optimize vulkan global pooling (#5191)

Co-authored-by: nihui <nihui@users.noreply.github.com>
Co-authored-by: michaelcai <michaelcai@tencent.com>

src/layer/vulkan/pooling_vulkan.cpp

@@ -30,12 +30,20 @@ Pooling_vulkan::Pooling_vulkan()
     pipeline_pooling = 0;
     pipeline_pooling_pack4 = 0;
     pipeline_pooling_pack8 = 0;
-    pipeline_pooling_global = 0;
-    pipeline_pooling_global_pack4 = 0;
-    pipeline_pooling_global_pack8 = 0;
+
     pipeline_pooling_adaptive = 0;
     pipeline_pooling_adaptive_pack4 = 0;
     pipeline_pooling_adaptive_pack8 = 0;
+
+    pipeline_pooling_global_reduce_first = 0;
+    pipeline_pooling_global_reduce_first_pack4 = 0;
+    pipeline_pooling_global_reduce_first_pack8 = 0;
+    pipeline_pooling_global_reduce = 0;
+    pipeline_pooling_global_reduce_pack4 = 0;
+    pipeline_pooling_global_reduce_pack8 = 0;
+    pipeline_pooling_global_reduce_last = 0;
+    pipeline_pooling_global_reduce_last_pack4 = 0;
+    pipeline_pooling_global_reduce_last_pack8 = 0;
 }
 
 int Pooling_vulkan::create_pipeline(const Option& _opt)
@@ -156,49 +164,129 @@ int Pooling_vulkan::create_pipeline(const Option& _opt)
 
     if (global_pooling)
     {
-        std::vector<vk_specialization_type> specializations(1 + 10);
-        specializations[0].i = pooling_type;
-        specializations[1 + 0].i = shape_bordered_packed.dims;
-        specializations[1 + 1].i = shape_bordered_packed.w;
-        specializations[1 + 2].i = shape_bordered_packed.h;
-        specializations[1 + 3].i = shape_bordered_packed.c;
-        specializations[1 + 4].i = shape_bordered_packed.cstep;
-        specializations[1 + 5].i = out_shape_packed.dims;
-        specializations[1 + 6].i = out_shape_packed.w;
-        specializations[1 + 7].i = out_shape_packed.h;
-        specializations[1 + 8].i = out_shape_packed.c;
-        specializations[1 + 9].i = out_shape_packed.cstep;
-
-        Mat local_size_xyz(64, 1, 1, (void*)0);
-        if (out_shape_packed.dims != 0)
+        // reduce first
         {
-            local_size_xyz.w = std::min(64, out_shape_packed.w);
-            local_size_xyz.h = 1;
-            local_size_xyz.c = 1;
-        }
+            std::vector<vk_specialization_type> specializations(6);
+            specializations[0].i = shape_bordered_packed.w;
+            specializations[1].i = shape_bordered_packed.h;
+            specializations[2].i = shape_bordered_packed.c;
+            specializations[3].i = shape_bordered_packed.cstep;
+            specializations[4].i = 0;
+            specializations[5].i = 0;
+
+            Mat local_size_xyz(64, 1, 1, (void*)0);
+
+            // pack1
+            if (shape.dims == 0 || elempack == 1)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_first : LayerShaderType::pooling_global_reduce_sum_first;
 
-        // pack1
-        if (shape.dims == 0 || elempack == 1)
-        {
-            pipeline_pooling_global = new Pipeline(vkdev);
-            pipeline_pooling_global->set_optimal_local_size_xyz(local_size_xyz);
-            pipeline_pooling_global->create(LayerShaderType::pooling_global, opt, specializations);
+                pipeline_pooling_global_reduce_first = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_first->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_first->create(layer_shader_type, opt, specializations);
+            }
+
+            // pack4
+            if (shape.dims == 0 || elempack == 4)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_first_pack4 : LayerShaderType::pooling_global_reduce_sum_first_pack4;
+
+                pipeline_pooling_global_reduce_first_pack4 = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_first_pack4->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_first_pack4->create(layer_shader_type, opt, specializations);
+            }
+
+            // pack8
+            if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_first_pack8 : LayerShaderType::pooling_global_reduce_sum_first_pack8;
+
+                pipeline_pooling_global_reduce_first_pack8 = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_first_pack8->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_first_pack8->create(layer_shader_type, opt, specializations);
+            }
         }
 
-        // pack4
-        if (shape.dims == 0 || elempack == 4)
+        // reduce more
         {
-            pipeline_pooling_global_pack4 = new Pipeline(vkdev);
-            pipeline_pooling_global_pack4->set_optimal_local_size_xyz(local_size_xyz);
-            pipeline_pooling_global_pack4->create(LayerShaderType::pooling_global_pack4, opt, specializations);
+            std::vector<vk_specialization_type> specializations(5);
+            specializations[0].i = 0;
+            specializations[1].i = shape_bordered_packed.c;
+            specializations[2].i = 0;
+            specializations[3].i = 0;
+            specializations[4].i = 0;
+
+            Mat local_size_xyz(64, 1, 1, (void*)0);
+
+            // pack1
+            if (shape.dims == 0 || elempack == 1)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max : LayerShaderType::pooling_global_reduce_sum;
+
+                pipeline_pooling_global_reduce = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce->create(layer_shader_type, opt, specializations);
+            }
+
+            // pack4
+            if (shape.dims == 0 || elempack == 4)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_pack4 : LayerShaderType::pooling_global_reduce_sum_pack4;
+
+                pipeline_pooling_global_reduce_pack4 = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_pack4->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_pack4->create(layer_shader_type, opt, specializations);
+            }
+
+            // pack8
+            if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_pack8 : LayerShaderType::pooling_global_reduce_sum_pack8;
+
+                pipeline_pooling_global_reduce_pack8 = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_pack8->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_pack8->create(layer_shader_type, opt, specializations);
+            }
         }
 
