add more pack8 shaders

src/layer/vulkan/shader/instancenorm_coeffs_pack8.comp

@@ -0,0 +1,65 @@
+// 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 eps = 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) writeonly buffer coeffs_blob { sfpvec8 coeffs_blob_data[]; };
+layout (binding = 1) readonly buffer mean_blob { sfpvec8 mean_data[]; };
+layout (binding = 2) readonly buffer var_blob { sfpvec8 var_data[]; };
+layout (binding = 3) readonly buffer gamma_blob { sfpvec8 gamma_data[]; };
+layout (binding = 4) readonly buffer beta_blob { sfpvec8 beta_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int w;
+} 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 >= 1 || gz >= 1)
+        return;
+
+    afpvec8 mean = buffer_ld8(mean_data, gx);
+    afpvec8 var = buffer_ld8(var_data, gx);
+    afpvec8 gamma = buffer_ld8(gamma_data, gx);
+    afpvec8 beta = buffer_ld8(beta_data, gx);
+
+    afpvec8 a;
+    afpvec8 b;
+    a[0] = gamma[0] / (sqrt(var[0] + afp(eps)));
+    a[1] = gamma[1] / (sqrt(var[1] + afp(eps)));
+    b[0] = - mean[0] * a[0] + beta[0];
+    b[1] = - mean[1] * a[1] + beta[1];
+
+    buffer_st8(coeffs_blob_data, gx*2, a);
+    buffer_st8(coeffs_blob_data, gx*2 +1, b);
+}

src/layer/vulkan/shader/instancenorm_norm_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) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer coeffs_blob { sfpvec8 coeffs_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 a = buffer_ld8(coeffs_blob_data, gz*2);
+    afpvec8 b = buffer_ld8(coeffs_blob_data, gz*2 +1);
+
+    v[0] = v[0] * a[0] + b[0];
+    v[1] = v[1] * a[1] + b[1];
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}

src/layer/vulkan/shader/instancenorm_reduce_mean_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) readonly buffer bottom_top_blob { mat2x4 bottom_top_blob_data[]; };
+layout (binding = 1) writeonly buffer mean_blob { sfpvec8 mean_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int size;
+    int c;
+    int cstep;
+    float area;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= p.c || gy >= 1 || gz >= 1)
+        return;
+
+    afpvec8 sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
+
+    int v_offset = gx * p.cstep;
+
+    for (int i = 0; i < p.size; i++)
+    {
+        sum += afpvec8(bottom_top_blob_data[v_offset]);
+        v_offset += 1;
+    }
+
+    afpvec8 mean;
+    mean[0] = sum[0] / afp(p.area);
+    mean[1] = sum[1] / afp(p.area);
+
+    buffer_st8(mean_data, gx, mean);
+}

src/layer/vulkan/shader/instancenorm_reduce_sum4_fp16_to_fp32_pack8.comp

@@ -0,0 +1,92 @@
+// 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_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) writeonly buffer sum_blob { mat2x4 sum_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int size;
+    int c;
+    int cstep;
+
+    int outsize;
+    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.outsize || gy >= 1 || gz >= p.outc)
+        return;
+
+    int gi = gz * p.outcstep + gx;
+
+    int sx = gx * 4;
+
+    int v_offset = gz * p.cstep + sx;
+
+    if (sx == p.size - 1)
+    {
+        sum_blob_data[gi] = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset));
+    }
+    else if (sx == p.size - 2)
+    {
+        mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset));
+        mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset + 1));
+
+        mat2x4 sum = v0 + v1;
+
+        sum_blob_data[gi] = sum;
+    }
+    else if (sx == p.size - 3)
+    {
+        mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset));
+        mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset + 1));
+        mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset + 2));
+
+        mat2x4 sum = v0 + v1 + v2;
+
+        sum_blob_data[gi] = sum;
+    }
+    else
+    {
+        mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset));
+        mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset + 1));
+        mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset + 2));
+        mat2x4 v3 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset + 3));
+
+        mat2x4 sum = v0 + v1 + v2 + v3;
+
+        sum_blob_data[gi] = sum;
+    }
+}

