer_st8(top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/sigmoid_pack8.comp b/src/layer/vulkan/shader/sigmoid_pack8.comp
new file mode 100644
index 000000000..e9bcc6d5f
--- /dev/null
+++ b/src/layer/vulkan/shader/sigmoid_pack8.comp
@@ -0,0 +1,57 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    v[0] = afp(1.f) / (afp(1.f) + exp(-v[0]));
+    v[1] = afp(1.f) / (afp(1.f) + exp(-v[1]));
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/slice_pack1to8.comp b/src/layer/vulkan/shader/slice_pack1to8.comp
new file mode 100644
index 000000000..0a9e5b91b
--- /dev/null
+++ b/src/layer/vulkan/shader/slice_pack1to8.comp
@@ -0,0 +1,75 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int offset;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    ivec3 gxyz = ivec3(gx, gy, gz);
+
+    gxyz[p.dims - 1] *= 8;
+    gxyz[p.dims - 1 - axis] += p.offset;
+
+    int v_offset_0 = gxyz.z * p.outcstep + gxyz.y * p.outw + gxyz.x;
+
+    ivec3 gxyz4 = ivec3(1, p.outw, p.outcstep);
+
+    ivec4 v_offset = v_offset_0 + ivec4(0, 1, 2, 3) * gxyz4[p.dims - 1 - axis];
+    ivec4 vv_offset = v_offset + 4 * gxyz4[p.dims - 1 - axis];
+
+    buffer_cp1to8(top_blob_data, gi, bottom_blob_data, v_offset, vv_offset);
+}
diff --git a/src/layer/vulkan/shader/slice_pack8.comp b/src/layer/vulkan/shader/slice_pack8.comp
new file mode 100644
index 000000000..39c1a722c
--- /dev/null
+++ b/src/layer/vulkan/shader/slice_pack8.comp
@@ -0,0 +1,69 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+
+    int offset;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    ivec3 gxyz = ivec3(gx, gy, gz);
+
+    gxyz[p.dims - 1 - axis] += p.offset;
+
+    int v_offset = gxyz.z * p.outcstep + gxyz.y * p.outw + gxyz.x;
+
+    buffer_cp8(top_blob_data, gi, bottom_blob_data, v_offset);
+}
diff --git a/src/layer/vulkan/shader/softmax_div_sum_pack8.comp b/src/layer/vulkan/shader/softmax_div_sum_pack8.comp
new file mode 100644
index 000000000..48a15da3b
--- /dev/null
+++ b/src/layer/vulkan/shader/softmax_div_sum_pack8.comp
@@ -0,0 +1,93 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2019 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer sum_workspace { sfpvec8 sum_workspace_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    afpvec8 sum;
+
+    if (p.dims == 1) // axis == 0
+    {
+        sum = buffer_ld8(sum_workspace_data, 0);
+    }
+    else if (p.dims == 2 && axis == 0)
+    {
+        sum = buffer_ld8(sum_workspace_data, gx);
+    }
+    else if (p.dims == 2 && axis == 1)
+    {
+        sum = buffer_ld8(sum_workspace_data, gy);
+    }
+    else if (p.dims == 3 && axis == 0)
+    {
+        sum = buffer_ld8(sum_workspace_data, gy * p.w + gx);
+    }
+    else if (p.dims == 3 && axis == 1)
+    {
+        sum = buffer_ld8(sum_workspace_data, gz * p.w + gx);
+    }
+    else if (p.dims == 3 && axis == 2)
+    {
+        sum = buffer_ld8(sum_workspace_data, gz * p.h + gy);
+    }
+
+    v[0] /= sum[0];
+    v[1] /= sum[1];
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/softmax_exp_sub_max_pack8.comp b/src/layer/vulkan/shader/softmax_exp_sub_max_pack8.comp
new file mode 100644
index 000000000..d292cf5ed
--- /dev/null
+++ b/src/layer/vulkan/shader/softmax_exp_sub_max_pack8.comp
@@ -0,0 +1,93 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer max_workspace { sfpvec8 max_workspace_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    afpvec8 max_value;
+
+    if (p.dims == 1) // axis == 0
+    {
+        max_value = buffer_ld8(max_workspace_data, 0);
+    }
+    else if (p.dims == 2 && axis == 0)
+    {
+        max_value = buffer_ld8(max_workspace_data, gx);
+    }
+    else if (p.dims == 2 && axis == 1)
+    {
+        max_value = buffer_ld8(max_workspace_data, gy);
+    }
+    else if (p.dims == 3 && axis == 0)
+    {
+        max_value = buffer_ld8(max_workspace_data, gy * p.w + gx);