-        // pack8
-        if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
+        // reduce last
         {
-            pipeline_pooling_global_pack8 = new Pipeline(vkdev);
-            pipeline_pooling_global_pack8->set_optimal_local_size_xyz(local_size_xyz);
-            pipeline_pooling_global_pack8->create(LayerShaderType::pooling_global_pack8, opt, specializations);
+            std::vector<vk_specialization_type> specializations(3);
+            specializations[0].i = 0;
+            specializations[1].i = shape_bordered_packed.c;
+            specializations[2].i = 0;
+
+            Mat local_size_xyz(1, 1, 64, (void*)0);
+
+            // pack1
+            if (shape.dims == 0 || elempack == 1)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_last : LayerShaderType::pooling_global_reduce_sum_last;
+
+                pipeline_pooling_global_reduce_last = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_last->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_last->create(layer_shader_type, opt, specializations);
+            }
+
+            // pack4
+            if (shape.dims == 0 || elempack == 4)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_last_pack4 : LayerShaderType::pooling_global_reduce_sum_last_pack4;
+
+                pipeline_pooling_global_reduce_last_pack4 = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_last_pack4->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_last_pack4->create(layer_shader_type, opt, specializations);
+            }
+
+            // pack8
+            if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
+            {
+                int layer_shader_type = pooling_type == 0 ? LayerShaderType::pooling_global_reduce_max_last_pack8 : LayerShaderType::pooling_global_reduce_sum_last_pack8;
+
+                pipeline_pooling_global_reduce_last_pack8 = new Pipeline(vkdev);
+                pipeline_pooling_global_reduce_last_pack8->set_optimal_local_size_xyz(local_size_xyz);
+                pipeline_pooling_global_reduce_last_pack8->create(layer_shader_type, opt, specializations);
+            }
         }
     }
     else if (adaptive_pooling)
@@ -331,15 +419,6 @@ int Pooling_vulkan::destroy_pipeline(const Option& _opt)
     delete pipeline_pooling_pack8;
     pipeline_pooling_pack8 = 0;
 
-    delete pipeline_pooling_global;
-    pipeline_pooling_global = 0;
-
-    delete pipeline_pooling_global_pack4;
-    pipeline_pooling_global_pack4 = 0;
-
-    delete pipeline_pooling_global_pack8;
-    pipeline_pooling_global_pack8 = 0;
-
     delete pipeline_pooling_adaptive;
     pipeline_pooling_adaptive = 0;
 
@@ -349,6 +428,33 @@ int Pooling_vulkan::destroy_pipeline(const Option& _opt)
     delete pipeline_pooling_adaptive_pack8;
     pipeline_pooling_adaptive_pack8 = 0;
 
+    delete pipeline_pooling_global_reduce_first;
+    pipeline_pooling_global_reduce_first = 0;
+
+    delete pipeline_pooling_global_reduce_first_pack4;
+    pipeline_pooling_global_reduce_first_pack4 = 0;
+
+    delete pipeline_pooling_global_reduce_first_pack8;
+    pipeline_pooling_global_reduce_first_pack8 = 0;
+
+    delete pipeline_pooling_global_reduce;
+    pipeline_pooling_global_reduce = 0;
+
+    delete pipeline_pooling_global_reduce_pack4;
+    pipeline_pooling_global_reduce_pack4 = 0;
+
+    delete pipeline_pooling_global_reduce_pack8;
+    pipeline_pooling_global_reduce_pack8 = 0;
+
+    delete pipeline_pooling_global_reduce_last;
+    pipeline_pooling_global_reduce_last = 0;
+
+    delete pipeline_pooling_global_reduce_last_pack4;
+    pipeline_pooling_global_reduce_last_pack4 = 0;
+
+    delete pipeline_pooling_global_reduce_last_pack8;
+    pipeline_pooling_global_reduce_last_pack8 = 0;
+
     return 0;
 }
 
@@ -372,31 +478,101 @@ int Pooling_vulkan::forward(const VkMat& bottom_blob, VkMat& top_blob, VkCompute
 
     if (global_pooling)
     {
-        top_blob.create(channels, elemsize, elempack, opt.blob_vkallocator);
-        if (top_blob.empty())
-            return -100;
-
-        std::vector<VkMat> bindings(2);
-        bindings[0] = bottom_blob;
-        bindings[1] = top_blob;
-
-        std::vector<vk_constant_type> constants(10);
-        constants[0].i = bottom_blob.dims;
-        constants[1].i = bottom_blob.w;
-        constants[2].i = bottom_blob.h;
-        constants[3].i = bottom_blob.c;
-        constants[4].i = bottom_blob.cstep;
-        constants[5].i = top_blob.dims;
-        constants[6].i = top_blob.w;
-        constants[7].i = top_blob.h;
-        constants[8].i = top_blob.c;
-        constants[9].i = top_blob.cstep;
+        // reduce first
+        VkMat reduced_blob;
+        {
+            int reduced_size = (w * h + 7) / 8;
+            size_t reduced_elemsize = pooling_type == 0 ? elemsize : 4u * elempack;
+            reduced_blob.create(reduced_size, 1, channels, reduced_elemsize, elempack, opt.workspace_vkallocator);
+            if (reduced_blob.empty())
+                return -100;
+
+            std::vector<VkMat> bindings(2);
+            bindings[0] = bottom_blob;
+            bindings[1] = reduced_blob;
+
+            std::vector<vk_constant_type> constants(6);
+            constants[0].i = bottom_blob.w;
+            constants[1].i = bottom_blob.h;
+            constants[2].i = bottom_blob.c;
+            constants[3].i = bottom_blob.cstep;
+            constants[4].i = reduced_blob.w;
+            constants[5].i = reduced_blob.cstep;
+
+            const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_reduce_first_pack8
+                                       : elempack == 4 ? pipeline_pooling_global_reduce_first_pack4
+                                       : pipeline_pooling_global_reduce_first;
+
+            VkMat dispatcher;
+            dispatcher.w = reduced_blob.w;
+            dispatcher.h = 1;
+            dispatcher.c = bottom_blob.c;
+
+            cmd.record_pipeline(pipeline, bindings, constants, dispatcher);
+        }
 