src/layer/vulkan/shader/instancenorm_reduce_sum4_fp32_pack8.comp

@@ -0,0 +1,92 @@
+// 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 square_blob { mat2x4 square_blob_data[]; };
+layout (binding = 1) writeonly buffer sqsum_blob { mat2x4 sqsum_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int size;
+    int c;
+    int cstep;
+
+    int outsize;
+    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.outsize || gy >= 1 || gz >= p.outc)
+        return;
+
+    int gi = gz * p.outcstep + gx;
+
+    int sx = gx * 4;
+
+    int v_offset = gz * p.cstep + sx;
+
+    if (sx == p.size - 1)
+    {
+        sqsum_blob_data[gi] = square_blob_data[v_offset];
+    }
+    else if (sx == p.size - 2)
+    {
+        mat2x4 v0 = square_blob_data[v_offset];
+        mat2x4 v1 = square_blob_data[v_offset + 1];
+
+        mat2x4 sum = v0 + v1;
+
+        sqsum_blob_data[gi] = sum;
+    }
+    else if (sx == p.size - 3)
+    {
+        mat2x4 v0 = square_blob_data[v_offset];
+        mat2x4 v1 = square_blob_data[v_offset + 1];
+        mat2x4 v2 = square_blob_data[v_offset + 2];
+
+        mat2x4 sum = v0 + v1 + v2;
+
+        sqsum_blob_data[gi] = sum;
+    }
+    else
+    {
+        mat2x4 v0 = square_blob_data[v_offset];
+        mat2x4 v1 = square_blob_data[v_offset + 1];
+        mat2x4 v2 = square_blob_data[v_offset + 2];
+        mat2x4 v3 = square_blob_data[v_offset + 3];
+
+        mat2x4 sum = v0 + v1 + v2 + v3;
+
+        sqsum_blob_data[gi] = sum;
+    }
+}

src/layer/vulkan/shader/instancenorm_sub_mean_square_pack8.comp

@@ -0,0 +1,62 @@
+// 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_top_blob { sfpvec8 bottom_top_blob_data[]; };
+layout (binding = 1) readonly buffer mean_blob { sfpvec8 mean_data[]; };
+layout (binding = 2) writeonly buffer square_blob { mat2x4 square_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 mean = buffer_ld8(mean_data, gz);
+
+    v[0] = v[0] - mean[0];
+    v[1] = v[1] - mean[1];
+    v[0] = v[0] * v[0];
+    v[1] = v[1] * v[1];
+
+    square_blob_data[gi] = mat2x4(v);
+}

src/layer/vulkan/shader/normalize_coeffs_pack8.comp

@@ -0,0 +1,79 @@
+// 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 eps = 0.f;
+layout (constant_id = 1) const int eps_mode = 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 sqsum_blob { mat2x4 sqsum_blob_data[]; };
+layout (binding = 1) writeonly buffer coeffs_blob { sfpvec8 coeffs_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int size;
+    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.size || gy >= 1 || gz >= p.c)
+        return;
+
+    int gi = gz * p.cstep + gx;
+
+    mat2x4 sqsum_8 = sqsum_blob_data[gi];
+
+    vec4 sqsum_4 = sqsum_8[0] + sqsum_8[1];
+
+    afp sqsum = afp(sqsum_4.r + sqsum_4.g + sqsum_4.b + sqsum_4.a);
+
+    afp a;
+
+    if (eps_mode == 0) // caffe/mxnet
+    {
+        a = afp(1.f) / sqrt(sqsum + afp(eps));
+    }
+
+    if (eps_mode == 1) // pytorch
+    {
+        a = afp(1.f) / max(sqrt(sqsum), afp(eps));
+    }
+
+    if (eps_mode == 2) // tensorflow
+    {
+        a = afp(1.f) / sqrt(max(sqsum, afp(eps)));
+    }
+
+    afpvec8 a8 = afpvec8(afpvec4(a), afpvec4(a));
+
+    buffer_st8(coeffs_blob_data, gi, a8);
+}

