Unified elempack activation function vulkan shader (#6175)

src/layer/vulkan/celu_vulkan.cpp

@@ -12,8 +12,6 @@ CELU_vulkan::CELU_vulkan()
     support_vulkan = true;
 
     pipeline_celu = 0;
-    pipeline_celu_pack4 = 0;
-    pipeline_celu_pack8 = 0;
 }
 
 int CELU_vulkan::create_pipeline(const Option& opt)
@@ -41,63 +39,15 @@ int CELU_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(1 + 5);
+    std::vector<vk_specialization_type> specializations(1 + 1);
     specializations[0].f = alpha;
-    specializations[1 + 0].i = shape_packed.dims;
-    specializations[1 + 1].i = shape_packed.w;
-    specializations[1 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[1 + 3].i = shape_packed.c;
-    specializations[1 + 4].i = shape_packed.cstep;
-
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_celu = new Pipeline(vkdev);
-        pipeline_celu->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_celu->create(LayerShaderType::celu, opt, specializations);
-    }
+    specializations[1 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_celu_pack4 = new Pipeline(vkdev);
-        pipeline_celu_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_celu_pack4->create(LayerShaderType::celu_pack4, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_celu_pack8 = new Pipeline(vkdev);
-        pipeline_celu_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_celu_pack8->create(LayerShaderType::celu_pack8, opt, specializations);
-    }
+    pipeline_celu = new Pipeline(vkdev);
+    pipeline_celu->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_celu->create(LayerShaderType::celu, opt, specializations);
 
     return 0;
 }
@@ -107,34 +57,24 @@ int CELU_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_celu;
     pipeline_celu = 0;
 
-    delete pipeline_celu_pack4;
-    pipeline_celu_pack4 = 0;
-
-    delete pipeline_celu_pack8;
-    pipeline_celu_pack8 = 0;
-
     return 0;
 }
 
 int CELU_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_celu_pack8
-                               : elempack == 4 ? pipeline_celu_pack4
-                               : pipeline_celu;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_celu, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/celu_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_celu;
-    Pipeline* pipeline_celu_pack4;
-    Pipeline* pipeline_celu_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/clip_vulkan.cpp

@@ -12,8 +12,6 @@ Clip_vulkan::Clip_vulkan()
     support_vulkan = true;
 
     pipeline_clip = 0;
-    pipeline_clip_pack4 = 0;
-    pipeline_clip_pack8 = 0;
 }
 
 int Clip_vulkan::create_pipeline(const Option& opt)
@@ -41,64 +39,16 @@ int Clip_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(2 + 5);
+    std::vector<vk_specialization_type> specializations(2 + 1);
     specializations[0].f = min;
     specializations[1].f = max;
-    specializations[2 + 0].i = shape_packed.dims;
-    specializations[2 + 1].i = shape_packed.w;
-    specializations[2 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[2 + 3].i = shape_packed.c;
-    specializations[2 + 4].i = shape_packed.cstep;
-
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_clip = new Pipeline(vkdev);
-        pipeline_clip->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_clip->create(LayerShaderType::clip, opt, specializations);
-    }
+    specializations[2 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_clip_pack4 = new Pipeline(vkdev);
-        pipeline_clip_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_clip_pack4->create(LayerShaderType::clip_pack4, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_clip_pack8 = new Pipeline(vkdev);
-        pipeline_clip_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_clip_pack8->create(LayerShaderType::clip_pack8, opt, specializations);
-    }
+    pipeline_clip = new Pipeline(vkdev);
+    pipeline_clip->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_clip->create(LayerShaderType::clip, opt, specializations);
 
     return 0;
 }
@@ -108,34 +58,24 @@ int Clip_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_clip;
     pipeline_clip = 0;
 
-    delete pipeline_clip_pack4;
-    pipeline_clip_pack4 = 0;
-
-    delete pipeline_clip_pack8;
-    pipeline_clip_pack8 = 0;
-
     return 0;
 }
 
 int Clip_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_clip_pack8
-                               : elempack == 4 ? pipeline_clip_pack4
-                               : pipeline_clip;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_clip, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/clip_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_clip;
-    Pipeline* pipeline_clip_pack4;
-    Pipeline* pipeline_clip_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/dropout_vulkan.cpp

@@ -12,8 +12,6 @@ Dropout_vulkan::Dropout_vulkan()
     support_vulkan = true;
 
     pipeline_dropout = 0;
-    pipeline_dropout_pack4 = 0;
-    pipeline_dropout_pack8 = 0;
 }
 
 int Dropout_vulkan::create_pipeline(const Option& opt)
@@ -40,57 +38,15 @@ int Dropout_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 2) shape_packed = Mat(shape.w, shape.h / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(1 + 5);
+    std::vector<vk_specialization_type> specializations(1 + 1);
     specializations[0].f = scale;
-    specializations[1 + 0].i = shape_packed.dims;
-    specializations[1 + 1].i = shape_packed.w;
-    specializations[1 + 2].i = shape_packed.h;
-    specializations[1 + 3].i = shape_packed.c;
-    specializations[1 + 4].i = shape_packed.cstep;
-
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_dropout = new Pipeline(vkdev);
-        pipeline_dropout->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_dropout->create(LayerShaderType::dropout, opt, specializations);
-    }
+    specializations[1 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_dropout_pack4 = new Pipeline(vkdev);
-        pipeline_dropout_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_dropout_pack4->create(LayerShaderType::dropout_pack4, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_dropout_pack8 = new Pipeline(vkdev);
-        pipeline_dropout_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_dropout_pack8->create(LayerShaderType::dropout_pack8, opt, specializations);
-    }
+    pipeline_dropout = new Pipeline(vkdev);
+    pipeline_dropout->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_dropout->create(LayerShaderType::dropout, opt, specializations);
 
     return 0;
 }
@@ -100,12 +56,6 @@ int Dropout_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_dropout;
     pipeline_dropout = 0;
 