-        const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_pack8
-                                   : elempack == 4 ? pipeline_pooling_global_pack4
-                                   : pipeline_pooling_global;
+        // reduce more
+        while (reduced_blob.w > 32)
+        {
+            int reduced_size = (reduced_blob.w + 7) / 8;
+            size_t reduced_elemsize = pooling_type == 0 ? elemsize : 4u * elempack;
+            VkMat reduced_blob2;
+            reduced_blob2.create(reduced_size, 1, channels, reduced_elemsize, elempack, opt.workspace_vkallocator);
+            if (reduced_blob2.empty())
+                return -100;
+
+            std::vector<VkMat> bindings(2);
+            bindings[0] = reduced_blob;
+            bindings[1] = reduced_blob2;
+
+            std::vector<vk_constant_type> constants(5);
+            constants[0].i = reduced_blob.w;
+            constants[1].i = reduced_blob.c;
+            constants[2].i = reduced_blob.cstep;
+            constants[3].i = reduced_blob2.w;
+            constants[4].i = reduced_blob2.cstep;
+
+            const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_reduce_pack8
+                                       : elempack == 4 ? pipeline_pooling_global_reduce_pack4
+                                       : pipeline_pooling_global_reduce;
+
+            VkMat dispatcher;
+            dispatcher.w = reduced_blob2.w;
+            dispatcher.h = 1;
+            dispatcher.c = reduced_blob2.c;
+
+            cmd.record_pipeline(pipeline, bindings, constants, dispatcher);
+
+            reduced_blob = reduced_blob2;
+        }
 
-        cmd.record_pipeline(pipeline, bindings, constants, top_blob);
+        // reduce last
+        {
+            top_blob.create(channels, elemsize, elempack, opt.blob_vkallocator);
+            if (top_blob.empty())
+                return -100;
+
+            std::vector<VkMat> bindings(2);
+            bindings[0] = reduced_blob;
+            bindings[1] = top_blob;
+
+            std::vector<vk_constant_type> constants(4);
+            constants[0].i = reduced_blob.w;
+            constants[1].i = reduced_blob.c;
+            constants[2].i = reduced_blob.cstep;
+            constants[3].i = w * h;
+
+            const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_reduce_last_pack8
+                                       : elempack == 4 ? pipeline_pooling_global_reduce_last_pack4
+                                       : pipeline_pooling_global_reduce_last;
+
+            VkMat dispatcher;
+            dispatcher.w = 1;
+            dispatcher.h = 1;
+            dispatcher.c = top_blob.w;
+
+            cmd.record_pipeline(pipeline, bindings, constants, dispatcher);
+        }
 
         return 0;
     }
@@ -588,31 +764,101 @@ int Pooling_vulkan::forward(const VkImageMat& bottom_blob, VkImageMat& top_blob,
 
     if (global_pooling)
     {
-        top_blob.create(channels, elemsize, elempack, opt.blob_vkallocator);
-        if (top_blob.empty())
-            return -100;
-
-        std::vector<VkImageMat> bindings(2);
-        bindings[0] = bottom_blob;
-        bindings[1] = top_blob;
-
-        std::vector<vk_constant_type> constants(10);
-        constants[0].i = bottom_blob.dims;
-        constants[1].i = bottom_blob.w;
-        constants[2].i = bottom_blob.h;
-        constants[3].i = bottom_blob.c;
-        constants[4].i = 0; //bottom_blob.cstep;
-        constants[5].i = top_blob.dims;
-        constants[6].i = top_blob.w;
-        constants[7].i = top_blob.h;
-        constants[8].i = top_blob.c;
-        constants[9].i = 0; //top_blob.cstep;
+        // reduce first
+        VkImageMat reduced_blob;
+        {
+            int reduced_size = (w * h + 7) / 8;
+            size_t reduced_elemsize = pooling_type == 0 ? elemsize : 4u * elempack;
+            reduced_blob.create(reduced_size, 1, channels, reduced_elemsize, elempack, opt.workspace_vkallocator);
+            if (reduced_blob.empty())
+                return -100;
+
+            std::vector<VkImageMat> bindings(2);
+            bindings[0] = bottom_blob;
+            bindings[1] = reduced_blob;
+
+            std::vector<vk_constant_type> constants(6);
+            constants[0].i = bottom_blob.w;
+            constants[1].i = bottom_blob.h;
+            constants[2].i = bottom_blob.c;
+            constants[3].i = 0; //bottom_blob.cstep;
+            constants[4].i = reduced_blob.w;
+            constants[5].i = 0; //reduced_blob.cstep;
+
+            const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_reduce_first_pack8
+                                       : elempack == 4 ? pipeline_pooling_global_reduce_first_pack4
+                                       : pipeline_pooling_global_reduce_first;
+
+            VkImageMat dispatcher;
+            dispatcher.w = reduced_blob.w;
+            dispatcher.h = 1;
+            dispatcher.c = bottom_blob.c;
+
+            cmd.record_pipeline(pipeline, bindings, constants, dispatcher);
+        }
 