src/layer/vulkan/shader/normalize_norm_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 across_spatial = 0;
+layout (constant_id = 1) const int across_channel = 0;
+layout (constant_id = 2) const int channel_shared = 0;
+layout (constant_id = 3) const int scale_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 coeffs_blob { sfpvec8 coeffs_blob_data[]; };
+layout (binding = 2) readonly buffer scale_blob { sfpvec8 scale_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 a;
+
+    if (across_spatial == 1 && across_channel == 1)
+    {
+        a = buffer_ld8(coeffs_blob_data, 0);
+    }
+
+    if (across_spatial == 1 && across_channel == 0)
+    {
+        a = buffer_ld8(coeffs_blob_data, gz);
+    }
+
+    if (across_spatial == 0 && across_channel == 1)
+    {
+        a = buffer_ld8(coeffs_blob_data, gy * p.w + gx);
+    }
+
+    v[0] = v[0] * a[0];
+    v[1] = v[1] * a[1];
+
+    if (scale_term == 1)
+    {
+        int si = channel_shared == 1 ? 0 : gz;
+
+        afpvec8 scale = buffer_ld8(scale_blob_data, si);
+
+        v[0] = v[0] * scale[0];
+        v[1] = v[1] * scale[1];
+    }
+
+    buffer_st8(bottom_top_blob_data, gi, v);
+}

src/layer/vulkan/shader/normalize_reduce_sum4_fp16_to_fp32_pack4.comp

@@ -21,12 +21,6 @@
 #extension GL_EXT_shader_explicit_arithmetic_types_float16: require
 #endif
 
-#if NCNN_fp16_packed
-#define sfp2floatvec4(v) vec4(unpackHalf2x16(v.x), unpackHalf2x16(v.y))
-#else
-#define sfp2floatvec4(v) vec4(v)
-#endif
-
 layout (constant_id = 0) const int across_spatial = 0;
 layout (constant_id = 1) const int across_channel = 0;
 
@@ -71,7 +65,7 @@ void main()
         {
             if (sx == p.size - 1)
             {
-                vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset0.r]);
+                vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset0.r));
 
                 vec4 sqsum = v0 * v0;
 
@@ -79,8 +73,8 @@ void main()
             }
             else
             {
-                vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset0.r]);
-                vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset0.g]);
+                vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset0.r));
+                vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset0.g));
 
                 vec4 sqsum = v0 * v0 + v1 * v1;
 
@@ -91,8 +85,8 @@ void main()
         {
             if (sx == p.size - 1)
             {
-                vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset0.r]);
-                vec4 v2 = sfp2floatvec4(bottom_top_blob_data[v_offset1.r]);
+                vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset0.r));
+                vec4 v2 = vec4(buffer_ld4(bottom_top_blob_data, v_offset1.r));
 
                 vec4 sqsum = v0 * v0 + v2 * v2;
 
@@ -100,10 +94,10 @@ void main()
             }
             else
             {
-                vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset0.r]);
-                vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset0.g]);
-                vec4 v2 = sfp2floatvec4(bottom_top_blob_data[v_offset1.r]);
-                vec4 v3 = sfp2floatvec4(bottom_top_blob_data[v_offset1.g]);
+                vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset0.r));
+                vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset0.g));
+                vec4 v2 = vec4(buffer_ld4(bottom_top_blob_data, v_offset1.r));
+                vec4 v3 = vec4(buffer_ld4(bottom_top_blob_data, v_offset1.g));
 
                 vec4 sqsum = v0 * v0 + v1 * v1 + v2 * v2 + v3 * v3;
 
@@ -121,7 +115,7 @@ void main()
 
         if (sx == p.size - 1)
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
 
             vec4 sqsum = v0 * v0;
 
@@ -129,8 +123,8 @@ void main()
         }
         else if (sx == p.size - 2)
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
-            vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset.g]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
+            vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.g));
 
             vec4 sqsum = v0 * v0 + v1 * v1;
 
@@ -138,9 +132,9 @@ void main()
         }
         else if (sx == p.size - 3)
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
-            vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset.g]);
-            vec4 v2 = sfp2floatvec4(bottom_top_blob_data[v_offset.b]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
+            vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.g));
+            vec4 v2 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.b));
 
             vec4 sqsum = v0 * v0 + v1 * v1 + v2 * v2;
 
@@ -148,10 +142,10 @@ void main()
         }
         else
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
-            vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset.g]);
-            vec4 v2 = sfp2floatvec4(bottom_top_blob_data[v_offset.b]);
-            vec4 v3 = sfp2floatvec4(bottom_top_blob_data[v_offset.a]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
+            vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.g));
+            vec4 v2 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.b));
+            vec4 v3 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.a));
 
             vec4 sqsum = v0 * v0 + v1 * v1 + v2 * v2 + v3 * v3;
 