-    delete pipeline_dropout_pack4;
-    pipeline_dropout_pack4 = 0;
-
-    delete pipeline_dropout_pack8;
-    pipeline_dropout_pack8 = 0;
-
     return 0;
 }
 
@@ -116,23 +66,19 @@ int Dropout_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, cons
         return 0;
     }
 
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_dropout_pack8
-                               : elempack == 4 ? pipeline_dropout_pack4
-                               : pipeline_dropout;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_dropout, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/dropout_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_dropout;
-    Pipeline* pipeline_dropout_pack4;
-    Pipeline* pipeline_dropout_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/elu_vulkan.cpp

@@ -12,8 +12,6 @@ ELU_vulkan::ELU_vulkan()
     support_vulkan = true;
 
     pipeline_elu = 0;
-    pipeline_elu_pack4 = 0;
-    pipeline_elu_pack8 = 0;
 }
 
 int ELU_vulkan::create_pipeline(const Option& opt)
@@ -41,63 +39,15 @@ int ELU_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(1 + 5);
+    std::vector<vk_specialization_type> specializations(1 + 1);
     specializations[0].f = alpha;
-    specializations[1 + 0].i = shape_packed.dims;
-    specializations[1 + 1].i = shape_packed.w;
-    specializations[1 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[1 + 3].i = shape_packed.c;
-    specializations[1 + 4].i = shape_packed.cstep;
-
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_elu = new Pipeline(vkdev);
-        pipeline_elu->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_elu->create(LayerShaderType::elu, opt, specializations);
-    }
+    specializations[1 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_elu_pack4 = new Pipeline(vkdev);
-        pipeline_elu_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_elu_pack4->create(LayerShaderType::elu_pack4, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_elu_pack8 = new Pipeline(vkdev);
-        pipeline_elu_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_elu_pack8->create(LayerShaderType::elu_pack8, opt, specializations);
-    }
+    pipeline_elu = new Pipeline(vkdev);
+    pipeline_elu->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_elu->create(LayerShaderType::elu, opt, specializations);
 
     return 0;
 }
@@ -107,34 +57,24 @@ int ELU_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_elu;
     pipeline_elu = 0;
 
-    delete pipeline_elu_pack4;
-    pipeline_elu_pack4 = 0;
-
-    delete pipeline_elu_pack8;
-    pipeline_elu_pack8 = 0;
-
     return 0;
 }
 
 int ELU_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_elu_pack8
-                               : elempack == 4 ? pipeline_elu_pack4
-                               : pipeline_elu;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_elu, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/elu_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_elu;
-    Pipeline* pipeline_elu_pack4;
-    Pipeline* pipeline_elu_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/erf_vulkan.cpp

@@ -12,8 +12,6 @@ Erf_vulkan::Erf_vulkan()
     support_vulkan = true;
 
     pipeline_erf = 0;
-    pipeline_erf_pack4 = 0;
-    pipeline_erf_pack8 = 0;
 }
 
 int Erf_vulkan::create_pipeline(const Option& opt)
@@ -40,56 +38,14 @@ int Erf_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 2) shape_packed = Mat(shape.w, shape.h / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(0 + 5);
-    specializations[0 + 0].i = shape_packed.dims;
-    specializations[0 + 1].i = shape_packed.w;
-    specializations[0 + 2].i = shape_packed.h;
-    specializations[0 + 3].i = shape_packed.c;
-    specializations[0 + 4].i = shape_packed.cstep;
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0].u32 = shape_packed.total() * elempack / 4;
 
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_erf = new Pipeline(vkdev);
-        pipeline_erf->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_erf->create(LayerShaderType::erf, opt, specializations);
-    }
-
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_erf_pack4 = new Pipeline(vkdev);
-        pipeline_erf_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_erf_pack4->create(LayerShaderType::erf_pack4, opt, specializations);
-    }
-
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_erf_pack8 = new Pipeline(vkdev);
-        pipeline_erf_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_erf_pack8->create(LayerShaderType::erf_pack8, opt, specializations);
-    }
+    pipeline_erf = new Pipeline(vkdev);
+    pipeline_erf->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_erf->create(LayerShaderType::erf, opt, specializations);
 
     return 0;
 }
@@ -99,34 +55,24 @@ int Erf_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_erf;
     pipeline_erf = 0;
 
-    delete pipeline_erf_pack4;
-    pipeline_erf_pack4 = 0;
-
-    delete pipeline_erf_pack8;
-    pipeline_erf_pack8 = 0;
-
     return 0;
 }
 
 int Erf_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_erf_pack8
-                               : elempack == 4 ? pipeline_erf_pack4
-                               : pipeline_erf;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_erf, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/erf_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_erf;
-    Pipeline* pipeline_erf_pack4;
-    Pipeline* pipeline_erf_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/gelu_vulkan.cpp

@@ -12,8 +12,6 @@ GELU_vulkan::GELU_vulkan()
     support_vulkan = true;
 
     pipeline_gelu = 0;
-    pipeline_gelu_pack4 = 0;
-    pipeline_gelu_pack8 = 0;
 }
 
 int GELU_vulkan::create_pipeline(const Option& opt)
@@ -41,62 +39,14 @@ int GELU_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(0 + 5);
-    specializations[0 + 0].i = shape_packed.dims;
-    specializations[0 + 1].i = shape_packed.w;
-    specializations[0 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[0 + 3].i = shape_packed.c;
-    specializations[0 + 4].i = shape_packed.cstep;
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0].u32 = shape_packed.total() * elempack / 4;
 