-        const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_pack8
-                                   : elempack == 4 ? pipeline_pooling_global_pack4
-                                   : pipeline_pooling_global;
+        // reduce more
+        while (reduced_blob.w > 32)
+        {
+            int reduced_size = (reduced_blob.w + 7) / 8;
+            size_t reduced_elemsize = pooling_type == 0 ? elemsize : 4u * elempack;
+            VkImageMat reduced_blob2;
+            reduced_blob2.create(reduced_size, 1, channels, reduced_elemsize, elempack, opt.workspace_vkallocator);
+            if (reduced_blob2.empty())
+                return -100;
+
+            std::vector<VkImageMat> bindings(2);
+            bindings[0] = reduced_blob;
+            bindings[1] = reduced_blob2;
+
+            std::vector<vk_constant_type> constants(5);
+            constants[0].i = reduced_blob.w;
+            constants[1].i = reduced_blob.c;
+            constants[2].i = 0; //reduced_blob.cstep;
+            constants[3].i = reduced_blob2.w;
+            constants[4].i = 0; //reduced_blob2.cstep;
+
+            const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_reduce_pack8
+                                       : elempack == 4 ? pipeline_pooling_global_reduce_pack4
+                                       : pipeline_pooling_global_reduce;
+
+            VkImageMat dispatcher;
+            dispatcher.w = reduced_blob2.w;
+            dispatcher.h = 1;
+            dispatcher.c = reduced_blob2.c;
+
+            cmd.record_pipeline(pipeline, bindings, constants, dispatcher);
+
+            reduced_blob = reduced_blob2;
+        }
 
-        cmd.record_pipeline(pipeline, bindings, constants, top_blob);
+        // reduce last
+        {
+            top_blob.create(channels, elemsize, elempack, opt.blob_vkallocator);
+            if (top_blob.empty())
+                return -100;
+
+            std::vector<VkImageMat> bindings(2);
+            bindings[0] = reduced_blob;
+            bindings[1] = top_blob;
+
+            std::vector<vk_constant_type> constants(4);
+            constants[0].i = reduced_blob.w;
+            constants[1].i = reduced_blob.c;
+            constants[2].i = 0; //reduced_blob.cstep;
+            constants[3].i = w * h;
+
+            const Pipeline* pipeline = elempack == 8 ? pipeline_pooling_global_reduce_last_pack8
+                                       : elempack == 4 ? pipeline_pooling_global_reduce_last_pack4
+                                       : pipeline_pooling_global_reduce_last;
+
+            VkImageMat dispatcher;
+            dispatcher.w = 1;
+            dispatcher.h = 1;
+            dispatcher.c = top_blob.w;
+
+            cmd.record_pipeline(pipeline, bindings, constants, dispatcher);
+        }
 
         return 0;
     }

src/layer/vulkan/pooling_vulkan.h

@@ -39,12 +39,20 @@ public:
     Pipeline* pipeline_pooling;
     Pipeline* pipeline_pooling_pack4;
     Pipeline* pipeline_pooling_pack8;
-    Pipeline* pipeline_pooling_global;
-    Pipeline* pipeline_pooling_global_pack4;
-    Pipeline* pipeline_pooling_global_pack8;
+
     Pipeline* pipeline_pooling_adaptive;
     Pipeline* pipeline_pooling_adaptive_pack4;
     Pipeline* pipeline_pooling_adaptive_pack8;
+
+    Pipeline* pipeline_pooling_global_reduce_first;
+    Pipeline* pipeline_pooling_global_reduce_first_pack4;
+    Pipeline* pipeline_pooling_global_reduce_first_pack8;
+    Pipeline* pipeline_pooling_global_reduce;
+    Pipeline* pipeline_pooling_global_reduce_pack4;
+    Pipeline* pipeline_pooling_global_reduce_pack8;
+    Pipeline* pipeline_pooling_global_reduce_last;
+    Pipeline* pipeline_pooling_global_reduce_last_pack4;
+    Pipeline* pipeline_pooling_global_reduce_last_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/shader/pooling_global.comp

@@ -1,126 +0,0 @@
-// 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
-#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;
-
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
-
-#if NCNN_image_shader
-layout (binding = 0) uniform unfp sampler3D bottom_blob;
-layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
-#else
-layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
-layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
-#endif
-
-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 >= psc(outw) || gy >= 1 || gz >= 1)
-        return;
-
-    int size = psc(w) * psc(h);
-    int v_offset = gx * psc(cstep);
-
-    afp res;
-
-    if (pooling_type == 0)
-    {
-        res = afp(-FLT_MAX);
-
-#if NCNN_image_shader
-        for (int y = 0; y < psc(h); y++)
-        {
-            for (int x = 0; x < psc(w); x++)
-            {
-                afp v = image3d_ld1(bottom_blob, ivec3(x, y, gx));
-                res = max(res, v);
-            }
-        }
-#else
-        for (int i = 0; i < size; i++)
-        {
-            afp v = buffer_ld1(bottom_blob_data, v_offset + i);
-            res = max(res, v);
-        }
-#endif
-    }
-    if (pooling_type == 1)
-    {
-        float sum = 0.f;
-
-#if NCNN_image_shader
-        for (int y = 0; y < psc(h); y++)
-        {
-            for (int x = 0; x < psc(w); x++)
-            {
-                sum += float(image3d_ld1(bottom_blob, ivec3(x, y, gx)));
-            }
-        }
-#else
-        for (int i = 0; i < size; i++)
-        {
-            sum += float(buffer_ld1(bottom_blob_data, v_offset + i));
-        }
-#endif
-
-        res = afp(sum / size);
-    }
-
-#if NCNN_image_shader
-    image3d_st1(top_blob, ivec3(gx, 0, 0), res);
-#else
-    buffer_st1(top_blob_data, gx, res);
-#endif
-}