@@ -168,7 +162,7 @@ void main()
 
         if (sz == p.c - 1)
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
 
             vec4 sqsum = v0 * v0;
 
@@ -176,8 +170,8 @@ void main()
         }
         else if (sz == p.c - 2)
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
-            vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset.g]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
+            vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.g));
 
             vec4 sqsum = v0 * v0 + v1 * v1;
 
@@ -185,9 +179,9 @@ void main()
         }
         else if (sz == p.c - 3)
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
-            vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset.g]);
-            vec4 v2 = sfp2floatvec4(bottom_top_blob_data[v_offset.b]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
+            vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.g));
+            vec4 v2 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.b));
 
             vec4 sqsum = v0 * v0 + v1 * v1 + v2 * v2;
 
@@ -195,10 +189,10 @@ void main()
         }
         else
         {
-            vec4 v0 = sfp2floatvec4(bottom_top_blob_data[v_offset.r]);
-            vec4 v1 = sfp2floatvec4(bottom_top_blob_data[v_offset.g]);
-            vec4 v2 = sfp2floatvec4(bottom_top_blob_data[v_offset.b]);
-            vec4 v3 = sfp2floatvec4(bottom_top_blob_data[v_offset.a]);
+            vec4 v0 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.r));
+            vec4 v1 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.g));
+            vec4 v2 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.b));
+            vec4 v3 = vec4(buffer_ld4(bottom_top_blob_data, v_offset.a));
 
             vec4 sqsum = v0 * v0 + v1 * v1 + v2 * v2 + v3 * v3;
 

src/layer/vulkan/shader/normalize_reduce_sum4_fp16_to_fp32_pack8.comp

@@ -0,0 +1,227 @@
+// 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 across_spatial = 0;
+layout (constant_id = 1) const int across_channel = 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 sqsum_blob { mat2x4 sqsum_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int size;
+    int c;
+    int cstep;
+
+    int outsize;
+    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.outsize || gy >= 1 || gz >= p.outc)
+        return;
+
+    int gi = gz * p.outcstep + gx;
+
+    if (across_spatial == 1 && across_channel == 1)
+    {
+        int sz = gz * 2;
+        int sx = gx * 2;
+
+        ivec2 v_offset0 = sz * p.cstep + sx + ivec2(0, 1);
+        ivec2 v_offset1 = v_offset0 + p.cstep;
+
+        if (sz == p.c - 1)
+        {
+            if (sx == p.size - 1)
+            {
+                mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset0.r));
+
+                mat2x4 sqsum;
+                sqsum[0] = v0[0] * v0[0];
+                sqsum[1] = v0[1] * v0[1];
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+            else
+            {
+                mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset0.r));
+                mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset0.g));
+
+                mat2x4 sqsum;
+                sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0];
+                sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1];
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+        }
+        else
+        {
+            if (sx == p.size - 1)
+            {
+                mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset0.r));
+                mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset1.r));
+
+                mat2x4 sqsum;
+                sqsum[0] = v0[0] * v0[0] + v2[0] * v2[0];
+                sqsum[1] = v0[1] * v0[1] + v2[1] * v2[1];
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+            else
+            {
+                mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset0.r));
+                mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset0.g));
+                mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset1.r));
+                mat2x4 v3 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset1.g));
+
+                mat2x4 sqsum;
+                sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0] + v2[0] * v2[0] + v3[0] * v3[0];
+                sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1] + v2[1] * v2[1] + v3[1] * v3[1];
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+        }
+    }
+
+    if (across_spatial == 1 && across_channel == 0)
+    {
+        int sz = gz;
+        int sx = gx * 4;
+
+        ivec4 v_offset = sz * p.cstep + sx + ivec4(0, 1, 2, 3);
+
+        if (sx == p.size - 1)
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0];
+            sqsum[1] = v0[1] * v0[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sx == p.size - 2)
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+            mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.g));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0];
+            sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sx == p.size - 3)
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+            mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.g));
+            mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.b));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0] + v2[0] * v2[0];
+            sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1] + v2[1] * v2[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+            mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.g));
+            mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.b));
+            mat2x4 v3 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.a));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0] + v2[0] * v2[0] + v3[0] * v3[0];
+            sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1] + v2[1] * v2[1] + v3[1] * v3[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+    }
+
+    if (across_spatial == 0 && across_channel == 1)
+    {
+        int sz = gz * 4;
+        int sx = gx;
+
+        ivec4 v_offset = (sz + ivec4(0, 1, 2, 3)) * p.cstep + sx;
+
+        if (sz == p.c - 1)
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0];
+            sqsum[1] = v0[1] * v0[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sz == p.c - 2)
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+            mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.g));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0];
+            sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sz == p.c - 3)
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+            mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.g));
+            mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.b));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0] + v2[0] * v2[0];
+            sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1] + v2[1] * v2[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else
+        {
+            mat2x4 v0 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.r));
+            mat2x4 v1 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.g));
+            mat2x4 v2 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.b));
+            mat2x4 v3 = mat2x4(buffer_ld8(bottom_top_blob_data, v_offset.a));
+
+            mat2x4 sqsum;
+            sqsum[0] = v0[0] * v0[0] + v1[0] * v1[0] + v2[0] * v2[0] + v3[0] * v3[0];
+            sqsum[1] = v0[1] * v0[1] + v1[1] * v1[1] + v2[1] * v2[1] + v3[1] * v3[1];
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+    }
+}