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_gelu = new Pipeline(vkdev);
-        pipeline_gelu->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_gelu->create(LayerShaderType::gelu, opt, specializations);
-    }
-
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_gelu_pack4 = new Pipeline(vkdev);
-        pipeline_gelu_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_gelu_pack4->create(LayerShaderType::gelu_pack4, opt, specializations);
-    }
-
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_gelu_pack8 = new Pipeline(vkdev);
-        pipeline_gelu_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_gelu_pack8->create(LayerShaderType::gelu_pack8, opt, specializations);
-    }
+    pipeline_gelu = new Pipeline(vkdev);
+    pipeline_gelu->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_gelu->create(LayerShaderType::gelu, opt, specializations);
 
     return 0;
 }
@@ -106,34 +56,24 @@ int GELU_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_gelu;
     pipeline_gelu = 0;
 
-    delete pipeline_gelu_pack4;
-    pipeline_gelu_pack4 = 0;
-
-    delete pipeline_gelu_pack8;
-    pipeline_gelu_pack8 = 0;
-
     return 0;
 }
 
 int GELU_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_gelu_pack8
-                               : elempack == 4 ? pipeline_gelu_pack4
-                               : pipeline_gelu;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_gelu, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/gelu_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_gelu;
-    Pipeline* pipeline_gelu_pack4;
-    Pipeline* pipeline_gelu_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/hardsigmoid_vulkan.cpp

@@ -12,8 +12,6 @@ HardSigmoid_vulkan::HardSigmoid_vulkan()
     support_vulkan = true;
 
     pipeline_hardsigmoid = 0;
-    pipeline_hardsigmoid_pack4 = 0;
-    pipeline_hardsigmoid_pack8 = 0;
 }
 
 int HardSigmoid_vulkan::create_pipeline(const Option& opt)
@@ -40,58 +38,16 @@ int HardSigmoid_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 2) shape_packed = Mat(shape.w, shape.h / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(2 + 5);
+    std::vector<vk_specialization_type> specializations(2 + 1);
     specializations[0].f = alpha;
     specializations[1].f = beta;
-    specializations[2 + 0].i = shape_packed.dims;
-    specializations[2 + 1].i = shape_packed.w;
-    specializations[2 + 2].i = shape_packed.h;
-    specializations[2 + 3].i = shape_packed.c;
-    specializations[2 + 4].i = shape_packed.cstep;
-
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_hardsigmoid = new Pipeline(vkdev);
-        pipeline_hardsigmoid->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_hardsigmoid->create(LayerShaderType::hardsigmoid, opt, specializations);
-    }
+    specializations[2 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_hardsigmoid_pack4 = new Pipeline(vkdev);
-        pipeline_hardsigmoid_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_hardsigmoid_pack4->create(LayerShaderType::hardsigmoid_pack4, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_hardsigmoid_pack8 = new Pipeline(vkdev);
-        pipeline_hardsigmoid_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_hardsigmoid_pack8->create(LayerShaderType::hardsigmoid_pack8, opt, specializations);
-    }
+    pipeline_hardsigmoid = new Pipeline(vkdev);
+    pipeline_hardsigmoid->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_hardsigmoid->create(LayerShaderType::hardsigmoid, opt, specializations);
 
     return 0;
 }
@@ -101,34 +57,24 @@ int HardSigmoid_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_hardsigmoid;
     pipeline_hardsigmoid = 0;
 
-    delete pipeline_hardsigmoid_pack4;
-    pipeline_hardsigmoid_pack4 = 0;
-
-    delete pipeline_hardsigmoid_pack8;
-    pipeline_hardsigmoid_pack8 = 0;
-
     return 0;
 }
 
 int HardSigmoid_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_hardsigmoid_pack8
-                               : elempack == 4 ? pipeline_hardsigmoid_pack4
-                               : pipeline_hardsigmoid;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_hardsigmoid, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/hardsigmoid_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_hardsigmoid;
-    Pipeline* pipeline_hardsigmoid_pack4;
-    Pipeline* pipeline_hardsigmoid_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/hardswish_vulkan.cpp

@@ -12,8 +12,6 @@ HardSwish_vulkan::HardSwish_vulkan()
     support_vulkan = true;
 
     pipeline_hardswish = 0;
-    pipeline_hardswish_pack4 = 0;
-    pipeline_hardswish_pack8 = 0;
 }
 
 int HardSwish_vulkan::create_pipeline(const Option& opt)
@@ -40,58 +38,16 @@ int HardSwish_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 2) shape_packed = Mat(shape.w, shape.h / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(2 + 5);
+    std::vector<vk_specialization_type> specializations(2 + 1);
     specializations[0].f = alpha;
     specializations[1].f = beta;
-    specializations[2 + 0].i = shape_packed.dims;
-    specializations[2 + 1].i = shape_packed.w;
-    specializations[2 + 2].i = shape_packed.h;
-    specializations[2 + 3].i = shape_packed.c;
-    specializations[2 + 4].i = shape_packed.cstep;
-
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_hardswish = new Pipeline(vkdev);
-        pipeline_hardswish->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_hardswish->create(LayerShaderType::hardswish, opt, specializations);
-    }
+    specializations[2 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_hardswish_pack4 = new Pipeline(vkdev);
-        pipeline_hardswish_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_hardswish_pack4->create(LayerShaderType::hardswish_pack4, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_hardswish_pack8 = new Pipeline(vkdev);
-        pipeline_hardswish_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_hardswish_pack8->create(LayerShaderType::hardswish_pack8, opt, specializations);
-    }
+    pipeline_hardswish = new Pipeline(vkdev);
+    pipeline_hardswish->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_hardswish->create(LayerShaderType::hardswish, opt, specializations);
 
     return 0;
 }
@@ -101,34 +57,24 @@ int HardSwish_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_hardswish;
     pipeline_hardswish = 0;
 
-    delete pipeline_hardswish_pack4;
-    pipeline_hardswish_pack4 = 0;
-
-    delete pipeline_hardswish_pack8;
-    pipeline_hardswish_pack8 = 0;
-
     return 0;
 }
 
 int HardSwish_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_hardswish_pack8
-                               : elempack == 4 ? pipeline_hardswish_pack4
-                               : pipeline_hardswish;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_hardswish, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/hardswish_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_hardswish;
-    Pipeline* pipeline_hardswish_pack4;
-    Pipeline* pipeline_hardswish_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/mish_vulkan.cpp