src/layer/vulkan/shader/pooling_global_pack4.comp

@@ -1,126 +0,0 @@
-// 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
-#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;
-
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
-
-#if NCNN_image_shader
-layout (binding = 0) uniform unfp sampler3D bottom_blob;
-layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
-#else
-layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
-layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
-#endif
-
-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 >= psc(outw) || gy >= 1 || gz >= 1)
-        return;
-
-    int size = psc(w) * psc(h);
-    int v_offset = gx * psc(cstep);
-
-    afpvec4 res;
-
-    if (pooling_type == 0)
-    {
-        res = afpvec4(-FLT_MAX);
-
-#if NCNN_image_shader
-        for (int y = 0; y < psc(h); y++)
-        {
-            for (int x = 0; x < psc(w); x++)
-            {
-                afpvec4 v = image3d_ld4(bottom_blob, ivec3(x, y, gx));
-                res = max(res, v);
-            }
-        }
-#else
-        for (int i = 0; i < size; i++)
-        {
-            afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + i);
-            res = max(res, v);
-        }
-#endif
-    }
-    if (pooling_type == 1)
-    {
-        vec4 sum = vec4(0.f);
-
-#if NCNN_image_shader
-        for (int y = 0; y < psc(h); y++)
-        {
-            for (int x = 0; x < psc(w); x++)
-            {
-                sum += vec4(image3d_ld4(bottom_blob, ivec3(x, y, gx)));
-            }
-        }
-#else
-        for (int i = 0; i < size; i++)
-        {
-            sum += vec4(buffer_ld4(bottom_blob_data, v_offset + i));
-        }
-#endif
-
-        res = afpvec4(sum / size);
-    }
-
-#if NCNN_image_shader
-    image3d_st4(top_blob, ivec3(gx, 0, 0), res);
-#else
-    buffer_st4(top_blob_data, gx, res);
-#endif
-}

src/layer/vulkan/shader/pooling_global_pack8.comp

@@ -1,134 +0,0 @@
-// 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;
-
-#define shape_constant_id_offset 1
-layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
-layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
-layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
-layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
-layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
-
-layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
-layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
-layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
-layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
-layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
-
-#if NCNN_image_shader
-layout (binding = 0) uniform unfp sampler3D bottom_blob;
-layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
-#else
-layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
-layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
-#endif
-
-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 >= psc(outw) || gy >= 1 || gz >= 1)
-        return;
-
-    int size = psc(w) * psc(h);
-    int v_offset = gx * psc(cstep);
-
-    afpvec8 res;
-
-    if (pooling_type == 0)
-    {
-        res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
-
-#if NCNN_image_shader
-        for (int y = 0; y < psc(h); y++)
-        {
-            for (int x = 0; x < psc(w); x++)
-            {
-                afpvec8 v = image3d_ld8(bottom_blob, ivec3(x, y, gx));
-                res[0] = max(res[0], v[0]);
-                res[1] = max(res[1], v[1]);
-            }
-        }
-#else
-        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]);
-        }
-#endif
-    }
-    if (pooling_type == 1)
-    {
-        mat2x4 sum = mat2x4(0.f);
-
-#if NCNN_image_shader
-        for (int y = 0; y < psc(h); y++)
-        {
-            for (int x = 0; x < psc(w); x++)
-            {
-                afpvec8 v = image3d_ld8(bottom_blob, ivec3(x, y, gx));
-                sum[0] += vec4(v[0]);
-                sum[1] += vec4(v[1]);
-            }
-        }
-#else
-        for (int i = 0; i < size; i++)
-        {
-            afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + i);
-            sum[0] += vec4(v[0]);
-            sum[1] += vec4(v[1]);
-        }
-#endif
-
-        res[0] = afpvec4(sum[0] / size);
-        res[1] = afpvec4(sum[1] / size);
-    }
-
-#if NCNN_image_shader
-    image3d_st8(top_blob, ivec3(gx, 0, 0), res);
-#else
-    buffer_st8(top_blob_data, gx, res);
-#endif
-}

src/layer/vulkan/shader/pooling_global_reduce_max.comp

@@ -0,0 +1,84 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int size_1 = psc(w) - 1;
+
+    const int v_offset = gz * psc(cstep);
+
+    afp res = afp(-FLT_MAX);
+
+    for (int ii = 0; ii < 8; ii++)
+    {
+        int i = min(gx + ii * psc(outw), size_1);
+
+#if NCNN_image_shader
+        afp v = image3d_ld1(bottom_blob, ivec3(i, 0, gz));
+#else
+        afp v = buffer_ld1(bottom_blob_data, v_offset + i);
+#endif
+        res = max(res, v);
+    }
+
+#if NCNN_image_shader
+    image3d_st1(top_blob, ivec3(gx, 0, gz), res);
+#else
+    buffer_st1(top_blob_data, gz * psc(outcstep) + gx, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_first.comp

@@ -0,0 +1,88 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int size_1 = psc(w) * psc(h) - 1;
+
+    const int v_offset = gz * psc(cstep);
+
+    afp res = afp(-FLT_MAX);
+
+    for (int ii = 0; ii < 8; ii++)
+    {
+        int i = min(gx + ii * psc(outw), size_1);
+
+#if NCNN_image_shader
+        int y = i / psc(w);
+        int x = i % psc(w);
+        afp v = image3d_ld1(bottom_blob, ivec3(x, y, gz));
+#else
+        afp v = buffer_ld1(bottom_blob_data, v_offset + i);
+#endif
+        res = max(res, v);
+    }
+
+#if NCNN_image_shader
+    image3d_st1(top_blob, ivec3(gx, 0, gz), res);
+#else
+    buffer_st1(top_blob_data, gz * psc(outcstep) + gx, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_first_pack4.comp