src/layer/vulkan/shader/normalize_reduce_sum4_fp32_pack8.comp

@@ -0,0 +1,203 @@
+// 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 across_spatial = 0;
+layout (constant_id = 1) const int across_channel = 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 square_blob { mat2x4 square_blob_data[]; };
+layout (binding = 1) writeonly buffer sqsum_blob { mat2x4 sqsum_blob_data[]; };
+
+layout (push_constant) uniform parameter
+{
+    int size;
+    int c;
+    int cstep;
+
+    int outsize;
+    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.outsize || gy >= 1 || gz >= p.outc)
+        return;
+
+    int gi = gz * p.outcstep + gx;
+
+    if (across_spatial == 1 && across_channel == 1)
+    {
+        int sz = gz * 2;
+        int sx = gx * 2;
+
+        ivec2 v_offset0 = sz * p.cstep + sx + ivec2(0, 1);
+        ivec2 v_offset1 = v_offset0 + p.cstep;
+
+        if (sz == p.c - 1)
+        {
+            if (sx == p.size - 1)
+            {
+                mat2x4 v0 = square_blob_data[v_offset0.r];
+
+                mat2x4 sqsum = v0;
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+            else
+            {
+                mat2x4 v0 = square_blob_data[v_offset0.r];
+                mat2x4 v1 = square_blob_data[v_offset0.g];
+
+                mat2x4 sqsum = v0 + v1;
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+        }
+        else
+        {
+            if (sx == p.size - 1)
+            {
+                mat2x4 v0 = square_blob_data[v_offset0.r];
+                mat2x4 v2 = square_blob_data[v_offset1.r];
+
+                mat2x4 sqsum = v0 + v2;
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+            else
+            {
+                mat2x4 v0 = square_blob_data[v_offset0.r];
+                mat2x4 v1 = square_blob_data[v_offset0.g];
+                mat2x4 v2 = square_blob_data[v_offset1.r];
+                mat2x4 v3 = square_blob_data[v_offset1.g];
+
+                mat2x4 sqsum = v0 + v1 + v2 + v3;
+
+                sqsum_blob_data[gi] = sqsum;
+            }
+        }
+    }
+
+    if (across_spatial == 1 && across_channel == 0)
+    {
+        int sz = gz;
+        int sx = gx * 4;
+
+        ivec4 v_offset = sz * p.cstep + sx + ivec4(0, 1, 2, 3);
+
+        if (sx == p.size - 1)
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+
+            mat2x4 sqsum = v0;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sx == p.size - 2)
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+            mat2x4 v1 = square_blob_data[v_offset.g];
+
+            mat2x4 sqsum = v0 + v1;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sx == p.size - 3)
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+            mat2x4 v1 = square_blob_data[v_offset.g];
+            mat2x4 v2 = square_blob_data[v_offset.b];
+
+            mat2x4 sqsum = v0 + v1 + v2;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+            mat2x4 v1 = square_blob_data[v_offset.g];
+            mat2x4 v2 = square_blob_data[v_offset.b];
+            mat2x4 v3 = square_blob_data[v_offset.a];
+
+            mat2x4 sqsum = v0 + v1 + v2 + v3;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+    }
+
+    if (across_spatial == 0 && across_channel == 1)
+    {
+        int sz = gz * 4;
+        int sx = gx;
+
+        ivec4 v_offset = (sz + ivec4(0, 1, 2, 3)) * p.cstep + sx;
+
+        if (sz == p.c - 1)
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+
+            mat2x4 sqsum = v0;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sz == p.c - 2)
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+            mat2x4 v1 = square_blob_data[v_offset.g];
+
+            mat2x4 sqsum = v0 + v1;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else if (sz == p.c - 3)
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+            mat2x4 v1 = square_blob_data[v_offset.g];
+            mat2x4 v2 = square_blob_data[v_offset.b];
+
+            mat2x4 sqsum = v0 + v1 + v2;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+        else
+        {
+            mat2x4 v0 = square_blob_data[v_offset.r];
+            mat2x4 v1 = square_blob_data[v_offset.g];
+            mat2x4 v2 = square_blob_data[v_offset.b];
+            mat2x4 v3 = square_blob_data[v_offset.a];
+
+            mat2x4 sqsum = v0 + v1 + v2 + v3;
+
+            sqsum_blob_data[gi] = sqsum;
+        }
+    }
+}