@@ -12,8 +12,6 @@ Mish_vulkan::Mish_vulkan()
     support_vulkan = true;
 
     pipeline_mish = 0;
-    pipeline_mish_pack4 = 0;
-    pipeline_mish_pack8 = 0;
 }
 
 int Mish_vulkan::create_pipeline(const Option& opt)
@@ -41,62 +39,14 @@ int Mish_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(0 + 5);
-    specializations[0 + 0].i = shape_packed.dims;
-    specializations[0 + 1].i = shape_packed.w;
-    specializations[0 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[0 + 3].i = shape_packed.c;
-    specializations[0 + 4].i = shape_packed.cstep;
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0].u32 = shape_packed.total() * elempack / 4;
 
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_mish = new Pipeline(vkdev);
-        pipeline_mish->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_mish->create(LayerShaderType::mish, opt, specializations);
-    }
-
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_mish_pack4 = new Pipeline(vkdev);
-        pipeline_mish_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_mish_pack4->create(LayerShaderType::mish_pack4, opt, specializations);
-    }
-
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_mish_pack8 = new Pipeline(vkdev);
-        pipeline_mish_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_mish_pack8->create(LayerShaderType::mish_pack8, opt, specializations);
-    }
+    pipeline_mish = new Pipeline(vkdev);
+    pipeline_mish->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_mish->create(LayerShaderType::mish, opt, specializations);
 
     return 0;
 }
@@ -106,34 +56,24 @@ int Mish_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_mish;
     pipeline_mish = 0;
 
-    delete pipeline_mish_pack4;
-    pipeline_mish_pack4 = 0;
-
-    delete pipeline_mish_pack8;
-    pipeline_mish_pack8 = 0;
-
     return 0;
 }
 
 int Mish_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_mish_pack8
-                               : elempack == 4 ? pipeline_mish_pack4
-                               : pipeline_mish;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_mish, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/mish_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_mish;
-    Pipeline* pipeline_mish_pack4;
-    Pipeline* pipeline_mish_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/relu_vulkan.cpp

@@ -12,8 +12,6 @@ ReLU_vulkan::ReLU_vulkan()
     support_vulkan = true;
 
     pipeline_relu = 0;
-    pipeline_relu_pack4 = 0;
-    pipeline_relu_pack8 = 0;
 }
 
 int ReLU_vulkan::create_pipeline(const Option& opt)
@@ -41,63 +39,15 @@ int ReLU_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(1 + 5);
+    std::vector<vk_specialization_type> specializations(1 + 1);
     specializations[0].f = slope;
-    specializations[1 + 0].i = shape_packed.dims;
-    specializations[1 + 1].i = shape_packed.w;
-    specializations[1 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[1 + 3].i = shape_packed.c;
-    specializations[1 + 4].i = shape_packed.cstep;
-
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_relu = new Pipeline(vkdev);
-        pipeline_relu->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_relu->create(LayerShaderType::relu, opt, specializations);
-    }
+    specializations[1 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_relu_pack4 = new Pipeline(vkdev);
-        pipeline_relu_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_relu_pack4->create(LayerShaderType::relu_pack4, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_relu_pack8 = new Pipeline(vkdev);
-        pipeline_relu_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_relu_pack8->create(LayerShaderType::relu_pack8, opt, specializations);
-    }
+    pipeline_relu = new Pipeline(vkdev);
+    pipeline_relu->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_relu->create(LayerShaderType::relu, opt, specializations);
 
     return 0;
 }
@@ -107,34 +57,24 @@ int ReLU_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_relu;
     pipeline_relu = 0;
 
-    delete pipeline_relu_pack4;
-    pipeline_relu_pack4 = 0;
-
-    delete pipeline_relu_pack8;
-    pipeline_relu_pack8 = 0;
-
     return 0;
 }
 
 int ReLU_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_relu_pack8
-                               : elempack == 4 ? pipeline_relu_pack4
-                               : pipeline_relu;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_relu, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/relu_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_relu;
-    Pipeline* pipeline_relu_pack4;
-    Pipeline* pipeline_relu_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/shader/celu.comp

@@ -6,37 +6,25 @@
 layout (constant_id = 0) const float alpha = 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 + 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    const uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    v = max(v, afpvec4(0.0f)) + min(alpha * (exp(v / afpvec4(alpha)) - afpvec4(1.0f)), afpvec4(0.0f));
 
-    v = max(v, afp(0.0f)) + min(alpha * (exp(v / alpha) - 1.0f), afp(0.0f));
-
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/celu_pack4.comp

@@ -1,42 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v = max(v, afp(0.0f)) + min(alpha * (exp(v / alpha) - 1.0f), afp(0.0f));
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/celu_pack8.comp

@@ -1,43 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = max(v[0], afp(0.0f)) + min(alpha * (exp(v[0] / alpha) - 1.0f), afp(0.0f));
-    v[1] = max(v[1], afp(0.0f)) + min(alpha*(exp(v[1]/alpha)-1), afp(0.0f));
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/clip.comp

@@ -7,37 +7,25 @@ layout (constant_id = 0) const float const_min = 0;
 layout (constant_id = 1) const float const_max = 0;
 
 #define shape_constant_id_offset 2
-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 + 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    const uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
+    if (gi >= psc(n))
+    return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    v = clamp(v, afpvec4(const_min), afpvec4(const_max));
 
-    v = clamp(v, afp(const_min), afp(const_max));
-
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/clip_pack4.comp

@@ -1,43 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float const_min = 0;
-layout (constant_id = 1) const float const_max = 0;
-
-#define shape_constant_id_offset 2
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v = clamp(v, afp(const_min), afp(const_max));
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/clip_pack8.comp