@@ -0,0 +1,88 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int size_1 = psc(w) * psc(h) - 1;
+
+    const int v_offset = gz * psc(cstep);
+
+    afpvec4 res = afpvec4(-FLT_MAX);
+
+    for (int ii = 0; ii < 8; ii++)
+    {
+        int i = min(gx + ii * psc(outw), size_1);
+
+#if NCNN_image_shader
+        int y = i / psc(w);
+        int x = i % psc(w);
+        afpvec4 v = image3d_ld4(bottom_blob, ivec3(x, y, gz));
+#else
+        afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + i);
+#endif
+        res = max(res, v);
+    }
+
+#if NCNN_image_shader
+    image3d_st4(top_blob, ivec3(gx, 0, gz), res);
+#else
+    buffer_st4(top_blob_data, gz * psc(outcstep) + gx, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_first_pack8.comp

@@ -0,0 +1,90 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int size_1 = psc(w) * psc(h) - 1;
+
+    const int v_offset = gz * psc(cstep);
+
+    afpvec8 res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
+
+    for (int ii = 0; ii < 8; ii++)
+    {
+        int i = min(gx + ii * psc(outw), size_1);
+
+#if NCNN_image_shader
+        int y = i / psc(w);
+        int x = i % psc(w);
+        afpvec8 v = image3d_ld8(bottom_blob, ivec3(x, y, gz));
+#else
+        afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + i);
+#endif
+        res[0] = max(res[0], v[0]);
+        res[1] = max(res[1], v[1]);
+    }
+
+#if NCNN_image_shader
+    image3d_st8(top_blob, ivec3(gx, 0, gz), res);
+#else
+    buffer_st8(top_blob_data, gz * psc(outcstep) + gx, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_last.comp

@@ -0,0 +1,75 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+    int size;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= 1 || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int v_offset = gz * psc(cstep);
+
+    afp res = afp(-FLT_MAX);
+
+    for (int i = 0; i < psc(w); i++)
+    {
+#if NCNN_image_shader
+        afp v = image3d_ld1(bottom_blob, ivec3(i, 0, gz));
+#else
+        afp v = buffer_ld1(bottom_blob_data, v_offset + i);
+#endif
+        res = max(res, v);
+    }
+
+#if NCNN_image_shader
+    image3d_st1(top_blob, ivec3(gz, 0, 0), res);
+#else
+    buffer_st1(top_blob_data, gz, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_last_pack4.comp

@@ -0,0 +1,75 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+    int size;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= 1 || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int v_offset = gz * psc(cstep);
+
+    afpvec4 res = afpvec4(-FLT_MAX);
+
+    for (int i = 0; i < psc(w); i++)
+    {
+#if NCNN_image_shader
+        afpvec4 v = image3d_ld4(bottom_blob, ivec3(i, 0, gz));
+#else
+        afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + i);
+#endif
+        res = max(res, v);
+    }
+
+#if NCNN_image_shader
+    image3d_st4(top_blob, ivec3(gz, 0, 0), res);
+#else
+    buffer_st4(top_blob_data, gz, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_last_pack8.comp

@@ -0,0 +1,77 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+    int size;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= 1 || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int v_offset = gz * psc(cstep);
+
+    afpvec8 res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
+
+    for (int i = 0; i < psc(w); i++)
+    {
+#if NCNN_image_shader
+        afpvec8 v = image3d_ld8(bottom_blob, ivec3(i, 0, gz));
+#else
+        afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + i);
+#endif
+        res[0] = max(res[0], v[0]);
+        res[1] = max(res[1], v[1]);
+    }
+
+#if NCNN_image_shader
+    image3d_st8(top_blob, ivec3(gz, 0, 0), res);
+#else
+    buffer_st8(top_blob_data, gz, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_pack4.comp

@@ -0,0 +1,84 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define FLT_MAX 3.402823466e+38
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int size_1 = psc(w) - 1;
+
+    const int v_offset = gz * psc(cstep);
+
+    afpvec4 res = afpvec4(-FLT_MAX);
+
+    for (int ii = 0; ii < 8; ii++)
+    {
+        int i = min(gx + ii * psc(outw), size_1);
+
+#if NCNN_image_shader
+        afpvec4 v = image3d_ld4(bottom_blob, ivec3(i, 0, gz));
+#else
+        afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + i);
+#endif
+        res = max(res, v);
+    }
+
+#if NCNN_image_shader
+    image3d_st4(top_blob, ivec3(gx, 0, gz), res);
+#else
+    buffer_st4(top_blob_data, gz * psc(outcstep) + gx, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_max_pack8.comp

@@ -0,0 +1,86 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int size_1 = psc(w) - 1;
+
+    const int v_offset = gz * psc(cstep);
+
+    afpvec8 res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
+
+    for (int ii = 0; ii < 8; ii++)
+    {
+        int i = min(gx + ii * psc(outw), size_1);
+
+#if NCNN_image_shader
+        afpvec8 v = image3d_ld8(bottom_blob, ivec3(i, 0, gz));
+#else
+        afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + i);
+#endif
+        res[0] = max(res[0], v[0]);
+        res[1] = max(res[1], v[1]);
+    }
+
+#if NCNN_image_shader
+    image3d_st8(top_blob, ivec3(gx, 0, gz), res);
+#else
+    buffer_st8(top_blob_data, gz * psc(outcstep) + gx, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum.comp