+    }
+    else if (p.dims == 3 && axis == 1)
+    {
+        max_value = buffer_ld8(max_workspace_data, gz * p.w + gx);
+    }
+    else if (p.dims == 3 && axis == 2)
+    {
+        max_value = buffer_ld8(max_workspace_data, gz * p.h + gy);
+    }
+
+    v[0] = exp(v[0] - max_value[0]);
+    v[1] = exp(v[1] - max_value[1]);
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/softmax_reduce_max_pack8.comp b/src/layer/vulkan/shader/softmax_reduce_max_pack8.comp
new file mode 100644
index 000000000..ef766f22b
--- /dev/null
+++ b/src/layer/vulkan/shader/softmax_reduce_max_pack8.comp
@@ -0,0 +1,148 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) writeonly buffer max_workspace { sfpvec8 max_workspace_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 max_value = afpvec8(afpvec4(-99999999.f), afpvec4(-99999999.f));
+
+    if (p.dims == 1) // axis == 0
+    {
+        for (int i = 0; i < p.w; i++)
+        {
+            afpvec8 v = buffer_ld8(bottom_top_blob_data, i);
+            max_value[0] = max(max_value[0], v[0]);
+            max_value[1] = max(max_value[1], v[1]);
+        }
+        afpvec4 max4 = max(max_value[0], max_value[1]);
+        afpvec2 max2 = max(max4.rg, max4.ba);
+        afp max1 = max(max2.r, max2.g);
+        max_value = afpvec8(afpvec4(max1), afpvec4(max1));
+        buffer_st8(max_workspace_data, 0, max_value);
+        return;
+    }
+
+    if (p.dims == 2 && axis == 0)
+    {
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = i * p.w + gx;
+            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
+            max_value[0] = max(max_value[0], v[0]);
+            max_value[1] = max(max_value[1], v[1]);
+        }
+        afpvec4 max4 = max(max_value[0], max_value[1]);
+        afpvec2 max2 = max(max4.rg, max4.ba);
+        afp max1 = max(max2.r, max2.g);
+        max_value = afpvec8(afpvec4(max1), afpvec4(max1));
+        buffer_st8(max_workspace_data, gx, max_value);
+        return;
+    }
+
+    if (p.dims == 2 && axis == 1)
+    {
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gx * p.w + i;
+            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
+            max_value[0] = max(max_value[0], v[0]);
+            max_value[1] = max(max_value[1], v[1]);
+        }
+        buffer_st8(max_workspace_data, gx, max_value);
+        return;
+    }
+
+    if (p.dims == 3 && axis == 0)
+    {
+        for (int i = 0; i < p.c; i++)
+        {
+            int v_offset = i * p.cstep + gy * p.w + gx;
+            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
+            max_value[0] = max(max_value[0], v[0]);
+            max_value[1] = max(max_value[1], v[1]);
+        }
+        afpvec4 max4 = max(max_value[0], max_value[1]);
+        afpvec2 max2 = max(max4.rg, max4.ba);
+        afp max1 = max(max2.r, max2.g);
+        max_value = afpvec8(afpvec4(max1), afpvec4(max1));
+        buffer_st8(max_workspace_data, gy * p.w + gx, max_value);
+        return;
+    }
+
+    if (p.dims == 3 && axis == 1)
+    {
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = gy * p.cstep + i * p.w + gx;
+            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
+            max_value[0] = max(max_value[0], v[0]);
+            max_value[1] = max(max_value[1], v[1]);
+        }
+        buffer_st8(max_workspace_data, gy * p.w + gx, max_value);
+        return;
+    }
+
+    if (p.dims == 3 && axis == 2)
+    {
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gy * p.cstep + gx * p.w + i;
+            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
+            max_value[0] = max(max_value[0], v[0]);
+            max_value[1] = max(max_value[1], v[1]);
+        }
+        buffer_st8(max_workspace_data, gy * p.h + gx, max_value);
+        return;
+    }
+}