@@ -1,44 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float const_min = 0;
-layout (constant_id = 1) const float const_max = 0;
-
-#define shape_constant_id_offset 2
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = clamp(v[0], afp(const_min), afp(const_max));
-    v[1] = clamp(v[1], afp(const_min), afp(const_max));
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/dropout.comp

@@ -1,4 +1,4 @@
-// Copyright 2018 Tencent
+// Copyright 2019 Tencent
 // SPDX-License-Identifier: BSD-3-Clause
 
 #version 450
@@ -6,37 +6,26 @@
 layout (constant_id = 0) const float scale = 1;
 
 #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 + 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
     v *= afp(scale);
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/dropout_pack4.comp

@@ -1,42 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float scale = 1;
-
-#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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v *= afp(scale);
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/dropout_pack8.comp

@@ -1,43 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float scale = 1;
-
-#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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = v[0] * afp(scale);
-    v[1] = v[1] * afp(scale);
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/elu.comp

@@ -6,37 +6,25 @@
 layout (constant_id = 0) const float alpha = 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 + 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    const uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    v = mix(afpvec4(alpha) * afpvec4(exp(v) - afpvec4(1.0f)), v, greaterThan(v, afpvec4(0.0f)));
 
-    v = v > afp(0.0f) ? v : afp(alpha * (exp(v) - 1.0f));
-
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/elu_pack4.comp

@@ -1,42 +0,0 @@
-// Copyright 2022 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v = mix(afpvec4(alpha) * afpvec4(exp(v) - afpvec4(1.0f)), v, greaterThan(v, afpvec4(0.0f)));
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/elu_pack8.comp

@@ -1,43 +0,0 @@
-// Copyright 2022 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = mix(afpvec4(alpha) * afpvec4(exp(v[0]) - afpvec4(1.0f)), v[0], greaterThan(v[0], afpvec4(0.0f)));
-    v[1] = mix(afpvec4(alpha) * afpvec4(exp(v[1]) - afpvec4(1.0f)), v[1], greaterThan(v[1], afpvec4(0.0f)));
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/erf.comp

@@ -3,53 +3,41 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
-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 = 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
-float erf(float x)
+afpvec4 erf(afpvec4 x)
 {
-    float a1 =  0.254829592f;
-    float a2 = -0.284496736f;
-    float a3 =  1.421413741f;
-    float a4 = -1.453152027f;
-    float a5 =  1.061405429f;
-    float p  =  0.3275911f;
-    float s = sign(x);
-    float x_abs = abs(x);
-    float t = 1.0f/(1.0f + p*x_abs);
-    float y = 1.0f - (((((a5*t + a4)*t) + a3)*t + a2)*t + a1)*t*exp(-x_abs*x_abs);
+    afpvec4 a1 = afpvec4(0.254829592f);
+    afpvec4 a2 = afpvec4(-0.284496736f);
+    afpvec4 a3 = afpvec4(1.421413741f);
+    afpvec4 a4 = afpvec4(-1.453152027f);
+    afpvec4 a5 = afpvec4(1.061405429f);
+    afpvec4 p = afpvec4(0.3275911f);
+    afpvec4 s = sign(x);
+    afpvec4 x_abs = abs(x);
+    afpvec4 t = 1.0f / (1.0f + p * x_abs);
+    afpvec4 y = 1.0f - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * exp(-x_abs * x_abs);
     return s * y;
 }
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    const uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
 
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
     v = erf(v);
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/erf_pack4.comp

@@ -1,55 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-afpvec4 erf(afpvec4 x)
-{
-    afpvec4 a1 = afpvec4(0.254829592f);
-    afpvec4 a2 = afpvec4(-0.284496736f);
-    afpvec4 a3 = afpvec4(1.421413741f);
-    afpvec4 a4 = afpvec4(-1.453152027f);
-    afpvec4 a5 = afpvec4(1.061405429f);
-    afpvec4 p = afpvec4(0.3275911f);
-    afpvec4 s = sign(x);
-    afpvec4 x_abs = abs(x);
-    afpvec4 t = 1.0f / (1.0f + p * x_abs);
-    afpvec4 y = 1.0f - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * exp(-x_abs * x_abs);
-    return s * y;
-}
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v = erf(v);
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/erf_pack8.comp

@@ -1,56 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-afpvec4 erf(afpvec4 x)
-{
-    afpvec4 a1 = afpvec4(0.254829592f);
-    afpvec4 a2 = afpvec4(-0.284496736f);
-    afpvec4 a3 = afpvec4(1.421413741f);
-    afpvec4 a4 = afpvec4(-1.453152027f);
-    afpvec4 a5 = afpvec4(1.061405429f);
-    afpvec4 p = afpvec4(0.3275911f);
-    afpvec4 s = sign(x);
-    afpvec4 x_abs = abs(x);
-    afpvec4 t = 1.0f / (1.0f + p * x_abs);
-    afpvec4 y = 1.0f - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1) * t * exp(-x_abs * x_abs);
-    return s * y;
-}
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = erf(v[0]);
-    v[1] = erf(v[1]);
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/gelu.comp

@@ -4,43 +4,32 @@
 #version 450
 
 #define shape_constant_id_offset 0
-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 = 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
 
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
     // y = 0.5x * (1 + tanh(sqrt(2/Pi) * (x + 0.044715x^3)))
 
 #if NCNN_moltenvk
-    v = afp(0.5f) * v * (afp(1.0f) + afp(tanh(float(afp(0.79788452f) * (v + afp(0.044715f) * v * v * v)))));
+    v = afpvec4(0.5f) * v * (afpvec4(1.0f) + afpvec4(tanh(vec4(afpvec4(0.79788452f) * (v + afpvec4(0.044715f) * v * v * v)))));
 #else
-    v = afp(0.5f) * v * (afp(1.0f) + tanh(afp(0.79788452f) * (v + afp(0.044715f) * v * v * v)));
+    v = afpvec4(0.5f) * v * (afpvec4(1.0f) + tanh(afpvec4(0.79788452f) * (v + afpvec4(0.044715f) * v * v * v)));
 #endif
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/gelu_pack4.comp