@@ -0,0 +1,82 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform highp sampler3D bottom_blob;
+layout (binding = 1, r32f) writeonly uniform highp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { float bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { float top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int end = min(8, (psc(w) - gx - 1) / psc(outw) + 1);
+
+    const int v_offset = gz * psc(cstep);
+
+    float sum = 0.f;
+
+    for (int ii = 0; ii < end; ii++)
+    {
+        int i = gx + ii * psc(outw);
+
+#if NCNN_image_shader
+        float v = texelFetch(bottom_blob, ivec3(i, 0, gz), 0).r;
+#else
+        float v = bottom_blob_data[v_offset + i];
+#endif
+        sum += v;
+    }
+
+#if NCNN_image_shader
+    imageStore(top_blob, ivec3(gx, 0, gz), vec4(sum));
+#else
+    top_blob_data[gz * psc(outcstep) + gx] = sum;
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_first.comp

@@ -0,0 +1,86 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, r32f) writeonly uniform highp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { float top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int end = min(8, (psc(w) * psc(h) - gx - 1) / psc(outw) + 1);
+
+    const int v_offset = gz * psc(cstep);
+
+    float sum = 0.f;
+
+    for (int ii = 0; ii < end; ii++)
+    {
+        int i = gx + ii * psc(outw);
+
+#if NCNN_image_shader
+        int y = i / psc(w);
+        int x = i % psc(w);
+        afp v = image3d_ld1(bottom_blob, ivec3(x, y, gz));
+#else
+        afp v = buffer_ld1(bottom_blob_data, v_offset + i);
+#endif
+        sum += float(v);
+    }
+
+#if NCNN_image_shader
+    imageStore(top_blob, ivec3(gx, 0, gz), vec4(sum));
+#else
+    top_blob_data[gz * psc(outcstep) + gx] = sum;
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_first_pack4.comp

@@ -0,0 +1,86 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, rgba32f) writeonly uniform highp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { vec4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int end = min(8, (psc(w) * psc(h) - gx - 1) / psc(outw) + 1);
+
+    const int v_offset = gz * psc(cstep);
+
+    vec4 sum = vec4(0.f);
+
+    for (int ii = 0; ii < end; ii++)
+    {
+        int i = gx + ii * psc(outw);
+
+#if NCNN_image_shader
+        int y = i / psc(w);
+        int x = i % psc(w);
+        afpvec4 v = image3d_ld4(bottom_blob, ivec3(x, y, gz));
+#else
+        afpvec4 v = buffer_ld4(bottom_blob_data, v_offset + i);
+#endif
+        sum += vec4(v);
+    }
+
+#if NCNN_image_shader
+    imageStore(top_blob, ivec3(gx, 0, gz), sum);
+#else
+    top_blob_data[gz * psc(outcstep) + gx] = sum;
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_first_pack8.comp

@@ -0,0 +1,89 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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 shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 5) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform unfp sampler3D bottom_blob;
+layout (binding = 1, rgba32f) writeonly uniform highp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { mat2x4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int h;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int end = min(8, (psc(w) * psc(h) - gx - 1) / psc(outw) + 1);
+
+    const int v_offset = gz * psc(cstep);
+
+    mat2x4 sum = mat2x4(0.f);
+
+    for (int ii = 0; ii < end; ii++)
+    {
+        int i = gx + ii * psc(outw);
+
+#if NCNN_image_shader
+        int y = i / psc(w);
+        int x = i % psc(w);
+        afpvec8 v = image3d_ld8(bottom_blob, ivec3(x, y, gz));
+#else
+        afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + i);
+#endif
+        sum[0] += vec4(v[0]);
+        sum[1] += vec4(v[1]);
+    }
+
+#if NCNN_image_shader
+    imageStore(top_blob, ivec3(gx * 2, 0, gz), sum[0]);
+    imageStore(top_blob, ivec3(gx * 2 + 1, 0, gz), sum[1]);
+#else
+    top_blob_data[gz * psc(outcstep) + gx] = sum;
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_last.comp

@@ -0,0 +1,75 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform highp sampler3D bottom_blob;
+layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { float bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+    int size;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= 1 || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int v_offset = gz * psc(cstep);
+
+    float sum = 0.f;
+
+    for (int i = 0; i < psc(w); i++)
+    {
+#if NCNN_image_shader
+        float v = texelFetch(bottom_blob, ivec3(i, 0, gz), 0).r;
+#else
+        float v = bottom_blob_data[v_offset + i];
+#endif
+        sum += v;
+    }
+
+    afp res = afp(sum / p.size);
+
+#if NCNN_image_shader
+    image3d_st1(top_blob, ivec3(gz, 0, 0), res);
+#else
+    buffer_st1(top_blob_data, gz, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_last_pack4.comp

@@ -0,0 +1,75 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform highp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { vec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+    int size;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= 1 || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int v_offset = gz * psc(cstep);
+
+    vec4 sum = vec4(0.f);
+
+    for (int i = 0; i < psc(w); i++)
+    {
+#if NCNN_image_shader
+        vec4 v = texelFetch(bottom_blob, ivec3(i, 0, gz), 0);
+#else
+        vec4 v = bottom_blob_data[v_offset + i];
+#endif
+        sum += v;
+    }
+
+    afpvec4 res = afpvec4(sum / p.size);
+
+#if NCNN_image_shader
+    image3d_st4(top_blob, ivec3(gz, 0, 0), res);
+#else
+    buffer_st4(top_blob_data, gz, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_last_pack8.comp