diff --git a/src/layer/vulkan/shader/softmax_reduce_sum_pack8.comp b/src/layer/vulkan/shader/softmax_reduce_sum_pack8.comp
new file mode 100644
index 000000000..ea70e42f3
--- /dev/null
+++ b/src/layer/vulkan/shader/softmax_reduce_sum_pack8.comp
@@ -0,0 +1,136 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int axis = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) readonly buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) writeonly buffer sum_workspace { sfpvec8 sum_workspace_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outdims;
+    int outw;
+    int outh;
+    int outc;
+    int outcstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.outw || gy >= p.outh || gz >= p.outc)
+        return;
+
+    afpvec8 sum_value = afpvec8(afpvec4(0.f), afpvec4(0.f));
+
+    if (p.dims == 1) // axis == 0
+    {
+        for (int i = 0; i < p.w; i++)
+        {
+            sum_value += buffer_ld8(bottom_top_blob_data, i);
+        }
+        afpvec4 sum4 = sum_value[0] + sum_value[1];
+        afpvec2 sum2 = sum4.rg + sum4.ba;
+        afp sum1 = sum2.r + sum2.g;
+        sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
+        buffer_st8(sum_workspace_data, 0, sum_value);
+        return;
+    }
+
+    if (p.dims == 2 && axis == 0)
+    {
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = i * p.w + gx;
+            sum_value += buffer_ld8(bottom_top_blob_data, v_offset);
+        }
+        afpvec4 sum4 = sum_value[0] + sum_value[1];
+        afpvec2 sum2 = sum4.rg + sum4.ba;
+        afp sum1 = sum2.r + sum2.g;
+        sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
+        buffer_st8(sum_workspace_data, gx, sum_value);
+        return;
+    }
+
+    if (p.dims == 2 && axis == 1)
+    {
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gx * p.w + i;
+            sum_value += buffer_ld8(bottom_top_blob_data, v_offset);
+        }
+        buffer_st8(sum_workspace_data, gx, sum_value);
+        return;
+    }
+
+    if (p.dims == 3 && axis == 0)
+    {
+        for (int i = 0; i < p.c; i++)
+        {
+            int v_offset = i * p.cstep + gy * p.w + gx;
+            sum_value += buffer_ld8(bottom_top_blob_data, v_offset);
+        }
+        afpvec4 sum4 = sum_value[0] + sum_value[1];
+        afpvec2 sum2 = sum4.rg + sum4.ba;
+        afp sum1 = sum2.r + sum2.g;
+        sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
+        buffer_st8(sum_workspace_data, gy * p.w + gx, sum_value);
+        return;
+    }
+
+    if (p.dims == 3 && axis == 1)
+    {
+        for (int i = 0; i < p.h; i++)
+        {
+            int v_offset = gy * p.cstep + i * p.w + gx;
+            sum_value += buffer_ld8(bottom_top_blob_data, v_offset);
+        }
+        buffer_st8(sum_workspace_data, gy * p.w + gx, sum_value);
+        return;
+    }
+
+    if (p.dims == 3 && axis == 2)
+    {
+        for (int i = 0; i < p.w; i++)
+        {
+            int v_offset = gy * p.cstep + gx * p.w + i;
+            sum_value += buffer_ld8(bottom_top_blob_data, v_offset);
+        }
+        buffer_st8(sum_workspace_data, gy * p.h + gx, sum_value);
+        return;
+    }
+}
diff --git a/src/layer/vulkan/shader/tanh_pack8.comp b/src/layer/vulkan/shader/tanh_pack8.comp
new file mode 100644
index 000000000..96eb050f1
--- /dev/null
+++ b/src/layer/vulkan/shader/tanh_pack8.comp
@@ -0,0 +1,57 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    v[0] = tanh(v[0]);
+    v[1] = tanh(v[1]);
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}
diff --git a/src/layer/vulkan/shader/unaryop_pack8.comp b/src/layer/vulkan/shader/unaryop_pack8.comp
new file mode 100644
index 000000000..bbc8f08ee
--- /dev/null
+++ b/src/layer/vulkan/shader/unaryop_pack8.comp
@@ -0,0 +1,144 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2020 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.