@@ -1,46 +0,0 @@
-// Copyright 2023 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    // y = 0.5x * (1 + tanh(sqrt(2/Pi) * (x + 0.044715x^3)))
-
-#if NCNN_moltenvk
-    v = afpvec4(0.5f) * v * (afpvec4(1.0f) + afpvec4(tanh(vec4(afpvec4(0.79788452f) * (v + afpvec4(0.044715f) * v * v * v)))));
-#else
-    v = afpvec4(0.5f) * v * (afpvec4(1.0f) + tanh(afpvec4(0.79788452f) * (v + afpvec4(0.044715f) * v * v * v)));
-#endif
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/gelu_pack8.comp

@@ -1,48 +0,0 @@
-// Copyright 2023 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    // y = 0.5x * (1 + tanh(sqrt(2/Pi) * (x + 0.044715x^3)))
-    
-#if NCNN_moltenvk
-    v[0] = afpvec4(0.5f) * v[0] * (afpvec4(1.0f) + afpvec4(tanh(vec4(afpvec4(0.79788452f) * (v[0] + afpvec4(0.044715f) * v[0] * v[0] * v[0])))));
-    v[1] = afpvec4(0.5f) * v[1] * (afpvec4(1.0f) + afpvec4(tanh(vec4(afpvec4(0.79788452f) * (v[1] + afpvec4(0.044715f) * v[1] * v[1] * v[1])))));
-#else
-    v[0] = afpvec4(0.5f) * v[0] * (afpvec4(1.0f) + tanh(afpvec4(0.79788452f) * (v[0] + afpvec4(0.044715f) * v[0] * v[0] * v[0])));
-    v[1] = afpvec4(0.5f) * v[1] * (afpvec4(1.0f) + tanh(afpvec4(0.79788452f) * (v[1] + afpvec4(0.044715f) * v[1] * v[1] * v[1])));
-#endif
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/hardsigmoid.comp

@@ -1,4 +1,4 @@
-// Copyright 2018 Tencent
+// Copyright 2019 Tencent
 // SPDX-License-Identifier: BSD-3-Clause
 
 #version 450
@@ -7,37 +7,26 @@ layout (constant_id = 0) const float alpha = 0;
 layout (constant_id = 1) const float beta = 0;
 
 #define shape_constant_id_offset 2
-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 + 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
     v = clamp(v * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/hardsigmoid_pack4.comp

@@ -1,43 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 0;
-layout (constant_id = 1) const float beta = 0;
-
-#define shape_constant_id_offset 2
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v = clamp(v * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/hardsigmoid_pack8.comp

@@ -1,44 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 0;
-layout (constant_id = 1) const float beta = 0;
-
-#define shape_constant_id_offset 2
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = clamp(v[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    v[1] = clamp(v[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/hardswish.comp

@@ -1,4 +1,4 @@
-// Copyright 2018 Tencent
+// Copyright 2019 Tencent
 // SPDX-License-Identifier: BSD-3-Clause
 
 #version 450
@@ -7,37 +7,25 @@ layout (constant_id = 0) const float alpha = 0;
 layout (constant_id = 1) const float beta = 0;
 
 #define shape_constant_id_offset 2
-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 + 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
     v = v * clamp(v * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/hardswish_pack4.comp

@@ -1,43 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 0;
-layout (constant_id = 1) const float beta = 0;
-
-#define shape_constant_id_offset 2
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v = v * clamp(v * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/hardswish_pack8.comp

@@ -1,44 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float alpha = 0;
-layout (constant_id = 1) const float beta = 0;
-
-#define shape_constant_id_offset 2
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = v[0] * clamp(v[0] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-    v[1] = v[1] * clamp(v[1] * afp(alpha) + afp(beta), afp(0.f), afp(1.f));
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/mish.comp

@@ -1,44 +1,32 @@
-// Copyright 2020 Tencent
+// Copyright 2019 Tencent
 // SPDX-License-Identifier: BSD-3-Clause
 
 #version 450
 
-#define shape_constant_id_offset 0
-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 = 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
 #if NCNN_moltenvk
-    v = v * afp(tanh(float(log(exp(v) + afp(1.f)))));
+    v = v * afpvec4(tanh(vec4(log(exp(v) + afpvec4(1.f)))));
 #else
-    v = v * tanh(log(exp(v) + afp(1.f)));
+    v = v * tanh(log(exp(v) + afpvec4(1.f)));
 #endif
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/mish_pack4.comp

@@ -1,44 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-#if NCNN_moltenvk
-    v = v * afpvec4(tanh(vec4(log(exp(v) + afpvec4(1.f)))));
-#else
-    v = v * tanh(log(exp(v) + afpvec4(1.f)));
-#endif
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/mish_pack8.comp

@@ -1,46 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-#if NCNN_moltenvk
-    v[0] = v[0] * afpvec4(tanh(vec4(log(exp(v[0]) + afpvec4(1.f)))));
-    v[1] = v[1] * afpvec4(tanh(vec4(log(exp(v[1]) + afpvec4(1.f)))));
-#else
-    v[0] = v[0] * tanh(log(exp(v[0]) + afpvec4(1.f)));
-    v[1] = v[1] * tanh(log(exp(v[1]) + afpvec4(1.f)));
-#endif
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/relu.comp

@@ -1,4 +1,4 @@
-// Copyright 2018 Tencent
+// Copyright 2019 Tencent
 // SPDX-License-Identifier: BSD-3-Clause
 