@@ -0,0 +1,81 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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 shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform highp sampler3D bottom_blob;
+layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { mat2x4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+    int size;
+} p;
+
+void main()
+{
+    int gx = int(gl_GlobalInvocationID.x);
+    int gy = int(gl_GlobalInvocationID.y);
+    int gz = int(gl_GlobalInvocationID.z);
+
+    if (gx >= 1 || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int v_offset = gz * psc(cstep);
+
+    mat2x4 sum = mat2x4(0.f);
+
+    for (int i = 0; i < psc(w); i++)
+    {
+#if NCNN_image_shader
+        mat2x4 v;
+        v[0] = texelFetch(bottom_blob, ivec3(i * 2, 0, gz), 0);
+        v[1] = texelFetch(bottom_blob, ivec3(i * 2 + 1, 0, gz), 0);
+#else
+        mat2x4 v = bottom_blob_data[v_offset + i];
+#endif
+        sum[0] += v[0];
+        sum[1] += v[1];
+    }
+
+    afpvec8 res;
+    res[0] = afpvec4(sum[0] / p.size);
+    res[1] = afpvec4(sum[1] / p.size);
+
+#if NCNN_image_shader
+    image3d_st8(top_blob, ivec3(gz, 0, 0), res);
+#else
+    buffer_st8(top_blob_data, gz, res);
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_pack4.comp

@@ -0,0 +1,82 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform highp sampler3D bottom_blob;
+layout (binding = 1, rgba32f) writeonly uniform highp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { vec4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { vec4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int end = min(8, (psc(w) - gx - 1) / psc(outw) + 1);
+
+    const int v_offset = gz * psc(cstep);
+
+    vec4 sum = vec4(0.f);
+
+    for (int ii = 0; ii < end; ii++)
+    {
+        int i = gx + ii * psc(outw);
+
+#if NCNN_image_shader
+        vec4 v = texelFetch(bottom_blob, ivec3(i, 0, gz), 0);
+#else
+        vec4 v = bottom_blob_data[v_offset + i];
+#endif
+        sum += v;
+    }
+
+#if NCNN_image_shader
+    imageStore(top_blob, ivec3(gx, 0, gz), sum);
+#else
+    top_blob_data[gz * psc(outcstep) + gx] = sum;
+#endif
+}

src/layer/vulkan/shader/pooling_global_reduce_sum_pack8.comp

@@ -0,0 +1,86 @@
+// Tencent is pleased to support the open source community by making ncnn available.
+//
+// Copyright (C) 2023 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
+#endif
+#if NCNN_fp16_arithmetic
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#endif
+
+#define shape_constant_id_offset 0
+layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
+layout (constant_id = shape_constant_id_offset + 1) const int c = 0;
+layout (constant_id = shape_constant_id_offset + 2) const int cstep = 0;
+
+layout (constant_id = shape_constant_id_offset + 3) const int outw = 0;
+layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
+
+#if NCNN_image_shader
+layout (binding = 0) uniform highp sampler3D bottom_blob;
+layout (binding = 1, rgba32f) writeonly uniform highp image3D top_blob;
+#else
+layout (binding = 0) readonly buffer bottom_blob { mat2x4 bottom_blob_data[]; };
+layout (binding = 1) writeonly buffer top_blob { mat2x4 top_blob_data[]; };
+#endif
+
+layout (push_constant) uniform parameter
+{
+    int w;
+    int c;
+    int cstep;
+
+    int outw;
+    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 >= psc(outw) || gy >= 1 || gz >= psc(c))
+        return;
+
+    const int end = min(8, (psc(w) - gx - 1) / psc(outw) + 1);
+
+    const int v_offset = gz * psc(cstep);
+
+    mat2x4 sum = mat2x4(0.f);
+
+    for (int ii = 0; ii < end; ii++)
+    {
+        int i = gx + ii * psc(outw);
+
+#if NCNN_image_shader
+        mat2x4 v;
+        v[0] = texelFetch(bottom_blob, ivec3(i * 2, 0, gz), 0);
+        v[1] = texelFetch(bottom_blob, ivec3(i * 2 + 1, 0, gz), 0);
+#else
+        mat2x4 v = bottom_blob_data[v_offset + i];
+#endif
+        sum[0] += v[0];
+        sum[1] += v[1];
+    }
+
+#if NCNN_image_shader
+    imageStore(top_blob, ivec3(gx * 2, 0, gz), sum[0]);
+    imageStore(top_blob, ivec3(gx * 2 + 1, 0, gz), sum[1]);
+#else
+    top_blob_data[gz * psc(outcstep) + gx] = sum;
+#endif
+}

tests/test_pooling.cpp

@@ -126,6 +126,10 @@ static int test_pooling_2()
            || test_pooling(7, 8, 8, 1, 1, 1, 0, 1, 0, 0, 0, 0)
            || test_pooling(11, 13, 16, 0, 1, 1, 0, 1, 0, 0, 0, 0)
            || test_pooling(13, 11, 16, 1, 1, 1, 0, 1, 0, 0, 0, 0)
+           || test_pooling(110, 103, 106, 0, 1, 1, 0, 1, 0, 0, 0, 0)
+           || test_pooling(130, 101, 106, 1, 1, 1, 0, 1, 0, 0, 0, 0)
+           || test_pooling(80, 93, 128, 0, 1, 1, 0, 1, 0, 0, 0, 0)
+           || test_pooling(80, 91, 128, 1, 1, 1, 0, 1, 0, 0, 0, 0)
            || test_pooling(48, 48, 4, 0, 2, 2, 0, 0, 0, 0, 0, 0)
            || test_pooling(48, 48, 15, 0, 2, 2, 1, 0, 0, 0, 0, 0);
 }