+
+#version 450
+
+#if NCNN_fp16_storage
+#extension GL_EXT_shader_16bit_storage: require
+struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+layout (constant_id = 0) const int op_type = 0;
+
+layout (local_size_x_id = 233) in;
+layout (local_size_y_id = 234) in;
+layout (local_size_z_id = 235) in;
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int dims;
+    int w;
+    int h;
+    int c;
+    int cstep;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.w || gy >= p.h || gz >= p.c)
+        return;
+
+    const int gi = gz * p.cstep + gy * p.w + gx;
+
+    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
+
+    afpvec8 res;
+
+    if (op_type == 0)
+    {
+        res[0] = abs(v[0]);
+        res[1] = abs(v[1]);
+    }
+    if (op_type == 1)
+    {
+        res[0] = -v[0];
+        res[1] = -v[1];
+    }
+    if (op_type == 2)
+    {
+        res[0] = floor(v[0]);
+        res[1] = floor(v[1]);
+    }
+    if (op_type == 3)
+    {
+        res[0] = ceil(v[0]);
+        res[1] = ceil(v[1]);
+    }
+    if (op_type == 4)
+    {
+        res[0] = v[0] * v[0];
+        res[1] = v[1] * v[1];
+    }
+    if (op_type == 5)
+    {
+        res[0] = sqrt(v[0]);
+        res[1] = sqrt(v[1]);
+    }
+    if (op_type == 6)
+    {
+        res[0] = inversesqrt(v[0]);
+        res[1] = inversesqrt(v[1]);
+    }
+    if (op_type == 7)
+    {
+        res[0] = exp(v[0]);
+        res[1] = exp(v[1]);
+    }
+    if (op_type == 8)
+    {
+        res[0] = log(v[0]);
+        res[1] = log(v[1]);
+    }
+    if (op_type == 9)
+    {
+        res[0] = sin(v[0]);
+        res[1] = sin(v[1]);
+    }
+    if (op_type == 10)
+    {
+        res[0] = cos(v[0]);
+        res[1] = cos(v[1]);
+    }
+    if (op_type == 11)
+    {
+        res[0] = tan(v[0]);
+        res[1] = tan(v[1]);
+    }
+    if (op_type == 12)
+    {
+        res[0] = asin(v[0]);
+        res[1] = asin(v[1]);
+    }
+    if (op_type == 13)
+    {
+        res[0] = acos(v[0]);
+        res[1] = acos(v[1]);
+    }
+    if (op_type == 14)
+    {
+        res[0] = atan(v[0]);
+        res[1] = atan(v[1]);
+    }
+    if (op_type == 15)
+    {
+        res[0] = afp(1.f) / v[0];
+        res[1] = afp(1.f) / v[1];
+    }
+    if (op_type == 16)
+    {
+        res[0] = tanh(v[0]);
+        res[1] = tanh(v[1]);
+    }
+
+    buffer_st8(bottom_top_blob_data, gi, res);
+}