 #version 450
@@ -6,40 +6,29 @@
 layout (constant_id = 0) const float slope = 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 + 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
     if (slope == 0)
         v = max(v, afp(0.f));
     else
-        v = v < afp(0.f) ? v * afp(slope) : v;
+        v = mix(v, v * afp(slope), lessThan(v, afpvec4(0.f)));
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/relu_pack4.comp

@@ -1,45 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float slope = 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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    if (slope == 0)
-        v = max(v, afp(0.f));
-    else
-        v = mix(v, v * afp(slope), lessThan(v, afpvec4(0.f)));
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/relu_pack8.comp

@@ -1,51 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-layout (constant_id = 0) const float slope = 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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    if (slope == 0)
-    {
-        v[0] = max(v[0], afp(0.f));
-        v[1] = max(v[1], afp(0.f));
-    }
-    else
-    {
-        v[0] = mix(v[0], v[0] * afp(slope), lessThan(v[0], afpvec4(0.f)));
-        v[1] = mix(v[1], v[1] * afp(slope), lessThan(v[1], afpvec4(0.f)));
-    }
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/swish.comp

@@ -1,40 +1,27 @@
-// Copyright 2020 Tencent
+// Copyright 2019 Tencent
 // SPDX-License-Identifier: BSD-3-Clause
 
 #version 450
 
-#define shape_constant_id_offset 0
-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 = 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    v = v / (afpvec4(1.f) + exp(-v));
 
-    v = v / (afp(1.f) + exp(-v));
-
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/swish_pack4.comp

@@ -1,40 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-    v = v / (afpvec4(1.f) + exp(-v));
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/swish_pack8.comp

@@ -1,41 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-    v[0] = v[0] / (afpvec4(1.f) + exp(-v[0]));
-    v[1] = v[1] / (afpvec4(1.f) + exp(-v[1]));
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/tanh.comp

@@ -3,42 +3,30 @@
 
 #version 450
 
-#define shape_constant_id_offset 0
-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 = 0) const uint n = 0;
 
-layout (binding = 0) buffer bottom_top_blob { sfp bottom_top_blob_data[]; };
+
+layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
 
 layout (push_constant) uniform parameter
 {
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
+    uint n;
 } p;
 
 void main()
 {
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
+    uint gi = gl_GlobalInvocationID.x;
 
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
+    if (gi >= psc(n))
         return;
 
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afp v = buffer_ld1(bottom_top_blob_data, gi);
+    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
 
 #if NCNN_moltenvk
-    v = afp(tanh(float(v)));
+    v = afpvec4(tanh(vec4(v)));
 #else
     v = tanh(v);
 #endif
 
-    buffer_st1(bottom_top_blob_data, gi, v);
+    buffer_st4(bottom_top_blob_data, gi, v);
 }

src/layer/vulkan/shader/tanh_pack4.comp

@@ -1,44 +0,0 @@
-// Copyright 2019 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
-
-#if NCNN_moltenvk
-    v = afpvec4(tanh(vec4(v)));
-#else
-    v = tanh(v);
-#endif
-
-    buffer_st4(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/shader/tanh_pack8.comp

@@ -1,46 +0,0 @@
-// Copyright 2020 Tencent
-// SPDX-License-Identifier: BSD-3-Clause
-
-#version 450
-
-#define shape_constant_id_offset 0
-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 (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
-
-layout (push_constant) uniform parameter
-{
-    int dims;
-    int w;
-    int h;
-    int c;
-    int cstep;
-} p;
-
-void main()
-{
-    int gx = int(gl_GlobalInvocationID.x);
-    int gy = int(gl_GlobalInvocationID.y);
-    int gz = int(gl_GlobalInvocationID.z);
-
-    if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
-        return;
-
-    const int gi = gz * psc(cstep) + gy * psc(w) + gx;
-
-    afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
-
-#if NCNN_moltenvk
-    v[0] = afpvec4(tanh(vec4(v[0])));
-    v[1] = afpvec4(tanh(vec4(v[1])));
-#else
-    v[0] = tanh(v[0]);
-    v[1] = tanh(v[1]);
-#endif
-
-    buffer_st8(bottom_top_blob_data, gi, v);
-}

src/layer/vulkan/swish_vulkan.cpp

@@ -12,8 +12,6 @@ Swish_vulkan::Swish_vulkan()
     support_vulkan = true;
 
     pipeline_swish = 0;
-    pipeline_swish_pack4 = 0;
-    pipeline_swish_pack8 = 0;
 }
 
 int Swish_vulkan::create_pipeline(const Option& opt)
@@ -41,62 +39,14 @@ int Swish_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(0 + 5);
-    specializations[0 + 0].i = shape_packed.dims;
-    specializations[0 + 1].i = shape_packed.w;
-    specializations[0 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[0 + 3].i = shape_packed.c;
-    specializations[0 + 4].i = shape_packed.cstep;
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_swish = new Pipeline(vkdev);
-        pipeline_swish->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_swish->create(LayerShaderType::swish, opt, specializations);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_swish_pack4 = new Pipeline(vkdev);
-        pipeline_swish_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_swish_pack4->create(LayerShaderType::swish_pack4, opt, specializations);
-    }
-
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_swish_pack8 = new Pipeline(vkdev);
-        pipeline_swish_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_swish_pack8->create(LayerShaderType::swish_pack8, opt, specializations);
-    }
+    pipeline_swish = new Pipeline(vkdev);
+    pipeline_swish->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_swish->create(LayerShaderType::swish, opt, specializations);
 
     return 0;
 }
@@ -106,36 +56,25 @@ int Swish_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_swish;
     pipeline_swish = 0;
 