src/layer/vulkan/shader/permute_pack8to1.comp

@@ -0,0 +1,123 @@
+// 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 order_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 { 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 == 2)
+    {
+        // order_type
+        // 0 = w h
+        // 1 = h w
+
+        if (order_type == 0)
+        {
+            v_offset = ivec4((gy * 8) * p.outw + gx) + ivec4(0, 1, 2, 3) * p.outw;
+            vv_offset = v_offset + 4 * p.outw;
+        }
+        if (order_type == 1)
+        {
+            v_offset = ivec4(gx * p.outw + gy * 8) + ivec4(0, 1, 2, 3);
+            vv_offset = v_offset + 4;
+        }
+    }
+    else if (p.dims == 3)
+    {
+        // order_type
+        // 0 = w h c
+        // 1 = h w c
+        // 2 = w c h
+        // 3 = c w h
+        // 4 = h c w
+        // 5 = c h w
+
+        if (order_type == 0)
+        {
+            v_offset = ivec4((gz * 8) * p.outcstep + gy * p.outw + gx) + ivec4(0, 1, 2, 3) * p.outcstep;
+            vv_offset = v_offset + 4 * p.outcstep;
+        }
+        if (order_type == 1)
+        {
+            v_offset = ivec4((gz * 8) * p.outcstep + gx * p.outw + gy) + ivec4(0, 1, 2, 3) * p.outcstep;
+            vv_offset = v_offset + 4 * p.outcstep;
+        }
+        if (order_type == 2)
+        {
+            v_offset = ivec4(gy * p.outcstep + (gz * 8) * p.outw + gx) + ivec4(0, 1, 2, 3) * p.outw;
+            vv_offset = v_offset + 4 * p.outw;
+        }
+        if (order_type == 3)
+        {
+            v_offset = ivec4(gy * p.outcstep + gx * p.outw + gz * 8) + ivec4(0, 1, 2, 3);
+            vv_offset = v_offset + 4;
+        }
+        if (order_type == 4)
+        {
+            v_offset = ivec4(gx * p.outcstep + (gz * 8) * p.outw + gy) + ivec4(0, 1, 2, 3) * p.outw;
+            vv_offset = v_offset + 4 * p.outw;
+        }
+        if (order_type == 5)
+        {
+            v_offset = ivec4(gx * p.outcstep + gy * p.outw + gz * 8) + ivec4(0, 1, 2, 3);
+            vv_offset = v_offset + 4;
+        }
+    }
+
+    int gi = gz * p.cstep + gy * p.w + gx;
+
+    buffer_cp8to1(top_blob_data, v_offset, vv_offset, bottom_blob_data, gi);
+}