-    delete pipeline_swish_pack4;
-    pipeline_swish_pack4 = 0;
-
-    delete pipeline_swish_pack8;
-    pipeline_swish_pack8 = 0;
-
     return 0;
 }
 
 int Swish_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    const Pipeline* pipeline = elempack == 8 ? pipeline_swish_pack8
-                               : elempack == 4 ? pipeline_swish_pack4
-                               : pipeline_swish;
-
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_swish, bindings, constants, dispatcher);
 
     return 0;
 }
-
 } // namespace ncnn

src/layer/vulkan/swish_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_swish;
-    Pipeline* pipeline_swish_pack4;
-    Pipeline* pipeline_swish_pack8;
 };
 
 } // namespace ncnn

src/layer/vulkan/tanh_vulkan.cpp

@@ -12,8 +12,6 @@ TanH_vulkan::TanH_vulkan()
     support_vulkan = true;
 
     pipeline_tanh = 0;
-    pipeline_tanh_pack4 = 0;
-    pipeline_tanh_pack8 = 0;
 }
 
 int TanH_vulkan::create_pipeline(const Option& opt)
@@ -41,62 +39,14 @@ int TanH_vulkan::create_pipeline(const Option& opt)
     if (shape.dims == 3) shape_packed = Mat(shape.w, shape.h, shape.c / elempack, (void*)0, elemsize, elempack);
     if (shape.dims == 4) shape_packed = Mat(shape.w, shape.h, shape.d, shape.c / elempack, (void*)0, elemsize, elempack);
 
-    std::vector<vk_specialization_type> specializations(0 + 5);
-    specializations[0 + 0].i = shape_packed.dims;
-    specializations[0 + 1].i = shape_packed.w;
-    specializations[0 + 2].i = shape_packed.h * shape_packed.d;
-    specializations[0 + 3].i = shape_packed.c;
-    specializations[0 + 4].i = shape_packed.cstep;
+    std::vector<vk_specialization_type> specializations(1);
+    specializations[0 + 0].u32 = shape_packed.total() * elempack / 4;
 
-    Mat local_size_xyz;
-    if (shape_packed.dims == 1)
-    {
-        local_size_xyz.w = std::min(64, shape_packed.w);
-        local_size_xyz.h = 1;
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 2)
-    {
-        local_size_xyz.w = std::min(8, shape_packed.w);
-        local_size_xyz.h = std::min(8, shape_packed.h);
-        local_size_xyz.c = 1;
-    }
-    if (shape_packed.dims == 3)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
-    if (shape_packed.dims == 4)
-    {
-        local_size_xyz.w = std::min(4, shape_packed.w);
-        local_size_xyz.h = std::min(4, shape_packed.h * shape_packed.d);
-        local_size_xyz.c = std::min(4, shape_packed.c);
-    }
+    const int local_size_x = vkdev->info.subgroup_size();
 
-    // pack1
-    if (shape.dims == 0 || elempack == 1)
-    {
-        pipeline_tanh = new Pipeline(vkdev);
-        pipeline_tanh->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_tanh->create(LayerShaderType::tanh, opt, specializations);
-    }
-
-    // pack4
-    if (shape.dims == 0 || elempack == 4)
-    {
-        pipeline_tanh_pack4 = new Pipeline(vkdev);
-        pipeline_tanh_pack4->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_tanh_pack4->create(LayerShaderType::tanh_pack4, opt, specializations);
-    }
-
-    // pack8
-    if ((opt.use_shader_pack8 && shape.dims == 0) || elempack == 8)
-    {
-        pipeline_tanh_pack8 = new Pipeline(vkdev);
-        pipeline_tanh_pack8->set_optimal_local_size_xyz(local_size_xyz);
-        pipeline_tanh_pack8->create(LayerShaderType::tanh_pack8, opt, specializations);
-    }
+    pipeline_tanh = new Pipeline(vkdev);
+    pipeline_tanh->set_optimal_local_size_xyz(local_size_x, 1, 1);
+    pipeline_tanh->create(LayerShaderType::tanh, opt, specializations);
 
     return 0;
 }
@@ -106,34 +56,24 @@ int TanH_vulkan::destroy_pipeline(const Option& /*opt*/)
     delete pipeline_tanh;
     pipeline_tanh = 0;
 
-    delete pipeline_tanh_pack4;
-    pipeline_tanh_pack4 = 0;
-
-    delete pipeline_tanh_pack8;
-    pipeline_tanh_pack8 = 0;
-
     return 0;
 }
 
 int TanH_vulkan::forward_inplace(VkMat& bottom_top_blob, VkCompute& cmd, const Option& /*opt*/) const
 {
-    int elempack = bottom_top_blob.elempack;
+    const size_t n = bottom_top_blob.total() * bottom_top_blob.elempack / 4;
 
     std::vector<VkMat> bindings(1);
     bindings[0] = bottom_top_blob;
 
-    std::vector<vk_constant_type> constants(5);
-    constants[0].i = bottom_top_blob.dims;
-    constants[1].i = bottom_top_blob.w;
-    constants[2].i = bottom_top_blob.h * bottom_top_blob.d;
-    constants[3].i = bottom_top_blob.c;
-    constants[4].i = bottom_top_blob.cstep;
-
-    const Pipeline* pipeline = elempack == 8 ? pipeline_tanh_pack8
-                               : elempack == 4 ? pipeline_tanh_pack4
-                               : pipeline_tanh;
+    std::vector<vk_constant_type> constants(1);
+    constants[0].u32 = n;
 
-    cmd.record_pipeline(pipeline, bindings, constants, bottom_top_blob);
+    VkMat dispatcher;
+    dispatcher.w = n;
+    dispatcher.h = 1;
+    dispatcher.c = 1;
+    cmd.record_pipeline(pipeline_tanh, bindings, constants, dispatcher);
 
     return 0;
 }

src/layer/vulkan/tanh_vulkan.h

@@ -21,8 +21,6 @@ public:
 
 public:
     Pipeline* pipeline_tanh;
-    Pipeline* pipeline_tanh_pack4;
-    Pipeline* pipeline_tanh_pack8;
 };
 
 } // namespace ncnn